text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'Coupling of photonic crystal (PC) linear three-hole defect cavities (L3) to PC waveguides is theoretically and experimentally investigated. The systems are designed to increase the overlap between the evanescent cavity field and the waveguide mode, and to operate in the linear dispersion region of the waveguide. Our simulations indicate increased coupling when the cavity is tilted by [$ 60 ^o $]{} with respect to the waveguide axis, which we have also confirmed by experiments. We obtained up to 90% coupling efficiency into the waveguide.'
author:
- 'Andrei Faraon, Dirk Englund, Ilya Fushman, Jelena Vučković'
- Edo Waks
bibliography:
- 'references.bib'
date: 'October 12, 2006'
title: 'Efficient Photonic Crystal Cavity-Waveguide Couplers'
---
Structures that consist of InGaAs/GaAs quantum dots (QDs) coupled to two-dimensional PC cavities are promising candidates for highly efficient single photon sources. They represent essential devices for quantum cryptography and quantum computation [@Vuckovic03APL; @EdoDIT; @05PRLEnglund; @VuckovicSingPhotDem]. Efficient implementation of quantum computation devices requires integration of photonic circuits directly on the chip. These circuits consist of single photon sources (SPSs) that inject single photons into the waveguides, which subsequently redirect them to other quantum nodes, i.e. other PC cavities containing QDs. Once the necessary quantum operations have been performed, photons need to be outcoupled from the waveguide either out-of-plane for vertical collection (e.g., by coupling the photons back into an “output cavity" that scatters them out of plane), or collected in the PC plane (e.g., by outcoupling to a fiber). The performance of this kind of circuit is limited by the coupling efficiency between the cavities and the waveguides. Our work investigates this coupling with the goal of improving the efficiency of single photon transmission from one cavity to another. The results are also relevant for channel drop filter applications in optical telecommunications.
In this paper we investigate the coupling of linear three-hole cavities (L3) [@NodaL3] into PC waveguides. We choose the L3 cavities for their high quality factor(Q) to mode volume (V) ratio and good matching between cavity and waveguide field patterns, which improves in-plane coupling efficiency [@EFV05OptExpr; @EdoCoupledModeTh]. The cavity mode we work with has magnetic field with even/odd symmetry with respect to the [**x/y**]{} axes. This mode, whose magnetic field configuration is depicted in Fig. 1(a), needs to be coupled to one of the guided modes in the PC waveguide. The field is computed using three-dimensional finite difference time domain simulations (3D FDTD). Of the possible waveguide bands inside the PC band gap [@EFV05OptExpr] the best choice for coupling the L3 cavity is the one with similar symmetry and frequency as the L3 cavity mode (Fig. 1(b)). To get efficient coupling, the cavity and waveguide modes need to be spatially overlapped and frequency matched. A closer look at the L3 cavity field profile (Fig. 1(a)) reveals that the evanescent field is strongest along a direction tilted with respect to the cavity axis and is weak along the cavity axis. A good approach for obtaining a larger overlap between the cavity and waveguide mode is to tilt the cavity with respect to the waveguide axis by an angle of [$ 60 ^{o} $]{} (Fig. 1(c)). The choice of this angle is determined by the symmetry constraints of the triangular lattice. Directional couplers with cavity axes non-parallel to waveguide axes have recently been studied by Kim et al [@NotomiHex] for coupling the hexapole modes of single hole defect cavities and by Shinya et al for coupling L3 and L4 cavities [@NotomiFlipFlop]. In contrast with previous work, we present here optimized designs of couplers as well as detailed theoretical and experimental data, confirming the advantage of the tilted configuration for coupling L3 cavities to PC waveguides.
![ (a) Magnetic field ($ B _z $ component) for the mode with the highest quality factor in a L3 cavity. (b) Magnetic field pattern of the even mode in a PC waveguide. (c) Fabricated tilted cavity coupled to a waveguide (four holes separation). In this experiment we shift the cavity with respect to the waveguide along the direction indicated by the arrow. (d) Fabricated straight cavity coupled to a waveguide (three holes separation).[]{data-label="figureone"}](figures/fig1_v8.eps){width="3in"}
To test the validity of our approach, we compare the coupling parameters for the tilted cavity configuration (Fig. 1(c)) to the standard approach where the cavity and the waveguide share the same axis (straight cavity configuration) (Fig. 1(d)). First, 3D FDTD simulations of coupled cavity waveguide systems were performed with both tilted and straight couplers. The frequency of the waveguide band was lowered with respect to the cavity frequency by reducing the size of the PC holes that bound the waveguide. In this way, coupling occurs in the dispersion-free linear region of the waveguide band. We directly simulated tilted and straight coupler configurations with spacing of two-to-five lattice holes separation between the cavity and the waveguide. An image of the simulated magnetic field profile for a tilted cavity coupled to a waveguide with three-hole separation is depicted in Fig. 2(inset) . In the tilted configuration, the separation between the cavity and the waveguide is changed along a direction indicated by the arrow in Fig. 1(c).
![ Simulation results for the cavity waveguide coupling expressed in terms of the quality factor. The coupling strength is proportional to ([$ 1/Q _{wg}$]{}). Simulated magnetic field of a cavity-waveguide coupler in tilted configuration with three hole separation (inset). ](figures/Fig2_v8.eps){width="3in"}
The energy transfer into the waveguide degrades the Q of the coupled cavity. The total Q of a coupled cavity relates to the uncoupled cavity quality factor ([$ Q _{c}$]{}) according to:
$${Q _{tot}} ^{-1} = {Q _{c}}^{-1} + {Q _{wg}} ^{-1},$$
where [$ Q _{wg} ^{-1} $]{} is the loss rate into the waveguide.
Different applications require different coupling. For high-efficiency single photon transfer, the in-plane coupling into the waveguide modes needs to be dominant so [$ Q _{wg}$]{} should be lower than [$ Q _{c} $]{}. On the other hand, the advanced single photon sources [@05PRLEnglund] require cavities with a quality factor on the order of thousands, which implies [$ Q _{wg}$]{} should also be in the same range. For other applications, single photons need to be scattered out of plane from a PC waveguide through an output cavity. To achieve high transfer efficiency from waveguides to the output cavities, the cavity-waveguide system needs to be in the critical coupling regime defined by [$ Q _{wg} = Q _{c} $]{}. In that case, we do not need the output cavity to have a high quality factor.
The coupling strength between the cavity and the waveguide is given by [$ 1/Q _{wg} $]{} which is proportional to the decay rate of the cavity field into the waveguide. The quality factor [$ Q _{wg} $]{} was computed from the 3D FDTD simulations, with results presented in Fig. 2. For the same cavity-waveguide separation, [$ Q _{wg} $]{} is generally smaller for the tilted than for the straight configuration. This is an indication of better cavity-waveguide coupling obtained by tilting the cavity. One peculiar aspect of the simulations is that for the tilted coupling configuration, the Q is actually larger for four-holes than for five-holes separation. This is unexpected because it is natural to assume that reducing the distance between the cavity and waveguide should improve the overlap integral between the two modes. However, this increase in the quality factor is observed under a large variety of different simulation parameters, suggesting that it is real, as opposed to a simulation artifact. We suspect that, at four hole separation, the anti-node of one of the modes overlaps with the node of the other resulting in an lower overlap integral. Further investigation is required in order to conclusively confirm this.
The coupling changes from [$ Q _{wg} \approx 500 $]{} for the tilted cavity with two-hole separation to [$ Q _{wg} \approx 10^6 $]{} for four and five-holes separation (both configurations). For single photon sources based on PC cavities with InGaAs QDs operating at [$ 900 nm - 1000 nm$]{}, the experimental out-of-plane quality factor is limited to about [$ Q _{c} = 10^4 $]{} because of material loss and fabrication imperfections [@Englund05_OptExp]. On the other hand, to get efficient photon transfer into the waveguide, [$ Q _{wg} $]{} needs to be lower than [$ Q _{c} $]{} therefore, only the coupling configurations with two- and three-hole separation represent good options. Experimentally we expect the total Q to be independent of the waveguide coupling in the case of four and five holes separation.
To test the validity of our simulation results, the couplers were fabricated on a 165 nm thick freestanding GaAs membrane containing a InGaAs QD layer. Structures with two- to five-hole separation in both tilted and straight configuration (Fig. 1(c, d)) were fabricated. We made seven structures of each kind. The spectrum of each cavity was measured using the InGaAs QDs embedded in the GaAs membrane as an internal light source. The fabrication and measurement procedures are similar with those reported in [@05PRLEnglund]. The mean value of the quality factor for each configuration is plotted in Fig. 3(a), where the error bars are given by the standard deviation in Q due to fabrication fluctuations between the seven structures of each kind.
![ Comparison between simulations and experimental data for cavity-waveguide couplers. (a) The measured value of total Q (mean) (b) The value of Q inferred from simulations by combining simulated [$Q _{wg}$]{} and measured [$Q _{c}$]{}. (c) The coupling efficiency from the PC cavity into the PC waveguide. (d) Measured spectrum of a closed waveguide coupled to a L3 cavity. The Fabry-Perot fringes are equidistant in the linear region of the waveguide dispersion relation (where the cavity is also located) and they get closer next to the waveguide band-edge (970nm).](figures/Fig3_v6.eps){width="3.5in"}
As expected from simulations, the experimental data show that for the same cavity-waveguide separation, the total quality factor is lower for the tilted than for the straight configuration. This result is a consequence of higher coupling for tilted cavities.
Since a more efficient coupling between the cavity and the waveguide degrades the cavity quality factor, when designing a PC network one should choose the configuration that gives the optimum trade-off between transfer efficiency and high Q. One advantage of using the tilted cavity is that the same set of parameters can be obtained with the cavity further spaced from the waveguide.
As mentioned before, the cavity coupling was designed to couple in the linear region of the waveguide-band dispersion relation. To test the position of the cavity with respect to the waveguide band, we fabricated longer waveguides closed at the ends. These waveguides act as Fabry-Perot resonators. Fringes can be observed using the broad distribution of the QDs [@viktorovitch]. In the linear region of the dispersion relation the fringes are equally spaced, and get closer together as the frequency approaches the band edge. Since the cavity resonance was positioned in the region with equidistant fringes, we concluded that the coupling occurs in the linear region (Fig. 4(d)).
For a direct comparison between simulation and experiment, [$ Q _{c}$]{} of the uncoupled cavity needs to be known. The upper bound for [$ Q _{c}$]{} is limited by fabrication imperfections and material loss. Our simulation results indicate that in the case of coupled cavities with four hole separation the coupling into the waveguide is very small so the total Q is well approximated by [$ Q _{c}$]{}. For this reason, the average value of the measured Q for the tilted configuration with four hole separation was used as [$ Q _{c}$]{}. By plugging [$ Q _{c}$]{} and the simulated value for [$ Q _{wg}$]{} into expression (1), the predicted value for the total Q ([$ Q _{tot}$]{}) was computed. The values for the Q inferred from simulations are plotted in Fig. 3(b) and show good agreement with the experimental data (Fig. 3(a)). Some inconsistency is observed in the case of five-hole separation. These inconsistencies result from fabrication errors.
The coupling efficiency into the waveguide was computed by taking the ratio $Q / Q_{wg}$ and the results are plotted in Fig. 3(c). The coupling efficiency is up to 90% in the case of tilted configuration with two holes separation and up to 40% for straight configuration with two holes separation.
In conclusion we have designed PC cavity-waveguide couplers with optimized coupling efficiency and operating in the linear waveguide dispersion region. We have shown both theoretically and experimentally that the coupling between a L3 PC cavity and PC waveguides can be improved by tilting the cavity with respect to the waveguide. The coupling is more efficient because the evanescent tails of the cavity field are not oriented along the cavity axis but at a [$ 30^{o} $]{} angle. Understanding and controlling the coupling mechanism is essential for on-chip single photon transfer and the implementation of on-chip quantum networks.
Financial support was provided by the MURI Center for photonic quantum information systems (ARO/DTO program No. DAAD19-03-1-0199), ONR Young Investigator Award and NSF Grant No. CCF-0507295.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show a uniqueness theorem for charged rotating black holes in the bosonic sector of five-dimensional minimal supergravity. More precisely, under the assumptions of the existence of two commuting axial isometries and spherical topology of horizon cross-sections, we prove that an asymptotically flat, stationary charged rotating black hole with finite temperature in five-dimensional Einstein-Maxwell-Chern-Simons theory is uniquely characterized by the mass, charge, and two independent angular momenta and therefore is described by the five-dimensional Cveti[č]{}-Youm solution with equal charges. We also discuss a generalization of our uniqueness theorem for spherical black holes to the case of black rings.'
author:
- Shinya Tomizawa
title: ' Uniqueness theorem for charged rotating black holes in five-dimensional minimal supergravity '
---
[ ]{}
Introduction {#sec:intro}
============
In string theory and various related contexts, higher dimensional black holes and other extended black objects have played an important role. In particular, physics of black holes in five-dimensional Einstein-Maxwell-Chern-Simons (EMCS) theory has recently been the subject of increased attention, as the five-dimensional EMCS theory describes the bosonic sector of five-dimensional minimal supergravity, a low-energy limit of string theory. Various types of black hole solutions in EMCS-theory [@Gauntlett0; @BMPV; @CY96; @CLP; @CCLP; @Bena2; @EEF; @Elvang; @EEMR; @Elvang3; @Gaiotto; @IM; @IKMT; @IKMT2; @NIMT; @TIMN; @MINT; @TI; @TYM; @GS; @TIKM; @T] have so far been found, with the help of, in part, recent development of solution generating techniques [@Belinskii; @solitonbook; @Tomizawa; @Tomizawa2; @Tomizawa3; @Tomizawa4; @Iguchiboost; @Pomeransky:2005sj; @Koikawa; @Azuma; @MishimaIguchi; @Iguchi; @EK07; @CAM; @BCCGSW; @GS; @Yaza06a; @Yaza06b; @Yaza08]. However, the classification of those black hole solutions has not been achieved yet. The purpose of this paper is to show a uniqueness theorem for charged rotating black holes in five-dimensional EMCS theory, as a partial solution to the black hole classification problem in string theory.
It is now evident that even within the framework of vacuum Einstein gravity, there is a much richer variety of black hole solutions in higher dimensions [@Myers:1986un; @Emparan:2001wn; @MishimaIguchi; @Pom; @diring; @saturn; @Izumi; @bi], the classification of which still remains a major open issue. As shown by Emparan and Reall [@Emparan:2001wn], five-dimensional vacuum Einstein gravity admits the co-existence of a rotating spherical hole and two rotating rings with the same conserved charges, illustrating explicitly the non-uniqueness property in higher dimensions. However, it is possible to show type of uniqueness theorems for some restricted cases in which certain additional conditions are imposed on some parameters/properties, other than the global conserved charges. For example, restricting attention to static solutions, Gibbons [*et al.*]{} [@shiromizu] showed that the only asymptotically flat, static vacuum black hole is the Schwarzschild-Tangherlini solution [@T-schwarzschild]. For rotating case, by assuming the existence of two axial Killing symmetries and spherical topology of the event horizon, Morisawa and Ida [@MI] succeeded in proving that five-dimensional asymptotically flat, stationary vacuum rotating black holes must be in the Myers-Perry family. Their theorem was recently generalized to a class of asymptotically flat solutions with non-spherical horizon topology [@Hollands; @MTY]. For other cases (such as cases including Maxwell-field), see [@HY; @R; @R2; @R3; @R4; @R5; @R6; @R7; @HY08].
In this paper, we generalize the boundary value analysis of Morisawa and Ida [@MI] performed in vacuum Einstein gravity to the case of the bosonic sector of five-dimensional minimal supergravity. We are concerned with stationary black hole spacetimes that are asymptotically flat in the standard sense: Namely, we demand that the exterior region of the black hole is globally hyperbolic, having a spherical spatial infinity, and that the metric and other physical fields, such as Maxwell field, fall-off in a certain manner at large distances. (The asymptotic fall-off conditions are given later.) Furthermore, for simplicity, we focus on the single black hole case, that is, the event horizon is connected. Then, we note that in five-dimensional EMCS theory the Cveti[č]{}-Youm black hole solution with equal charges [@CY96] appears to be the most general such solutions that describe an asymptotically flat, stationary charged rotating black hole with spherical horizon topology, characterized by four conserved charges, i.e., the mass, two independent angular momenta, and electric charge, and that encompass the known asymptotically flat, spherical black hole solutions in a subclass of EMCS theory, such as the Myers-Perry solution [@Myers:1986un], in a certain limit. Thus, we wish to show the following theorem.
[**Theorem.**]{} [*Consider, in five-dimensional Einstein-Maxwell-Chern-Simons theory \[given by eq. (\[action\]) below\], a stationary charged rotating black hole with finite temperature that is regular on and outside the event horizon and asymptotically flat in the standard sense with spherical spatial infinity. If (1) the black hole spacetime admits, besides the stationary Killing vector field, two mutually commuting axial Killing vector fields so that the isometry group is ${\Bbb R}\times U(1)\times U(1)$ and (2) the topology of the horizon cross-sections is spherical, ${\rm S^3}$, and the topology of the black hole exterior region is ${\Bbb R} \times \{ {\Bbb R}^4 \setminus {\Bbb B}^4 \}$, then the black hole spacetime is uniquely characterized by its mass, electric charge, and two independent angular momenta, and hence must be isometric to the Cveti[č]{}-Youm solution with equal charges.* ]{}
Before presenting our proof, we would like to make a few comments concerning the assumptions made in our theorem. In order to obtain global results, we need the symmetry-condition (1), which, in particular makes it possible to reduce five-dimensional minimal supergravity to a non-linear sigma model with certain symmetries as shown in [@MO; @MO2]. Since all known exact black hole solutions in higher dimensions admit multiple axial isometries, our additional symmetry-condition (1) does not appear to be too restrictive. However, we should note that the rigidity theorem [@HIW] (see also [@MI08]) in higher dimensions—which is recently shown to be applicable also to EMCS theory [@HI]—only guarantees the existence of a single rotational isometry (provided the spacetime metric and other fields are real, analytic), and therefore at present, the condition (1) is not yet fully justified. In this respect, note also that the possibility for higher-dimensional black holes with fewer isometries than ${\Bbb R} \times U(1)\times U(1)$ has been suggested [@Reall03]. Since the rigidity theorem yields that the event horizon is a Killing horizon, the notion of surface gravity is well-defined. Then, by [*finite temperature*]{} we mean that the event horizon is of non-degenerate type, having non-vanishing surface gravity and a bifurcate surface [@RW92; @RW96]. For extremal (zero-temperature) black holes with vanishing surface gravity, the event horizon is of degenerated type and does not possess a bifurcate surface. Then, our boundary conditions to be imposed on target space fields at the event horizon would not appear to straightforwardly apply to such a case that the horizon has no bifurcate surface. It would be of great interest to consider the classification problem of such extremal (zero-temperature) black holes. In this respect, there have recently appeared some attempts to classify near-horizon geometries of extremal black objects, rather than extremal black objects themselves (see e.g., [@KL08a; @KL08b; @FKLR08; @KLR07; @AGJST06; @AY08] and references therein).
We also need to additionally impose the topology-condition (2), in order to explicitly specify boundary conditions on target space variables at the event horizon, in terms of certain coordinates, globally defined over the black hole exterior region. The topological censorship, together with our assumption of asymptotic flatness described above, immediately implies that the exterior region is topologically ${\Bbb R} \times V^{(4)}$ with $V^{(4)}$ being some four-dimensional simply connected Riemannian manifold. However, the simple connectedness by itself does not completely determine the topology of $V^{(4)}$. Therefore, in the present theorem, we simply demand that $V^{(4)} \approx \{{\Bbb R}^4 \setminus {\Bbb B}^4 \}$, which is in accordance with the topology of the Cveti[č]{}-Youm solution with equal charges. Our boundary conditions—in particular, the [*rod structure*]{}, which was first introduced by Harmark [@Harmark] based on earlier work for static solutions [@weyl]—are accordingly specified in the manner discussed in Sec. \[sec:coincidence\]. The topology theorem [@Cai; @Helfgott; @galloway] yields that in five-dimensions, cross-sections of the event horizon must be topologically either a sphere, a ring, or a lens-space. The requirement (2) excludes some interesting class of solutions to be dealt with. It would be interesting to consider generalization of our uniqueness theorem to include solutions with non-spherical horizon topology. We would like to emphasize that even under these restrictive assumptions (1) and (2), still it is not at all obvious whether black holes in EMCS-theory are uniquely specified by their global charges. In fact, it has been shown by numerical studies [@Kunz] that when the value of the Chern-Simons coupling is larger than some critical value, spherical black holes in such a general EMCS-theory no longer enjoy the uniqueness property. In the present paper, motivated from sting theory, we restrict attention to a special class of EMCS-theory, that is, five-dimensional minimal supergravity and then are able to show the above uniqueness theorem. It would be interesting to find the precise onset of this non-uniqueness property in general EMCS theory, using the formulas developed in this paper.
The rest of the paper is devoted to prove the above uniqueness theorem. In the next section, we present the metric and the gauge potential in Einstein-Maxwell-Chern-Simons theory with three Killing symmetries, introduce the Weyl-Papapetrou coordinates, and reduce the system to a non-linear sigma model with certain symmetries. In Sec. \[sec:mazur\], using the matrix representation of the sigma model, we derive a divergence identity/Mazur identity associated with our nonlinear sigma model. A good part of the material in Sec. \[sec:metric\] and the first part of Sec. \[sec:mazur\] concerning the matrix representation is discussed in [@BCCGSW]. Then, in Sec. \[sec:coincidence\], presenting our boundary conditions for our sigma model fields and using the Mazur identity, we show that if two asymptotically flat black hole solutions have the same conserved charges, i.e., the mass, electric charge, and two angular momenta, then they must coincide with each other, and complete our proof of the uniqueness theorem. In Sec. \[sec:conclusion\], we summarize our results and discuss possible generalization of our theorem to include non-spherical black objects. We discuss that in order to have a uniqueness theorem for black ring solutions in EMCS-theory, we need to specify rod-data, besides global charges and horizon topology. In Appendix \[sec:A\], we explicitly compute relevant components of the Maxwell-field. In Appendix \[sec:solution\], we provide the five-dimensional Cveti[č]{}-Youm solution with equal charges, and study, in terms of the Weyl-Papapetrou coordinates, the limiting behavior of the solution near relevant boundaries.
Einstein-Maxwell-Chern-Simons system with symmetries {#sec:metric}
====================================================
We consider the bosonic sector of five-dimensional minimal supergravity theory, which can be obtained by a suitable truncation of eleven-dimensional supergravity. The five-dimensional action is given by $$\begin{aligned}
S=\frac{1}{16\pi}
\left[
\int dx^5\sqrt{-g}\left(R-\frac{1}{4}F^2\right)
-\frac{1}{3\sqrt{3}} \int F\wedge F\wedge A
\right] \,,
\label{action} \end{aligned}$$ where we set a Newton constant to be unity and $F=dA$. Varying this action (\[action\]), we derive the Einstein equation $$\begin{aligned}
R_{\mu \nu } -\frac{1}{2} R g_{\mu \nu }
= \frac{1}{2} \left( F_{\mu \lambda } F_\nu^{ ~ \lambda }
- \frac{1}{4} g_{\mu \nu } F_{\rho \sigma } F^{\rho \sigma } \right) \,,
\label{Eineq}\end{aligned}$$ and the Maxwell equation $$\begin{aligned}
d*F+\frac{1}{\sqrt{3}}F\wedge F=0 \,.
\label{Maxeq}\end{aligned}$$ The purpose of this section is to reduce the above five-dimensional Einstein-Maxwell-Chern-Simons system to a non-linear sigma model with certain target space symmetries. We first consider consequences of the existence of commuting Killing vector fields in our spacetime and identify the target space variables in Subsec. \[subsec:II-A\]. Then, having another (stationary) Killing vector field, we introduce the Weyl-Papapetrou coordinates and write down explicitly the desired non-linear sigma model action in Subsec. \[subsec:II-B\].
Two Killing system {#subsec:II-A}
------------------
Let $\xi_a \ (a=1,2)$ be two mutually commuting Killing vector fields, so that $[\xi_a,\xi_b]=0$, ${\cal L}_{\xi_a} g=0$, and ${\cal L}_{\xi_a}F=0$. Then, introducing the coordinates $x^a$ as Killing parameters of $\xi_a$ (so that $\xi_a = \partial/\partial x^a$), one can express the metric $g$ and the gauge potential one-form $A$, respectively, as $$\begin{aligned}
ds^2 = \lambda_{ab}(dx^a+a^a{}_idx^i)(dx^b+a^b{}_jdx^j)
+|\tau|^{-1}h_{ij}dx^idx^j \,, \end{aligned}$$ $$\begin{aligned}
A = A_adx^a + A_i dx^i \,, \end{aligned}$$ where the functions $\tau:=-{\rm det}(\lambda_{ab})$, $a^a{_i}$, $h_{ij}$, $A_a$, and $A_i$ ($i=3,4,5$) are independent of the coordinates $x^a$.
Let us define the electric one-form $E_a$ with respect to $\xi_a$ by $$\begin{aligned}
&&E_a=-i_{\xi_a}F \,.\end{aligned}$$ Then the exterior derivatives of the electric one-forms yield $$\begin{aligned}
dE_a=i_{\xi_a}dF-{\cal L}_{\xi_a}F=0 \,,\end{aligned}$$ where $F=dA$ is used. Hence there exist locally the potentials $\psi_a$ such that $$\begin{aligned}
d\psi_a=-\frac{1}{\sqrt{3}}i_{\xi_a}F \,. \end{aligned}$$ Hence, the gauge potential can be written as $$\begin{aligned}
A = \sqrt{3}\psi_adx^a + A_i dx^i \,,\end{aligned}$$ where $\psi_a$ is also independent of the coordinates $x^a$. Next, define the magnetic one-form $B$ by $$\begin{aligned}
&&B=*(\xi_1\wedge \xi_2\wedge F) \,. \end{aligned}$$ Noting that $B$ can be rewritten as $
B=*(\xi_1\wedge \xi_2\wedge F)=-i_{\xi_2}*(\xi_1\wedge F)=i_{\xi_2}i_{\xi_1}*F
$ and using the identity $
di_{\xi_2}i_{\xi_1}=i_{\xi_2}i_{\xi_1}d+i_{\xi_1}{\cal L}_{\xi_2}-i_{\xi_2}{\cal L}_{\xi_1}
$, we can write the exterior derivative of $B$ as $$\begin{aligned}
dB=i_{\xi_2}i_{\xi_1}d*F \,. \end{aligned}$$ Then, using the Maxwell equation (\[Maxeq\]), we find that $$\begin{aligned}
dB= -\frac{1}{\sqrt{3}}i_{\xi_2}i_{\xi_1}F\wedge F
= \frac{2}{\sqrt{3}} E_1\wedge E_2
= 2\sqrt{3} d\psi_1\wedge d\psi_2=\sqrt{3} d(\psi_1d\psi_2-\psi_2d\psi_1) \,.\end{aligned}$$ This immediately implies that there exists the magnetic potential $\mu$ such that $$\begin{aligned}
d\mu=\frac{1}{\sqrt{3}}B-\epsilon^{ab}\psi_ad\psi_b \,,
\label{eq:mu} \end{aligned}$$ where $\epsilon^{12}=-\epsilon^{21}=1$. We also introduce the twist one-form by $$V_a=*(\xi_1\wedge \xi_2\wedge d\xi_a) \,.$$ Using the Einstein-equation, we can write the exterior derivative of $V_a$ as $$\begin{aligned}
dV_a &=& 2*(\xi_1\wedge \xi_2\wedge R(\xi_a)) \nonumber \\
&=&-\tau^{-1}i_{\xi_2}i_{\xi_1}*^2(\xi_1\wedge\xi_2\wedge E_a\wedge B)
\nonumber \\
&=&-E_a\wedge B \nonumber \\
&=&-3d\psi_a\wedge (d\mu+\epsilon^{bc}\psi_bd\psi_c) \nonumber \\
&=&-3d[\psi_ad\mu]-d[\psi_a\epsilon^{bc}\psi_bd\psi_c] \,, \end{aligned}$$ where $R(\xi_a)$ in the first line is the Ricci one-form. Therefore, there exists the twist potentials $\omega_a$ that satisfy $$\begin{aligned}
d\omega_a=V_a+\psi_a(3d\mu+\epsilon^{bc}\psi_bd\psi_c)\,.
\label{eq:twistpotential} \end{aligned}$$
Thus, as a consequence of the existence of isometries $\xi_a$, we have eight scalar fields $\lambda_{ab},\omega_a,\psi_a,\mu$ $(a=1,2)$, which we denote collectively by coordinates $\Phi^A=(\lambda_{ab},\omega_a,\psi_a,\mu)$. As we will see soon, other components, such as $a^a{}_i$, $A_i$ are determined by $\Phi^A$. Then, we can find that the equations of motion, eqs. (\[Eineq\]) and (\[Maxeq\]), are cast into a set of equations derived from the following action for sigma-model $\Phi^A$ coupled with three-dimensional gravity with respect to the metric $h_{ij}$, $$\begin{aligned}
S=\int_\Sigma\left({\cal R}^h
-G_{AB}\frac{\partial \Phi^A}{\partial x^i}
\frac{\partial \Phi^B}{\partial x^j}h^{ij}\right)\sqrt{|h|}dx^3 \,,\end{aligned}$$ where the target space metric, $G_{AB}$, is given by $$\begin{aligned}
G_{AB}d\Phi^Ad\Phi^B
&=& \frac{1}{4}{\rm Tr}(\lambda^{-1}d\lambda\lambda^{-1}d\lambda )
+ \frac{1}{4}\tau^{-2}d\tau^2
+ \frac{3}{2}d\psi^T \lambda^{-1}d\psi \nonumber \\
& &-\frac{1}{2}\tau^{-1}V^T\lambda^{-1}V
-\frac{3}{2}\tau^{-1}(d\mu+\epsilon^{ab}\psi_ad\psi_b)^2 \,,\end{aligned}$$ where $\lambda=(\lambda_{ab})$, $\psi=(\psi_1,\psi_2)^T$, $\omega=(\omega_1,\omega_2)^T$ and $V=d\omega-\psi(3d\mu+\epsilon^{bc}\psi_bd\psi_c)$. Varying the action by $h_{ij}$, we obtain the equations $$\begin{aligned}
R^h_{ij} = G_{AB}\frac{\partial \Phi^A}{\partial x^i}
\frac{\partial \Phi^B}{\partial x^j} \,,
\label{eq:Rij}\end{aligned}$$ where $R^h_{ij}$ denotes the Ricci tensor with respect to $h_{ij}$. Next varying the action by $\Phi^A$, we derive the equation $$\begin{aligned}
\Delta_h\Phi^A+h^{ij}\Gamma^A_{BC}\frac{\partial \Phi^B}{\partial x^i}
\frac{\partial \Phi^C}{\partial x^j}
=0,\end{aligned}$$ where $\Delta_h$ is the Laplacian with respect to the three-dimensional metric $h_{ij}$ and $\Gamma^A_{BC}$ is the Christoffel symbol with respect to the target space metric $G_{AB}$.
Weyl-Papapetrou form {#subsec:II-B}
--------------------
Now we consider another Killing vector field $\xi_3$ which is assumed to commute with the other Killing vectors $\xi_a$ and will be identified below as the asymptotic time-translation Killing vector field. Let us consider the condition that the two-dimensional distribution orthogonal to three Killing vector fields $\xi_I\ (I=1,2,3)$ becomes integrable. The commutativity of Killing vector fields, $[\xi_I, \xi_J]=0$, enables us to find coordinate system $x_I$ $(I=1, 2, 3)$, so that $\xi_I=\partial/\partial {x^I}$ and the coordinate components of the metric become independent of $x^I$. We now recall the following theorem about the integrability of two-planes orthogonal to Killing vector fields [@weyl; @Harmark]:
[**Proposition.**]{}
*If three mutually commuting Killing vector fields $\xi_I\ (I=1,2,3)$ in a five-dimensional spacetime satisfy the following two conditions*
1. $\xi_{1}^{[\mu _1}\xi_{2}^{\mu _2}
\xi_{3}^{\mu _{2}}D^\nu \xi_{I}^{\rho ]}=0$ holds at at least one point of the spacetime for a given $I=1,2,3$,
2. $\xi_I^\nu R_\nu ^{[\rho}
\xi_{1}^{\mu _1}\xi_{2}^{\mu _2}
\xi_{3}^{\mu _{2}]}=0 $ holds for all $I=1,2,3$,
then the two-planes orthogonal to the Killing vector fields $\xi_I\ (I=1,2,3)$ are integrable.
Note here that one can replace a pair of Killing vector fields $(\xi_1,\xi_2)$ above by another pair $(\xi_2,\xi_3)$. We denote the corresponding quantities in the choice $(\xi_2,\xi_3)$ with [*tilde*]{} $\ \tilde{}\ $. For example, we denote the twist one-forms with respect to $(\xi_2,\xi_3)$ by $$\begin{aligned}
\tilde V_{\tilde a}=*(\xi_2\wedge\xi_3\wedge d\xi_{\tilde a}) \,,\end{aligned}$$ where $\tilde a=2,3$. Then, using $
i_{\xi_I}d\psi_a=i_{\xi_I}d\mu=0, \ i_{\xi_I}d\tilde\psi_a=i_{\xi_I}d\tilde\mu=0
$, and eq. (\[eq:twistpotential\]), we show $
i_{\xi_I}dV_a=i_{\xi_I}d\tilde V_{\tilde a}=0
$, and hence have $$\begin{aligned}
*(\xi_1\wedge \xi_2\wedge \xi_3\wedge R(\xi_a))
&=& -i_{\xi_3}*(\xi_1\wedge\xi_2\wedge R(\xi_a))
= -\frac{1}{2}i_{\xi_3}dV_a =0 \,, \end{aligned}$$ and $$\begin{aligned}
*(\xi_1\wedge \xi_2\wedge \xi_3\wedge R(\xi_3))
&=& -i_{\xi_1}*(\xi_2\wedge\xi_3\wedge R(\xi_{3}))
= -\frac{1}{2}i_{\xi_1}d\tilde V_{3}
=0 \,. \end{aligned}$$ This implies that the condition $2$ holds in our present system (\[eq:action\]) with three commuting Killing vector fields. Furthermore, the axial symmetry of at least one of $\xi_I$ $(I=1,2,3)$ implies that the condition $1$ also holds on the axis of rotation. Therefore, the two-dimensional surface orthogonal to three $\xi_I$ is integrable.
Now, without loss of generality, we choose our three coordinates $(x^1,x^2,x^3)$ as the three Killing parameters, so that $\xi_3 = \partial/\partial t$ denotes the stationary (asymptotic time-translation) Killing vector field in our spacetimes and $\xi_1= \partial/ \partial \phi$ and $\xi_2 = \partial/\partial \psi$ are two independent axial Killing symmetries. Then, from the above observation, we can express the three dimensional metric $h_{ij}$ by $h=h_{pq}dx^pdx^q- \rho^2dt^2\ (p,q=4,5)$, where $\rho^2=-{\rm det}(g_{IJ})$. Note that the function $\rho$ is globally well-defined [@Chru08]. That $\rho$ is a harmonic function can be seen by looking at the $(tt)$-component of eq. (\[eq:Rij\]), which is written $$\begin{aligned}
R_{tt}=\rho \hat D^2\rho=0 \,,\end{aligned}$$ where $\hat D_p$ is the covariant derivative associated with the two-dimensional metric $h_{pq}$. Let $z$ be harmonic function conjugate to $\rho$ which satisfies $\hat D^2 z=0$, $\hat D_p \rho\hat D^pz=0$, $\hat D_p \rho\hat D^p\rho=\hat D_p z\hat D^p z$. Choose the coordinates $(x^4,x^5)$ as $x^4=\rho$ and $x^5=z$. Then, the metric can be written in the Weyl-Papapetrou type form as $$\begin{aligned}
ds^2&=&\lambda_{\phi\phi}(d\phi+a^\phi{}_tdt)^2+\lambda_{\psi\psi}(d\psi+a^\psi{}_tdt)^2 \nonumber\\
&&+2\lambda_{\phi\psi}(d\phi+a^\phi{}_tdt)(d\psi+a^\psi{}_tdt)+|\tau|^{-1}[e^{2\sigma}(d\rho^2+dz^2)-\rho^2 dt^2] \,, \end{aligned}$$ where all the metric components depend only on $\rho$ and $z$.
In this coordinate system, $\Phi^A$ are determined by the equations of motion $$\begin{aligned}
\Delta_\gamma \Phi^A +
\Gamma^A_{BC}[\Phi^B_{,\rho}\Phi^C_{,\rho}+\Phi^C_{,z}\Phi^C_{,z}]=0 \,,
\label{eq:scalar}\end{aligned}$$ where $\Delta_\gamma$ is the Laplacian with respect to the abstract three-dimensional metric $\gamma=d\rho^2+dz^2+\rho^2d\varphi^2$. On the other hand, once $\Phi^A$ are given, one can completely determine $\sigma$, $a^\phi{}_t$, $a^\psi{}_t$, $A_i$. In fact, the function $\sigma$ is determined by $$\begin{aligned}
\frac{2}{\rho}\sigma_{,\rho}&=&R^h_{\rho\rho}-R^h_{zz}\nonumber\\
&=&G_{AB}[\Phi^A_{,\rho}\Phi^B_{,\rho}-\Phi^A_{,z}\Phi^B_{,z}] \,,\\
\frac{1}{\rho}\sigma_{,z}&=&R^h_{\rho z}\nonumber\\
&=&G_{AB}\Phi^A_{,\rho}\Phi^B_{,z} \,.\end{aligned}$$ The integrability $\sigma_{,\rho z}=\sigma_{,z \rho }$ is assured by eq. (\[eq:scalar\]). From eq.(\[eq:twistpotential\]), the metric functions $a^a{_t}$ are determined by $$\begin{aligned}
a^a{}_{t,\rho}
&=&\rho \tau^{-1} \lambda^{ab}(\omega_{b,z}-3\psi_b\mu_{,z}-\psi_b\epsilon^{cd}\psi_c\psi_{d,z}) \\
a^a{}_{t,z}&=&-\rho\tau^{-1}\lambda^{ab}(\omega_{b,\rho}-3\psi_b\mu_{,\rho}-\psi_b\epsilon^{cd}\psi_c\psi_{d,\rho}) .\end{aligned}$$ As shown in Appendix \[sec:A\], we can set $A_{\rho}=A_{z}=0$. Therefore it follows from eq. (\[eq:mu\]) that the $t$-component of the gauge potential $A$ is determined by $$\begin{aligned}
A_{t,\rho}&=&\sqrt{3}\left[a^a{}_t\psi_{a,\rho}-\rho \tau^{-1}(\mu_{,z}+\epsilon^{bc}\psi_b\psi_{c,z})\right],\\
A_{t,z}&=&\sqrt{3}\left[a^a{}_t\psi_{a,z}+\rho\tau^{-1}(\mu_{,\rho}+\epsilon^{bc}\psi_b\psi_{c,\rho})\right].\end{aligned}$$ Thus, once we determine $\Phi^A= (\lambda_{ab},\omega_a,\psi_a,\mu)$, we can specify the solutions of the system given originally by the action, eq. (\[action\]), with our Killing symmetry assumption. It turns out that the above equations of motion, eq. (\[eq:scalar\]), for $\Phi^A$ are derived from the following action $$\begin{aligned}
S&=&\int d\rho dz \rho\left[ G_{AB}(\partial\Phi^A)(\partial\Phi^B)\right]
\nonumber \\
&=& \int d\rho dz \rho
\biggl[\:
\frac{1}{4}{\rm Tr}(\lambda^{-1}\partial\lambda\lambda^{-1}
\partial\lambda )
+ \frac{1}{4}\tau^{-2}\partial\tau^2
+ \frac{3}{2}\partial\psi^T \lambda^{-1}\partial\psi
\nonumber \\
& &{} \qquad \qquad
- \frac{1}{2}\tau^{-1}v^T\lambda^{-1}v
- \frac{3}{2}\tau^{-1}(\partial\mu
+\epsilon^{ab}\psi_a\partial\psi_b)^2
\biggr] \,,
\label{eq:action} \end{aligned}$$ where $v=\partial \omega-\psi(3\partial \mu+\epsilon^{bc}\psi_b\partial \psi_c)$. This action is invariant under the global $G_{2(2)}$ transformation.
Mazur identity {#sec:mazur}
==============
In the proof of uniqueness theorems for four-dimensional charged rotating black holes, a key role was played by a certain global identity—called the Mazur identity. This is also the case for five-dimensional charged rotating black holes. In this section, we present the Mazur type identity for our non-linear sigma models derived in the previous section. The derivation parallels that for the vacuum Einstein case given in other literature, e.g., Morisawa and Ida [@MI], and therefore we present here only some key formulas. Following [@BCCGSW], we introduce the $G_{2(2)}/SO(4)$ coset matrix, $M$, defined by $$\begin{aligned}
M= \left(
\begin{array}{ccc}
\hat A&\hat B&\sqrt{2}\hat U\\
\hat B^T&\hat C&\sqrt{2}\hat V\\
\sqrt{2}\hat U^T&\sqrt{2}\hat V^T&\hat S\\
\end{array}
\right) \,,\end{aligned}$$ where $\hat A$ and $\hat C$ are symmetric $3\times 3$ matrices, $\hat B$ is a $3\times 3$ matrix, $\hat U$ and $\hat V$ are 3-component column matrices, and $\hat S$ is a scalar, defined, respectively, by $$\begin{aligned}
&&\hat A=\left(
\begin{array}{ccc}
[(1-y)\lambda+(2+x)\psi \psi^T-\tau^{-1}\tilde\omega\tilde\omega^T+\mu(\psi \psi^T\lambda^{-1}\hat J-\hat J\lambda^{-1}\psi\psi^T)]&\tau^{-1}\tilde\omega\\
\tau^{-1}\tilde\omega^T& -\tau^{-1}
\end{array}
\right) \,,\nonumber\\
&&\hat B=\left(
\begin{array}{ccc}
(\psi\psi^T-\mu \hat J)\lambda^{-1}-\tau^{-1}\tilde\omega \psi^T \hat J&[(-(1+y)\lambda \hat J-(2+x)\mu+\psi^T\lambda^{-1}\tilde\omega)\psi+(z-\mu \hat J\lambda^{-1}\tilde)\omega] \\
\tau^{-1}\psi^T \hat J&-z\\
\end{array}
\right) \,, \nonumber\\
&&\hat C=\left(
\begin{array}{ccc}
(1+x)\lambda^{-1}-\lambda^{-1}\psi\psi^T\lambda^{-1}&\lambda^{-1}\tilde\omega-\hat J(z-\mu \hat J \lambda^{-1})\psi\\
\tilde\omega^T\lambda^{-1}+\psi^T(z+\mu \lambda^{-1}\hat J)\hat J&[\tilde\omega^T\lambda^{-1}\tilde\omega-2\mu\psi^T\lambda^{-1}\tilde\omega-\tau(1+x-2y-xy+z^2)]\\
\end{array}
\right) \,,\nonumber\\
&&\hat U=
\left(
\begin{array}{cc}
(1+x-\mu \hat J\lambda^{-1})\psi-\mu \tau^{-1}\tilde\omega\\
\mu\tau^{-1}\\
\end{array}
\right) \,,\nonumber\\
&&\hat V=\left(
\begin{array}{ccc}
(\lambda^{-1}+\mu\tau^{-1}\hat J)\psi\\
\psi^T\lambda^{-1}\tilde\omega-\mu(1+x-z)\nonumber\\
\end{array}
\right) \,,\nonumber\\
&&\hat S=1+2(x-y) \,, \nonumber\end{aligned}$$ with $$\begin{aligned}
&&\tilde \omega=\omega-\mu\psi \,,\end{aligned}$$ $$\begin{aligned}
&&x=\psi^T\lambda^{-1}\psi,\quad y
=\tau^{-1}\mu^2,\quad z=y-\tau^{-1}\psi^T\hat J\tilde\omega \,,\end{aligned}$$ and the $2\times 2$ matrix, $$\begin{aligned}
\hat J= \left(
\begin{array}{ccccccc}
0&1\\
-1&0\\
\end{array}
\right) \,. \end{aligned}$$ We note that this $7\times7$ matrix $M$ is symmetric, $M^T=M$, and unimodular, $\det(M)= 1$. Since we choose the Killing vector fields $\xi_\phi$ and $\xi_\psi$ to be spacelike, all the eigenvalues of $M$ are real and positive. Therefore, there exists an $G_{2(2)}$ matrix $\hat g$ such that $$\begin{aligned}
M = \hat g \hat g^T \,.\end{aligned}$$ We define a current matrix as $$\begin{aligned}
J_i = M^{-1} \partial_i M \,,\end{aligned}$$ which is conserved if the scalar fields are the solutions of the equation of motion derived by the action (\[eq:action\]). Then, the action (\[eq:action\]) can be written in terms of $J$ and $M$ as follows $$\begin{aligned}
S&=&\frac{1}{4}\int d\rho dz \rho {\rm tr}(J_iJ^i) \nonumber \\
&=&\frac{1}{4}\int d\rho dz
\rho {\rm tr}(M^{-1}\partial_iMM^{-1}\partial^iM)\,. \end{aligned}$$ Thus, the matrix $M$ completely specify the solutions to our system.
Let us now consider two sets of field configurations, $M_{[0]}$ and $M_{[1]}$, that satisfy the equations of motion derived from the action, eq. (\[eq:action\]). We denote the difference between the value of the functional obtained from the field configuration $M_{[1]}$ and the value obtained from $M_{[0]}$ as a bull’s eye $\stackrel{\odot}{}$, e.g., $$\begin{aligned}
\stackrel{\odot}J{}^i=J^i_{[1]}-J^i_{[0]} \,,\end{aligned}$$ where the subscripts ${}_{[0]}$ and ${}_{[1]}$ denote, respectively, the quantities associated with the field configurations $M_{[0]}$ and $M_{[1]}$. The deviation matrix, $\Psi$, is then defined by $$\begin{aligned}
\Psi=\stackrel{\odot}MM^{-1}_{[0]}=M_{[1]}M^{-1}_{[0]}-{\bf 1} \,,\end{aligned}$$ where ${\bf 1}$ is the unit matrix. Taking the derivative of this, we have the relation between the derivative of the deviation matrix and $\stackrel{\odot}J{}^i$, $$\begin{aligned}
D^i\Psi=M_{[1]}\stackrel{\odot}J{}^iM_{[0]}^{-1} \,,
\label{eq:deriv}\end{aligned}$$ where $D_i$ is a covariant derivative associated with the abstract three-metric $\gamma$. Taking, further, the divergence of the above formula and also the trace of the matrix elements, we have the following divergence identity $$\begin{aligned}
D_i D^i {\rm tr} \Psi
= {\rm tr} \left( \stackrel{\odot}J{}^{Ti} M_{[1]}\stackrel{\odot}J{}^iM_{[0]}^{-1} \right) \,,
\label{id:global-divergence}\end{aligned}$$ where we have also used the conservation equation $D_i J{}^i =0$. Then, integrating this divergence identity over the region $\Sigma=\{(\rho,z)|\rho\ge 0,\ -\infty<z<\infty \}$, we obtain the Mazur identity, $$\begin{aligned}
\int_{\partial \Sigma}\rho \partial_p {\rm tr} \Psi dS^p
= \int_{\Sigma}\rho \hat h_{pq}{\rm tr}
({\cal M}^{Tp} \: {\cal M}^q)d\rho dz \,,
\label{eq:id} \end{aligned}$$ where $\hat h_{pq}$ is the two-dimensional flat metric $$\begin{aligned}
\hat h=d\rho^2+dz^2 \,, \end{aligned}$$ and the matrix ${\cal M}$ is defined by $$\begin{aligned}
{\cal M}^p=\hat g_{[0]}^{-1}\: \stackrel{\odot}J{}^{Tp} \: \hat g_{[1]} \,. \end{aligned}$$
Now we note that the right-hand side of the identity, (\[eq:id\]), is non-negative. Therefore, if we impose the boundary conditions at $\partial\Sigma$, under which the left-hand side of Eq.(\[eq:id\]) vanishes, then we must have $\stackrel{\odot}J{}^i=0$. In that case, it follows from eq. (\[eq:deriv\]) that $\Psi$ must be a constant matrix over the region $\Sigma$. Therefore, in particular, if $\Psi$ is shown to be zero on some part of the boundary $\partial \Sigma$, it immediately follows that $\Psi$ must be identically zero over the base space $\Sigma$, implying that the two solutions $M_{[0]}$ and $M_{[1]}$ must coincide with each other. This is indeed the case under our boundary conditions discussed in the next section.
Boundary value problems {#sec:coincidence}
=======================
In this section, we derive necessary boundary conditions for determining the scalar fields $\Phi^A=(\lambda_{ab},\omega_a,\psi_a,\mu)$, requiring asymptotic flatness at infinity, regularity on the two rotation axes (i.e., the $\phi$-invariant plane and the $\psi$-invariant plane), and on the event horizon (of which cross-sections are assumed to be topologically spherical). Note that by [*asymptotically flat*]{}, we mean that the spacetime metric has the following fall off behavior at large distances, $$\begin{aligned}
ds^2 &\simeq& \left(-1+\frac{8M_{ADM}}{3\pi r^2}+{\cal O}(r^{-3})\right)dt^2
- \left(
\frac{8J_\phi\sin^2\theta}{\pi r^2}+{\cal O}(r^{-3})
\right)dtd\phi\nonumber\\
& &- \left(
\frac{8J_\psi\cos^2\theta}{\pi r^2}+{\cal O}(r^{-3})
\right)dtd\psi
\nonumber\\
& & + \left( 1 + {\cal O}(r^{-1}) \right)
\left(dr^2+
r^2\left(d\theta^2 + \sin^2\theta d\phi^2
+ \cos^2\theta d\psi^2 \right)
\right) \,,
\label{condi:AF}\end{aligned}$$ having the spherical spatial infinity, $S^3_\infty$. Here the constants $M_{ADM}$ and $J_a$ are the asymptotic conserved mass and angular momenta. Since we are concerned with stationary, axisymmetric spacetimes with Killing symmetries $\xi_I$, the conserved charges $M_{ADM}$ and $J_a$ are defined, respectively, by $$\begin{aligned}
M_{ADM} &=& -\frac{3}{32\pi}
\int_{S^3_\infty}
dS^{\mu \nu} \nabla_\mu (\xi_3)_\nu \,,
\label{def:Mass}
\\
J_a &=& \frac{1}{16\pi}\int_{S^3_\infty}
dS^{\mu \nu} \nabla_\mu (\xi_a)_\nu \,.
\label{def:Ja} \end{aligned}$$ We write below our boundary conditions for $\Phi^A$ in terms of the Weyl-Papapetrou coordinates. Therefore, in particular, relevant conditions at infinity—see below eqs. (\[gtt\]) – (\[grhorho\])—are derived from the above fall-off behavior, eq. (\[condi:AF\]), by the coordinate transformation $$\begin{aligned}
\rho = \frac{1}{2} r^2 \sin 2\theta \,, \quad
z = \frac{1}{2} r^2 \cos 2\theta \,. \end{aligned}$$ Then, we can find that the boundary conditions given in this section are, in fact, the same as the limiting behavior of $\Phi^A$ for the exact solution of Cvetič-Youm [@CY96] at the corresponding boundaries, which we discuss in Appendix \[sec:solution\].
In terms of the Weyl-Papapetrou coordinate system introduced in Sec. \[subsec:II-B\] and the rod-structure [@Harmark], the boundary $\partial \Sigma$ of the base space $\Sigma=\{(\rho,z)|\ \rho>0,\ -\infty<z<\infty \}$ is described as a set of three rods and the infinity: Namely,
1. the $\phi$-invariant plane: $\partial \Sigma_\phi=\{(\rho,z)|\rho=0,k^2<z<\infty \}$ with the rod vector $v=(0,1,0)$ ,
2. the horizon: $\partial \Sigma_{\cal H}=\{(\rho,z)|\ \rho=0,-k^2<z<k^2\}$ ,
3. the $\psi$-invariant plane: $\partial \Sigma_\psi=\{(\rho,z)|\rho=0,-\infty<z<-k^2\}$ with the rod vector $v=(0,0,1)$ ,
4. the infinity: $\partial \Sigma_\infty
= \{(\rho,z)|\sqrt{\rho^2+z^2}\to
\infty\ {\rm with}\ z/\sqrt{\rho^2+z^2}\ {\rm finite} \}$ ,
where here and hereafter ${\cal H}$ denotes a spatial cross-section of the event horizon. Accordingly, the boundary integral in the left-hand side of the Mazur identity, eq. (\[eq:id\]), is decomposed into the integrals over the three rods (i)–(iii), and the integral at infinity (iv), as $$\begin{aligned}
\int_{\partial \Sigma}\rho\partial_p{\rm tr}\Psi dS^p
&=& \int_{-\infty}^{-k^2}\rho\frac{\partial {\rm tr}\Psi }{\partial z}dz
+\int_{-k^2}^{k^2}\rho\frac{\partial {\rm tr}\Psi }{\partial z}dz
\nonumber\\
&& +\int_{k^2}^{\infty}\rho\frac{\partial {\rm tr}\Psi }{\partial z}dz
+\int_{\partial\Sigma_\infty}\rho\partial_a{\rm tr}\Psi dS^a \,.
\label{eq:integral} \end{aligned}$$
In order to evaluate this boundary integral, let us first consider the integrals of the twist one-forms $d\omega_a$ along the $z$-axis. By definition, the partial derivatives with respect to $z$ of the twist potentials $\omega_a$ vanish on both rotation axes. This means that the twist potentials $\omega_a$ are constant over the $\phi$-invariant plane and the $\psi$-invariant plane. Therefore, the integral can be written as $$\begin{aligned}
\int_{-\infty}^{\infty}\omega_{a,z}dz
&=& \int_{-k^2}^{k^2}\omega_{a,z}dz \\
&=& \biggl[\omega_a\biggr]_{z=-k^2}^{z=k^2} \,.
\label{eq:int} \end{aligned}$$ On the other hand, by Stokes’s theorem, the integral of $d\omega_a$ on the horizon is evaluated as $$\begin{aligned}
\int_{\partial\Sigma_{\cal H}}d\omega_a
&=& \int_{\partial\Sigma_\infty}d\omega_a\nonumber \\
&=& \int_{\partial\Sigma_\infty}V_a
+\int_{\partial\Sigma_\infty}\psi_a(3d\mu+\epsilon^{bc}\psi_bd\psi_c) \,.
\label{eq:int3}\end{aligned}$$ We find that the first integral in the right-hand side of eq. (\[eq:int3\]) is proportional to the angular momenta $J_a$, defined by eq. (\[def:Ja\]) above. As will be seen later, the second integral vanishes at infinity. Hence, using the degrees of freedom in adding a constant to $\omega_a$, we can always set the value of $\omega_a$ on the two rotation axes to be $$\begin{aligned}
\omega_a(z)=-\frac{2J_a}{\pi} \,, \label{eq:twist}\end{aligned}$$ for $z\in [k^2,\infty]$, and $$\begin{aligned}
\omega_a(z)=\frac{2J_a}{\pi} \,, \label{eq:twist2} \end{aligned}$$ for $z\in [-\infty,-k^2]$.
Next, consider the integral of $\mu_{,z}$ on the horizon $\partial \Sigma_{\cal H}$. The derivative of the potential, $d\mu$, vanishes on the two rotation axes by definition. Hence the integral along the $z$-axis becomes $$\begin{aligned}
\int^{\infty}_{-\infty}\mu_{,z}dz=\int^{k^2}_{-k^2}\mu_{,z}dz=\biggl[\ \mu\ \biggr]_{z=-k^2}^{z=k^2}.\end{aligned}$$ We find that this integral is proportional to the electric charge $Q$ defined by $$\begin{aligned}
Q&=&\frac{1}{16\pi}\int_{\cal H}\left(*F+\frac{1}{\sqrt{3}}A\wedge F\right) \,. \end{aligned}$$ In fact, straightforward calculation shows $$\begin{aligned}
&&\frac{1}{16\pi}\int_{\cal H}\left(*F+\frac{1}{\sqrt{3}}A\wedge F\right)\nonumber\\
&=&\frac{\pi}{4}\int^{k^2}_{-k^2}
\left[
\frac{\tau}{\rho}(A_{t,\rho}-a^\phi{}_tA_{\phi,\rho}
-a^\psi{}_tA_{\psi,\rho})
-\frac{1}{\sqrt{3}}(A_\phi A_{\psi,z}-A_\psi A_{\phi,z})
\right]dz
\nonumber \\
&=&\frac{\pi\sqrt{3}}{4}\int^{k^2}_{-k^2} \mu_{,z}dz \,. \end{aligned}$$ Hence, without loss of generality, $\mu$ can be set to be $$\begin{aligned}
\mu=-\frac{2Q}{\sqrt{3}\pi} \,,
\label{eq:charge1}\end{aligned}$$ for $\rho=0,\ z\in [-\infty,-k^2]$, and $$\begin{aligned}
\mu=\frac{2Q}{\sqrt{3}\pi} \,,
\label{eq:charge2} \end{aligned}$$ for $\rho=0,\ z\in [k^2,\infty]$.
Now we would like to show that the boundary integral, eq. (\[eq:integral\]), indeed vanishes under our preferable boundary conditions that require the regularity on the three rods and asymptotic flatness at infinity. For this purpose, in the following we evaluate the limiting behavior of the integrand, $\rho\ \partial_z{\rm tr}\ \Psi$, of eq. (\[eq:integral\]), separately on each boundary (i)–(iv).
\(i) $\phi$-invariant plane: $\partial\Sigma_\phi=\{(\rho,z)|\rho=0,\ k^2<z<\infty\}$. The regularity on the $\phi$-invariant plane requires that for $\rho\to 0$, the scalar fields behave as $$\begin{aligned}
&&\lambda_{\phi\phi}\simeq {\cal O}(\rho^2) \,,\label{eq:a1}\\
&&\lambda_{\psi\psi}\simeq {\cal O}(1) \,,\label{eq:a2}\\
&&\lambda_{\phi\psi}\simeq {\cal O}(\rho^2) \,, \label{eq:a3}\\
&&\omega_{\phi}\simeq -\frac{2J_\phi}{\pi}+{\cal O}(\rho^2) \,, \label{eq:a4}\\
&&\omega_{\psi}\simeq -\frac{2J_\psi}{\pi}+{\cal O}(\rho^2) \,,\label{eq:a5} \end{aligned}$$ and $$\begin{aligned}
&&\psi_{\phi}\simeq {\cal O}(\rho^2)\,, \label{eq:a6}\\
&&\psi_{\psi}\simeq {\cal O}(1) \,,\label{eq:a7}\\
&&\mu\simeq \frac{2Q}{\sqrt{3}\pi}+{\cal O}(\rho^2) \,, \label{eq:a8} \end{aligned}$$ where the boundary conditions, eqs. (\[eq:a1\])-(\[eq:a3\]) and eqs. (\[eq:a6\])-(\[eq:a7\]), come from the requirement that $\partial\Sigma_\phi$ is the $\phi$-invariant plane, i.e., the plane invariant under the rotation with respect to the axial Killing vector $\partial/\partial \phi$. The conditions, eqs. (\[eq:a4\])-(\[eq:a5\]), are derived from eq. (\[eq:twist\]). In the derivation of the condition (\[eq:a8\]), eq. (\[eq:charge2\]) is used. Then for two solutions, $M_{[0]}$ and $M_{[1]}$, with the same mass, the same angular momenta, and the same electric charge, $\rho\ {\rm tr}\Psi$ behaves as $$\begin{aligned}
\rho\ \partial_z{\rm tr}\ \Psi\simeq O(\rho) \,. \end{aligned}$$
\(ii) Horizon: $\partial \Sigma_{\cal H}=\{(\rho,z)|\ \rho=0, \ -k^2<z<k^2 \}$. The regularity on the horizon requires that for $\rho \to 0$, $$\begin{aligned}
&&\lambda_{ab}\simeq {\cal O}(1),\quad \omega_{a}\simeq {\cal O}(1) \,,\\
&&\psi_{a}\simeq {\cal O}(1), \quad \mu\simeq {\cal O}(1) \,. \end{aligned}$$ Therefore, for $\rho \to 0$, $\rho\ {\rm tr}\Psi$ behaves as $$\begin{aligned}
\rho\ \partial_z{\rm tr}\ \Psi\simeq O(\rho) \,. \end{aligned}$$
\(iii) $\psi$-invariant plane: $\partial\Sigma_\psi=\{(\rho,z)|\rho=0,\ -\infty<z<-k^2\}$. Similarly to the case (i), the regularity on the $\phi$-invariant plane requires $$\begin{aligned}
&&\lambda_{\psi\psi}\simeq {\cal O}(\rho^2) \, ,\label{eq:b1}\\
&&\lambda_{\phi\phi}\simeq {\cal O}(1) \,, \label{eq:b2}\\
&&\lambda_{\phi\psi}\simeq {\cal O}(\rho^2) \,, \label{eq:b3}\\
&&\omega_{\phi}\simeq \frac{2J_\phi}{\pi}+{\cal O}(\rho^2) \,, \label{eq:b4}\\
&&\omega_{\psi}\simeq \frac{2J_\psi}{\pi}+{\cal O}(\rho^2) \,, \label{eq:b5}\end{aligned}$$ and $$\begin{aligned}
&&\psi_{\phi}\simeq {\cal O}(1) \,, \label{eq:b6} \\
&&\psi_{\psi}\simeq {\cal O}(\rho^2) \,,\label{eq:b7} \\
&&\mu\simeq -\frac{2Q}{\sqrt{3}\pi}+{\cal O}(\rho^2) \,. \label{eq:b8}\end{aligned}$$ Therefore, for $\rho \to 0$, $\rho\ {\rm tr}\Psi$ behaves as $$\begin{aligned}
\rho\ \partial_z{\rm tr}\ \Psi \simeq O(\rho) \,. \end{aligned}$$
\(iv) Infinity: $\partial\Sigma_\infty=\{(\rho,z)|\ \sqrt{\rho^2+z^2}\to\infty$ with $z/\sqrt{\rho^2+z^2}$ finite $\}$. Recall that the three-dimensional metric $g=(g_{IJ})\ (I,J=t,\phi,\psi)$ is subject to the constraint $$\begin{aligned}
{\rm det}(g)=-\rho^2 \,. \end{aligned}$$ Therefore, using the constraint and the formula, $$\begin{aligned}
{\rm det}(g+\delta g)
&=&{\rm det}[g(1+g^{-1}\delta g)]
\nonumber \\
&=&-\rho^2\left(1+{\rm tr}(g^{-1}\delta g)+{\rm det}(g^{-1}\delta g)\right)
\nonumber \\
&\simeq&-\rho^2\left(1+{\rm tr}(g^{-1}\delta g)\right) \,, \end{aligned}$$ we can see in the next order that the metric has to satisfy the constraint $$\begin{aligned}
\sum_{I=t,\phi,\psi}\frac{\delta g_{II}}{g_{II}}=0 \,,\end{aligned}$$ which is the same constraint as in the vacuum case [@Harmark]. Then, the asymptotic flatness, eq. (\[condi:AF\]), requires that the limiting behavior of the metric be $$\begin{aligned}
&&g_{tt} \simeq -1 + \frac{4M_{ADM}}{3\pi}\frac{1}{\sqrt{\rho^2+z^2}}
+ {\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,,
\label{gtt}
\\
&&g_{t\phi} \simeq -\frac{J_\phi}{\pi}\frac{\sqrt{\rho^2+z^2}-z}{\rho^2+z^2}
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,,
\label{gtphi}
\\
&&g_{t\psi}\simeq - \frac{J_\psi}{\pi}\frac{\sqrt{\rho^2+z^2}+z}{\rho^2+z^2}
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,,
\label{gtpsi}
\\
&&\lambda_{\phi\phi}
\simeq (\sqrt{\rho^2+z^2}-z)
\left(1+\frac{2(M_{ADM}+\eta)}{3\pi\sqrt{\rho^2+z^2}}
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right)
\right) \,,
\label{gphiphi}
\\
&&\lambda_{\psi\psi}\simeq
(\sqrt{\rho^2+z^2}+z)
\left(1+\frac{2(M_{ADM}-\eta)}{3\pi\sqrt{\rho^2+z^2}}+{\cal O}
\left(\frac{1}{\rho^2+z^2}\right)
\right) \,,
\label{gpsipsi}
\\
&&\lambda_{\phi\psi}\simeq
\zeta\frac{\rho^2}{(\rho^2+z^2)^{3/2}}
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,,
\\
&&g_{\rho\rho}=g_{zz}\simeq \frac{1}{2\sqrt{\rho^2+z^2}}
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,,
\label{grhorho}\end{aligned}$$ where the constant $M_{ADM}$ denotes the conserved mass defined by eq. (\[def:Mass\]) and $J_\phi$ and $J_\psi$ the angular momenta, defined by eq. (\[def:Ja\]). Here $\eta$ is a constant that comes from gauge degrees of freedom in the choice of the coordinate $z$, i.e., degrees of freedom with respect to shift translation $z\to z+\alpha$. (This gauge freedom exists even after the gauge freedom of the conjugate coordinate, $\rho$, is fixed at infinity.) Since in our proof we choose the coordinate $z$ such that the horizons are located at the interval $[-k^2,k^2]$ for two configurations $M_{[0]}$ and $M_{[1]}$, we choose the same values of $\eta$ for the two solutions.
The left-hand side of the Einstein-Maxwell equation behaves as ${\cal O}((\rho^2+z^2)^{-1})$ in a neighborhood of the infinity. The energy-momentum tensor of the Maxwell field must also behave as ${\cal O}((\rho^2+z^2)^{-1})$. Hence from the asymptotic flatness, the gauge potential must behave as $$\begin{aligned}
&&A_t\simeq \frac{2Q}{\pi\sqrt{\rho^2+z^2}}
+ {\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,,
\label{eq:At} \\
&&\psi_\phi\simeq {\cal O}\left(\frac{1}{\sqrt{\rho^2+z^2}}\right) \,,
\label{eq:Aphi} \\
&&\psi_\psi \simeq {\cal O}\left(\frac{1}{\sqrt{\rho^2+z^2}}\right) \,.
\label{eq:Apsi} \end{aligned}$$
Next, we derive the behavior of $\mu$ and $\omega_a$ near infinity. The magnetic potential, $\mu$, is determined by eq. (\[eq:mu\]). From eqs. (\[eq:Aphi\]) and (\[eq:Apsi\]), the second term in the right-hand side of eq. (\[eq:mu\]) behaves as ${\cal O}((\rho^2+z^2)^{-1})$. The leading term $\mu^{(0)}$, where $\mu\simeq\mu^{(0)}+{\cal O}((\rho^2+z^2)^{-1/2})$, is derived from the equations $$\begin{aligned}
\mu_{,z}^{(0)}\simeq -\frac{\rho}{\sqrt{3}}A_{t,\rho} \,, \quad
\mu_{,\rho}^{(0)}\simeq \frac{\rho}{\sqrt{3}}A_{t,z} \,. \end{aligned}$$ Using the asymptotic behavior (\[eq:At\]) of the gauge field $A_t$, we obtain $$\begin{aligned}
\mu^{(0)}= \frac{2Qz}{\pi\sqrt{3}\sqrt{\rho^2+z^2}} \,. \end{aligned}$$ The twist potential, $\omega_a$, is determined by eq. (\[eq:twistpotential\]). The second term behaves as ${\cal O}((\rho^2+z^2)^{-1})$. Hence, the leading term $\omega_{a}^{(0)}$, where $\omega_a\simeq\omega_a^{(0)}+{\cal O}((\rho^2+z^2)^{-1/2})$, is derived from the equations $$\begin{aligned}
\omega_{a,z}^{(0)}&\simeq&\frac{\tau}{\rho}\lambda_{ab}a^b{}_{t,\rho} \,,
\label{eq:aa1}\\
\omega_{a,\rho}^{(0)}&\simeq&-\frac{\tau}{\rho}\lambda_{ab}a^b{}_{t,z} \,.
\label{eq:aa2}\end{aligned}$$ The functions $a^a{}_t$ behaves as $$\begin{aligned}
%&&\tau=\lambda_{\psi\phi}^2-\lambda_{\phi\phi}\lambda_{\psi\psi}\simeq -\rho^2,%\\
&&a^\phi{}_t=\frac{\lambda_{\phi\psi}g_{t\psi}
-\lambda_{\psi\psi}g_{t\phi}}{\tau}\simeq
-\frac{J_\phi}{\pi}\frac{1}{\rho^2+z^2} \,,
\\
&&a^\psi{}_t=\frac{\lambda_{\phi\psi}g_{t\phi}
-\lambda_{\phi\phi}g_{t\psi}}{\tau}
\simeq
-\frac{J_\psi}{\pi}\frac{1}{\rho^2+z^2} \,. \end{aligned}$$ Therefore, solving eqs. (\[eq:aa1\]) and (\[eq:aa2\]), we obtain $$\begin{aligned}
\omega_{\phi}^{(0)}&=& \frac{J_\phi}{\pi}
\left(\frac{\rho^2}{\rho^2+z^2}
-\frac{2z}{\sqrt{\rho^2+z^2}}
\right) \,, \\
\omega_{\psi}^{(0)}&=& \frac{J_\psi}{\pi}
\left(\frac{\rho^2}{\rho^2+z^2}
-\frac{2z}{\sqrt{\rho^2+z^2}}
\right) \,. \end{aligned}$$ Then, for $\sqrt{\rho^2+z^2} \to \infty$, $\rho\ {\rm tr}\Psi$ behaves as $$\begin{aligned}
\rho\ {\rm tr}\Psi
%&\simeq&\left[\left(\stackrel{\odot}{\lambda_{\phi\phi}^{(2)}}\right)^2+\left(\%stackrel{\odot}{\lambda_{\psi\psi}^{(2)}}\right)^2+\stackrel{\odot}{\lambda_{\p%hi\phi}^{(2)}}\stackrel{\odot}{\lambda_{\psi\psi}^{(2)}}\right]\frac{\rho}{2(\r%ho^2+z^2)}+\frac{6\left(\stackrel{\odot}{\mu^{(0)}}\right)^2}{\sqrt{\rho^2+z^2}%} +{\cal O}\left(\frac{1}{\rho^2+z^2}\right)\nonumber\\
&\simeq&{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,. \end{aligned}$$ Therefore, $$\begin{aligned}
\rho\ \partial_p{\rm tr}\Psi
dS^p\simeq {\cal O}\left(\frac{1}{\sqrt{\rho^2+z^2}}\right) \,.\end{aligned}$$
Thus, we find from (i)–(iv) that the boundary integral, eq. (\[eq:integral\]), vanishes on each rod and the infinity. The deviation matrix, $\Psi$, is constant and has the asymptotic behavior, $\Psi\to 0$. Therefore, $\Psi$ vanishes over $\Sigma$, and the two configurations, $M_{[0]}$ and $M_{[1]}$, coincide with each other. Furthermore, as shown in Appendix \[sec:solution\], the boundary conditions derived above are the same as the limiting behavior of the Cveti[č]{}-Youm solution with equal charges at each corresponding boundary. Therefore, the data $M_{[0]}$ (and now equivalently $M_{[1]}$) must also be the same as the corresponding matrix to the five-dimensional Cveti[č]{}-Youm solution with equal charges. This completes our proof for the uniqueness theorem.
Summary {#sec:conclusion}
=======
We have shown the uniqueness theorem which states that in five-dimensional Einstein-Maxwell-Chern-Simons theory, an asymptotically flat, stationary charged rotating black hole with finite temperature is uniquely specified by its asymptotic conserved charges and therefore is described by the five-dimensional Cveti[č]{}-Youm solution with equal charges, if (1) it admits two independent axial Killing symmetries and (2) the topology of the event horizon cross-section is spherical. Our theorem generalizes the uniqueness theorem for spherical black holes in five-dimensional vacuum Einstein gravity [@MI] to the case of EMCS theory. In our proof, in addition to the symmetry-assumption (1), the Chern-Simons term in the theory, eq. (\[eq:action\]), plays an important role to reduce the system into a non-linear sigma model with desired symmetry property, $G_{2(2)}/SO(4)$, as discussed in [@MO; @MO2]. Then, having this symmetry property on the target space, we have obtained the matrix representation of [@BCCGSW], in which our system is completely determined by $G_{2(2)}/SO(4)$ coset matrix $M$. We then derived the Mazur identity, and used the identity to show that if two solutions, i.e., two matrices, $M_{[0]}$ and $M_{[1]}$, satisfy the same boundary conditions (imposed at infinity, on two rotational axis, and on the horizon), then the solutions $M_{[0]}$ and $M_{[1]}$ must coincide with each other. We have shown that our boundary conditions (the asymptotic flatness and the regularity) are the same as the limiting behavior of the Cveti[č]{}-Youm solution.
In the present theorem, we restrict attention to topologically spherical black holes by the assumption (2). Our theorem can be generalized to the case of charged rotating black ring solutions by imposing certain additional conditions. We first note that under the same symmetry condition (1), the analysis in Sec. \[sec:metric\], \[sec:mazur\] and \[sec:coincidence\] apply also for black ring solutions (if exist) in EMCS theory. (See [@EEF] for such a ring solution.) The only difference from the spherical black hole case arises in the boundary value analysis. Now we also note that asymptotically flat, five-dimensional black ring solutions that satisfy the symmetry assumption (1) have the following rod structure: (i) $[c,\infty], \ v=(0,1,0)$, (ii) $[ck^2,c],\ v=(0,0,1)$, (iii) $[-ck^2,ck^2]$, and (iv) $[-\infty,-ck^2],\ v=(0,0,1)$, where $c>0, k^2<1$ and $v$’s are eigenvectors with respect to a zero eigenvalue of the three-dimensional matrix $g_{IJ}$ for each segment. It should be noted that we are not concerned with a lens space throughout discussion here, and therefore the only non-trivial rod data are given by rod intervals. Then, after fixing the scale $c$, one can completely specify the rod data in terms of $k^2$. The finite spacelike rod (ii) is the main difference from the rod structure for topologically spherical black holes considered in Sec. \[sec:coincidence\]. We believe that by appropriately specifying rod structure, one can determine the topology of the horizon, as well as the topology of black hole exterior region. In this respect, it has recently been shown [@HY08] that the topology and symmetry structure of the black hole spacetime can be completely determined in terms of [*rod-intervals*]{}, which is similar to but somewhat different from the rod-structure of Harmark [@Harmark]. In the charged black ring case, a dipole charge may also play a role. These issues deserve further study.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank P. Figueras for comments and discussions on possible generalization of the present results to the black ring case. We also would like to thank S. Hollands for discussions concerning the topology of the black hole exterior region. ST is supported by the JSPS under Contract No. 20-10616.
Maxwell field with symmetries {#sec:A}
=============================
Let $F$ denote the stationary and axisymmetric Maxwell field, i.e., that satisfies $$\begin{aligned}
{\cal L}_{\xi_I} F=0 \,, \end{aligned}$$ with $\xi_I\ (I=\phi,\psi,t)$ being commuting Killing vector fields for the axial-symmetries and the stationary symmetry, discussed in sec. \[subsec:II-B\]. From the Maxwell equation, $dF=0$, and the identity $$\begin{aligned}
di_{\xi_I}i_{\xi_J}&=&-i_{\xi_I}{\cal L}_{\xi_J}
+i_{\xi_J}{\cal L}_{\xi_I}+i_{\xi_I}i_{\xi_J}d \,, \end{aligned}$$ we have $$\begin{aligned}
di_{\xi_I}i_{\xi_J}F &=& -i_{\xi_I}{\cal L}_{\xi_J}F
+i_{\xi_J}{\cal L}_{\xi_I}F
+i_{\xi_I}i_{\xi_J}dF
\nonumber \\
&=&0 \,. \end{aligned}$$ Similarly, using the identity $$\begin{aligned}
di_{\xi_I}i_{\xi_J}i_{\xi_K}=i_{\xi_I}i_{\xi_J}{\cal L}_{\xi_K}-i_{\xi_I}i_{\xi_K}{\cal L}_{\xi_J}+i_{\xi_J}i_{\xi_K}{\cal L}_{\xi_I}-i_{\xi_I}i_{\xi_J}i_{\xi_K}d \,,\end{aligned}$$ we have $$\begin{aligned}
di_{\xi_I}i_{\xi_J}i_{\xi_K}*F
&=&i_{\xi_I}i_{\xi_J}{\cal L}_{\xi_K}*F
-i_{\xi_I}i_{\xi_K}{\cal L}_{\xi_J}*F
+i_{\xi_J}i_{\xi_K}{\cal L}_{\xi_I}*F
-i_{\xi_I}i_{\xi_J}i_{\xi_K}d*F
\nonumber \\
&=&i_{\xi_I}i_{\xi_J}*{\cal L}_{\xi_K}F
-i_{\xi_I}i_{\xi_K}*{\cal L}_{\xi_J}F
+i_{\xi_J}i_{\xi_K}*{\cal L}_{\xi_I}F
-i_{\xi_I}i_{\xi_J}i_{\xi_K}d*F
\nonumber \\
&=&\frac{1}{\sqrt{3}}i_{\xi_I}i_{\xi_J}i_{\xi_K} F\wedge F
\nonumber \\
&=& 0 \,. \end{aligned}$$ Therefore, $F(\xi_I,\xi_J)$ and $(*F)(\xi_I,\xi_J,\xi_K)$, are constant. Since they vanish, at least, on rotation axes, these imply $$\begin{aligned}
&&F(\xi_I,\xi_J)=0 \,,\\
&&(*F)(\xi_I,\xi_J,\xi_K)=0 \,.\end{aligned}$$ In terms of the coordinates $(t,\phi,\psi,\rho,z)$, these can be written as $$\begin{aligned}
&&F_{t\phi}=F_{t\psi}=F_{\phi\psi}=0 \,, \\
&&F_{\rho z}=0 \,. \label{eq:Frz}\end{aligned}$$ Then, from (\[eq:Frz\]), using the gauge degrees of freedom, $A \to A-d\chi$, with the function $\chi$ satisfying $A_\rho=\chi_{,\rho} \,, \quad A_\theta=\chi_{,\theta}$ we can show $$\begin{aligned}
A_\rho=A_z=0 \,. \end{aligned}$$
Cveti[č]{}-Youm solution with equal charges {#sec:solution}
===========================================
Here we present the asymptotically flat stationary charged rotating black hole solution in five-dimensional Einstein-Maxwell-Chern-Simons theory, found by Cvetič et al [@CY96]. The solution has three mutually commuting Killing vectors that generate isometries ${\Bbb R} \times U(1) \times U(1)$, and spherical topology of the horizon cross-sections. We observe that the limiting behavior of relevant scalar functions of the solution, which correspond to $\Phi^A$, are in perfect accordance with our general boundary conditions discussed in Sec. \[sec:coincidence\].
The metric and the gauge potential in [@CY96; @CLP; @CCLP] are given, respectively, by $$\begin{aligned}
ds^2&=&
-dt^2-\frac{2q}{\tilde \rho^2}\nu(dt-\omega)
+\frac{f}{\tilde\rho^4}(dt-\omega)^2
+\frac{\tilde \rho^2r^2}{\Delta}dr^2+\tilde\rho^2d\theta^2
\nonumber \\
&+&(r^2+a^2)\sin^2\theta d\phi^2+(r^2+b^2)\cos^2\theta d\psi^2 \,,
\label{eq:solution}\end{aligned}$$ and $$\begin{aligned}
A=\frac{\sqrt{3}q}{\tilde\rho^2}(dt-\omega) \,,\label{eq:gauge}\end{aligned}$$ where $$\begin{aligned}
&&\nu=b\sin^2\theta d\phi+a\cos^2\theta d\psi \,,\\
&&\omega=a\sin^2d\phi+b\cos^2\theta d\psi \,, \\
&&f=2m\tilde\rho^2-q^2 \,, \\
&&\Delta=(r^2+a^2)(r^2+b^2)+q^2+2ab q-2mr^2 \,, \\
&&\tilde\rho^2=r^2+a^2\cos^2\theta+b^2\sin^2\theta \, . \end{aligned}$$ The scalar fields $\Phi^A=(\lambda_{ab},\omega_a,\psi_a,\mu)$ for the solution (\[eq:solution\]) and (\[eq:gauge\]) are computed as $$\begin{aligned}
&&\lambda_{\phi\phi}=\frac{2q}{\tilde\rho^2}ab \sin^4\theta
+\frac{f}{\tilde\rho^4}a^2\sin^4\theta+(r^2+a^2)\sin^2\theta \,,
\\
&&\lambda_{\psi\psi}=\frac{2q}{\tilde\rho^2}ab \cos^4\theta
+\frac{f}{\tilde\rho^4}b^2\cos^4\theta+(r^2+b^2)\cos^2\theta \,,
\\
&&\lambda_{\phi\psi}=\frac{q}{\tilde\rho^2}(a^2+b^2)
\cos^2\theta\sin^2\theta+\frac{f}{\tilde\rho^4}ab\cos^2\theta\sin^2\theta \,,\end{aligned}$$ $$\begin{aligned}
\omega_\phi&=&\frac{(2am+bq)(-4\cos2\theta+\cos4\theta)}{8}
\nonumber\\
&&-\frac{2(a^2-b^2)(2aq^2+(2am+bq)F)\cos^2\theta\sin^4\theta}{F^2} \,,
\\
\omega_\psi&=&-\frac{(2bm+aq)(4\cos2\theta+\cos4\theta)}{8}
\nonumber\\
&&-\frac{2(a^2-b^2)(2bq^2+(2bm+aq)F)\cos^4\theta\sin^2\theta}{F^2} \,,\end{aligned}$$ $$\begin{aligned}
\psi_\phi&=&-\frac{qa\sin^2\theta}{\tilde\rho^2} \,,\\
\psi_\psi&=&-\frac{qb\cos^2\theta}{\tilde\rho^2} \,, \\
\mu&=&\frac{1}{2}q\cos2\theta-\frac{2(b^2-a^2)q\cos^2\theta\sin^2\theta}{F} \,,\end{aligned}$$ where the function $F$ is defined by $$\begin{aligned}
F=a^2+b^2+2r^2+(a^2-b^2)\cos2\theta \,. \end{aligned}$$
Let us introduce the coordinates $(\rho,z)$ defined by $$\begin{aligned}
\rho=\frac{1}{2}\sqrt{\Delta}\sin2\theta \,,\quad
z =\frac{2r^2+a^2+b^2-2m}{4}\cos2\theta \,. \end{aligned}$$ Then, the base space $\Sigma=\{(\rho,z)|\ \rho>0,\ -\infty<z<\infty \}$ has four boundaries, which exactly correspond to the four boundaries discussed in Sec. \[sec:coincidence\]: Namely, (i) $\phi$-invariant plane, i.e., the plane which is invariant under the rotation with respect to the Killing vector field $\partial/\partial \phi$: $\partial \Sigma_\phi=\{(\rho,z)|\rho=0,k^2<z<\infty \}$, (ii) Horizon: $\partial \Sigma_{\cal H}=\{(\rho,z)|\ \rho=0,-k^2<z<k^2\}$, (iii) $\psi$-invariant plane, i.e., the plane which is invariant under the rotation with respect to the Killing vector field $\partial /\partial \psi$: $\partial \Sigma_\psi=\{(\rho,z)|\rho=0,-\infty<z<-k^2\}$, and (iv) Infinity: $\partial \Sigma_\infty=\{(\rho,z)|\sqrt{\rho^2+z^2}\to\infty
\ {\rm with}\ z/\sqrt{\rho^2+z^2}\ {\rm finite} \}$, where the constant $k^2$ is given by $$\begin{aligned}
k^2=\frac{\sqrt{(2m-a^2-b^2)^2-4(ab+q)^2}}{4} \,. \end{aligned}$$
Let us examine the behavior of the scalar fields on each boundary.
\(i) Near the [$\phi$-invariant plane]{} $\partial \Sigma_\phi$, each scalar field behaves as $$\begin{aligned}
&&\lambda_{\phi\phi}\simeq {\cal O}(\rho^2)\,, \quad
\lambda_{\psi\psi}\simeq {\cal O}(1) \,, \quad
\lambda_{\phi\psi}\simeq {\cal O}(\rho^2) \,,
\\
&& \omega_\phi \simeq -\frac{3}{8}(2am+bq)
+ {\cal O}(\rho^2),\, \quad
\omega_\psi \simeq -\frac{5}{8}(2bm+aq)+{\cal O}(\rho^2) \,,
\\
&& \psi_\phi \simeq {\cal O}(\rho^2) \,, \quad
\psi_\psi \simeq {\cal O}(1) \,, \quad
\mu \simeq \frac{1}{2}q+{\cal O}(\rho^2) \,.\end{aligned}$$
\(ii) Near the horizon $\partial \Sigma_{\cal H}$, the scalar fields behave as $$\begin{aligned}
&&\lambda_{ab} \simeq {\cal O}(1) \,, \quad
\omega_{a} \simeq {\cal O}(1) \,, \\
&&\psi_{a} \simeq {\cal O}(1) \,, \quad
\mu \simeq {\cal O}(1) \,.\end{aligned}$$
\(iii) Near the [$\psi$-invariant plane]{} $\partial \Sigma_\psi$, each potential behaves as $$\begin{aligned}
&&\lambda_{\phi\phi}\simeq {\cal O}(1) \,, \quad
\lambda_{\psi\psi}\simeq {\cal O}(\rho^2) \,, \quad
\lambda_{\phi\psi}\simeq {\cal O}(\rho^2) \,,
\\
&&\omega_\phi \simeq \frac{5}{8}(2am+bq)+{\cal O}(\rho^2)\,, \quad
\omega_\psi \simeq\frac{3}{8}(2bm+aq)+{\cal O}(\rho^2)\,,
\\
&&\psi_\phi \simeq {\cal O}(1) \,, \quad
\psi_\psi \simeq {\cal O}(\rho^2) \,, \quad
\mu \simeq -\frac{q}{2}+{\cal O}(\rho^2) \,. \end{aligned}$$
\(iv) In the neighborhood of infinity $\partial\Sigma_\infty$, the behavior of the potentials becomes $$\begin{aligned}
&&\lambda_{\phi\phi}\simeq
(\sqrt{\rho^2+z^2}-z)
\left(1+\frac{a^2}{2\sqrt{\rho^2+z^2}}\right)
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,, \\
&&\lambda_{\psi\psi}\simeq
(\sqrt{\rho^2+z^2}+z)
\left(1+\frac{2m-a^2}{2\sqrt{\rho^2+z^2}}\right)
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,, \\
&&\lambda_{\psi\phi}\simeq
\frac{(a^2q+b^2q+2abm)\rho^2}{8(\rho^2+z^2)^{3/2}}
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,, \\
&& \omega_{\phi}\simeq
\frac{1}{8}(2am+bq)(-4\cos2\theta+\cos4\theta)
+{\cal O}\left(\frac{1}{\sqrt{\rho^2+z^2}}\right) \,, \\
&&\omega_{\psi}\simeq-\frac{1}{8}(2bm+aq)(4\cos2\theta+\cos4\theta)
+{\cal O}\left(\frac{1}{\sqrt{\rho^2+z^2}}\right) \,, \\
&&\psi_\phi\simeq
-\frac{qa(\sqrt{\rho^2+z^2}-z)}{4(\rho^2+z^2)}
+ {\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,,\\
&&\psi_\psi\simeq
-\frac{qb(\sqrt{\rho^2+z^2}+z)}{4(\rho^2+z^2)}
+{\cal O}\left(\frac{1}{\rho^2+z^2}\right) \,, \\
&&\mu \simeq
\frac{qz}{2\sqrt{\rho^2+z^2}}
+{\cal O}\left(\frac{1}{\sqrt{\rho^2+z^2}}\right) \,. \end{aligned}$$
0$$\begin{aligned}
&&\lambda_{00}\simeq r^2\sin^2\theta\left(1+\frac{a^2}{r^2}\right),\\
&&\lambda_{01}\simeq \frac{a^2q+b^2q+2abm}{r^2},\\
&&\lambda_{11}\simeq r^2\cos^2\theta\left(1+\frac{b^2}{r^2}\right),\\
&&\psi_0\simeq \frac{-q\sin^2\theta}{r^2},\\
&&\psi_1\simeq \frac{-q\cos^2\theta}{r^2},\\
&&\omega_0\simeq-\frac{1}{8}(2am+bq)(-4\cos2\theta+\cos4\theta),\\
&&\omega_1\simeq-\frac{1}{8}(2bm+aq)(-4\cos2\theta+\cos4\theta),\\
&&\mu\simeq -\frac{1}{2}q\cos\theta.\end{aligned}$$
[99]{}
J.P. Gauntlett, J.B. Gutowski, C.M. Hull, S. Pakis and H.S. Reall, Class. Quant. Grav. [**20**]{}, 4587 (2003). J.C. Breckenridge, R.C. Myers, A.W. Peet and C. Vafa, Phys. Lett. B [**391**]{}, 93 (1997).
M. Cvetič and D. Youm, Nucl. Phys. B [**476**]{}, 118 (1996).
M. Cvetič, H. Lü and C.N. Pope, Phys. Lett B [**598**]{}, 273 (2004).
Z.-W. Chong, M. Cvetič, H. Lü and C.N. Pope, Phys. Rev. Lett. [**95**]{} 161301 (2005).
I. Bena, P. Kraus and R. Warner, Phys. Rev. D [**72**]{}, 084019 (2005). H. Elvang, R. Emparan and P. Figueras, JHEP [**0502**]{}, 031 (2005). H. Elvang, R. Emparan, D. Mateos and H. S. Reall, Phys. Rev. Lett. [**93**]{}, 211302 (2004). H. Elvang, R. Emparan, D. Mateos and H. S. Reall, Phys. Rev. D [**71**]{}, 024033 (2005). H. Elvang, R. Emparan, D. Mateos and H. S. Reall, JHEP [**08**]{}, 042, (2005). D. Gaiotto, A. Strominger and X. Yin, JHEP [**02**]{}, 023, (2006).
H. Ishihara and K. Matsuno, Prog. Theor. Phys. [**116**]{}, 417 (2006). H. Ishihara, M. Kimura, K. Matsuno and S. Tomizawa, Class. Quant. Grav. [**23**]{}, 6919 (2006). H. Ishihara, M. Kimura, K. Matsuno and S. Tomizawa, Phys. Rev. D [**74**]{}, 047501 (2006). T. Nakagawa, H. Ishihara, K. Matsuno and S. Tomizawa, Phys. Rev. D [**77**]{}, 044040 (2008). S. Tomizawa, H. Ishihara, K. Matsuno and T. Nakagawa, e-Print: arXiv:0803.3873 \[hep-th\]; to be publihed in Prog. Theor. Phys. K. Matsuno, H. Ishihara, T. Nakagawa and S. Tomizawa, Phys. Rev. D [**78**]{}, 064016 (2008). S. Tomizawa and A. Ishibashi, Class. Quant. Grav. [**25**]{}, 245007 (2008). S. Tomizawa, Y. Yasui and Y. Morisawa, e-Print: arXiv:0809.2001 \[hep-th\]. D. V. Gal’tsov and N. G. Scherbluk, e-Print: arXiv:0812.2336 \[hep-th\]. S. Tomizawa, H. Ishihara, M. Kimura and K. Matsuno, Class. Quant. Grav. [**24**]{}, 5609 (2007). S. Tomizawa, Class. Quant. Grav. [**25**]{}, 145014 (2008).
V. A. Belinsky and V. E. Zakharov, Sov. Phys. JETP [**48**]{}, 985 (1978);\
V. A. Belinsky and V. E. Zakharov, Sov. Phys. JETP [**50**]{}, 1 (1979). V. A. Belinski and E. Verdaguer, [*Gravitational Solitons*]{}, (Cambridge University Press, 2001).
S. Tomizawa, Y. Morisawa and Y. Yasui, Phys. Rev. D [**73**]{}, 064009 (2006). S. Tomizawa and M. Nozawa, Phys. Rev. D [**73**]{}, 124034 (2006). S. Tomizawa, H. Iguchi and T. Mishima, Phys. Rev. D [**74**]{}, 104004 (2006). S. Tomizawa, H. Iguchi and T. Mishima, Phys. Rev. D [**78**]{}, 084001 (2008). H. Iguchi and T. Mishima and S. Tomizawa, Phys. Rev. D [**76**]{}, 124019 (2007). A. Pomeransky, Phys. Rev. D [**73**]{}, 044004 (2006). T. Koikawa, Prog. Theor. Phys. [**114**]{}, 793 (2005). T. Azuma and T. Koikawa, Prog. Theor. Phys. [**116**]{}, 319 (2006). H. Iguchi and T. Mishima, Phys. Rev. D [**73**]{}, 121501 (2006).
J. Evslin and C. Krishnan, e-Print: arXiv:0706.1231 \[hep-th\]
J. Castejon-Amenedo and V. S. Manko, Phys. Rev. D [**41**]{}, 2018 (1990). A. Bouchareb, G. Clement, C-M Chen, D. V. Gal’tsov, N. G. Scherbluk and T. Wolf. Phys. Rev. D [**76**]{},104032 (2007); Erratum-ibid. D [**78**]{}, 029901 (2008). T. Mishima and H. Iguchi, Phys. Rev. D [**73**]{}, 044030 (2006).
S.S. Yazadjiev, Phys. Rev. D [**73**]{}, 104007 (2006).
S.S. Yazadjiev, JHEP 0607 (2006) 036 S.S. Yazadjiev, Phys. Rev. D [**78**]{}, 064032 (2008).
R. C. Myers and M. J. Perry, Annals Phys. [**172**]{}, 304 (1986). R. Emparan and H. S. Reall, Phys. Rev. Lett. [**88**]{}, 101101 (2002). A.A. Pomeransky and R.A. Sen’kov, e-Print: arXiv:hep-th/0612005. H. Iguchi and T. Mishima, Phys. Rev. D [**75**]{}, 064018 (2007). H. Elvang and P. Figueras, JHEP [**0705**]{}, 050 (2007). K. Izumi, Prog. Theor. Phys. [**119**]{}, 757 (2008). H. Elvang and M. J. Rodriguez, JHEP [**0804**]{} ,045 (2008).
G. W. Gibbons, D. Ida and T. Shiromizu, Prog. Theor. Phys. Suppl. [**148**]{}, 284 (2002);\
G. W. Gibbons, D. Ida and T. Shiromizu, Phys. Rev. Lett. [**89**]{}, 041101 (2002). F. R. Tangherlini, Nuovo. Cimento, [**27**]{}, 636 (1963). Y. Morisawa and D. Ida, Phys. Rev. D [**69**]{}, 124005 (2004).
S. Hollands and S. Yazadjiev, Commun. Math. Phys. [**283**]{}, 749 (2008). Y. Morisawa, S. Tomizawa and Y. Yasui, Phys. Rev. D [**77**]{}, 064019 (2008). S. Hollands and S. Yazadjiev, Class. Quantum Grav. [**25**]{}, 095010 (2008).
M. Rogatko, Class. Quant. Grav. [**19**]{}, 875 (2002). M. Rogatko, Class. Quant. Grav. [**19**]{}, L151 (2002). M. Rogatko, Phys. Rev. D [**67**]{}, 084025 (2003). M. Rogatko, Phys. Rev. D [**70**]{}, 044023 (2004). M. Rogatko, Phys. Rev. D [**70**]{}, 084025 (2004). M. Rogatko, Phys. Rev. D[**73**]{}, 124027 (2006). M. Rogatko, Phys. Rev. D [**77**]{}, 124037 (2008).
S. Hollands and S.S. Yazadjiev, Arxiv: 0812.3036 \[gr-qc\]
S. Mizoguchi and N. Ohta, Phys. Lett. B [**441**]{}, 123 (1998). S. Mizoguchi and G. Schröder, Class. Quan. Grav, [**17**]{}, 835 (2000).
S. Hollands, A. Ishibashi and R. M. Wald, Commun. Math. Phys. [**271**]{}, 699 (2007).
V. Moncrief and J. Isenberg, Class. Quan. Grav, [**25**]{}, 195015 (2008). S. Hollands and A. Ishibashi, e-Print: arXiv:0809.2659\[gr-qc\].
H.S. Reall, Phys. Rev. D [**68**]{}, 024024 (2003).
I. Racz and R.M. Wald, Class. Quantum Grav. [**9**]{} 2643-2656 (1992).
I. Racz and R.M. Wald, Class. Quantum Grav. [**13**]{} 539-552 (1996).
H.K. Kunduri and J. Lucietti, arXiv:0812.1576
H.K. Kunduri and J. Lucietti, arXiv:0806.2051
P. Figueras, H.K. Kunduri, J. Lucietti, and M. Rangamani, Phys. Rev. D [**78**]{} 044042 (2008).
H.K. Kunduri, J. Lucietti, and H.S. Reall, Class. Quantum Grav. [**24**]{}, 4169-4190 (2007).
D. Astefanesei, K. Goldstein, R. P. Jena, A. Sen, and S. P. Trivedi, JHEP 0610: 058 (2006).
D. Astefanesei, H. Yavartanoo Nucl. Phys. B [**794**]{} 13 (2008).
M. I. Cai and G. J. Galloway, Class Quant. Grav. [**18**]{}, 2707 (2001). C. Helfgott, Y. Oz and Y. Yanay, JHEP 0602, 025 (2006). G. J. Galloway and R. Schoen, Commun. Math. Phys. [**266**]{}, 571 (2006).
T. Harmark, Phys. Rev. D [**70**]{}, 124002 (2004); T. Harmark and P. Olesen, Phys. Rev. D [**72**]{}, 124017 (2005).
R. Emparan and H. S. Reall, Phys. Rev. D [**65**]{}, 084025 (2002).
J. Kunz and F. Navarro-Lerida, Phys. Rev. Lett. [**96**]{} 081101, (2006).
B. Carter, Phys. Rev. Lett. [**26**]{}, 331 (1971); P.T. Chrusciel, arXive:0812.3424 \[gr-qc\].
G.L. Bunting, PhD Thesis, (1983) Univ. of New England, Armidale, N.S.W.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A general experimental technique for high resolution studies of nucleon transfer reactions using radioactive beams is briefly described, together with the first new physics results that have been obtained with the new TIARA array. These first results from TIARA are for the reaction $^{24}$Ne(d,p)$^{25}$Ne, studied in inverse kinematics with a pure radioactive beam of $10^5$ pps from the SPIRAL facility at GANIL. The reaction probes the energies of neutron orbitals relevant to very neutron rich nuclei in this mass region and the results highlight the emergence of the N=16 magic number for neutrons and the associated disappearance of the N=20 neutron magic number for the very neutron rich neon isotopes.'
address: |
$^1$Department of Physics, University of Surrey, Guildford, Surrey GU2 7XH, UK\
$^2$CCLRC Daresbury Laboratory, Daresbury, Warrington WA4 4AD, UK\
$^3$University of Paisley, Paisley, Scotland PA1 2BE, UK\
$^4$LPC, IN2P3-CNRS, ISMRA and Université de Caen, F-14050 Caen, France\
$^5$IFC, CSIC-Universidad de Valencia, E-46071 Valencia, Spain\
$^6$Department of Physics, The University of Liverpool, Liverpool L69 7ZE, UK\
$^7$GANIL, BP 55027, 14076 Caen Cedex 5, France\
E-mail: [email protected]
author:
- 'W. N. CATFORD$^1$, C. N. TIMIS$^1$, R. C. LEMMON$^2$, M. LABICHE$^3$, N. A. ORR$^4$, L. CABALLERO$^5$, R. CHAPMAN$^3$, M. CHARTIER$^6$, M. REJMUND$^7$, H. SAVAJOLS$^7$ and the TIARA COLLABORATION'
title: 'Neutron Transfer Studied with a Radioactive beam of $^{24}$Ne, using TIARA at SPIRAL'
---
Introduction {#sec:Intro}
============
A great hope for the future in radioactive beam experiments is to be able to map out the changing shell structure for very exotic nuclei, away from stability, where this arises from effects such as the monopole migration of orbital energies and the changes brought about by alterations in the surface environment and spin-orbit splitting[@grawe]. Nucleon transfer reactions such as (d,p), (p,d) etc. are an established means of populating and studying nuclear levels that have a substantial single-particle structure. The development of techniques to use such reactions with radioactive beams, across a wide range of beam energies and masses and with high energy resolution, will open the way to exploit transfer across new regions of the nuclear chart and hence to study the new nuclear structure effects that evolve.
The technique that is described here, and implemented via the new TIARA array used in association with the VAMOS spectrometer and the EXOGAM gamma-ray array, is designed to achieve excitation energy resolution of better than 20-40 keV in the final nucleus. This is an order of magnitude better than can be achieved in a reasonable experimental setup that uses charged-particle observations only[@winfield]. The complete kinematical detection of the binary transfer reaction products specifies the reaction channel cleanly, where the identification of the heavy (beam-like) particle at least in Z is required, and the light (target-like)-particle detection allows angular distributions to be measured for any mass of projectile.
The present paper updates and extends results of the analysis in progress, reported elsewhere[@nustar; @enam].
The TIARA Array
===============
The requirement to use inverse kinematics in order to study nucleon transfer reactions, induced on radioactive species by protons and deuterons, imposes certain rather general requirements on the detection system to be used. The kinematics turns out to lead to particular reactions always appearing in the same characteristic range of laboratory angles and with similar energies, regardless of the mass or velocity of the incident beam[@rnb; @enam]. This allows a general purpose transfer apparatus to be designed.
The design philosophy and detailed description of TIARA has been discussed elsewhere[@nustar; @denton] but briefly the aim was to surround the target with a charged particle array that approached $4\pi$ coverage, with reasonable energy measurment and an angular resolution of 1 or 2 degrees. This array needed to be very compact so that a high gamma-ray efficiency of $>$15% (at 1.3 MeV) could be achieved, whilst avoiding the exposure of gamma-ray detectors to decay radiation from beam particles scattered in the forward 40$^\circ$. In the present setup, the TIARA array covers 82% of $4\pi$ with active silicon and the gamma-ray detectors are in a close cube geometry and thus subtend 67% of $4\pi$. The setup is mounted in front of a magnetic spectrometer which is used to separate physically the direct beam and the transfer reaction products, after the target. The region around the target is shown in fig. \[TIARA-pic\].
Experimental Details
====================
The TIARA system was set up in front of the VAMOS spectrometer at GANIL[@vamos], which was operated in dispersive mode at zero degrees. Direct beam was intercepted just in front of the focal plane detectors. The support frame and four detectors of EXOGAM[@exogam] surrounded the TIARA chamber. All events in which a particle was detected in TIARA were recorded. The gamma ray parameters were recorded via the VAMOS acquisition system and events were correlated with TIARA in real time via an event stamping method developed at GANIL.
An isotopically pure beam of $^{24}$Ne was supplied at 10 MeV/nucleon after reacceleration in the CIME cyclotron connected to the SPIRAL facility at GANIL. The beam intensity of $10^5$ pps was a factor of two lower than the maximum due to a limitation placed on the emittance, which limited the beam spot on target to a diameter of approximately 2mm base width. The target was 1 mg/cm$^2$ of CD$_2$ self supporting on a thin 25mm diameter frame.
A test experiment was performed with a stable beam of $^{14}$N at similar intensity and beam quality, in order to verify that normal kinematics (d,p) results from the literature could be reproduced with the TIARA setup. Good agreement was found[@labiche].
Results
=======
The isotopic identification for beam-like particles recorded at the focal plane of VAMOS is shown in fig.\[PID\]. This is derived from measured $\Delta E$, $E$ and time-of-flight parameters plus a ray tracing calculation that used the horizontal and vertical angles and positions measured at the focal plane. The ray tracing algorithm employed a neural network that was trained using a set of theoretically calculated rays obtained by numerical integration of their trajectories through VAMOS[@cozmin] and this gives results identical to an algebraic algorithm developed at GANIL.
=4.0in
By selecting the $^{24}$Ne ions in fig. \[PID\], the scattered deuterons recorded in TIARA could be analysed. The $^{24}$Ne momentum changes sufficiently quickly with scattering angle that very forward scattered elastics can still avoid the beam stop. The energy of the deuterons changes rapidly with their angle[@rnb] and, by using energy cuts, the elastic angular distribution could be extracted (see fig. \[elastics\]). A good fit was obtained using the optical potential measured for d+$^{26}$Mg at a similar energy[@meurders]. The normalisation obtained using these elastic data allowed absolute transfer cross sections to be extracted with confidence.
=4.1in
From the measured energy and angle recorded for protons from the (d,p) reaction to make $^{25}$Ne (cf. ref. [@nustar]) the excitation energy spectrum for states in $^{25}$Ne could be deduced. Different peaks in this spectrum could be used to gate the spectrum of associated gamma rays. Example data are included in fig, \[gammas\]. An important result of this analysis was that the excitation energies of the populated states could be fixed with an accuracy of order 30 keV. The limiting factor in this accuracy was the poor statistics of the gamma ray spectrum. This was in fact severely compromised in the present experiment by an intermittent fault in an electronic discriminator unit, and the eventual aim in this type of experiment will be to use individual gamma ray peaks to apply gates in the analysis. In the present case, however, it was still of vital importance that the gamma ray data could fix the energies and the number of peaks to be fitted to the (poorer resolution) excitation energy spectra derived from the particle energies. These fits are shown in the inset of fig. \[gammas\]. The data are just sufficient to allow limited gamma-gamma coincidence analysis. In the case of the state near 4 MeV it can be seen that the 1.7 MeV and 2.4 MeV gamma rays seen in its decay (fig. \[gammas\](b)) appear to be in coincidence (fig. \[gammas\] (c) and (d)).
=4.1in
The excitation energy spectrum (derived from the proton energy and angle) was fitted with 5 peaks. The widths of these peaks depended on the experimental resolution of the system, and this was in turn dependent on the detection angle of the proton. Thus, the data were binned for angular regions chosen to be 8$^\circ$ wide in order to give sufficient statistics for fitting. In order to fit the angular distributions, different optical model potentials were investigated, taken from (d,p) reactions on neighbouring nuclei $^{26}$Mg[@meurders] and $^{22}$Ne[@howard]. A systematic comparison with adiabatic calculations according to the prescription of Johnson and Soper[@johnson] was also performed. The adiabatic analysis was adopted for the extraction of spectroscopic factors, which were determined by normalising the theoretical curve to the data for each state, with particular emphasis placed on the data for the smallest center of mass angles (closest to 180$^\circ$ in the laboratory)[@tsang].
The results of the analysis are included in Table \[specfacs\]. The identifications of the spins are discussed below. In general, spectroscopic factors extracted in this fashion have an uncertainty of order 20% arising from the assumptions in the reaction theory, and this is the dominant source of uncertainty in the quoted results.
------------------------------ ------------------- -------------------- --------- --------- ------ -------------------
\[-1.5ex\] E$_{\rm x}$ (MeV) E$_{\rm x}$ (MeV) $\ell$ ( $\hbar$ ) $J^\pi$ $S$ $S$ E$_{\rm x}$ (MeV)
present previous transfer present USD USD
\[1ex\]
\[-1.5ex\] 0 0 0 1/2$^+$ 0.80 0.63 0
\[1ex\] 1.680 1.703 2 5/2$^+$ 0.15 0.10 1.779
\[1ex\] 2.030 - 2 3/2$^+$ 0.44 0.49 1.687
\[1ex\] 3.330 - 1 3/2$^-$ 0.75 - -
\[1ex\] 4.030 - (3) 7/2$^-$ 0.73 - -
\[1ex\]
------------------------------ ------------------- -------------------- --------- --------- ------ -------------------
: Results for states in $^{25}$Ne identified as being populated in neutron transfer on $^{24}$Ne. Previous E$_{\rm x}$ is from Reed [*et al.*]{} and USD refers to a $1s0d$ shell model calculation.
\[specfacs\]\
Discussion
==========
The key feature emerging from Table \[specfacs\] is that the state identified as the first 3/2$^+$ state in $^{25}$Ne, which reflects most directly the single particle energy of the $0d_{3/2}$ shell model orbital, is substantially higher than predicted. The identification rests on both the relative strength of this “particle” state compared to the $0d_{5/2}$ “hole” state and the observed gamma decay pathways. The shift of order 350 keV is presumably due to matrix elements in the USD effective interaction that are not well determined from data on less neutron rich nuclei. The shift can be understood very naturally in the monopole shift picture[@otsuka; @grawe], in which the emptying of the $d_{5/2}$ proton orbital in the more neutron rich N=15 isotones removes an attractive interaction that lowers the neutron $0d_{3/2}$ energy for nuclei closer to stability. This tends to make N=16 a magic number for neutron rich nuclei. Simultaneously, the gap to the negative parity orbitals $0f_{7/2}$ and $1p_{3/2}$ is reduced and N=20 loses its magicity[@utsuno].
The state identified as the 5/2$^+$ is almost certainly the state seen in beta decay[@reed] at 1.703 MeV and has also been seen recently in neutron knockout from $^{26}$Ne[@terry]. This latter observation also supports the identification of the 1.703 MeV level as the hole state and the newly observed level at 2.03 MeV as the 3/2$^+$ state. The further implications of these results are still under investigation.
Acknowledgments {#acknowledgments .unnumbered}
===============
We acknowledge with thanks the support of the GANIL and LPC Caen technicians, during the installation and commissioning of the TIARA array. Mr Geoffrey Moores (University of Paisley) and the Daresbury design staff are thanked for their vital contributions. This work was supported in the UK by EPSRC grants held at Surrey, Paisley, Daresbury and Birmingham.
[0]{} H. Grawe [*et al.*]{}, [*Eur. Phys. J.*]{} [**A**]{}, DOI:10.1140/epjad/i2005-06-025-1 (2005).
J.S. Winfield, W.N. Catford, N.A. Orr, [*Nucl. Instr. Meths.*]{} [**A396**]{}, 147 (1997).
W.N. Catford [*et al.*]{}, [*J. Phys.*]{} [**G**]{}, in press.
W.N. Catford [*et al.*]{}, [*Eur. Phys. J.*]{}, DOI: 10.1140/epjad/i2005-06-171-4 (2005).
W.N. Catford [*et al.*]{}, [*Nucl. Phys.*]{} [**A701**]{}, 1 (2002).
W.N. Catford [*et al.*]{}, [*AIP Conference Proc.*]{} [**704**]{}, 185 (2004).
H. Savajols [*et al.*]{}, [*Nucl. Instr. Meths.*]{} [**B204**]{}, 146 (2003).
J. Simpson [*et al.*]{}, [*Acta Physica Hungaria: Heavy Ions*]{} [**11**]{}, 159 (2000).
M. Labiche [*et al.*]{}, [*J. Phys.*]{} [**G**]{}, in press.
C.N. Timis, University of Surrey.
F. Meurders and A. Van Der Steld, [*Nucl. Phys.*]{} [**A230**]{}, 317 (1974).
A.J. Howard, J.G. Pronko, C.A. Whitten, Jr., [*Nucl. Phys.*]{} [**A152**]{}, 317 (1970).
R.C. Johnson and P.J.R. Soper, [*Phys. Rev.*]{} [**C1**]{}, 976 (1970).
X.D. Liu [*et al.*]{}, [*Phys. Rev.*]{} [**C69**]{}, 064313 (2004).
T. Otsuka [*et al.*]{}, [*Prog. Part. Nucl. Phys.*]{} [**47**]{}, 319 (2001); [*Phys. Rev. Lett.*]{} [**87**]{}, 082502 (2001).
Y. Utsuno [*et al.*]{}, [*Phys. Rev.*]{} [**C64**]{}, 011301R (2001); [**C70**]{}, 044307 (2004).
A.T. Reed [*et al.*]{}, [*Phys. Rev.*]{} [**C60**]{}, 024311 (1999).
J.R. Terry and J.L. Lecouey, [*Nucl. Phys.*]{} [**A734**]{}, 469 (2004).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We study average case approximation of Euler and Wiener integrated processes of $d$ variables which are almost surely $r_k$-times continuously differentiable with respect to the $k$-th variable and $0\le r_k\le r_{k+1}$. Let $n(\eps,d)$ denote the minimal number of continuous linear functionals which is needed to find an algorithm that uses $n$ such functionals and whose average case error improves the average case error of the zero algorithm by a factor $\eps$. Strong polynomial tractability means that there are nonnegative numbers $C$ and $p$ such that $$n(\eps,d)\le C\eps^{-p}\ \ \ \ \
\mbox{for all}\ \ \ d\in\N=\{1,2,\dots\},\ \ \mbox{and}\ \ \eps\in(0,1).$$ We prove that the Wiener process is much more difficult to approximate than the Euler process. Namely, strong polynomial tractability holds for the Euler case iff $$\liminf_{k\to \infty}\ \frac {r_k}{\ln\,k}> \frac1{2\ln\,3},$$ whereas it holds for the Wiener case iff $$\liminf_{k\to \infty}\ \frac {r_k}{k^s}> 0\ \ \ \ \ \mbox{for some}\ \
s>\tfrac12.$$ Other types of tractability are also studied.
author:
- 'M. A. Lifshits, A. Papageorgiou, H. Woźniakowski'
title: 'Tractability of Multi-Parametric Euler and Wiener Integrated Processes'
---
Introduction {#s1}
============
Tractability of multivariate problems has been recently an active research area. The reader may see the current progress on tractability in [@NW08; @NW10; @NW12]. Tractability has been studied in various settings and for various error criteria.
This field deals with problems defined on spaces of $d$-variate functions. For many practical computational problems $d$ is large. This holds for problems in mathematical finance, statistics and physics. We usually want to solve multivariate problems to within an error threshold $\eps$ by algorithms that use finitely many function values or, more generally, finitely many continuous linear functionals. Let $n(\eps,d)$ be the information complexity or shortly the complexity, denoting the minimal number of function values or continuous linear functionals that are needed to find an algorithm approximating the solution of a multivariate problem to within $\eps$.
Many multivariate problems suffer from the *curse of dimensionality*. That is, $n(\eps,d)$ is exponentially large in $d$. One of the goals of tractability is to determine under which conditions the curse of dimensionality is *not* present. Even more, we would like to have the complexity bounded by some non-exponential function of $d$ and $\eps^{-1}$. In particular, we have
- *weak* tractability if the complexity is *not* exponential in $d$ or $\eps^{-1}$,
- *quasi-polynomial* tractability if the complexity is of order $\exp(\,t\,(1+\ln\,d)(1+\ln\,\eps^{-1}))$,
- *polynomial* tractability if the complexity is of order $d^{\,q}\,\eps^{-p}$,
- *strong* polynomial tractability if the complexity is of order $\eps^{-p}$.
All bounds above hold for all $d$ and all $\eps\in(0,1)$ with the parameters $t,q,p$ and the factors multiplying the corresponding complexity bounds independent of $d$ and $\eps^{-1}$.
The strong polynomial tractability is the most challenging property. If this holds then the complexity has a bound independent of $d$. One may think that this property may hold only for trivial problems. Luckily, the opposite is true.
The curse of dimensionality often holds for multivariate problems for which all variables and groups of variables play the same role. One way to vanquish the curse is to shrink the class of functions by introducing weights that monitor the influence of successive variables and groups of variables. For sufficiently fast decaying weights we not only vanquish the curse but even obtain strong polynomial tractability; a survey of such results may be found in [@NW08; @NW10; @NW12].
The other way to vanquish the curse is by increasing the smoothness of functions with respect to successive variables. This approach was taken recently in [@PW10]. It was done for multivariate approximation defined over Korobov spaces in the worst case setting. The current paper can be viewed as a continuation of [@PW10]. We consider multivariate approximation but now in the average case setting with the normalized error criterion. This error criterion is defined as follows. We first take the zero algorithm and find its average case error for multivariate approximation; this is called the initial error. The initial error tells us how the problem scales and what can be achieved without sampling the functions. The normalized error criterion means that we want to improve the initial error by a factor $\eps$. We analyze algorithms that use arbitrary continuous linear functionals. We stress that the same results hold for algorithms that use only function values. This is due to general relations between these two classes of algorithms established in [@HWW08] and in Chapter 24 of [@NW12].
In this paper we analyze two multivariate approximation problems defined for the Euler and Wiener integrated processes, whereas in [@LPW2] we consider average case approximation for general non-homogeneous tensor products. More precisely, here we take the space of continuous real functions defined on the $d$-dimensional unit cube $[0,1]^d$. We stress that $d$ can be an arbitrary positive integer, however, our emphasis is on large $d$. We equip this space with a zero-mean Gaussian measure whose covariance kernel is denoted by $K_d$. We study two such kernels. The first one is $K_d=K_d^{\ee}$ for the Euler integrated process, whereas the second one $K_d=K_d^{\ww}$ is for the Wiener integrated process. These processes are precisely defined in the next section. Here we only mention that for both of them we know that almost surely the functions are $r_k$ times continuously differentiable with respect to the $k$-th variable for $k=1,2,\dots,d$.
The information complexity is then denoted by $n^{\ee}(\eps,d)$ and $n^{\ww}(\eps,d)$ for the Euler and Wiener integrated processes, respectively. Obviously, it depends on the the smoothness parameters $\{r_k\}$. Our main goal in this paper is to find necessary and sufficient conditions in terms of $\{r_k\}$ such that the four notions of tractability are satisfied.
We now briefly describe the results obtained in this paper. For both processes we prove that weak tractability holds iff $\lim_{k\to\infty}r_k=\infty$. Otherwise, if $r=\lim_{k\to\infty}r_k<\infty$ then we have the curse of dimensionality. This means that if all $r_k\le r<\infty$ then both $n^{\ee}(\eps,d)$ and $n^{\ww}(\eps,d)$ depend exponentially on $d$ and this holds for all $\eps\in(0,1)$. Hence, the function $n^{\xx}(\cdot,d)$ is discontinuous at $1$. Indeed, $n^{\xx}(1,d)=0$ although for $\eps$ pathologically close to one $n^{\rm x}(\eps,d)$ depends exponentially on $d$; here $\xx \in\{ \textrm{\sc E,\sc W} \}$.
We stress that weak tractability does not depend on the rate of convergence of $r_k$ to infinity. However, if we want to obtain other types of tractability we must require a certain convergence rate for the $r_k$, although the rate is different for the Euler and the Wiener case. For simplicity, let us consider $$r_k=\left\lceil 1+ a\,\ln(1+k\ln\,k)\right\rceil$$ for some positive number $a$. Then for the Euler case we have:
- $a<\frac1{2\,\ln\,3}$ no quasi-polynomial tractability,
- $a=\frac1{2\,\ln\,3}$ quasi-polynomial tractability but no polynomial tractability,
- $a>\frac1{2\,\ln\,3}$ strong polynomial tractability.
For the Wiener case we have to assume much more since for $r_k$ given above only weak tractability holds. For $$r_k=\left\lceil k^s\,\ln^2(1+k)\right\rceil$$ we have
- $s<\frac1{2}$ no quasi-polynomial tractability,
- $s=\frac1{2}$ quasi-polynomial tractability but no polynomial tractability,
- $s>\frac1{2}$ strong polynomial tractability.
For general $\{r_k\}$, we prove that quasi-polynomial tractability holds iff $$\begin{aligned}
\mbox{For the Euler case}\ &:&\ \ \ \limsup_{d\to\infty}\ \frac1{\ln\,d}
\,\sum_{k=1}^d (1+ r_k)\,3^{-2r_k}<\infty,\\
\mbox{For the Wiener case:}\ &:&\ \ \ \limsup_{d\to\infty}\ \frac1{\ln\,d}
\,\sum_{k=1}^d (1+r_k)^{-2}\,\max(1,\ln\,r_k)<\infty.\end{aligned}$$ Furthermore, for both processes polynomial tractability is equivalent to strong polynomial tractability and holds iff $$\begin{aligned}
\mbox{For the Euler case}\ &:&\ \ \
a_{\ee}:=\liminf_{d\to\infty}\ \frac{r_k}{\ln\,k}>\frac1{2\,\ln\,3},\\
\mbox{For the Wiener case:}\ &:&\ \ \
\liminf_{d\to\infty}\ \frac{r_k}{k^{s}}>0
\ \ \ \ \ \mbox{for some}\ \ \ s>\tfrac12.\end{aligned}$$ We also study the exponent $p^{\,\rm str-avg-x}$ of strong polynomial tractability which is defined as the infimum of $p$ for which the complexity is of order $\eps^{-p}$. For the Euler case we have $$p^{\,\rm str-avg-\ee}
=\max\left(\frac2{2r_1+1},\frac2{2 a_{\ee}\,\ln\,3 -1}\right)$$ For the Wiener case and $r_k=k^s$ for some $s>\tfrac12$ we have $$\max\left(\frac 2{2r_1+1},\frac2{2s-1}\right)\le
p^{\,\rm str-avg-\ww}\le \max\left(\frac2{2s-1},3\right).$$ Hence, for $s\in(\tfrac12,\tfrac56]$ we know that $$p^{\,\rm str-avg-\ww}=\frac2{2s-1}\, ;$$ otherwise our bounds are too weak to provide the exact value of the exponent.
Our results solve a special case of Open Problem 11 in [@NW08], where $r_{d,k}=r_k$, $k=1,\dots,d$, and, with slightly modified proofs, they also solve Open Problem 10 in [@NW08].
The Euler, Wiener and other univariate integrated processes can be characterized as follows. Consider $$X^{\a}(t):=
(-1)^{\a_1+\dots+\a_r} \underbrace{
\int_{\a_r}^t \int_{\a_{r-1}}^{t_{r-1}} \dots \int_{\a_1}^{t_1} W(s)\
{\rm d}s {\rm d}t_1 \dots {\rm d}t_{r-1}
}_{ r\ {\rm times}},
\qquad \qquad 0\le t\le 1,$$ where $\a=(\a_1,\a_2,\dots, \a_r)$ is a multi-index with components $\a_i\in\{ 0,1\}$ for $i=1,2,\dots,r$, $W(s)$ is the standard Wiener process for $0\le s\le 1$, and $r\in\N$. Then $X^{(0,\dots,0)}$ is the integrated Wiener process and $X^{(1,0,1,0,1,\dots)}$ is the integrated Euler process. It is an open problem to consider the integrated processes resulting from the different values of the multi-index $\a$ and to compare the necessary and sufficient conditions on $\{ r_k\}$ for weak, quasi-polynomial and polynomial tractability, respectively, with those obtained for the Euler and Wiener processes. In particular, it seems of interest to verify whether the Euler process is the easiest and the Wiener process is the most difficult among all of these $2^r$ processes.
The paper is organized as follows. In Section 2 we present the precise definitions of the average case approximation problem, the Euler and Wiener integrated processes and tractability notions. In Section 3 we present results for the Euler and in Section 4 for the Wiener integrated processes. The proofs of three theorems are presented in Sections 5 to 7.
Preliminaries {#s2}
=============
In this section we precisely define the Euler and Wiener processes, multivariate approximation in the average case setting, and we cite known results that will be needed for our analysis.
Euler and Wiener Processes {#s2.1}
--------------------------
Let $F_d=C([0,1]^d)$ be the space of real continuous functions defined on $[0,1]^d$ with the max norm, $$\|f\|_{F_d}=\max_{x\in [0,1]^d}|f(x)|\ \ \ \ \
\mbox{for all}\ \ \ f\in F_d.$$ We equip the space $F_d$ with a zero-mean Gaussian measure $\mu_d$ defined on Borel sets of $F_d$. The covariance kernel $K_d$ related to $\mu_d$ is defined by $$K_d(x,t)=\int_{F_d}f(x)\,f(t)\,\mu_d({\rm d}f)\ \ \ \ \
\mbox{for all}\ \ \ x,t\in [0,1]^d.$$ We refer to [@L96] for extensive theory of Gaussian measures in linear spaces and their covariance kernels.
By $\{r_k\}$ we mean a sequence of non-negative non-decreasing integers $$0\le r_1\le r_2\le\cdots\le r_d\le \cdots\ .$$ The Euler and Wiener integrated processes differ in the choice of the covariance kernel $K_d$. Our presentation of the Euler integrated processes is based on [@CH01] and [@GHT03]. The Wiener integrated process is more standard and can be found in many books and papers.
- *Euler integrated process.* We now have $K_d=K_d^{\ee}$ given by $$K_d^{\ee}(x,y)=\prod_{k=1}^d K_{1,r_k}^{\ee}(x_k,y_k)\ \ \
\mbox{for all}\ \ \ x,y\in[0,1]^d,$$ where $$K_{1,r}^{\ee}(x,y)=
\int_{[0,1]^r}\min(x,s_1)\,\min(s_1,s_2)\,\cdots\,\min(s_r,y)\,{\rm d}s_1
\,{\rm d}s_2\cdots{\rm d}s_r$$ for all $x,y\in [0,1]$. This kernel is equal to $$K_{1,r}^{\ee}(x,y)=
(-1)^{r+1}\frac{2^{2r}}{(2r+1)!}
\bigg(E_{2r+1}(\tfrac12|x-y|)-E_{2r+1}(\tfrac12(x+y))\bigg)$$ for all $x,y\in [0,1]$. Here, $E_n$ is the $n$-th degree Euler polynomial which can be defined as the coefficient of the generating function $$\frac{2\exp(x\,t)}{\exp(x)+1}=\sum_{n=0}^{\infty} E_n(x)\frac{t^n}{n!}\
\ \ \ \ \mbox{for all}\ \ \ x,t\in{{\mathbb R}}.$$ In particular, we have $E_0=1$, $E_1(x)=x-\tfrac12$ and $E_2(x)=x^2-x$.
The process is called *Euler* due to the fact that the covariance kernel is expressed by Euler polynomials.
- *Wiener integrated process.* We now have $K_d=K_d^{\ww}$ given by $$K_d^{\ww}(x,y)=\prod_{k=1}^d K_{1,r_k}^{\ww}(x_k,y_k)\ \ \
\mbox{for all}\ \ \ x,y\in[0,1]^d,$$ where $$K_{1,r}^{\ww}(x,y)=\int_0^{\min(x,y)}
\frac{(x-u)^r}{r!}\,
\frac{(y-u)^r}{r!}\,{\rm d}u
=\int_0^1\frac{(x-u)_+^r}{r!}\,
\frac{(y-u)_+^r}{r!}\,{\rm d}u$$ for all $x,y\in [0,1]$ and with the standard notation $t_+=\max(t,0)$.
Let us stress that the univariate Euler and Wiener processes are close relatives since they emerge from very similar integration schemes. Indeed, let $W(t), t\in[0,1]$, be a standard Wiener process, i.e. a Gaussian random process with zero mean and covariance $$K_{1,0}^\ee(s,t)=K_{1,0}^\ww(s,t)=\min(s,t).$$ Consider two sequences of integrated processes $X_r, Y_r$ on $[0,1]$ defined by $X_0=Y_0=W$, and for $r=0,1,2,\dots$ $$\begin{aligned}
X_{r+1}(t)&=& \int_0^t X_{r}(s) {\rm d} s
\\
Y_{r+1}(t)&=& \int_{1-t}^1 Y_{r}(s) {\rm d} s, \end{aligned}$$ Then the covariance kernel of $X_r$ is $K_{1,r}^\ww$ while the covariance kernel of $Y_r$ is $K_{1,r}^\ee$. Clearly, $X_r$ and $Y_r$ have the same smoothness properties. That is why different tractability results are surprising.
On the other hand, there are some differences between the two processes. The Gaussian measure $\mu_d$ on $F_d$ corresponding to the covariance kernel $K_d^{\ee}$ or $K_d^{\ww}$ is concentrated on functions that are almost surely $r_k$-times continuously differentiable with respect to the $k$-th variable for $k=1,2,\dots,d$, and satisfy certain boundary conditions which are different for the Euler and Wiener cases.
For the Euler case, we have $$\label{E1}
\frac{\partial^{k_1+k_2+\cdots +k_d}}{\partial\,x_1^{k_1}\,\partial\,
x_2^{k_2}\cdots\partial\,x_d^{k_d}}\, f(x) = 0$$ if for some $i$ we have $x_i=0$ and $k_i$ is even, or $x_i=1$ and $k_i$ is odd. Here, $k_i=0,1,\dots,r_i$.
For the Wiener case, we have $$\label{W1}
\frac{\partial^{k_1+k_2+\cdots +k_d}}{\partial\,x_1^{k_1}\,\partial\,
x_2^{k_2}\cdots\partial\,x_d^{k_d}}\, f(x) = 0$$ if one of the components of $x$ is zero. As before, $k_i=0,1,\dots,r_i$.
To see the difference between and more explicitly, we take $d=1$. Then for the Euler case for all $0\le k\le r_1$ we have $$f^{(k)}(0)=0 \ \ \ \mbox{if $k$ is even \ \ \ and}\ \ \
\ \ f^{(k)}(1)=0 \ \ \mbox{if $k$ is odd},$$ whereas for the Wiener case we have $$f^{(k)}(0)=0\ \ \ \mbox{for all $k\le r_1$}.$$
Finally, note that Nazarov and Nikitin studied in [@Naz; @NazNik] a slightly different version $W_r^N$ of Euler integrated process. The processes $W_r^N$ and $W_r^\ee$ coincide for even $r$ but $W_r^N(t)=W_r^\ee(1-t)$ for odd $r$. The covariance spectra of both processes are the same but the boundary conditions are different. Since the spectra are the same, the tractability results for the Nazarov and Nikitin process are the same as for the Euler process.
Multivariate Approximation {#s2.2}
--------------------------
Multivariate approximation is defined by the embedding ${\rm APP}_d:F_d\to L_2$ given by $${\rm APP}_df=f\ \ \ \ \ \mbox{for all}\ \ \ f\in F_d.$$ Here, $L_2=L_2([0,1]^d)$ is the standard $L_2$ space with the norm $$\|f\|_{L_2}=\bigg(\int_{[0,1]^d}f^{\,2}(t)\,{\rm d}t\bigg)^{1/2}.$$ We approximate functions $f$ from $F_d$ by algorithms $A_n$ that use $n$ function values or arbitrary continuous linear functionals. We only consider the case of arbitrary continuous functionals since it is known that the results are the same for function values, see [@HWW08] and Chapter 24 of [@NW12]. In the average case setting, without essential loss of generality, see e.g., [@TWW88] as well as [@NW08], we can restrict ourselves to linear algorithms $A_n$ of the form $$A_n(f)=\sum_{j=1}^nL_j(f)\,g_j \ \ \ \ \
\mbox{with} \ \ \ \ \ L_j\in F^{\,*}_d,\ \ g_j\in L_2.$$ The average case error of $A_n$ is defined as $$e^{\rm avg}(A_n)=\bigg(\int_{F_d}\|{\rm APP}_df-A_n(f)\|_{L_2}^2\,
\mu_d({\rm d}f)\bigg)^{1/2},$$ where $\mu_d$ is a zero-mean Gaussian measure with a covariance kernel $K_d$ and $$\|{\rm APP}_df-A_n(f)\|_{L_2}^2=\int_{[0,1]^d}(f(t)-A_n(f)(t))^2\,{\rm d}t.$$ Then $\nu_d=\mu_d{\rm APP}_d^{-1}$ is a zero-mean Gaussian measure defined on Borel sets of $L_2$. The covariance operator $C_{\nu_d}:L_2\to L_2$ of $\nu_d$ is given by $$C_{\nu_d}f= \int_{[0,1]^d}K_d(\cdot,t)\,f(t)\,{\rm d}t\ \ \ \ \
\mbox{for all}\ \ \ f\in L_2.$$ The operator $C_{\nu_d}$ is a self-adjoint, nonnegative definite, and has finite trace. Let $(\l_{d,j},\eta_{d,j})_{j=1,2,\dots}$ denote its eigenpairs $$C_{\nu_d}\eta_{d,j}=\lambda_{d,j}\,\eta_{d,j} \ \ \ \ \
\mbox{with} \ \ \ \lambda_{d,1}\ge\lambda_{d,2}\ge\cdots,$$ and $$\sum_{j=1}^\infty\lambda_{d,j}=\int_{[0,1]^d}K_d(t,t)\,{\rm d}t<\infty.$$
For a given $n$, it is well known that the algorithm $A_n$ that minimizes the average case error is of the form $$\label{optalg}
A_n(f)=\sum_{j=1}^n\il f,\eta_{d,j}\ir_{L_2}\eta_{d,j},$$ and its average case error is $$\label{avgerror}
e^{\rm avg}(A_n)=\bigg(\sum_{j=n+1}^\infty\lambda_{d,j}\bigg)^{1/2}.$$ For $n=0$ we obtain the zero algorithm $A_0=0$. Its average case error is called the initial error, and is given by the square-root of the trace of the operator $C_{\nu_d}$, i.e., by with $n=0$.
We now define the average case information complexity $n(\eps,d)$ as the minimal $n$ for which there is an algorithm whose average case error reduces the initial error by a factor $\eps$, $$\label{infocomp}
n(\eps,d)
=\min\bigg\{\,n\ \bigg| \ \ \sum_{j=n+1}^\infty\lambda_{d,j}\le
\eps^2\,\sum_{j=1}^\infty\lambda_{d,j}\bigg\}.$$
From it is clear that all notions of tractability depend only on the eigenvalues $\lambda_{d,j}$. Therefore the more we know about the eigenvalues $\lambda_{d,j}$ the more we can say about various notions of tractability.
Eigenvalues for the Euler and Wiener Integrated Processes {#s3.1}
---------------------------------------------------------
For both processes the corresponding covariance kernel is of product form. Therefore the eigenvalues for the $d$-variate case are products of the eigenvalues of the univariate cases which depend on the smoothness parameters $r_k$ for $k=1,2,\dots,d$. That is, if we denote by $\lambda_{d,j}^\xx$’s the eigenvalues of the Euler integrated process, $\xx={\rm E}$, or the eigenvalues of the Wiener integrated process, $\xx={\rm W}$, then $$\{\lambda_{d,j}^\xx\}_{j=1,2,\dots}
= \left\{\lambda_{j_1,r_1}^\xx \lambda_{j_2,r_2}^\xx \dots
\lambda_{j_d,r_d}^\xx \right\}_{j_1,j_2,\dots,j_d=1,2\dots},$$ with the $\lambda_{j_k,r_k}^\xx$’s denoting the eigenvalues of the univariate case with smoothness $r_k$.
For the Euler case, the $\lambda_{j_k,r_k}^{\ee}$’s are the eigenvalues of the operator $$(C_{1,r_k}^{\,\ee}f)(x)=\int_0^1K^{\ee}_{1,r_k}(x,t)\,f(t)\,\,{\rm d}t.$$ By successive differentiation of this equation with respect to $x$ and using the properties of the kernel $K_{1,r_k}^{\ee}$, it is easy to show that the eigenvalues of $C_{1,r_k}^{\,\ee}$ satisfy the Sturm-Liouville problem $$\label{SL}
\lambda\,f^{(2r_k+2)}(x)=(-1)^{r_k+1}f(x)\ \ \ \ \ \mbox{for all}
\ \ \ x\in(0,1),$$ with the boundary conditions $$f(t_0)
=f^{\prime}(t_1)=f^{\prime\prime}(t_2)=\cdots=f^{(2r_k+1)}(t_{2r_k+1})
=0,$$ where $t_i=0$ for even $i$ and $t_i=1$ for odd $i$. For the Euler case, we know the eigenvalues exactly, see [@CH01] and [@GHT03], and they are equal to $$\label{Eulereig}
\lambda_{j,r_k}^{\ee}=\left(\frac{1}{\pi(j-1/2)}\right)^{2r_k+2}\ \ \
\mbox{for}\ \ j=1,2,\dots.$$ Note that the eigenvalues are well separated. In particular, $$\frac{\lambda_{2,r_k}^{\ee}}{\lambda_{1,r_k}^{\ee}}=\frac1{3^{2r_k+2}}.$$
For the Wiener case, $\lambda_{j,r_k}^{\ww}$’s are the eigenvalues of the operator
$$(C_{1,r_k}^{\,\ww}f)(x)=\int_0^1K^{\ww}_{1,r_k}(x,t)\,f(t)\,\,{\rm d}t.$$ The eigenvalues $\lambda_{j,r_k}^{\ww}$ also satisfy the Sturm-Liouville problem but with different boundary conditions $$f(0)=f^\prime(0)=\cdots=f^{(r_k)}(0)=f^{(r_k+1)}(1)=f^{(r_k+2)}(1)=
\cdots=f^{(2r_k+1)}(1)=0.$$ The eigenvalues $\lambda_{j,r_k}^{\ww}$ are *not* exactly known. It is known [@GHT03] that they have the same asymptotic behavior as in (\[Eulereig\]), $$\label{Wiener-eig}
\lambda_{j,r_k}^{\ww}=\left(\frac{1}{\pi(j-1/2)}\right)^{2r_k+2}
+\OO\left(j^{-(2r_k+3)}\right)
\ \ \ \ \ \mbox{as}\ \ j\to\infty.$$ For tractability studies the asymptotic behavior is not enough and the two largest eigenvalues play an essential role. That is why we will prove that $$\begin{aligned}
\lambda_{1,r_k}^{\ww}&=&\frac1{(r_k!)^2}
\left(\frac1{(2r_k+2)(2r_k+1)}+\OO(r_k^{-4})\right),\\
\lambda_{2,r_k}^{\ww}&=&\Theta\left(\frac1{(r_k!)^2\,r_k^4}\right),\end{aligned}$$ where the factors in the big $\OO$ and $\Theta$ notations do not depend on $r_k$.
Note that the largest eigenvalues for the Euler case go to zero exponentially fast with $r_k$, whereas for the Wiener case they go to zero super exponentially fast due to the presence of factorials. However, the ratio of the two largest eigenvalues for the Wiener case, $$\frac{\lambda_{2,r_k}^{\ww}}{\lambda_{1,r_k}^{\ww}}=\Theta(r_k^{-2}),$$ is much larger than that for the Euler case.
Tractability
------------
We present the precise definitions of four notions of tractability. Let $n(\eps,d)$ denote the average case information complexity defined in , and let ${\rm APP}=\{{\rm APP}_d\}_{d=1,2,\dots}$ denote the sequence of multivariate approximation problems. We say that
- ${\rm APP}$ is *weakly tractable* iff $$\lim_{\eps^{-1}+ d\to\infty}\frac{\ln\,n(\eps,d)}{\eps^{-1}+d}=0,$$ with the convention that $\ln\,0=0$.
- ${\rm APP}$ is *quasi-polynomially tractable* iff there are positive numbers $C$ and $t$ such that $$n(\eps,d)\le C\,\exp\big(\,t\,(1+\ln\,d)\,(1+\ln\,\eps^{-1}\,\big)
\ \ \ \ \ \mbox{for all}\ \ \ d=1,2,\dots, \ \ \eps\in(0,1).$$
- ${\rm APP}$ is *polynomially tractable* iff there are non-negative numbers $C,q$ and $p$ such that $$n(\eps,d)\le C\,d^{\,q}\,\eps^{-p}
\ \ \ \ \ \mbox{for all}\ \ \ d=1,2,\dots, \ \ \eps\in(0,1).$$
- ${\rm APP}$ is *strongly polynomially tractable* iff there are positive numbers $C$ and $p$ such that $$n(\eps,d)\le C\,\eps^{-p}
\ \ \ \ \ \mbox{for all}\ \ \ d=1,2,\dots, \ \ \eps\in(0,1).$$ The infimum of $p$ satisfying the last bound is called the exponent of strong polynomial tractability and denoted by $p^{\rm \,str-avg}$. For the Euler and Wiener case, we use the notation $p^{\rm\,str-avg-x}$ with $\xx\in\{{\rm E},{\rm W}\}$.
Tractability can be fully characterized in terms of the eigenvalues $\lambda_{d,j}$. Necessary and sufficient conditions on weak, quasi-polynomial, polynomial and strong polynomial tractability can be found in Chapter 6 of [@NW08] and Chapter 24 of [@NW12] as well as in [@LPW2] for non-homogeneous tensor products. For the Euler and Wiener integrated processes we need such conditions that are based on the sums of some power of the eigenvalues $\lambda_{d,j}$. We will cite these conditions when they are needed for specific tractability results.
Euler Integrated Process
========================
We now analyze the Euler integrated process for which the eigenvalues in the univariate cases are given by . Our aim is to express tractability conditions in terms of the smoothness parameters $\{r_k\}$.
\[Euler-thm\] Consider the approximation problem ${\rm APP}$ for the Euler integrated process.
- ${\rm APP}$ is weakly tractable iff $$\label{weakEuler}
\lim_{k\to\infty}r_k=\infty.$$ Furthermore, if does not hold then we have the curse of dimensionality since $n^{\ee}(\eps,d)$ depends exponentially on $d$ for each $\eps<1$.
- ${\rm APP}$ is quasi-polynomially tractable iff $$\label{Eulerquasi}
\sup_{d\in\N}\ \frac1{\ln_+d}\ \sum_{k=1}^d(1+r_k)\,3^{-2r_k}<\infty,$$ where $\ln_+d=\max(1,\ln\,d)$.
- ${\rm APP}$ is polynomially tractable iff ${\rm APP}$ is strongly polynomially tractable iff $$\sum_{k=1}^\infty 3^{-2\,\tau\,r_k}<\infty \ \ \ \
\mbox{for some}\ \ \tau\in(0,1)$$ or equivalently iff $$a_{\ee}:=\liminf_{k\to \infty}\ \frac{r_k}{\ln\,k}>\frac1{2\ln\,3}.$$ If so, then the exponent[^1] of strong polynomial tractability is $$p^{\rm str-avg-E}
=\max\left(\frac2{2r_1+1},\frac2{2a_{\ee}\,\ln\,3-1}\right).$$
We briefly comment on Theorem \[Euler-thm\]. First of all, we stress that polynomial and strong tractability are equivalent. That is, these two notions coincide for the Euler integrated process: in this case a weaker" property of polynomial tractability implies a stronger property of strong polynomial tractability. Weak tractability requires that the smoothness parameters $r_k$ go to infinity, however, the speed of convergence is irrelevant. To obtain at least quasi-polynomial tractability, we need to assume that $r_k$ increases at least as $a_{\ee}\,\ln\,k$ with $a_{\ee} > 1/(2\ln\,3)$. Indeed, assume for simplicity that $$a_{\ee}:=\lim_{k\to\infty}\frac{r_k}{\ln\,k}.$$ exists. If $a_{\ee}<1/(2\,\ln\,3)$ then for any positive $\beta<1-2\,a_{\ee}\,\ln\,3$ we have $$\label{qpol111}
n^{\ee}(\eps,d)\ge c_1(\beta)\,(1-\eps^2)\,
\exp\left((c_2(\beta)\,d^{\,\beta}\right)$$ for some positive functions $c_1$ and $c_2$ of $\beta$. Note that contradicts quasi-polynomial tractability. The proof of goes like follows. We will show later that $$n^{\ee}(\eps,d)
\ge
(1-\eps^2)\prod_{k=1}^d \left(1+3^{-2(r_k+1)}\right).$$ Then each factor $1+3^{-2(r_k+1)}$ for large $j$ can be estimated from below by $\exp(-c(\beta)k^{-1+\beta})$. From this we easily obtain .
If $a_{\ee}=1/(2\,\ln\,3)$ then we can have quasi-polynomial tractability as illustrated by an example of $\{r_k\}$ in the introduction. Furthermore, for this example we do not have polynomial tractability. However, it may also happen that for $a_{\ee}=1/(2\,\ln\,3)$ we do not have quasi-polynomial tractability. For example, this is the case when $$r_k=\left\lceil\frac{\ln_+k}{2\,\ln\,3}\right\rceil,$$ which can be checked directly from .
On the other hand, if $a_{\ee}>1/(2\,\ln\,3)$ then we obtain strong polynomial tractability. This shows that there is a “thin” zone of $\{r_k\}$ that separates quasi-polynomial and strong polynomial tractabilities.
We now comment on the exponent of strong polynomial tractability. Note that for $a_{\ee}\ge (r_1+1)/\ln\,3$ we have $$p^{\,\rm str-avg-E}=\frac2{2r_1+1}.$$ In this case, the result is especially pleasing hence the complexity for any $d$ is roughly bounded by the complexity for the univariate case. Furthermore, this happens for all $r_k$’s that tend to infinity faster than $\ln\,k$. On the other hand, if $a_\ee\in(1/(2\,\ln\,3)),2(r_1+1)/(2\,\ln\,3))$ then we have $$p^{\,\rm str-avg-E}=\frac2{2a_{\ee}\,\ln\,3-1 },$$ and $p^{\,\rm str-avg-E}$ can be arbitrarily large when $a_{\ee}$ is close to $1/(2\,\ln\,3)$.
Wiener Integrated Process
=========================
We now turn to the Wiener integrated process for which the eigenvalues for the univariate cases $\lambda_{j,r_k}^{\ww}$ are only known asymptotically, see . To express tractability conditions in terms of the smoothness parameters $\{r_k\}$ we will need to prove the behavior of the two largest eigenvalues for large $r_k$.
\[Wiener-largest\] Consider the univariate Wiener process with the smoothness parameter $r$, and let $\lambda_{j,r}^{\ww}$’s denote the eigenvalues of the covariance operator $C_{1,r}^{\ww}$. Then
$$\begin{aligned}
\lambda_{1,r}^{\ww}&=&\frac1{(r!)^2}\left(\frac1{(2r+2)\,(2r+1)}
+\OO(r^{-4})\right),\\
\lambda_{2,r}^{\ww}&=&\Theta\left(\frac1{(r!)^2\,r^4}\right),\\
\sup_{\tau\in[\tau_0,1]}\
\frac{\sum_{j=3}^\infty\left[
\lambda_{j,r}^{\ww}\right]^\tau}{\left[\lambda_{2,r}^{\ww}\right]^\tau}
&=&\OO(r^{-h})
\ \ \ \ \ \ \ \ \ \mbox{for some}\ h>0\ \ \
\mbox{and for all}\ \ \ \tau_0\in(\tfrac35,1].\end{aligned}$$
Observe that the two largest eigenvalues for the Wiener case are much smaller than for the Euler case. On the other hand, their ratio for the Wiener case is much larger than for the Euler case. Therefore, the sequences $\{\lambda_{j,r}^{\ww}\}$ and $\{\lambda_{j,r}^{\ee}\}$ are quite different although they have the same asymptotic behavior.
The uniform convergence in the last assertion of Theorem \[Wiener-largest\] at the neighborhood of $\tau=1$ is needed when we deal with quasi-polynomial tractability. The convergence for a specific $\tau$ is needed for strong polynomial and polynomial tractability. The lower bound $\tfrac35$ for $\tau_0$ is surely not sharp. A possible improvement of this lower bound would improve the exponent of strong polynomial tractability.
Based on the estimates presented in Theorem \[Wiener-largest\] we will be able to express tractability conditions for the Wiener case in terms of $\{r_k\}$.
\[Wiener-thm\] Consider the approximation problem ${\rm APP}$ for the Wiener integrated process.
- ${\rm APP}$ is weakly tractable iff $$\label{weakWiener}
\lim_{k\to\infty}r_k=\infty.$$ Furthermore, if does not hold then we have the curse of dimensionality since $n^{\ww}(\eps,d)$ depends exponentially on $d$ for each $\eps<1$.
- ${\rm APP}$ is quasi-polynomially tractable iff $$\label{Wienerquasi}
\sup_{d\in\N}\ \frac1{\ln_+d}\ \sum_{k=1}^d
(1+r_k)^{-2}\,\ln_+r_k<\infty,$$ where, we use $\ln_+x=\max(1,\ln\,x)$ for $x>0$, and $\ln_+0=1$.
- ${\rm APP}$ is polynomially tractable iff ${\rm APP}$ is strongly polynomially tractable iff $$\liminf_{k\to \infty}\ \frac{r_k}{k^s}>0
\ \ \ \ \mbox{for some}\ \ \ \ \ s>\tfrac12.$$
We briefly comment on Theorem \[Wiener-thm\]. As for the Euler case, strong polynomial and polynomial tractability are equivalent, and weak tractability holds under the same condition $\lim_k r_k=\infty$. That ends the similarity between the Wiener and Euler cases since the conditions on quasi-polynomial and polynomial tractability are quite different. For the Wiener case, we must assume that $r_k$’s go to infinity at least as fast as $k^{-s}$ for some $s>\tfrac12$. However, the zone between quasi-polynomial and polynomial tractabilities is again thin, as for the Euler case.
It is worth to add that quasi-polynomial tractability plays a much more important role in the worst case setting. The difference with the average case setting is due to the fact that even for the constant sequence $r_k={\rm const}>0$ we have quasi-polynomial tractability in the worst case setting as shown in [@GW11].
We now discuss the exponent of strong tractability which is not addressed in Theorem \[Wiener-thm\]. For simplicity, let us assume that for some $s>\tfrac12$ we have $$r_k=k^s \ \ \ \ \mbox{for all}\ \ \ k\in\N.$$ Then we have strong polynomial tractability and the exponent $p^{\rm str-avg-\ww}$ is given in as the infimum of $2\tau/(1-\tau)$ for $\tau$ from $(0,1)$ which satisfies condition below with $q=0$. From the proof of Theorem \[Wiener-thm\] we know that $\tau>1/(2r_1+2)$. Furthermore, implies that $\tau>1/(2s)$. These two estimates yield lower bounds on the exponent. On the other hand, our proof of strong polynomial tractability is valid only for $\tau>\frac35$, and this effects an upper bound on the exponent. Hence, $$\max\left(\frac2{2r_1+1},\frac2{2s-1}\right)
\le p^{\rm str-avg-\ww}
\le \max\left(\frac2{2s-1},3\right).$$ We stress that only for $s\in(\tfrac12,\tfrac56]$ we know the exponent exactly, $
p^{\rm str-avg-\ww}=\tfrac{2}{2s-1}.
$ Note that $p^{\rm str-avg-\ww}$ can be arbitrarily large if $s$ is close to $\tfrac12$.
For $s>\tfrac56$, our bounds on the eigenvalues $\lambda_{j,r_k}^{\ww}$ are too weak to get the exact value of the exponent but sufficient to deduce strong polynomial tractability.
Proof of Theorem \[Euler-thm\]
==============================
It is convenient to deal first with polynomial tractability. Let PT stand for polynomial tractability and SPT for strong polynomial tractability. To prove this point of Theorem \[Euler-thm\] it is enough to show that $$\label{4stat}
a_\ee>\frac1{2\,\ln\,3} \
\Rightarrow\
\sum_{k=1}^\infty 3^{-2\tau\,r_k}<\infty\
\Rightarrow \
{\rm SPT}\
\Rightarrow\
{\rm PT}
\Rightarrow \
a_\ee>\frac1{2\,\ln\,3} \ .$$ The first claim, $a_\ee>1/(2\,\ln\,3)\ \Rightarrow
S_\tau:=\sum_{k=1}^\infty3^{-2\tau\,r_k}<\infty$ for some $\tau\in(0,1)$, is an easy calculus exercise. Indeed, let $a_\ee>1/(2\ln\,3)$. Then for some $\delta>0$ and all $k$ large enough we have $\tfrac{r_k}{\ln k}> \tfrac{1+\delta}{2\ln 3}$, hence $3^{-2\tau r_k} < k^{-(1+\delta)\tau}$ and $S_\tau<\infty$ whenever $\tfrac{1}{1+\delta}<\tau<1$.
Recall now the polynomial tractability criteria. We know from Chapter 6 of [@NW08] that ${\rm APP}$ is polynomially tractable iff there exist $q\ge0$ and $\tau\in(0,1)$ such that $$\label{poltract}
C:= \sup_{d\in\N}
\frac{\left(\sum_{j=1}^\infty\lambda_{d,j}^\tau\right)^{1/\tau}}
{\sum_{j=1}^\infty\lambda_{d,j}}\ d^{\, -q} <\infty.$$ If so then $$n(\eps,d)
\le \left(\left(\frac{\tau\,C}{1-\tau}\right)^{\tau/(1-\tau)}+1\right)\,
d^{\,q\,\tau/(1-\tau)}\,\eps^{-2\tau/(1-\tau)}$$ for all $d\in\N$ and $\eps\in(0,1)$.
Furthermore, ${\rm APP}$ is strongly polynomially tractable iff holds with $q=0$. The exponent of strong polynomial tractability is $$\label{expostpol}
p^{\,\rm str-avg}=\inf\left\{\frac{2\tau}{1-\tau}\ \bigg| \ \ \tau \
\mbox{satisfies~\eqref{poltract} with $q=0$}\right\}.$$
Motivated by condition (\[poltract\]) and based on the explicit knowledge of the univariate eigenvalues for the Euler integrated process , we take $\tau\in(0,1)$ and obtain $$\begin{aligned}
\frac{\left(\sum_{j=1}^\infty\lambda_{d,j}^\tau\right)^{1/\tau}}
{\sum_{j=1}^\infty\lambda_{d,j}}&=&
\prod_{k=1}^d\frac{\left(\sum_{j=1}^\infty\left(
\lambda_{j,r_k}^{\ee}\right)^\tau\right)^{1/\tau}}{\sum_{j=1}^\infty
\lambda_{j,r_k}^{\ee}}\\
&=&
\prod_{k=1}^d\frac{\left(\sum_{j=1}^\infty(2j-1)^{-(2r_k+2)\tau}
\right)^{1/\tau}}{\sum_{j=1}^\infty(2j-1)^{-(2r_k+2)}}\\
&=&\prod_{k=1}^d
\frac{\left(1+\sum_{j=2}^\infty(2j-1)^{-2\tau(r_k+1)}\right)^{1/\tau}}
{1+\sum_{j=2}^\infty(2j-1)^{-2(r_k+1)}}.\end{aligned}$$ Since $r_k\ge r_1$, note that the expression above is finite for all $\tau\in(1/(2r_1+2),1)$. Furthermore for such $\tau$ we have $$3^{-2\tau(r_k+1)}
\le\sum_{j=2}^\infty(2j-1)^{-2\tau(r_k+1)}
\le 3^{-2\tau(r_k+1)}+\sum_{j=5}^\infty j^{-2\tau(r_k+1)},$$ and $$\sum_{j=5}^\infty j^{-2\tau(r_k+1)}
\le \int_4^\infty x^{-2\tau(r_k+1)}\,{\rm d}x
= \frac{4^{1-2\tau(r_k+1)}}{2\tau(r_k+1)-1}
\le \frac{3}{2\tau(r_1+1)-1} \,3^{-2\tau(r_k+1)}.$$ Therefore $$\label{useful}
\frac{\left(\sum_{j=1}^\infty\lambda_{d,j}^\tau\right)^{1/\tau}}
{\sum_{j=1}^\infty\lambda_{d,j}}=
\prod_{k=1}^d\frac{\left(1 +a_k 3^{-2\tau(r_k+1)}\right)^{1/\tau}}
{1+b_k 3^{-2(r_k+1)}},$$ where $a_k\ge b_k$ and they are uniformly bounded, $$\label{useful2}
1\le a_k\le\frac{2\tau(r_1+1)+2}{2\tau(r_1+1)-1}\ \ \ \ \
\mbox{and}\ \ \ \ \ 1\le b_k\le \frac{2r_1+4}{2r_1+1}.$$
Assume now that $S_\tau<\infty$ for some $\tau<1$. By using (\[useful\]) and (\[useful2\]) we obtain $$\begin{aligned}
\sup_{d}
\frac{\left(\sum_{j=1}^\infty\lambda_{d,j}^\tau\right)^{1/\tau}}
{\sum_{j=1}^\infty\lambda_{d,j}}
&\le&
\prod_{k=1}^\infty \left(1 +a_k 3^{-2\tau(r_k+1)}\right)^{1/\tau}
\\
&\le&
\exp \left( \tau^{-1} \sup_k a_k \sum_{k=1}^\infty 3^{-2\tau(r_k+1)}\right)
\le
\exp \left( \tau^{-1} \sup_k a_k \ S_\tau \right) <\infty.\end{aligned}$$ Hence, the criterion (\[poltract\]) is verified with $q=0$, and we conclude that $S_\tau<\infty$ $\Rightarrow$ [SPT]{}.
Implication [SPT]{} $\Rightarrow$ [PT]{} is trivial.
Assume now that [PT]{} holds. By (\[poltract\]) and (\[useful\]) this implies that $$\prod_{k=1}^d
\frac{\left(1 +a_k 3^{-2\tau(r_k+1)}\right)^{1/\tau}}
{1+b_k 3^{-2(r_k+1)}}<C\,d^{\,q}$$ for some $C,q\ge 0$ and $\tau\in (0,1)$. Moreover, is easy to check that $$\frac{\left(1 +a_k 3^{-2\tau(r_k+1)}\right)^{1/\tau}}
{1+b_k 3^{-2(r_k+1)}}\ge 1+c_k 3^{-2\tau(r_k+1)}$$ for $c_k\ge a_k (1-3^{-2(r_k+1)(1-\tau)})/(1+b_k 3^{-2(r_k+1)})=\Omega(1)$. Taking logarithms we conclude that $$M:= \sup_d\ \frac1{\ln_+d}\ \sum_{k=1}^d 3^{-2\tau(r_k+1)}
<\infty.$$ The sum with respect to $k$ can be lower bounded by $d \cdot 3^{-2\tau(r_d+1)}$, as done at the beginning of the proof, and we obtain $d \cdot 3^{-2\tau(r_d+1)}\le M \ln_+d$, which is equivalent to $$\frac{r_d+1}{\ln d} \ge
\frac{1-\tfrac{\ln\ln_+d-\ln M}{\ln d}}{\tau\cdot2\ln 3},$$ and implies that $a_\ee\ge 1/(2\tau\,\ln\,3) >1/(2\,\ln\,3)$, as claimed. The equivalence of all statements in (\[4stat\]) is therefore verified.
We now consider the exponent $p^{\rm str-avg-E}$. Assume now that $a_\ee>\tfrac{1}{2\,\ln\,3}$. Then, as already shown, $\sum_{k=1}^\infty3^{-2\tau(r_k+1)}<\infty$ for all $\tau>\tfrac{1}{2a_\ee\ln\,3}$ and holds with $q=0$ if $\tau>\tfrac{1}{2r_1+2}$. Hence, we obtain strong polynomial tractability. Furthermore, $\tau$ can be taken in the limit as $\tau_*:= \max\left(\tfrac{1}{2r_1+2},\tfrac{1}{2a_\ee\ln\,3}\right)$, and yields that the exponent of strong polynomial tractability is at most $$p_*:= \frac{2\tau_*}{1-\tau_*}
= \max\left( \frac{2}{2r_1+1},\frac{2}{2a_\ee\ln\,3 -1}\right).$$ Conversely, assume that strong polynomial tractability holds. Then $$\prod_{k=1}^\infty
\frac{\left(1 +a_k 3^{-2\tau(r_k+1)}\right)^{1/\tau}}
{1+b_k 3^{-2(r_k+1)}}<\infty$$ for some $\tau\in(0,1)$. Clearly, we must take $\tau>1/(2r_1+2)$ and $\tau>1/(2a_\ee\ln\,3)$. This implies that the exponent is at least $p_*$. This completes the part of the proof related to polynomial and strong polynomial tractability.
We now turn to weak tractability. We know from [@LPW2] that ${\rm APP}$ is weakly tractable if there exists $\tau\in(0,1)$ such that $$\label{weakcond}
\lim_{d\to \infty}\
\frac1d\ \sum_{k=1}^d\ \sum_{j=2}^\infty
\left(\frac{\lambda_{j,r_k}^{\ee}}{\lambda_{1,r_k}^{\ee}}
\right)^{\tau}=0.$$ In our case, we have $$\frac{\lambda_{j,r_k}^{\ee}}{\lambda_{1,r_k}^{\ee}}= (2j-1)^{-2(r_k+1)}.$$ As before, for $\tau\in(\tfrac12,1)$ we have $$\sum_{j=2}^\infty(2j-1)^{-2\tau(r_k +1)}\le
\frac{2\tau(r_k+1)+2}{2\tau(r_k+1)-1}\,3^{-2\tau(r_k+1)}
\le \frac{2(1+\tau)}{2\tau-1}\,3^{-2\tau(r_k+1)}.$$ Assume that $\lim_{k\to\infty}r_k=\infty$. Then for an arbitrarily large $M$ there is an integer $k_M$ such that $r_k\ge M$ for all $k\ge k_M$. Hence, for $d\ge k_M$ we have $$\frac1d\ \sum_{k=1}^d\ \sum_{j=2}^\infty
\left(\frac{\lambda_{j,r_k}^{\ee}}{\lambda_{1,r_k}^{\ee}}
\right)^{\tau}
\le \frac{2(1+\tau)}{2\tau-1} \left(\frac{k_M}d+3^{-2\tau(M+1)}\right),$$ and we obtain (\[weakcond\]) by letting first $d$, and then $M$ go to infinity.
On the other hand, if $r=\lim_{k\to\infty}r_k<\infty$ then there is an integer $k_0$ such that $r_k=r$ for all $k\ge k_0$, and the limit in is not zero. In this case, we prove that $n=n^{\ee}(\eps,d)$ is an exponential function of $d$ and therefore weak tractability does not hold. Indeed, we have $$\sum_{j=1}^\infty\lambda_{d,j}-n\lambda_{d,1}\le
\sum_{j=n+1}^\infty\lambda_{d,j}
\le \eps^2\sum_{j=1}^\infty\lambda_{d,j},$$ and therefore $$\begin{aligned}
n&\ge& (1-\eps^2)\sum_{j=1}^\infty\frac{\lambda_{d,j}}{\lambda_{d,1}}
=(1-\eps^2)\left(\prod_{k=1}^{k_0-1}\sum_{j=1}^\infty
\frac{\lambda_{j,r_k}^{\ee}}
{\lambda_{1,r_k}^{\ee}}\right)
\left(1+\sum_{j=2}^\infty
\frac{\lambda_{j,r}^{\ee}}{\lambda_{1,r}^{\ee}} \right)^{d-k_0+1}
\\
&\ge&(1-\eps^2)\left(1+\sum_{j=2}^\infty
\frac{\lambda_{j,r}^{\ee}}{\lambda_{1,r}^{\ee}} \right)^{d-k_0+1}.\end{aligned}$$ This bound is an exponential function of $d$. It contradicts weak tractability and completes the part of the proof related to this notion.
We finally consider quasi-polynomial tractability. We know from [@LPW2] that ${\rm APP}$ is quasi-polynomially tractable iff there exists a positive $\delta$ such that $$\label{quasicondition}
\sup_{d\in\N}\ \frac{\sum_{j=1}^\infty\lambda_{d,j}^{1-\delta/\ln_+d}}
{\left(\sum_{j=1}^\infty\lambda_{d,j}\right)^{1-\delta/\ln_+d}}<\infty,$$ where $\ln_+d=\max(1,\ln\,d)$.
Sufficiency. We first prove that implies with $\delta=\tfrac12$. Let $$\lambda(j,k)=(2j-1)^{-2(r_k+1)}.$$ We have $$\sup_{d\in\N}\ \frac{\sum_{j=1}^\infty\lambda_{d,j}^{1-\frac1{2\ln_+d}}}
{\left(\sum_{j=1}^\infty\lambda_{d,j}\right)^{1-\frac1{2\ln_+d}}}=
\sup_{d\in\N}\ \prod_{k=1}^d
\frac{ \sum_{j=1}^\infty \lambda(j,k)^{1-\frac{1}{2\ln_+d}}}
{\left( \sum_{j=1}^\infty\lambda(j,k)\right)^{1-\frac1{2\ln_+d}}} .$$ We split the last product into two products $$\Pi_1(d) := \prod_{k=1}^d
\left(\sum_{j=1}^\infty \lambda(j,k)\right)^{\frac{1}{2\ln_+d}}$$ and $$\Pi_2(d):= \prod_{k=1}^d
\frac{ \sum_{j=1}^\infty \lam(j,k)^{1-\frac{1}{2\ln_+d}} }
{ \sum_{j=1}^\infty \lam(j,k) }.$$ In what follows we use $C$ to denote a positive number which is independent of $d$ and $\{r_k\}$, and whose value may change for successive estimates. For $\Pi_1(d)$ we simply have $$\begin{aligned}
\Pi_1(d) &=& \prod_{k=1}^d
\left( 1+ \sum_{j=2}^\infty \lam(j,k) \right)^{\frac{1}{2\ln_+d}}
\le \exp\left( \frac{1}{2\ln_+d} \sum_{k=1}^d
\sum_{j=2}^\infty \lam(j,k) \right)
\\
&\le& \exp\left( \frac C{\ln_+d} \sum_{k=1}^d \lam(2,k) \right)
= \exp\left( \frac C{\ln_+d} \sum_{k=1}^d 3^{-2(r_k+1)} \right). \end{aligned}$$ Clearly, implies that $\sup_{d\in\N} \Pi_1(d) <\infty$.
We now turn to the product $\Pi_2(d)$. We estimate each of its factors by $$\begin{aligned}
\nonumber
\frac{ \sum_{j=1}^\infty \lam(j,k)^{1-\frac{1}{2\ln_+d}} }
{ \sum_{k=1}^\infty \lam(j,k) }
&\le&
\frac{ 1+ \lam(2,k)^{1-\frac{1}{2\ln_+d}}+\sum_{j=3}^\infty
\lam(j,k)^{1-\frac{1}{2\ln_+d}} } { 1+ \lam(2,k) }
\\ \label{suf0}
&\le& \frac{ 1+ \lam(2,k)^{1-\frac{1}{2\ln_+d}} }{ 1+ \lam(2,k) }
+ \sum_{j=3}^\infty \lam(j,k)^{1-\frac{1}{2\ln_+d}} . \end{aligned}$$
Note that if $|\ln\lam(2,k)|\le 3\ln_+d$, then $$\begin{aligned}
\frac{ 1+ \lam(2,k)^{1-\frac{1}{2\ln_+d}} } { 1+ \lam(2,k)}
&=& \frac{ 1+\lam(2,k)
\exp\left({\frac{-\ln\lam(2,k)}{2\ln_+d}}\right) } { 1+\lam(2,k)}
\\
&\le& \frac{ 1+\lam(2,k)\left({1+ \frac{C|\ln\lam(2,k)|}{\ln_+d}}\right)}
{1+ \lam(2,k)}
\\
&\le& 1+ \frac{C\lam(2,k) |\ln\lam(2,k)|} {\ln_+d},\end{aligned}$$ while if $|\ln \lam(2,k)|\ge 3\ln_+d$, then $$\frac{ 1+ \lam(2,k)^{1-\frac{1}{2\ln_+d}} } { 1+ \lam(2,k) }
\le 1+ \lam(2,k)^{1-\frac{1}{2\ln_+d}}
\le 1+ \lam(2,k)^{1/2}
\le 1+ d^{-3/2}.$$ Thus, in any case $$\label{suf1}
\frac{ 1+ \lam(2,k)^{1-\frac{1}{2\ln_+d}} } { 1+ \lam(2,k) }
\le 1+ d^{-3/2} + \frac{C\lam(2,k) |\ln\lam(2,k)|} {\ln_+d}.$$ Next, we have $$\label{suf1a}
\sum_{j=3}^\infty \lam(j,k)^{1-\frac{1}{2\ln_+d}}
\le C \lam(3,k)^{1-\frac{1}{2\ln_+d}}
= C \lam(2,k)^{\frac{\ln 5}{\ln 3} (1-\frac{1}{2\ln_+d })}.$$ We now show that implies that $\lam(2,k)=3^{-2(r_k+1)}\le C/k$. First of all note that implies that $\lim_k r_k=\infty$, so that only finitely many initial $r_k$ may be zero. Assume that $d$ is so large that $r_d\ge 1$ and $d\ge 3$. Since $(1+r_k)3^{-2r_k}$ is non-increasing, we have $$r_d 3^{-2r_d}
\le \frac1d\,\sum_{k=1}^d(1+r_k)3^{-2r_k}\le\frac{C\,\ln\,d}{d},$$ so that $3^{2r_d}\ge 3^{2r_d}/r_d\ge d/(C\,\ln\,d)$ and $$r_d\ge \frac{\ln\,d-\ln(C\,\ln\,d)}{2\,\ln\,3}\ge C_1\,\ln\,d.$$ Hence, $$\lam(2,d)=3^{-2(r_d+1)} \le \frac{r_d3^{-2r_d}}{r_d}\le
\frac{C\,\ln\,d}{r_d\,d}\le \frac{C}{C_1\,d}\,$$ as claimed. By enlarging the constant, we obtain the same inequality for [*all*]{} $d$. For $k\le d$, we then have by (\[suf1a\]) $$\label{suf2}
\sum_{j=3}^\infty\lambda(j,k)^{1-\frac1{2\ln_+d}}
\le C\, k^{-\frac{\ln(5)}{\ln(3)}(1-\frac1{2\ln_+d})}
\le C\,k^{-\frac{\ln(5)}{\ln(3)}}$$
Using $1+x\le \exp(x)$, from (\[suf0\]), (\[suf1\]), and (\[suf2\]), we obtain $$\frac{ \sum_{j=1}^\infty \lam(j,k)^{1-\frac{1}{2\ln_+d}} }
{ \sum_{j=1}^\infty \lam(j,k) }
\le \exp\left( d^{-3/2} + \frac{C\lam(2,k) |\ln\lam(2,k)|} {\ln_+d}
+ C k^{-\frac{\ln 5}{\ln 3}} \right).$$ Then it follows that $$\begin{aligned}
\Pi_2(d) &\le& \exp\left ( \sum_{k=1}^d \left( d^{-3/2} +
\frac{C\lam(2,k) |\ln\lam(2,k)|} {\ln_+d}
+ C k^{-\frac{\ln 5} {\ln 3} } \right) \right)
\\
&\le& \exp\left(\sum_{k=1}^d \left( d^{-3/2}
+ \frac{C 3^{-2r_k} (r_k+1)} {\ln_+d}
+ C k^{-\frac{\ln 5}{\ln 3} } \right)\right),\end{aligned}$$ and implies that $\sup_{d\in\N} \Pi_2(d) <\infty$. Therefore, $$\sup_{d\in\N} \Pi_1(d)\,\Pi_2(d) \le \sup_{d\in\N} \Pi_1(d) \,
\ \sup_{d\in\N} \Pi_2(d) <\infty,$$ the required property is verified, so that the quasi-polynomial tractability is proved.
Necessity. Assume now that quasi-polynomial tractability holds. We prove in [@LPW2] that quasi-polynomial tractability implies $$\label{qpolnec}
\sup_{d\in\N}\, \frac1{\ln_+d}\,\sum_{k=1}^d\sum_{j=1}^\infty
\frac{\lambda(j,k)}{\Lambda(k)}\,\ln\left(\frac{\Lambda(k)}
{\lambda(j,k)}\right)<\infty,$$ where $\Lambda(k)=\sum_{j=1}^\infty\lambda(j,k)$. Clearly, $\Lambda(k)/\lambda(j,k)>1$ so that all terms in the sums over $j$ are positive. We simplify the last condition by omitting all terms for $j\not=2$, and obtain $$\label{conse1}
\sup_{d\ge\N} \ \frac{1}{\ln_+d} \
\sum_{k=1}^d \frac{\lam(2,k)}{\Lam(k)} \,
\ln\left(\frac{\Lam(k)} {\lam(2,k)} \right) <\infty.$$ Next, since $\Lam(k)>1$ we can also omit $\ln\, \Lambda(k)$ and obtain $$\sup_{d\in\N} \ \frac{1}{\ln_+d} \
\sum_{k=1}^d \frac{\lam(2,k)}{\Lam(k)}
\, \ln\left(\frac{1} {\lam(2,k)} \right) <\infty.$$ Furthermore, since $\{\Lam(k)\}$ is non-increasing, we have $$\sup_{d\in\N} \ \frac{1}{\ln_+d} \
\sum_{k=1}^d \lam(2,k) \, \ln\left(\frac{1} {\lam(2,k)} \right)
<\infty.$$ This is equivalent to , and completes the proof. [$\Box$]{}
Proof of Theorem \[Wiener-largest\]
===================================
We represent the $r$-times integrated Wiener process $W_r$ through a white-noise integral representation $$\label{wr}
W_r(t):= \int_0^1 \frac{(t-u)_+^r}{r!}\,{\rm d}\,W(u),$$ where the integration is carried over a standard Wiener process $W$ defined over $[0,1]$. Clearly, $$\begin{aligned}
\nonumber
\E \|W_r\|_2^2&=&\sum_{j=1}^\infty\lambda_{j,r}^{\ww}=
\int_0^1K_{1,r}^{\ww}(t,t)\,{\rm d}t
\\ \label{upperlargest}
&=&\int_0^1\left(\int_0^t\frac{(t-u)^{2r}}{r!^2}\,
{\rm d}u\right)\,{\rm d}t=\int_0^1\frac{t^{2r+1}}{(2r+1)r!^2}\,{\rm d}t
= \frac 1{(2r+2)(2r+1)r!^2}.\end{aligned}$$
We now supply a lower bound on the sum $\sum_{j=2}^\infty\lambda_{j,r}^{\ww}$. To do this, we approximate $W_r$ by $$V_{r,1}(t):= t^r \, W_r(1) = \frac1{r!}
\int_0^1 t^r (1-u)^r\,{\rm d}W(u)\ \ \ \ \
\mbox{for all}\ \ \ t\in[0,1].$$ The process $V_{r,1}$ is of rank $1$ since $V_{r,1}(t):=\xi_1(\omega) \psi_1(t)$, where $\psi_1(t)=t^r/r!$ and $\xi_1(\omega)= \int_0^1 (1-u)^r dW(u)$. We now prove the following lemma.
\[lemma1\] For any $r>1$ we have $$\label{WV_rt}
\E |W_r(t)-V_{r,1}(t)|^2 \le
\frac1{r!}\,\frac{3r^2}{(2r-2)^3} \, t^{2r-2} (1-t)^2
\ \ \ \ \ \mbox{for all}\ \ \ t\in[0,1],$$ and $$\label{WV_r}
\E ||W_r-V_{r,1}||_2^2 \le \frac1{r!^2}\, \frac{6r^2}{(2r-2)^6}.$$
Before we prove the lemma, we stress that the order of the right hand side in (\[WV\_r\]) is smaller than that of $\E||W_r||_2^2$. This means that $V_{r,1}$ incorporates the essential part of $W_r$ for large $r$.
[**Proof of Lemma \[lemma1\].**]{} Let $\ed{0\le u\le t}$ be the characteristic function of $[0,t]$, i.e., $\ed{0\le u\le t}=1$ for $u\in [0,t]$ and $\ed{0\le u\le t}=0$ for $u\notin [0,t]$. We have $$\begin{aligned}
\E |W_r(t)-V_{r,1}(t)|^2
&=& \frac1{r!^2} \int_0^1
\left[ t^r(1-u)^r -(t-u)^r\ed{0\le u\le t} \right]^2\,{\rm d}u
\\
&=& \frac{t^{2r}}{r!^2} \int_0^t (1-u)^{2r}
\left[ 1 -\left( \frac{t-u}{t(1-u)}\right)^r \right]^2\,{\rm d}u
+ \frac{t^{2r}}{r!^2} \int_t^1 (1-u)^{2r}\,{\rm d}u
\\
&=& \frac{t^{2r}}{r!^2} \int_0^t (1-u)^{2r}
\left[ 1 -\left( 1- \frac{(1-t)u}{t(1-u)}\right)^r \right]^2\,{\rm d}u
+ \frac{t^{2r}}{r!^2}\int_t^1 (1-u)^{2r}\,{\rm d}u
\\
&:=&
\frac{t^{2r}}{r!^2} [ I_1+ I_2].\end{aligned}$$
For $I_1$, we use an elementary bound $0\le 1-(1-h)^r\le rh$ and get
$$\begin{aligned}
I_1 &\le& \int_0^t (1-u)^{2r} r^2 \frac{(1-t)^2u^2}
{t^2(1-u)^2}\,{\rm d}u
\\
&=& r^2 (1-t)^2 t^{-2} \int_0^t (1-u)^{2r-2} u^2\,{\rm d}u
\\
&\le& r^2 (1-t)^2 t^{-2} \int_0^\infty \exp(-(2r-2)u)
u^2\,{\rm d}u
\\
&=& \frac{2r^2}{(2r-2)^3}\ (1-t)^2 t^{-2} . \end{aligned}$$
On the other hand, $$I_2= \int_0^{1-t} v^{2r}\,{\rm d}v =\frac{(1-t)^{2r+1}}{2r+1}
\le \frac{r^2}{(2r-2)^3}\ (1-t)^2 t^{-2}.$$ By summing up we obtain $$\E |W_r(t)-V_{r,1}(t)|^2
\le \frac1{r!^2}\,\frac{3r^2}{(2r-2)^3} t^{2r-2} (1-t)^2 ,$$ as claimed in the first estimate of the lemma. The second claim is obtained by a simple integration: $$\begin{aligned}
\E ||W_r-V_{r,1}||_2^2 &=& \int_0^1 \E |W_r(t)-V_{r,1}(t)|^2\,{\rm d}t
\\
&\le& \frac1{r!^2}\,\frac{3r^2}{(2r-2)^3}
\int_0^1 t^{2r-2}(1-t)^2\,{\rm d}t
\\
&=& \frac1{r!^2}\,\frac{3r^2}{(2r-2)^3}
\int_0^1 (1-t)^{2r-2} t^2\,{\rm d}t
\\
&\le& \frac1{r!^2}\,\frac{3r^2}{(2r-2)^3}
\int_0^\infty \exp(-(2r-2)t) t^2 \,{\rm d}t
\\
&=& \frac1{r!^2}\,\frac{6r^2}{(2r-2)^6}.\end{aligned}$$ as claimed. [$\Box$]{}
From Lemma \[lemma1\] we conclude that $$\sum_{j=2}^\infty \lam_{j,r}^{\ww}=
\inf_{V\ {\rm is \ rank \ one}} \E||W_r-V||_2^2
\le \E||W_r- V_r||_2^2 \le \frac{C}{r!^2\,r^{4}}.$$
This fact and yield $$\lam_{1,r}^{\ww}= \frac{1}{r!^2}
\left( \frac{1} {(2r+2)(2r+1)}+O(r^{-4})\right),$$ as claimed in Theorem \[Wiener-largest\].
We now proceed to estimates on the second largest eigenvalue $\lambda_{2,r}^{\ww}$ for large $r$. Obviously, $$\label{uppertwoeig}
\lambda_{2,r}^{\ww}\le
\sum_{j=2}^\infty \lam_{j,r}^{\ww}=
\OO\left(\frac{1}{r!^2\,r^4}\right).$$ We now show that the last bound is essentially sharp. To do this we approximate $W_r$ by $$V_{r,2}(t):= \frac1{r!}
\int_0^1 \left[ t^r (1-u)^r - rt^{r-1}(1-t) u(1-u)^{r-1}\right]
\,{\rm d}W(u)\ \ \ \ \ \mbox{for all}\ \ \ t\in[0,1].$$ The process $V_{r,2}$ is of rank $2$ since $$V_{r,2}(t) =\xi_1(\omega) \psi_1(t) -r \xi_2(\omega) \psi_2(t),$$ where $$\begin{aligned}
\xi_1(\omega)= \int_0^1 (1-u)^r\,{\rm d}W(u)\ \ \ &\mbox{and}&\ \ \
\psi_1(t)= \frac{t^r}{r!},\\
\xi_2(\omega)= \int_0^1 u\, (1-u)^{r-1}\,{\rm d}W(u)\ \ \
&\mbox{and}&\ \ \
\psi_2(t)=\frac{t^{r-1}(1-t)}{r!}.\end{aligned}$$ Note that the term $\xi_1 \psi_1 $ coming from rank 1 approximation is dominating in the rank 2 approximation, since $$\E\,\xi_1^2 ||\psi_1||_2^2
=\int_0^1(1-u)^{2r}\,{\rm d}u \cdot \frac1{r!^2}
\cdot \int_0^1 t^{2r} dt
= \frac1{r!^2} \frac{1}{(2r+1)^2} \approx \frac1{r!^2}\, r^{-2},$$ while for the correction term $r \xi_2 \psi_2$ we have $$r^2 \E\, \xi_2^2 ||\psi_2||_2^2
=r^2 \int_0^1u^2(1-u)^{2r-2}\,{\rm d}u \cdot \frac1{r!^2}
\cdot \int_0^1 t^{2r-2} (1-t)^2dt
\approx \frac1{r!^2} r^{-4}.$$ A careful analysis shows that the second eigenvalue of the covariance operator of $V_{2,r}$ is also of order $\frac1{r!^2} r^{-4}$. In other words, there exists a positive $C$ independent of $r$ such that $$\label{k2}
\inf_{V \ {\rm is \, rank \, one}} \E||V_{r,2}-V||_2^2
\ge \frac{C}{r!^2\,r^{4}}.$$
We now estimate how well $V_{r,2}$ approximates $W_r$.
\[lemma2\] For any $r>2$ we have $$\label{WV_rt2}
\E |W_r(t)-V_{r,2}(t)|^2 \le \frac1{r!^2}\,
\frac{14 r^2 (r-1)^2}{(2r-4)^{5}} \, t^{2r-4} (1-t)^4
\ \ \ \ \ \mbox{for all}\ \ \ t\in[0,1],$$ and $$\label{WV_r2}
\E ||W_r-V_{r,2}||_2^2 \le \frac1{r!^2}\,
\frac{24\cdot 14\cdot r^2(r-1)^2}{(2r-4)^{10}}
=\OO \left(\frac1{r!^2\,r^6}\right).$$
The proofs of and repeat (mostly, but not entirely) line by line those
of Lemma \[lemma1\] but we provide them for the sake of completeness. These proofs also clearly indicate how higher order approximations can be handled. As in Lemma \[lemma1\] we again stress that the the order of the right hand side in is smaller than the rank 1 approximation error computed in (\[WV\_r\]). Therefore, rank 2 approximation $V_{r,2}$ performs much better than rank 1 approximation $V_{r,1}$ for approximation of $W_r$ when $r$ is large.
[**Proof of Lemma \[lemma2\].**]{} Let $a:=\E |W_r(t)-V_{r,2}(t)|^2$. We have $$\begin{aligned}
a&=& \frac1{r!^2} \int_0^1 \left[ t^r(1-u)^r -
r t^{r-1}(1-t) u(1-u)^{r-1} -(t-u)^r\ed{0\le u\le t}\right]^2\,{\rm d}u
\\
&=& \frac{t^{2r}}{r!^2}\int_0^t (1-u)^{2r}
\left[ 1 - \frac{ r(1-t) u} {t(1-u)} -
\left( \frac{t-u}{t(1-u)}\right)^r \right]^2\,{\rm d}u
\\
&& \quad + \frac{t^{2r}}{r!^2}
\int_t^1 \left((1-u)^{r}-\frac{r(1-t)u}{t} (1-u)^{r-1}\right)^2\,{\rm d}u
\\
&=&
\frac{t^{2r}}{r!^2}
\int_0^t (1-u)^{2r} \left[ 1-
\frac{ r(1-t) u} {t(1-u)} -\left( 1 - \frac{(1-t)u}{t(1-u)}\right)^r
\right]^2 {\rm d}u
\\
&& \quad + \frac{t^{2r}}{r!^2} \int_t^1
\left((1-u)^{r} - \frac{r(1-t)u}{t} (1-u)^{r-1}
\right)^2\,{\rm d}u =: \frac{t^{2r}}{r!^2}\,[ I_1+ I_2].\end{aligned}$$
For $I_1$, we use an elementary bound $0\ge 1-rh - (1-h)^r\ge -\frac{r(r-1)}{2}h^2$ and get $$\begin{aligned}
I_1 &\le& \int_0^t (1-u)^{2r} \left(\frac{r(r-1)}{2} \cdot
\frac{(1-t)^2u^2} {t^2(1-u)^2}\right)^2\,{\rm d}u
\\
&=& \frac{r^2(r-1)^2}{4} (1-t)^4 t^{-4}
\int_0^t (1-u)^{2r-4} u^4\,{\rm d}u
\\
&\le& \frac{r^2(r-1)^2}{4} (1-t)^4 t^{-4}
\int_0^\infty \exp(-(2r-4)u) u^4\,{\rm d}u
\\
&=& \frac{6r^2(r-1)^2}{(2r-4)^5} (1-t)^4 t^{-4} . \end{aligned}$$ On the other hand, we can give the following, rather crude, estimate for $I_2$. Note that for $u>t$ and $r>1$ we have $$\frac{r(1-t)u}{t} (1-u)^{r-1} = r\cdot
\frac{(1-t)u}{t(1-u)} \cdot (1-u)^{r} \ge (1-u)^{r}.$$ Therefore, $$\begin{aligned}
I_2 &\le& \int_t^1
\left( \frac{r(1-t)u}{t} (1-u)^{r-1} \right)^2\,{\rm d}u
\\
&\le& \frac{r^2(1-t)^4}{t^4} \int_t^1 u^2
(1-u)^{2r-4}\,{\rm d}u
\\
&\le& \frac{r^2(1-t)^4}{t^4} \int_0^\infty u^2
\exp(-(2r-4)u)\,{\rm d}u
\\
&=& \frac{2r^2(1-t)^4}{(2r-4)^3t^4} \le
\frac{8r^2(r-1)^2}{(2r-4)^5} (1-t)^4 t^{-4} . \end{aligned}$$ By summing up, we obtain $$\E |W_r(t)-V_{r,2}(t)|^2
\le \frac1{r!^2}\,\frac{14 r^2(r-1)^2}{(2r-4)^5} (1-t)^4 t^{2r-4} ,$$ as claimed in the first estimate of the lemma. The second claim is obtained by a simple integration: $$\begin{aligned}
\E\, ||W_r-V_{r,2} ||_2^2 &=& \int_0^1\,
\E |W_r(t)-V_{r,2}(t)|^2\,{\rm d}t
\\
&\le& \frac1{r!^2}\,
\frac{14r^2(r-1)^2}{(2r-4)^5} \int_0^1 (1-t)^4 t^{2r-4}\,{\rm d}t
\\
&=& \frac1{r!^2}\,
\frac{14r^2(r-1)^2}{(2r-4)^5} \int_0^1 (1-t)^{2r-4} t^4\,{\rm d}t
\\
&\le& \frac1{r!^2}\,
\frac{14r^2(r-1)^2}{(2r-4)^5}
\int_0^\infty \exp(-(2r-4)t)t^4\,{\rm d}t
\\
&=& \frac1{r!^2}\,\frac{24 \cdot 14 r^2 (r-1)^2} {(2r-4)^{10}}, \end{aligned}$$ as claimed. [$\Box$]{}
From Lemma \[lemma2\] we easily estimate $\lam_{2,r}^{\ww}$. Let $\zeta \eta_1:=\zeta(\omega)\eta_1(t)$ be the first term of Karhunen-Loève expansion for $W_r$. Then $$\begin{aligned}
\frac{C}{r!^2\,r^{4}} &\stackrel{\textrm{by}\ (\ref{k2})} {\le}&
\E\, ||V_{r,2}- \zeta\eta_1 ||_2^2
\\
&=& \E\, ||(V_{r,2}-W_r) + (W_r-\zeta\eta_1) ||_2^2
\\
&\le& 2\,\E\, ||V_{r,2}-W_r||_2^2 + 2\,
\E\,||W_r-\zeta \eta_1)||_2^2
\\
&=& 2\,\E\, ||V_{r,2}-W_r||_2^2 + 2\lam_{2,r}^{\ww}
+2\sum_{j=3}^{\infty} \lam_{j,r}^{\ww} . \end{aligned}$$ Since $V_{r,2}$ is a process of rank 2, we also have $$\label{sum3le}
\sum_{j=3}^{\infty} \lam_{j,r}^{\ww}
= \inf_{V\ {\rm of \ rank\ two}} \E\,||W_r-V||_2^2
\le \E\,||W_r-V_{r,2}||_2^2.$$ For future use, we combine this with and get $$\label{sum3ler6}
\sum_{i=3}^{\infty} \lam_{i,r}^{\ww} \le \frac{C_1}{r!^2\,r^{6}}.$$ Furthermore, (\[sum3le\]) immediately yields $$\frac{C}{r!^2\,r^{4}} \le 4\E ||V_{r,2}-W_r||_2^2 + 2\lam_{2,r}
\stackrel{\textrm{by}\ (\ref{WV_r2})} {\le}
\frac{C_1}{r!^2\,r^{6}} + 2\lam_{2,r}^{\ww}.$$ This provides a lower bound for $\lam_{2,r}^{\ww}$ and together with proves that $$\label{secondeig}
\lam_{2,r}^{\ww}=\Theta\left(\frac1{r!^2\,r^{4}}\right),$$ as claimed.
We are ready to prove the last assertion of Theorem \[Wiener-largest\]. To simplify notation, let $\lambda_{j,r}=\lambda_{j,r}^{\ww}$. We split the series $\sum_{j=3}^\infty\lambda_{j,r}$ into two pieces - a long but finite initial part and a tail. Let $M>2$ and $\tau\in[\tau_0,1]$ with $\tau_0\in(\tfrac35,1]$. Consider the initial part including $j=3,4,\dots,\lceil r^M\rceil$. Using Hölder’s inequality we obtain $$\begin{aligned}
\nonumber
\sum_{j=3}^{\lceil r^M\rceil} \lam_{j,r}^\tau
&\le&
\left(\sum_{j=3}^{\lceil r^M\rceil} \lam_{j,r}\right)^\tau
\left(\sum_{j=3}^{\lceil r^M\rceil} 1 \right)^{1-\tau}
\\ \nonumber
&\stackrel{\textrm{by}\ (\ref{sum3ler6})} {\le}&
\left(\frac{C_1}{r!^2\,r^{6}} \right)^{\tau} r^{M(1-\tau)}
=
r^{-2\tau} \left(\frac{C_1}{r!^2\,r^{4}} \right)^{\tau} r^{M(1-\tau)}
\\ \nonumber
&\stackrel{\textrm{by}\ (\ref{secondeig})} {\le}&
C \lam_{2,r}^{\tau} r^{-2\tau + M(1-\tau)}\le
C \lam_{2,r}^{\tau} r^{-2\tau_0 + M(1-\tau_0)}.\end{aligned}$$ Since $C$ can be taken independent of $\tau$, for some $h>0$ we have $$\sup_{\tau\in[\tau_0,1]}\
\frac{\sum_{j=3}^{\lceil r^M\rceil}
\lam_{j,r}^\tau}{\lambda_{2,r}^\tau}=\OO(r^{-h}),
\quad \textrm{as}\ r\to\infty,$$ as long as $$\label{Mle}
M<\frac{2\tau_0}{1-\tau_0}.$$
For the tail estimation of the eigenvalue series $\sum_{j=\lceil r^M\rceil+1}^\infty\lambda_{j,r}$ we use [*approximation numbers*]{} (or [*linear widths*]{}, in other terminology).
We need to recall the definition and few basic properties which we will use in the sequel. Let $A:B_1\to B_2$ be a bounded linear operator acting between two Banach spaces. The approximation number $a_n(A)$ for $n\ge 1$ is defined as $$a_n(A):= \inf\left\{\, \|A-A_n\|\ \
\big|\ \ \ A_n:B_1\to B_2\ \ \ \mbox{with}\ \ \
{\rm rank}(A_n)<n \ \right\}.$$ The following properties of $a_n(A)$ are well known, see [@Pi87].
- the sequence $\{a_n(A)\}_{n\in\N}$ is non-increasing,
- for the adjoint operator $A^*$ we have $$\label{anstar}
a_n(A)=a_n(A^*),$$
- multiplicative property: for $A_1:B_1\to B_2$ and $A_2:B_2\to B_3$ we have $$\label{anmult}
a_{n+m-1} (A_2A_1) \le a_{n}(A_2)\,a_{m}(A_1)\ \ \
\ \ \mbox{for all}\ \ \ n,m\in \N,$$
- if $A:H\to H$ is a self-adjoint compact operator acting for a Hilbert space $H$ with the non-increasing eigenvalues $\{\lam_n\}$ then $$\label{anH}
a_n(A)=\lam_n.$$
We will study approximation numbers for integration operators. Let $I:L_2[0,1]\to L_2[0,1]$ be the conventional integration operator $$(Ix)(t):=\int_0^t x(s)\,{\rm d}s\ \ \ \ \ \mbox{for all}
\ \ \ t\in[0,1].$$ Let $I^r$ denote the $r$-th iteration of $I$ for $r\ge1$. It is easy to check by induction that $$\begin{aligned}
(I^rx)(t)&=&\int_0^t\frac{(t-s)^{r-1}}{(r-1)!}\,x(s)\,{\rm d}s
\ \
\ \ \ \ \ \mbox{for all}\ \ \ t\in[0,1],\\
([I^r]^*x)(t)&=&\int_t^1\frac{(s-t)^{r-1}}{(r-1)!}\,
x(s)\,{\rm d}s \ \ \ \ \ \ \ \mbox{for all}\ \ \ t\in[0,1],\\
(I^r\,[I^r]^*)(t)&=&\int_0^1\left(
\int_0^{\,\min(s,t)}
\frac{\ (s-u)_+^{r-1}}{(r-1)!}\,\frac{\ (t-u)_+^{r-1}}{(r-1)!}\,{\rm d}u
\right)\,x(s)\,{\rm d}s
\ \ \ \ \ \mbox{for all}\ \ \ t\in[0,1].\end{aligned}$$ This shows that $$C_{1,r}^{\ww}=I^{r+1}\,(I^{r+1})^*.$$ We are interested in the approximation numbers of $I^r$. For $r=0$, it is well known that for some positive $C$ we have $$\label{anI1}
a_n(I) \le C\, n^{-1}\ \ \ \ \mbox{for all}\ \ \ n\in\N,$$ see [@ET], pp. 118–119. We will extend this estimate for $I^r$ with an arbitrary $r$. Although the constant we get is certainly not optimal, it suffices for our needs.
\[lemma3\] We have $$\label{anIr}
a_n(I^r) \le C^{\,r}\, (2r)^{2r}\,n^{-r}\ \ \ \ \
\mbox{for all}\ \ \ n,r\in\N,$$ where $C$ is a constant from .
[**Proof of Lemma \[lemma3\].**]{} Let $$B_p := 2^{\,p \, 2^p}\ \ \ \ \ \mbox{for all}\ \ \ p=0,1,2,\dots\,.$$ We will first prove by induction on $p$ that for any integer $p\ge 0$ we have $$\label{anind}
a_n(I^r) \le C^{\,r}\,B_p\, n^{-r} \ \ \ \ \
\mbox{for all}\ \ \ n\ge 1\ \ \mbox{and}\ \ r\in[2^{p-1}, 2^{p}].$$ For $p=0$ this fact is equivalent to . Assume that (\[anind\]) holds for some integer $p$. Take any integer $r\in [2^p,2^{p+1}]$ and write it as $r=r'+r''$ with $2^{p-1}\le r_1,r_2\le 2^p$. By using $I^r=I^{r_1} I^{r_2}$ and the multiplicative property , we get for an odd index $2n-1$ $$\begin{aligned}
a_{2n-1} (I^r) &=& a_{2n-1} (I^{r_1} I^{r_2} )
\le a_n( I^{r_1})\, a_n(I^{r_2})
\\
&\le& C^{r_1} B_p n^{-r_1} \cdot C^{r_2} B_p n^{-r_2}
= C^{r}\,B_p^2\, n^{-r}
\\
&=& C^r\,B_p^2\, 2^r\, (2n)^{-r} \le C^r\,
[ B_p^2 2^{2^{p+1} } ]\, (2n)^{-r}
\\
&=& C^r 2^ {\,2 p 2^p +2^{p+1}}\, (2n)^{-r}
= C^r \, 2^ {(p+1) 2^{p+1}}\, (2n)^{-r}
\\
&=& C^r \, B_{p+1} \ (2n)^{-r} \le C^r B_{p+1} \ (2n-1)^{-r}. \end{aligned}$$ For an even index $2n$ we simply have $$a_{2n} (I^r) \le a_{2n-1} (I^r) \le C^r B_{p+1} \ (2n)^{-r} .$$ Therefore, (\[anind\]) is proved by induction.
For $r$ and $p$ as in (\[anind\]), we have $B_p= (2^p)^{2^p}\le (2r)^{2r}$. Hence, (\[anIr\]) follows from (\[anind\]). [$\Box$]{}
We now relate approximation numbers $a_n(I^r)$ to the eigenvalues $\lambda_{j,r}$ of the operator $C_{1,r}^{\ww}=I^{r+1}(I^{r+1})^*$. We have $$\begin{aligned}
\lam_{2j,r} &\le& \lam_{2j-1,r}
\stackrel{\textrm{by}\ (\ref{anH})} {=}
a_{2j-1} (I^{r+1} (I^{r+1})^{*})
\\
&\stackrel{\textrm{by}\ (\ref{anmult})} {\le}&
a_j(I^{r+1}) a_j((I^{r+1})^{*})
\stackrel{\textrm{by}\ (\ref{anstar})} {=} a_j(I^{r+1})^2
\\
&\stackrel{\textrm{by}\ (\ref{anIr})} {\le}&
C^{2(r+1)} (2(r+1))^{4(r+1)} j^{-2(r+1)}.\end{aligned}$$ This can be written as $$\lam_{j,r}\le C_1^{\,r} \,r^{4(r+1)}\, j^{-2(r+1)}\ \ \ \ \
\mbox{for all}\ \ \ r,j\in\N.$$
Take a (small) positive $\alpha$. Consider $r$ so large that $r\ge C_1^{1/\alpha}$ and $2(r+1)\tau>1$. Then again for $\tau\in[\tau_0,1]$ we can sum up $$\begin{aligned}
\sum_{j=\lceil r^M\rceil +1}^\infty \lam_{j,r}^\tau
&\le& C_1^{\,r\tau}\,r^{4(r+1)\tau}
\sum_{j=\lceil r^M\rceil +1}^\infty j^{-2(r+1)\tau}
\\
&\le& r^{(4+\alpha)r\tau+4\tau}\int_{r^M}^\infty
x^{-2(r+1)\tau}\,{\rm d}x\\
&=& \frac{r^{r\tau(4+\alpha-2M)}\ r^{4\tau+M(1-2\tau)}}{2(r+1)\tau-1}.\end{aligned}$$ We relate the last estimate to $\lambda_{2,r}=\Theta(1/(r!^2r^4))$. Since $r!=r^{r+1/2}\,e^{-r}\,\sqrt{2\pi}(1+o(1))$ by Stirling’s formula, we have $$\lambda_{2,r}=\frac{e^{2r}}{2\pi\,r^{2r+5}}\,(1+o(1)) \ \ \ \ \ \mbox{as}
\ \ \ r\to\infty.$$ Therefore $$\sum_{j=\lceil r^M\rceil +1}^\infty\lambda_{j,r}^\tau=
\OO\left(\lambda_{2,r}^\tau\,
\frac{r^{-r\tau(2M-6-\alpha)}\ r^{9\tau+M(1-2\tau)}}
{2(r+1)\tau-1}\,e^{-2r\tau}\right)
=\OO\left(\lambda_{2,r}^\tau\,r^{-r\tau(2M-6-\alpha)}\right),$$ where the factors in the big $\OO$ notation are independent of $r,\tau$ and $\alpha$.
Assume that $M>3$. Then we can take a positive $\alpha$ such that $2M-6-\alpha>0$ and get $$\sup_{\tau\in[\tau_0,1]}\ \frac{\sum_{j=\lceil r^M\rceil +1}^\infty
\lambda_{j,r}^\tau}{\lambda_{2,r}^\tau}=\OO(r^{-h}),
\quad \textrm{as}\ r\to\infty.$$ Hence, $$\sup_{\tau\in[\tau_0,1]}\
\frac{\sum_{j=1}^\infty \lambda_{j,r}^\tau}{\lambda_{2,r}^\tau}
=\OO(r^{-h}) \quad \textrm{as}\ r\to\infty,$$ assuming holds for some $M>3$. It is easy to see that such a number $M$ exists since $\tau>\tfrac35$. This completes the proof. [$\Box$]{}
Proof of Theorem \[Wiener-thm\]
===============================
As in the Euler case, we begin with polynomial tractability. We now need to show that $${\rm PT}\ \Rightarrow\ \liminf_k\frac{r_k}{k^s}>0\ \Rightarrow
\ {\rm SPT}\ \Rightarrow \ {\rm PT}.$$
Observe that for $\lambda_{d,j}=\lambda_{d,j}^{\ww}$ and $\tau\in(0,1)$, the expression in is now $$\label{ad}
a_d:=\frac{\left(\sum_{j=1}^\infty\lambda_{d,j}^\tau\right)^{1/\tau}}
{\sum_{j=1}^\infty\lambda_{d,j}}
= \prod_{k=1}^d
\frac{\left(1+(\lambda_{2,r_k}/\lambda_{1,r_k})^\tau
+\sum_{j=3}^\infty(\lambda_{j,r_k}/\lambda_{1,r_k})^\tau\right)^{1/\tau}}
{1+\lambda_{2,r_k}/\lambda_{1,r_k}+\sum_{j=3}^\infty
\lambda_{j,r_k}/\lambda_{1,r_k}}.$$ Since $\lambda_{j,r_k}=\Theta(j^{-2(r_k+1)})$ as $j\to \infty$, with the factors in the $\Theta$ notation depending on $r_k$, then $a_d$ is finite iff $2(r_k+1)\tau>1$ for all $r_k$. Then $r_k\ge r_1$ implies that we need to consider $\tau\in( \tfrac{1}{2r_1+2},1)$.
Assume that we have polynomial tractability. Then $a_d\le C\,d^{\,q}$. Each ratio in the product (\[ad\]) is strictly larger than one. This implies that $\lim_{k\to\infty}r_k=\infty$.
Note that we can estimate $a_d$ from below by dropping the sums over $j$. Then $$\prod_{k=1}^d
\frac{\left(1+(\lambda_{2,r_k}/\lambda_{1,r_k})^\tau\right)^{1/\tau}}
{1+2\lambda_{2,r_k}/\lambda_{1,r_k}}<C\,d^{\,q}\ .$$ Taking logarithms and using the asymptotic formulas for $\lambda_{1,r_k}$ and $\lambda_{2,r_k}$ from Theorem \[Wiener-largest\] yield $$\sup_d\, \frac1{\ln_+d}\ \sum_{k=1}^d r_k^{-2\tau}<\infty.$$ Since $d\,r_d^{-2\tau}\le \sum_{k=1}^dr_k^{-2\tau}$ we get $r_d^{-2\tau}=\OO(d^{-1}\,\ln_+d)$ and there exists $\delta>0$ such that $$r_d\ge \delta\,\left(\frac{d}{\ln_+d}\right)^{1/(2\tau)}\ \ \ \ \
\mbox{for all}\ \ \ d\in\N.$$ Letting $s\in(\tfrac 12,\tfrac{1}{2\tau})$ we obtain $$\label{rWiener}
\liminf_{k\to\infty}\ \frac{r_k}{k^s}>0,$$ as claimed.
Assume now that holds for some $s>\tfrac 12$. For $\tau\in(\max(\tfrac 35,\tfrac{1}{2s}),1]$ we can use the last assertion of Theorem \[Wiener-largest\] to conclude that $$\label{adstrong}
\sup_d a_d =\prod_{k=1}^\infty
\frac{\left(1+\OO(r_k^{-2\tau})\right)^{1/\tau}} {1+\OO(r_k^{-2})}
\le \exp\left\{ \OO\left(\sum_{k=1}^\infty r_k^{-2\tau}\right)\right\}
= \exp\left\{\OO\left(\sum_{k=1}^\infty k^{-2s\tau}\right)\right\}<\infty.$$ By criterion (\[poltract\]) this implies strong polynomial and obviously polynomial tractability.
We turn to weak tractability. Assume that $\lim_{k\to\infty}r_k=\infty$. We verify the analogue of for $\tau\in(\tfrac 35,1)$. From Theorem \[Wiener-largest\] we have $$b_d:=\frac1d\,\sum_{k=1}^d\sum_{j=2}^\infty\left(\frac{\lambda_{j,r_k}^{\ww}}
{\lambda_{1,r_k}^{\ww}}\right)^{\tau}
=\frac1d\,\sum_{k=1}^d\OO\left(r_k^{-2\tau}\right)
=\OO \left(\frac1d\,\sum_{k=1}^d r_k^{-2\tau}\right).$$ Clearly, $\lim_k r_k^{-2\tau}=0$ implies $\lim_d b_d=0$, which yields weak tractability.
Let $r=\lim_{k\to\infty}r_k<\infty$. Then proceeding exactly as for the Euler case, we can show that $n^{\ww}(\eps,d)$ is an exponential function of $d$ which contradicts weak tractability and completes this part of the proof.
We finally consider quasi-polynomial tractability. The proof is similar to the proof for the Euler case and we only sketch it. We need to study and for the Wiener eigenvalues. For we take $\delta=\tfrac12$ and $\tau_0\in(\tfrac35,1)$. Let us chose $d_0$ such that $1-\tfrac{1}{2\ln\,d_0} \in[\tau_0,1]$. Then for all such $d\ge d_0$ we have $\tau_d:=1-1/(2\ln\,d)\in[\tau_0,1]$ and we can use the result on the uniform convergence presented in the last assertion of Theorem \[Wiener-largest\] with respect now to $d$. Let denote $Q_k:= \frac{\lambda_{2,r_k}}{\lambda_{1,r_k}}$. We obtain $$\begin{aligned}
\frac{\sum_{j=1}^\infty\lambda_{d,j}^{1-\delta/\ln_+d}}
{\left(\sum_{j=1}^\infty\lambda_{d,j}\right)^{1-\delta/\ln_+d}}
&=& \prod_{k=1}^d
\frac{1+Q_k^{1-\frac1{2\,\ln\,d}}+
\sum_{j=3}^\infty\left(\frac{\lambda_{j,r_k}}{\lambda_{1,r_k}}\right)^{1-
\frac1{2\,\ln\,d}}}
{\left(1+Q_k+
\sum_{j=3}^\infty\frac{\lambda_{j,r_k}}{\lambda_{1,r_k}}\right)^{1-
\frac1{2\ln\,d}}}
\\
&\le&\OO(1)\ \prod_{k=d_0}^d
\frac{1+Q_k^{1-\frac1{2\,\ln\,d}} \left(1+o(r_k^{-h})\right)}
{\left(1+Q_k\right)^{1-\frac1{2\ln\,d}}},\end{aligned}$$ with absolute constants as pre-factors in the $\OO(\cdot)$ notation.
Suppose that holds. Then $\lim_k r_k=\infty$ and $$\prod_{k=d_0}^d \left(1+Q_k\right)^{\frac1{2\ln\,d}}
\le
\exp\left(\frac{2}{\ln\,d}\,\sum_{k=d_0}^d Q_k \right)
\le
\exp\left(\frac{C}{\ln\,d}\,\sum_{k=d_0}^d r_k^{-2}\right)$$ is uniformly bounded in $d$. The factor $\prod_{k=d_0}^d
\frac{1+Q_k^{1-\frac1{2\,\ln\,d}} \left(1+o(r_k^{-h})\right)}
{1+Q_k}
$ can be analyzed exactly as for the Euler case. By using $Q_k=\Theta (r_k^{-2})$, we have $$\frac{1+Q_k^{1-\frac1{2\,\ln\,d}} \left(1+o(r_k^{-h})\right)}
{1+Q_k}
\le
1+d^{-3/2}+C\,(1+r_k)^{-2} \left(\frac{\,\ln_+r_k}{\ln\,d}+o(r_k^{-h})\right).$$ Recall that assumption yields $r_k^{-2}=\OO(\frac{\ln k}{k})$, hence $$\prod_{k=d_0}^d
\frac{1+Q_k^{1-\frac1{2\,\ln\,d}} \left(1+o(r_k^{-h})\right)}
{1+Q_k}
\le \exp\left(\sum_{k=d_0}^d \left(d^{-3/2}+C\,(1+r_k)^{-2}
\left[ \frac{\ln_+r_k}{\ln\,d} + r_k^{-h}\right] \right)\right)$$ is also uniformly bounded in $d$. This means that implies quasi-polynomial tractability.
Suppose now that quasi-polynomial tractability holds. Then we use and its consequence , which is equivalent to . This completes the proof. [$\Box$]{}
Acknowledgment {#acknowledgment .unnumbered}
==============
The work of the first and the third authors was done while they participated in the Trimester Program “Analysis and Numerics for High Dimensional Problems”, May-August 2011, in Bonn, Germany, and enjoyed warm hospitality of the Hausdorff Research Institute for Mathematics.
The work of the first author was supported by RFBR grants 10-01-00154à, 11-01-12104-ofi-m, and by Federal Focused Program 2010-1.1.-111-128-033. The work of the second and third authors was partially supported by the National Science Foundation.
[50]{}
C.-H. Chang and C.-W. Ha, The Green’s functions of some boundary value problems via Bernoulli and Euler polynomials, *Arch. Math. (Basel)* [**76**]{}, 360–365, 2001.
D. E. Edmunds and H. Triebel, *Function Spaces, Entropy Numbers and Differential Operators* Cambridge University Press, Cambridge, 1996.
F. Gao, J. Hanning and F. Torcaso, Integrated Brownian motions and exact $L_2$-small balls, *Ann. Probab.* [**31**]{}, 1320–1337, 2003.
M. Gnewuch, H. Woźniakowski, Quasi-polynomial tractability, *J. Complexity*, [**27**]{}, 312–330, 2011.
F. J. Hickernell, G. W. Wasilkowski and H. Wo[ź]{}niakowski, Tractability of linear multivariate problems in the average case setting, in: *Monte Carlo and Quasi-Monte Carlo Methods 2006*, A. Keller, S. Heinrich, H. Niederreiter (eds.), 461–494, Springer, Berlin, 2008.
M. A. Lifshits, *Gaussian Random Functions*. Kluwer, Dordrecht, 1996.
M. A. Lifshits, A. Papageorgiou and H. Woźniakowski, Average Case Tractability of non-homogeneous tensor products problems, in preparation.
A.I. Nazarov, *On the sharp constant in the small ball asymptotics of some [G]{}aussian processes under [$L_2$]{}-norm*, [*J. Math. Sci.*]{} [**117**]{} (2003), no. 3, 4185–4210.
A.I. Nazarov and Ya.Yu. Nikitin, *Exact [$L_2$]{}-small ball behavior of integrated [G]{}aussian processes and spectral asymptotics of boundary value problems*, [*Probab. Theory Relat. Fields*]{} [**129**]{} (2004), 469–494.
E. Novak and H. Woźniakowski, *Tractability of Multivariate Problems, Volume I: Linear Information*, European Mathematical Society, Zürich, 2008.
E. Novak and H. Woźniakowski, *Tractability of Multivariate Problems, Volume II: Standard Information for Functionals*, European Mathematical Society, Zürich, 2010.
E. Novak and H. Woźniakowski, *Tractability of Multivariate Problems, Volume III: Standard Information for Operators*, European Mathematical Society, Zürich, to appear.
A. Papageorgiou and H. Woźniakowski, Tractability through increasing smoothness, [*J. Complexity*]{}, [**20**]{}, 409–421, 2010.
A. Pietsch, *Eigenvalues and $s$-Numbers*, Cambridge University Press, Cambridge, 1987.
J. F. Traub, G. W. Wasilkowski and H. Woźniakowski, *Information-Based Complexity*, Academic Press, New York, 1988.
[**Authors’ Addresses:**]{}
M. A. Lifshits, Department of Mathematics and Mechanics,\
St. Petersburg State University, 198504 St. Petersburg, Russia,\
email: [email protected]
A. Papageorgiou, Department of Computer Science, Columbia University,\
New York, NY 10027, USA, email: [email protected]
H. Woźniakowski, Department of Computer Science, Columbia University,\
New York, NY 10027, USA, and\
Institute of Applied Mathematics and Mechanics, University of Warsaw,\
ul. Banacha 2, 02-097 Warszawa, Poland, email: [email protected]
[^1]: It may happen that $a_{\ee}=\infty$. Then the second term in the maximum defining $p^{\,\rm str-avg-E}$ is zero.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the Distant Red Galaxy (DRG, $J-K_s>2.3$) neighbour population of Quasi Stellar Objects (QSOs) selected from the Sloan Digital Sky Survey (SDSS) in the redshift range $1\leq z\leq 2$. We perform a similar analysis for optically obscured AGNs (i.e. with a limiting magnitude $I > 24$) detected in the mid-infrared (24 $\mu$m) with the Spitzer Space Telescope and a mean redshift $z\sim 2.2$ in the Flamingos Extragalactic Survey (FLAMEX). Both QSOs and obscured AGN target samples cover 4.7 deg$^2$ in the same region of the sky. We find a significant difference in the environment of these two target samples. Neighbouring galaxies close to QSOs tend to be bluer than galaxies in optically obscured source environments. We also present results on the cross–correlation function of DRGs around QSOs and optically faint mid-infrared sources. The corresponding correlation length obtained for the QSO sample targets is $r_0$=$5.4\pm1.6$ Mpc h$^{-1}$ and a slope of $\gamma$=$1.94\pm0.10$ . For the optically obscured galaxy sample we find $r_0$=$8.9\pm1.4$ Mpc h$^{-1}$ and a slope of $\gamma$=$2.27\pm0.20$. These results indicate that optically faint obscured sources are located in denser environment of evolved red galaxies compare to QSOs.'
author:
- |
\
$^1$Grupo de Investigaciones en Astronomía Teórica y Experimental, IATE\
Observatorio Astronómico, Universidad Nacional de Córdoba\
Laprida 854, X5000BGR, Córdoba, Argentina.\
$^2$ Secretaría de Ciencia y Técnica de la Universidad Nacional de Córdoba.\
$^3$Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Avenida Rivadavia 1917, C1033AAJ, Buenos Aires, Argentina.\
title: 'The distant red galaxy neighbour population of $1\leq z \leq 2$ QSOs and optically obscured sources[^1]'
---
cosmology: large-scale structure of Universe–galaxies: galaxies: high-redshift–quasars: general
Introduction
============
Studies of large samples of distant galaxies are fundamental to provide a deeper insight in the formation and evolution of galaxies and systems of galaxies, such as clusters and groups. Using measurements of galaxy clustering at $z\sim 1-2$ one can test the predictions of cosmological models of structure formation and evolution [@kauff]. In the last years diverse photometric selection techniques have been developed in order to select high redshift galaxies. In particular the Lyman break technique is an ideal method for selecting a large number of distant galaxies and for studying the large-scale distribution and properties of star-forming systems at high redshifts from multicolour optical data [@steidel96; @madau] . This technique requires a high rest-frame far ultraviolet luminosity, and so will preferentially select galaxies with recent or unobscured active star-formation [@steidel99]. A new near-infrared selection technique has been developed in recent years to select samples of galaxies at high redshifts. The Distant Red Galaxies (DRGs, hereafter) colour–cut criterion ($J-K_s>2.3$, Vega system) [@saracco; @franx; @van] is expected to select galaxies with prominent rest frame optical breaks, caused by the 3625 Å Balmer-break or the 4000 Å Ca II H+K break. The Balmer discontinuity at 3625 Å is strongest in A-type stars and the 4000 Å break is characteristic of coolers stars with types later than G0 and strong in giant and supergiant stars [@forster]. van Dokkum et al. (2003) found that the DRG criterion selects galaxies with rest-frame optical colours similar to those of normal nearby galaxies. Using deep mid-infrared observations with IRAC on the Spitzer Space Telescope, @labbe found that 70% of the DRGs are best described by dust-reddened star forming models and 30% are very well fit with old stellar population, passive star evolution galaxy models. By comparison of the stellar populations of DGRs to those obtain in Lyman-break galaxies (LBGs), @labbe found that the average mass-to light ratios ($M/L_K$) of the DRGs are about three times higher than the LBGs sample, indicating that DGRs may represent massive and old galaxies, similar to those found in the local Universe. A similar study by @forster comparing LBGs and DRGs at similar redshifts and rest-frame $V$-band luminosities, shows that DRGs are older, more massive and more obscured for any given star formation history than LBGs. Recent works by @conse and @graziana show that this single near IR colour–cut selects a rather heterogeneous sample of galaxies, from distant luminous massive systems, to a significant fraction of less luminous dusty star-burst and galaxies with mixed morphology at redshifts $z\sim2$, with extended tails at $z=1$ and $z=4$.
At low redshifts, QSO environments show similar characteristics of those found in normal galaxies [@smith; @coldwell]. Moreover, @smith found that the cross-correlation function between low ($z < 0.3$) QSOs and galaxies, is consistent with the auto-correlation function of galaxies selected from the APM Galaxy Survey. Unlike radio-loud QSOs, @elli found that radio-quiet QSOs at $0.3<z<0.6$ are rarely found in high density environments as rich as Abell class 1. Recently, @serber studying the environment of $z<0.4$ luminous ($M_{i}<-$22) quasars in the SDSS area, found that they are located in higher local overdensity regions than are typical $L^*$ galaxies. However, the results for QSO environment at higher redshifts ($1\leq z\leq 2$) are contradictory. @yee found that some of the $z\sim0.6$ QSOs are located in environments as rich as those of Abell class 1 clusters. @hall reported a significant excess of faint galaxies in the fields of $z=1-2$ quasars. No excess galaxy population associated with radio-quiet QSOs are reported by @boyle at $z\sim1$. @croom found an anti-correlation between radio-quiet QSOs and galaxies at redshifts $z\sim 1-1.5$. @coil studied the clustering of galaxies around a sample of $0.7< z<1.4$ QSOs selected from the SDSS and DEEP2 surveys. They found, from a two-point cross-correlation analysis, that the local environment of QSOs is consistent with the mean environment of the full DEEP2 galaxy population and that they cluster similar to the blue, star-forming galaxies rather than the red galaxies. Their results imply that high redshift QSOs do not reside in particularly massive dark matter halos.
Using the DRGs criterion, @kaji reported a discovery of proto-clusters candidates around 6 high redshifts radio galaxies at $z\sim2.5$ on the basis of excess of $J-K_s$ colour–selected galaxies. Recently, a similar colour criterion was adopted to identify and to study the environment of very distant objects. This method uses a colour cut $i_{775}-z_{850}>1.3$ which select objects at $z\sim6$. An over-density of galaxies around the most distant radio–loud QSO (SDSS J0836+0054 at $z=5.8$) was reported by @wei, who found a surface density six times higher than the number expected from deep fields at similar redshifts. @stiavelli found an excess of $i_{775}-z_{850}>1.3$ galaxies around one of the most distant QSOs at $z=6.28$.
In this paper we investigate the population and colour distribution of DRGs, selected at bright near–IR magnitudes ($K_s<19.5$, Vega system) around high redshift QSOs ($1\leq z\leq 2$ ) and optical obscured sources ($z\sim2.2$) in the FLAMINGOS Extragalactic Survey (FLAMEX, @elston).
This paper is organised as follows: Section 2 describes the sample analysed, The DRG number counts are analysed in Section 3. In Section 4 we study the colour distribution of galaxies in different environments. We investigate the QSO–DRGs and Obscured–DRGs cross-correlation analysis in Section 5. Finally we discuss our results in Section 6.
In this work we assume a standard $\Lambda$CDM model Universe with cosmological parameters, $\Omega_{M}$=0.3, $\Omega_{\Lambda}$=0.7 and a Hubble constant of $H_0=$100 Km s$^{-1}$Mpc$^{-1}$. All magnitudes are expressed in the Vega system.
Observational Sample
====================
The FLAMINGOS Extragalactic Survey (FLAMEX, @elston) is a wide-area, near-infrared imaging survey in $J$ and $K_s$ bands within the NOAO Deep Wide-Field Survey (NDWFS) regions [@jannu]. This paper uses catalogues of the first Data Release DR1 of the northern part of the Survey (Bo[ö]{}tes field) that covers 4.7 deg$^2$ in both $J$ and $K_s$ bands [^2]. Using Monte Carlo simulations @elston found that more than 90 % of the survey region is complete to $K_s=$ 19.2, with 50 % being complete to $K_s=19.5$. A detailed description of the observing strategy and data reduction can be found in @elston. The catalogues were made using $K_s$–selected objects for each survey subset with the SExtractor package [@bertin], using dual image mode to measure the $J$–band magnitudes within the same regions. Detections in the different bands were matched if the centroids were within 1 $\arcsec$ of each other. $J-K_s$ colours were calculated using 4$\arcsec$ diameter aperture photometry, sufficiently large to avoid intrinsic aberrations that cause the PSF to vary significantly across the field and to produce robust photometry across entire subfields.
In order to reject spurious objects lying in the edges of the images, we selected objects that not overlap with another object using the SExtractor parameters ${\tt {FLAGS=0}}$ and ${\tt {WEIGHT\_MAP}}$ $<$ 0.7, in both $J$ and $K_s$ bands.
For two different reasons, no star-galaxy separation was performed. First, the variable PSF in the FLAMEX area precluded use of structural information to separate stars from extended sources and second, DRGs are of such small apparent size that considerable fraction of them would possibly be misclassified as stars and rejected from the final catalogues. As noted by @elston, and as can be seen in the colour-magnitude diagram (Figure 1), the $J-K_s$ colours alone provides a simple means of performing star-galaxy separation. The two horizontal sequences at $J-K_s$=0.4 and $J-K_s$=0.8 correspond to galactic stars with types later than G5 and earlier than K5, respectively [@finlator]. Red galaxies with $J-K_s>$2.3 are well separated from stars. We also show in Figure 1 the DRGs criterion adopted (Dashed lines). Our final catalogue consist of 7131 DRGs in a contiguous area of 4.7 degree$^2$, representing the largest sample of DRGs selected at bright magnitudes $K_s<19.5$.
![Left panel: $J-K_s>$ vs $K_s$ (Vega) colour-magnitude diagram for the FLAMEX field (Bo[ö]{}tes area). Right panel: Colour distribution. Dashed lines show the DRGs criterion adopted.[]{data-label="ic"}](fig1_low.ps){width="80mm"}
The QSO sample analysed in this work was obtained using the NED Database. [^3] We selected QSOs in the redshift range $1\leq z\leq 2$ in the FLAMEX area. The sample consists of twelve QSOs selected from the Sloan Digital Sky Survey (SDSS, Data Release 1 to 4) [@york; @richards] and one X–ray selected AGN [@puch] (See Table 2).
We also choose a sample of optically obscured galaxies discovered with the Spitzer Space Telescope, which are either optically very faint ($R>$24.5) or invisible ($R>$26) with no counterparts in the NOAO Deep Wide-Field Survey regions. The main goal of this survey is to select optically very faint galaxies but bright enough in the mid-infrared for redshift determination with the Infrared Spectrograph on Spitzer (IRS), which would not have been identified in previous optical studies. Sources with $I > 24$ mag were identified with the Multi-band Imaging Photometer (MIPS) on the Spitzer Space Telescope to a limiting 24 $\mu$m flux density above 0.75 mJy. 114 sources of the total sample met this optical criterion. Out of these, 17 sources have no optical or have very faint optical counterparts. Rejecting sources lying in the edges of the images, where noise level are high and signal to noise values are low, we finally selected 6 sources located within the FLAMEX area, with a mean redshift distribution $z\sim2.2$. The redshift range of the selected sources are similar to that of the redshift distribution of DRGs with $K_s<20.15$ (Vega magnitudes) found by @graziana. Redshifts from this sample were derived primarily from strong silicate absorption features [@houck]. We also selected another two sources with feasible redshift determinations derived by a weak silicate absorption feature, based on fits of redshifted spectral templates of local Ultra-luminous Infrared Galaxies (ULIRGS) from @weedman. Full details of the sample selection and redshift determination is given in @houck [@weedman].
As noted by @houck template fits suggest that most of the optically obscured sources are dominated by an AGN component, i.e. a steeply rising spectra and varying levels of silicate absorption, similar to the infrared spectra of known local AGNs. By comparing bolometric luminosities of submillimeter selected galaxies, such as those detected in SCUBA or MAMBO suryeys, @houck found that the obscured source sample seems to represent sources that have a hotter dust component than typical submillimeter-selected galaxies, assuming that mid-infrared luminosities scale to bolometric luminosities similar than in template sources which fit the mid-infrared spectra. Using this scaling, the obscured sources have implied infrared luminosities between 6$\times10^{12}$ L$_{\sun}$ and 6$\times10^{13}$ L$_{\sun}$, similar to bright QSOs at $z\sim2$ [@hop].
![Differential number counts as a function of $K_s$ magnitude for the DRG sample (Filled circles). The error bars plotted represent Poissonian errors (3 $\sigma$). We include the DRG number counts obtained in the literature. Open triangles represent data from the UKIDSS Ultra Deep Survey Early Data Release (Foucaud et al. 2006, private communication) and open circles are data obtained in the GOODS-MUSIC Sample, @graziana.[]{data-label="ic"}](fig2.ps){width="80mm"}
Galaxy number counts
====================
We have computed the differential number counts of DRGs per unit area as a function of $K_s$ magnitude which are shown in Figure 2. We use SExtractor ${\tt {MAG\_AUTO}}$ magnitudes for this calculation and estimate error bars using 3 $\sigma$ uncertainties estimated using Poissonian errors of the raw galaxy counts (See Table 1). We compare our determinations with $K-$band number counts for DRGs obtained by @graziana in the GOODS–MUSIC Sample and with estimates obtained in the UKIDSS Ultra Deep Survey Early Data Release (Foucaud et al. 2006, private communication). We have done this comparison using Vega magnitudes, assuming a transformation $K_{Vega}=K_{AB}-1.85$ [@cool]. For $K_s>$17.5 our counts are systematically higher than those obtain by Foucaud et al. This discrepancy can be explained by cosmic variance because of the small area (0.62 deg$^2$) of the Early Data Release of the UKIDSS Ultra Deep Survey.\
\[t0\] Table 1. DRG number counts as a function of $K_s$ band. The columns give the central bin magnitude, the raw counts, counts in units of arcmin$^{-2}$mag$^{-1}$ and 3 $\sigma$ uncertainties estimated using Poissonian errors.
------- ------ --------------- -------------------
(1) (2) (3) (4)
$bin$ $N$ log($\Sigma$) Error (3$\sigma$)
17.25 28 $-$2.466 0.0019
17.75 160 $-$1.709 0.0046
18.25 756 $-$1.035 0.0100
18.75 2248 $-$0.561 0.0173
19.25 3939 $-$0.318 0.0229
19.75 3944 $-$0.317 0.0230
------- ------ --------------- -------------------
![Percentage of galaxies with $J-K_s>2.3$ vs projected distance for QSOs (Open triangles) and optically obscured galaxies (Filled triangles). Error bars were estimated using jackknife re–sampling techniques.[]{data-label="ic"}](fig3.ps){width="80mm"}
Colour distribution of galaxies around QSOs and optically obscured sources
==========================================================================
In order to study the neighbour population of galaxies in these environments, we calculate the colour distribution of neighbouring galaxies according to the different target distances. In Figure 3, we can see the percentage of DRGs respect to $15.5<K_s<19.5$ galaxies vs projected distance for two different target samples. QSO targets are represented with open triangles and optically obscured galaxies are represent with filled triangles. Error bars were estimated using $jackknife$ re–sampling techniques [@efron]. As can be seen, in the vicinity of these targets ($r < 0.3$ Mpc h$^{-1}$) the distribution of colours of galaxies is significantly different. Neighbouring galaxies close to QSOs tend to be bluer than galaxies in optically obscured source environments. For these objects, the fraction of galaxies with $J-K_s>2.3$ decreases from 20% to 5%.
QSO-DRGs and OBSCURED-DRGs CROSS-CORRELATION ANALYSIS
=====================================================
In this section we analyse the relative spatial clustering of DRGs and QSOs and optically obscured galaxies. In order to obtain the cross–correlation length $r_0$, we first determine the projected cross–correlation function using QSOs as targets and the DRG population in their fields as tracer galaxies. We have done the same computation using optically obscured galaxies as targets and DRGs as tracer galaxies.
We use the Peebles estimator [@peebles80] of the projected cross-correlation function:
$$\omega(\sigma)=\frac{n_{R}}{n_{G}}\frac{QG(\sigma)}{QR(\sigma)}-1,$$
where $n_G$ and $n_R$ are the numbers of DRGs in the sample and in a random sample respectively, $QG(\sigma)$ is the number of real QSO-DRGs (or OBS-DRGs, using obscured galaxy targets) pairs separated by a projected distance in the range $\sigma$, $\sigma+\delta \sigma$, and $QR(\sigma)$ are the corresponding pairs when considering the random galaxy sample.
![Distance distribution of $K_s<19.5$ DRGs selected from the GOODS survey [@graziana] with photometric redshift measurements. Dashed lines shows the selection function obtained by fitting a Gaussian function.[]{data-label="ic"}](fig4.ps){width="80mm"}
We have estimated the cross-correlation length taking into account the selection function of the galaxy survey as a function of the distance to the targets (QSOs or optically obscured galaxies, respectively). Assuming a power-law model for the real space cross correlation function it is found:
$$\omega(\sigma)=B \sqrt{\pi} \frac{\Gamma \left[(\gamma-1)/2\right] }{\Gamma(\gamma/2)}\frac{r_{0}^{\gamma}}{\sigma^{\gamma-1}},
\label{lil}$$
where the constant $B$ depends on the differences of the selection function of the different targets and tracers.
When the distance of these targets are known, the constant $B$ is given by [@lilje]:
$$B=\frac{\sum_{i} N(y_i)} {\sum_{i}\frac{1}{y_{i}^2}\int_{0}^{\infty}N(x)x^2dx},
\label{b}$$
where $N(y_i)$ is the selection function of the galaxy survey, $y_i$ is the distance to target $i$ and the sum extends over all QSO (or obscured source) targets in the sample.
We model the selection function using the distance distribution of DRGs in the GOODS Survey ([@grazianb]) for galaxies with photometric redshifts and $K_s$ magnitudes $< 19.5$. We convert measured redshifts to comoving distances along the line of sight assuming a flat $\Lambda$CDM model Universe with the same parameters quoted in the Introduction. In Figure 4 we show the distance distribution of DRGs in the GOODS area with $K_s<19.5$ and with photometric redshift determinations. We model the distance distribution with a Gaussian function centred at $y$=2850 Mpc h$^{-1}$ and with a standard deviation of $\sigma_y$= 600 Mpc h$^{-1}$.
![Projected cross–correlation function between QSOs and DRGs in the redshift range $1\leq z\leq2$ (Filled circles) and between optically obscured galaxies with $z\sim2.2$ and DRGs with $K_s<19.5$ (Open circles). Error bars correspond to 1 $\sigma$ uncertainty estimated using the jackknife technique [@efron].[]{data-label="ic"}](fig5.ps){width="80mm"}
In Figure 5 we show the obtained projected cross–correlation function between QSOs and DRGs in the redshift range $1\leq z\leq2$ (Filled circles) and between optically obscured galaxies with $z\sim2.2$ and DRGs with $K_s<19.5$ (open circles). The error bars were estimated using the $jackknife$ technique [@efron]. The corresponding cross–correlation length derived for the QSO sample targets is $r_0$=$5.4\pm1.6$ Mpc h$^{-1}$ and a slope of $\gamma$=$1.94\pm0.10$. Using the optically obscured galaxy sample as targets we find $r_0$=$8.9\pm1.4$ Mpc h$^{-1}$ and a slope of $\gamma$=$2.27\pm0.20$. We tested the accuracy of this result by varying the $N(y_i)$ distribution over a reasonable range of values for $\bar y$ and $\sigma_y$. We find that the calculated $r_{0}$ is only weakly dependent on the $\bar y$ used, and is only affected at the 10% level.
Summary and discussions
=======================
We have analysed the QSOs and optically obscured sources environment using the DRG photometric technique ($J-K_s>2.3$). The sample of QSOs at $1\leq z\leq 2$ were selected from the Sloan Digital Sky Survey and the optically faint sources sample were detected in the mid-infrared with the Spitzer Space Telescope at z$\sim 2.2$ in the Flamingos Extragalactic Survey (FLAMEX). We find that there are significant differences in the environment and galaxy properties around these two different targets. We find that neighbouring galaxies next to QSOs tend to be bluer than galaxies in optically obscured source environments. The fraction of galaxies with $J-K_s>2.3$ in the vicinity ($r_p<0.3$ Mpc $h^{-1}$) of optically obscured sources decreases from 20% to 5%. These results are consistent with those of @coil who find that the local environment of $0.7<z<1.4$ QSOs (using the $3^{rd}$–nearest–neighbour surface density of surrounding DEEP2 galaxies) has a similar overdensity than that of blue galaxies, differing from red galaxy population at 2$\sigma$ significance. We also present results on the cross–correlation function of DRGs around QSOs and optically obscured sources. The corresponding cross–correlation length derived for the QSO and the DRGs is $r_0$=$5.4\pm1.6$ Mpc h$^{-1}$ with a slope of $\gamma$=$1.94\pm0.10$. Using the optically obscured galaxy sample as targets we find $r_0$=$8.9\pm1.4$ Mpc h$^{-1}$ and a slope of $\gamma$=$2.27\pm0.20$. We point out that the results of Sections 4 and 5 consistently indicate that obscured sources are located in higher density environment compared to QSOs, as shown by both, the colour distribution and the cross-correlation analysis.
It is possible that the optically obscured sources population are not rare objects. Using a QSO sample in the full Bo[ö]{}tes field, @weedman found that unobscured, classical type 1 QSOs and optically obscured sources with similar redshifts are similar in number. As noted by @weedman, sources selected at 24 $\mu$m, which are optically faint, are dominated by dusty sources with spectroscopic indicators of an obscured AGN rather than a star-burst. @houck found that these sources represent ULIRGS powered by a AGN and they have bolometric luminosities exceeding $10^{13}$ M$\sun$, if bolometric luminosities scale with mid–infrared luminosities as in the templates which fits the spectra. This value is similar to those found for bright QSOs at $z\sim2$ [@hop]. @elli have showed that luminous QSOs in rich galaxy cluster environments evolves much more quickly than those in poor environments which are present at all redshifts. It is possible that the QSO activity in optical obscured sources have vanished as a consequence of the rich environment where they are located or the luminous central engine is heavily obscured by the dust.
Few studies exists in the Literature about the spatial clustering of DRGs. @foucaud found a large correlation length of $r_0\sim12$ $h^{-1}$ Mpc for a sample of bright DRGs ($K_{AB}\sim$20.7, or $K_{Vega}$$\sim$ 18.8) in a small area of the Early data release of the UKIDSS survey. @graziana found $r_0=7.41_{-4.84}^{+3.45}$ $h^{-1}$ Mpc for DRGs with $1 < z < 2$ and $K_{AB}<$22 ($K_{Vega}< 20$). The larger correlation length found by @foucaud could be interpreted as evidence for luminosity segregation, where the most luminous DRGs are more strongly clustered. If we compare our values obtained in the cross-correlation function analysis with the mean value obtained for the DRGs in the field by @graziana, our results suggest that the sample of unobscured QSOs at high redshift analysed tend to be located in typical environment of DRGs with a transient phase in the evolution of normal galaxies. However, our cross-correlation results indicate that optically obscured sources probably represent massive galaxies located in rich environment of DRGs at high redshifts.
Acknowledgements
================
We are thankful to the referee Dr. Michael Strauss for his careful reading of the manuscript and a number of comments, which improved the paper.
This paper used public catalogues form the Flamingos Extragalactic Survey, with support from NSF grant AST97-31180 and Kitt Peak National Observatory This work was partially supported by the Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET) and the Secretaría de Ciencia y Técnica de la Universidad Nacional de Córdoba.
This research has made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society, and the Higher Education Funding Council for England. The SDSS Web Site is http://www.sdss.org/. The SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, Cambridge University, Case Western Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), The Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory, and the University of Washington.
Bertin, E., & Arnouts, S. 1996, [[*A&AS*]{}]{}, 117, 393
Boyle, B. J., & Couch, W. J. 1993, [[*MNRAS*]{}]{}, 264, 604
Coil et al. ApJ letters, in press
Coldwell, G. V., & Lambas, D. G. 2006, [[*MNRAS*]{}]{}, 371, 786
Conselice et al. ApJ letters, in press
Cool, R. J., et al. 2006, [[*AJ*]{}]{}, 132, 823
Croom, S. M., & Shanks, T. 1999, [[*MNRAS*]{}]{}, 303, 411
Ellingson, E., Yee, H. K. C., & Green, R. F. 1991, [[*ApJ*]{}]{}, 371, 49
Elston, R. J., et al. 2006, [[*ApJ*]{}]{}, 639, 816
Efron, B., 1982, The Jackknife, the Bootstrap and Other Resampling Plans, Philadelphia: SIAM.
Finlator, K., et al. 2000, [[*AJ*]{}]{}, 120, 2615
F[ö]{}rster Schreiber, N. M., et al. 2004, [[*ApJ*]{}]{}, 616, 40
Foucaud et al. 2006, [[*MNRAS*]{}]{}, submitted
Franx, M., et al. 2003, [[*ApJ*]{}]{}, 587, L79
Hall, P. B., & Green, R. F. 1998, [[*ApJ*]{}]{}, 507, 558
Grazian, A., et al. 2006a, [[*A&A*]{}]{}, 449, 951
Grazian, A., et al. 2006b, [[*A&A*]{}]{}, 453, 507
Hopkins, P. F., Richards, G. T., & Hernquist, L., 2006, ApJ submitted, astro-ph/0605678
Houck, J. R., et al. 2005, [[*ApJ*]{}]{}, 622, L105
Jannuzi, B. T., & Dey, A., 1999, in ASP Conf. Ser. 191, Photometric Redshifts and Detection of High Redshifts Galaxies, ed. R. Weynmann, L. Storrie Lombardi, M. Sawicki, & R. Brunner (San Francisco; ASP), 111-117
Kajisawa, M., Kodama, T., Tanaka, I., Yamada, T., & Bower, R. 2006, [[*MNRAS*]{}]{}, 371, 577
Kauffmann, G., Colberg, J. M., Diaferio, A., & White, S. D. M. 1999, [[*MNRAS*]{}]{}, 307, 529
Labb[é]{}, I., et al. 2005, [[*ApJ*]{}]{}, 624, L81
Lilje, P. B., & Efstathiou, G. 1988, [[*MNRAS*]{}]{}, 231, 635
Madau, P., Ferguson, H. C., Dickinson, M. E., Giavalisco, M., Steidel, C. C., & Fruchter, A. 1996, [[*MNRAS*]{}]{}, 283, 1388
, P. J. E., 1980, “The Large-Scale Structure of the Universe”, Princeton University Press.
Puchnarewicz, E. M., et al. 1997, [[*MNRAS*]{}]{}, 291, 177
Richards, G. T., et al. 2002, [[*AJ*]{}]{}, 123, 2945
Saracco, P., Giallongo, E., Cristiani, S., D’Odorico, S., Fontana, A., Iovino, A., Poli, F., & Vanzella, E. 2001, [[*A&A*]{}]{}, 375, 1
Serber, W., Bahcall, N., M[é]{}nard, B., & Richards, G. 2006, [[*ApJ*]{}]{}, 643, 68
Smith, R. J., Boyle, B. J., & Maddox, S. J. 1995, [[*MNRAS*]{}]{}, 277, 270
Steidel, C. C., Giavalisco, M., Pettini, M., Dickinson, M., & Adelberger, K. L. 1996, [[*ApJ*]{}]{}, 462, L17
Steidel, C. C., Adelberger, K. L., Giavalisco, M., Dickinson, M., & Pettini, M. 1999, [[*ApJ*]{}]{}, 519, 1
Stiavelli, M., et al. 2005, [[*ApJ*]{}]{}, 622, L1
van Dokkum, P. G., et al. 2003, [[*ApJ*]{}]{}, 587, L83
Weedman, D. W., Le Floc’h, E., Higdon, S. J. U., Higdon, J. L., & Houck, J. R. 2006, [[*ApJ*]{}]{}, 638, 613
Yee, H. K. C., & Green, R. F. 1987, [[*ApJ*]{}]{}, 319, 28
York, D. G., et al. 2000, [[*AJ*]{}]{}, 120, 1579
Zheng, W., et al. 2006, [[*ApJ*]{}]{}, 640, 574
\[t0\]
--------------------------- ---------------
(1) (2)
Name redshift
SST24 1435203.99+330706.8 2.59$\pm$0.34
SST24 J142804.12+332135.2 2.34$\pm$0.28
SST24 J143358.00+332607.1 1.96$\pm$0.34
SST24 J143001.91+334538.4 2.46$\pm$0.20
SST24 J143251.82+333536.3 1.78$\pm$0.14
SST24 J143520.75+340418.2 2.08$\pm$0.21
SST24 J143026.05+331516.4 1.90$^*$
SST24 J143429.56+343633.1 2.00$^*$
--------------------------- ---------------
Table 2. Optically obscured sources sample characteristics. SST24 source name derives from discovery with the MIPS 24 $\mu$m images, redshifts derived from strong ($^*$ weak) silicate absorption features.
\[t0\]
-------------------------- ----------
(1) (2)
Name redshift
SDSS J143106.78+340910.9 1.098000
SDSS J143132.13+341417.3 1.039840
SDSS J143331.80+341532.8 0.957354
SDSS J143421.33+340446.9 1.956500
SDSS J142912.88+340959.1 2.229540
SDSS J143307.89+342315.9 1.950550
SDSS J143201.75+343526.2 1.070690
RIXOS F110–50 1.335000
SDSS J142744.44+333828.7 1.237000
SDSS J143605.08+334242.6 1.983830
SDSS J143628.09+335524.3 0.903084
SDSS J143543.72+342906.4 2.547330
SDSS J143627.79+343416.8 1.883490
-------------------------- ----------
\
Table 3. QSOs sample characteristics. Designation in IAU format and spectroscopic redshift.
[^1]: Based on observations and/or data products by the Flamingos Extragalactic Survey. FLAMINGOS was designed and constructed by the IR instrumentation group (PI: R. Elston) at the University of Florida, Department of Astronomy, with support from NSF grant AST97-31180 and Kitt Peak National Observatory
[^2]: http://flamingos.astro.ufl.edu/extragalactic/
[^3]: http://nedwww.ipac.caltech.edu/ - the NASA-IPAC Extragalactic Database
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Knowing about optimal quantum measurements is important for many applications in quantum information and quantum communication. However, deriving optimal quantum measurements is often difficult. We present a collection of results for minimum-cost quantum measurements, and give examples of how they can be used. Among other results, we show that a minimum-cost measurement for a set of given pure states is formally equivalent to a minimum-error measurement for mixed states of those same pure states. For pure symmetric states it turns out that for a certain class of cost matrices, the minimum-cost measurement is the square-root measurement. That is, the optimal minimum-cost measurement is in this case the same as the minimum-error measurement. Finally, we consider sequences of individual “local" systems, and examine when the global minimum-cost measurement is a sequence of optimal local measurements. We also a consider an example where the global minimum-cost measurement is, perhaps counter-intuitively, not a sequence of local measurements, and discuss how this is related to related to the Pusey-Barrett-Rudolph argument for the nature of the wave function.'
author:
- Petros Wallden
- Vedran Dunjko
- Erika Andersson
title: 'Minimum-cost quantum measurements for quantum information'
---
Introduction
============
The problem of finding optimal quantum measurements which decode classical information stored in quantum states, with various optimization criteria, has been studied since the very beginnings of quantum information theory [@Helstrom]. A common scenario is minimum-error measurements. Here, given a known ensemble of quantum states $\{\rho_i,\eta_i \}_i$, where $\eta_i$ is the probability with which the state $\rho_i$ appears, the task is to find a measurement which minimizes the average error probability in the result. Somewhat more generally, different types of error in the result can carry different costs according to a so-called cost matrix. The measurement which minimizes the average cost is then called the minimum-cost measurement. In a quantum communication situation, classical information $i$ could first be encoded into a quantum state $\rho_i$, after which one may want to decode it back to classical information via a quantum measurement. For example, finding relevant optimal figures of merit often plays an important role in security proofs of quantum cryptographic protocols, where an adversary tries to obtain information about a quantum state. Optimal so-called generalised quantum measurements are certainly not only of theoretical interest, but have also been experimentally realized on photons, see for example [@Clarke], on NV centres [@NVPRL], and could be realized on trapped ions or atoms with existing experimental means [@atomPOM].
Finding optimal quantum measurements is in general hard. Optimal strategies have been obtained for some special cases, with various assumptions on the initial states. For minimum-error measurements, for instance, the input states usually have to possess some kind of symmetry [@Helstrom; @mult_sym; @mirror_sym; @Chou_Hsu; @Nakahira]. An exception is the minimum-error measurement for arbitrary pure qubit states, occurring with uniform probability, which was obtained by Hunter [@min_err_qubit]. A general geometric structure of the minimum-error problem was given only recently [@geom_gener]. Minimum-cost settings have been much less studied [@Helstrom; @ErikaImperfect].
In this paper, we study both minimum-error and minimum-cost measurements, and establish a link between minimum-cost measurements for pure states and minimum-error measurements for mixed states. We then apply the general results we obtain to symmetric states, and their natural generalization, states which are sequences of (that is, tensor products of) symmetric states. Symmetric states are ubiquitous in quantum information. Quantum key distribution (QKD) using the BB84 protocol [@BB84] or coherent states [@barbosa-2002; @Sych_Leuchs], universal blind quantum computing (UBQC) [@UBQC] and quantum digital signatures (QDS) [@QDS; @Expr; @QDS_VPE; @QDS_Exp], for instance, use trains of independent symmetric states, giving rise to a tensor product structure. Optimal measurements on whole trains of states, versus measurements on individual elements, are analogous to individual and collective/coherent attacks in QKD.
The outline of this paper is as follows. We begin by proving some general results concerning minimum-cost measurements, and establish a formal equivalence between minimum-cost measurements for pure states and minimum-error measurements for mixed states. Following this, we focus on the minimum-cost problem of the so-called symmetric states, for both mixed and pure states. Finally, we explore the minimum-cost problem for states which are tensor products of individual (local) states, motivated by situations which often appear in quantum cryptographic protocols. We analyse when the local measurements are the minimum-cost, give example that the minimum-cost measurement is global and highlight a connection with the Pusey, Barrett and Rudolph (PBR) argument for the nature of the wave function [@PBR] and quantum state elimination measurements [@stevebook; @OppenUSE]. We conclude with a brief discussion.
\[Section-general-results-cost\]General results for minimum-cost measurements
=============================================================================
Suppose that some quantum states $\rho_i$ each occur with probability $\eta_i$, and that we are making a quantum measurement described by the measurement operators $\Pi_j$. We will denote the measurement by $\Pi$, and also define $B_{i,j}(\Pi)=Tr(\Pi_j \rho_i)$ as the probability to obtain result $j$ given that the state was $\rho_i$. Because probabilities have to be positive, it follows that the operators $\Pi_i$ have to be positive semi-definite. Also, since probabilities for all possible outcomes (including not obtaining a result, if this may happen) should sum to one, it holds that $\sum_i \Pi_i=\one$.
Further, suppose that obtaining result $j$ when the state was $\rho_i$ carries a cost $C_{i,j}$. The average cost of the measurement $\Pi = \{\Pi_k\}_k$, with respect to the (real) cost matrix $C = [C_{i,j}]$ is denoted $\bar C(\Pi)$ and is given by The minimum cost is obtained by minimizing this average cost over all possible POVM’s $\{\Pi_i\}$, It is well established [@Helstrom] that a minimum-cost measurement is optimal if and only if the following criteria are met:
1. $\Gamma=\sum_j\Pi_jW_j=\sum_jW_j\Pi_j$ for $W_j=\sum_i\eta_i C_{i,j}\rho_i$.
2. $\Gamma=\Gamma^\dagger$.
3. $\Pi_j(W_j-\Gamma)=(W_j-\Gamma)\Pi_j=0$ for all $j$.
4. $(W_j-\Gamma)$ is positive semidefinite for all $j$.
It can be shown that the three first conditions are equivalent to This form of the conditions was first derived by Holevo [@Holevo] and Yuen et al. [@Yuen] independently. We will refer to the criteria above, as is usually done, as the Helstrom criteria.
For minimum-cost measurements we can prove the following general properties, which we first give informally. Keeping the states $\rho_i$ and prior probabilities $\eta_i$ the same,
1. The optimal measurement remains the same if the same column is added to or subtracted from each of the columns of the cost matrix. This means that the costs associated with different outcomes, for the same prior state $\rho_i$, all shift by the same amount. The average cost will also shift by a fixed amount.
2. The average minimal cost is superadditive with respect to the cost matrix. This means that the sum of the optimal minimal costs for some cost matrices $C^1,\ldots, C^n$ is lower than the minimal cost for the cost matrix $\sum_{k=1}^{n}C^k$.
3. Increasing (decreasing) each entry of the cost matrix by a varying amount increases (decreases) the optimal minimum cost of the problem. In other words, the minimum cost is monotone under the point-wise partial order of the cost matrices.
A special class of minimum-cost problems is the well-studied minimum-error problem. In the minimum-error problem the task is to, given some fixed set of states with some prior probabilities, find the measurement (and the ensuing success probability) which, on average, minimizes the probability of an error in the result. It is easy to see that this is a special class of minimum-cost problems, for a cost matrix with elements $C_{i,j} =A - \delta_{i,j}$ for any (real) constant $A$. If we choose $A=1$, then the minimum cost $\bar C_{min}$ is the minimum-error probability. At the end of this section, we will show that there is an additional one-to-one correspondence between minimum-error measurements on mixed states and minimum-cost measurements for pure states.
Next, we will formally state and prove the above claims for minimum-cost measurements. In this paper, whenever there is addition or subtraction in matrix indices, this is understood as modular addition or subtraction. For example, if $A$ is an $N\times M$ matrix, then $A_{i+N,j+M}=A_{i,j}$.
\[constant-row\] Assume a minimum-cost problem with the cost matrix with elements $C_{i,j}$, where the states $\rho_i$ appear with the frequencies $\eta_i$. If we add (subtract) a *constant-row cost matrix* with elements $C^r_{i,j} = C^r_i$, to (from) the original cost matrix, i.e. $C^t_{i,j}=C_{i,j}\pm C^r_{i,j}$, then the following two properties hold.
- The measurement that gives the minimum cost for the problem with the cost matrix $C^t_{i,j}$ also gives the minimum cost for $C_{i,j}$. That is, the measurement that gives the minimum cost is not altered.
- The minimum cost of $C^t$ is equal to the minimum cost of $C$, shifted by the cost of the constant row matrix $\bar C^r=\sum_i \eta_i C^r_i$.
First note that a cost matrix with fixed elements in each row (a *constant-row matrix*), i.e. $C_{i,j}=C_{i,j+k}=c_i\forall k$, gives the same cost for *every* measurement, and this cost is equal to $\bar C^r=\sum_i \eta_i c_i$. This follows from Therefore, all measurements are optimal for such a minimum-cost problem. For the situation in this lemma, the total cost is given by Now it is easy to see that our lemma holds as the second (additive) term on the rightmost side of the equation above is independent of the measurement $\Pi$. Thus the changed cost matrix $C_{i,j}^t$ yields the same optimal measurement, with the minimum shifted by $\sum_i \eta_i c_i=\bar C^r$.
\[sum of cost matrices\] Let $C$ be a cost matrix such that $C_{i,j}=\sum_k C^k_{i,j}$, for some individual cost matrices $C^k, k=1, \ldots n$. Then the minimum cost induced by the cost matrix $C$ is bounded from below by the sum of the individual minimum costs induced by the individual cost matrices appearing in the sum, i.e. $\bar C_{min}\geq \sum_k \bar C^k_{min}.$
For any measurement $\Pi$ it holds that Suppose that the measurement $\Pi'$ gives the minimum cost for the total cost matrix, and that the measurements $\Pi^k$ give the minimum costs for the cost matrices $C^k$, respectively. We then have
Assume that we have a cost matrix $C=[C_{i,j}]$, and an element-wise smaller cost matrix $C^l = [C^l_{i,j}]$, with $C^l_{i,j}\leq C_{i,j}$ for all $i,j$, and an element-wise larger cost matrix $C^{u} = [C^{u}_{i,j}]$ with $C^u_{i,j}\geq C_{i,j}$ for all $i,j$. Then the minimum cost induced by the cost matrix $C$ is bounded from below by the minimum cost induced by $C^l$ and from above by the minimum cost of $C^u$. In other words, We can write $C_{i,j}=C^l_{i,j}+C^s_{i,j}$, where $C^s$ is a strictly positive cost matrix. If the cost matrix has only non-negative real elements then for any measurement $\Pi$ we have that $\bar C^s(\Pi)\geq 0$. From Lemma \[sum of cost matrices\] it follows that
Similarly, by noting that $C_{i,j}+C^s_{i,j}=C^u_{i,j}$ for some positive cost matrix $C^s$ we conclude that $\bar C_{min}\leq \bar C^u_{min}$.
We will use these lemmas in the remainder of this paper.
\[Section-error-cost-general\]Minimum-error measurements of mixed states as minimum-cost measurements of pure states
====================================================================================================================
Here we point out an equivalence between minimum-error measurements for mixed states and minimum-cost measurements for pure states. Using the results in this subsection, we will then in the next section provide analytic bounds on the minimum-error probabilities of a wide class of mixed states, and also, for some special cases, give analytical expressions for the minimum-error probability.
As we have noted, a minimum-error measurement is simply a minimum-cost measurement for distinguishing between the same set of states, with a cost matrix given by $C_{i,j}=1-\delta_{i,j}$. Suppose that we are interested in the minimum-error problem where the input states are a collection of $N$ mixed states $\{\rho_i\}$, appearing with respective frequencies $\{ \eta_i \}_i$, of the form where $a_{i,m}$ are $N\times N$ coefficients such that $\sum_m a_{i,m}=1$, and $\{\ket{\psi_1},\cdots,\ket{\psi_N} \}$ are $N$ pure states. Then, the minimum-error measurement minimizes the expression This minimum-error problem for the $N$ mixed states in (\[eq:mixedstates\]), occurring with prior probabilities $\eta_i$, is equivalent to a minimum-cost problem for the $N$ equiprobable pure states $\{\ket{\psi_j} \}_j$, with the cost matrix (note the inverse order of indices in $C_{m,i}$). This can be seen from the following derivation, $$\begin{aligned}
P_{err}(\Pi)&=&1-\sum_i\eta_iTr(\Pi_i\rho_i)=1-\sum_i\eta_iTr(\Pi_i\sum_ma_{i,m}\ket{\psi_m}\bra{\psi_m})\nonumber\\
& = & 1/N\left(\sum_{i,m}Tr(\Pi_i\ket{\psi_m}\bra{\psi_m})-\sum_{i,m}N\eta_ia_{i,m}Tr(\Pi_i\ket{\psi_m}\bra{\psi_m})\right)\nonumber\\
& = & 1/N\sum_{i,m}C_{m,i}Tr(\Pi_i\ket{\psi_m}\bra{\psi_m})=C(\Pi),\end{aligned}$$ where $C(\Pi)$ is the cost of the measurement corresponding to the POVM $\{\Pi\}$ for the pure states $\{\ket{\psi_m}\}$ with equal prior probabilities $1/N$ and cost $C_{m,i}$ that is defined in Eq. (\[cost\_error\]). This shows that for *any* measurement (POVM) $\{\Pi\}$, the cost for the considered pure states is the same as the error probability for the mixed states. It follows that the minimum-cost measurement for the pure states will also be the minimum-error measurement for the mixed states, with prior probabilities as stated above. Another thing to note regarding Eq. (\[cost\_error\]) is that in the case where $a_{i,j}=\delta_{i,j}$, it reduces to the usual formula for a minimum-error measurement on pure states.
Minimum-cost measurements for pure symmetric states
===================================================
In this section we will consider minimum-cost measurements for pure symmetric states. In the section following this one, we will use these results to obtain the minimum-error probability for certain classes of mixed states, in particular, for mixed states that are mixtures of pure symmetric states. We will first consider the Square Root Measurement (SRM), which is known to be the minimum-error measurement for pure symmetric states. We will express the success probability of the SRM (that is, the minimum-error measurement) as a function of the eigenvalues of the Gram matrix of the states we are considering. Following this, we will extend the minimum-error problem to a minimum-cost problem and prove that for certain class of cost matrices, the SRM is the minimum-cost measurement. We will then apply the results of the previous sections to provide bounds for the minimum cost in an example, for four symmetric coherent states with equal amplitude but different phases.
Let $U$ be a unitary such that $U^N = I \label{p1}$. We define $\ket{\psi_i} = U^i \ket{\psi_0}$ for some $\ket{\psi_0}$. The $N$ states $\{\ket{\psi_0},\cdots,\ket{\psi_{N-1}}\}$ are called [*symmetric*]{}, and we will call $U$ the *symmetry unitary*. We furthermore assume that the prior probabilities for the states are equal i.e. $\eta_i=1/N$. We define which is the probability that outcome $j$ is obtained, using the measurement $\{\Pi\}$, if the state sent was $\rho_i$. We can then rewrite the cost as where we have used $\eta_i=1/N$.
SRM measurement of symmetric states
-----------------------------------
The square root measurement is known to be the minimum-error measurement for many cases, such as for pure symmetric states [@Helstrom], for pure multiply symmetric states [@mult_sym] and for a certain class of mixed states [@Chou_Hsu] where at least one state has strictly positive coefficients when written in the symmetry operator eigenbasis. In the present paper, we will show that this measurement is important for a much wider range of cases, involving minimum-cost measurements and minimum-error measurements for certain mixed states (exact conditions will be given later). We will also show how it is possible to bound the minimum cost and minimum-error probabilities for even more cases. If we define then the square root measurement is defined by where The Gram matrix of the states we are trying to distinguish between is defined as since $(U^{i})^\dagger$ is the unique inverse of $U^i$, and therefore $(U^{i})^\dagger = U^{N-i}=U^{-i}.$ A matrix is circulant if $A_{i,j}=A_{i+k,j+k}$ where the addition is taken modulo $N$. The Gram matrix of the symmetric states is circulant, since it depends only on the difference $(j-i)$.
We should also note that we can write $U$ as where $\{|\gamma_k\rangle\}_D$ is an orthonormal basis and $D$ is the dimension of the space spanned by the $\ket{\psi_i}$. We therefore have $\bra{\gamma_k}\gamma_{k'}\rangle=\delta_{k,k'}$. Note, that in general $N\neq D$, and it is important to keep track of in what range each index is defined. For the special case of linearly independent symmetric states, $N=D$ and the derivations simplify. By expressing $\ket{\psi_0}$ in terms of $\ket{\gamma_k}$, we obtain We can then express the Gram matrix $G$ which is $N\times N$ matrix, in terms of a matrix $M$ which is $D\times N$ matrix, where The columns of $M$ are representations of the $\ket{\psi_i}$’s in the $\ket{\gamma_k}$ basis. We have The Gram matrix, being circulant, can be diagonalised with the unitary discrete fourier transform $F$, and therefore where $\Lambda$ is a diagonal matrix with the eigenvalues $\lambda_k$ of $G$ on the diagonal. With the above definitions we can see that which leads to In the derivation above we used the fact that $\sum_j [ \exp (2 \pi I (i-k)j /N)]=N\delta_{i,k}$. We can now rewrite the initial states $\ket{\psi_i}$ in terms of the eigenvalues of the Gram matrix, In the basis of the $\ket{\gamma_k}$, the average operator $\Phi$ in Eq. (\[eq:phi\]) becomes $$\begin{aligned}
\Phi &=& 1/N \sum_{k=0}^{N-1} \sum_{i,j=0}^{D-1} \sqrt{\lambda_i \lambda_j} \exp (2 \pi I i k /N) \exp (-2 \pi I j k /N) \ket{\gamma_i}\bra{\gamma_j} \nonumber\\
&=& 1/N \sum_{k=0}^{N-1} \sum_{i,j=0}^{D-1} \sqrt{\lambda_i \lambda_j} \exp (2 \pi I (i-j)k /N) \ket{\gamma_i}\bra{\gamma_j}\nonumber\\&=&\sum_{i=0}^{D-1} \lambda_i \ket{\gamma_i}\bra{\gamma_i},\end{aligned}$$ where we in the last step used the fact that $\lambda_i$ are all non-negative. In this basis, the average operator is thus diagonal and the elements on the diagonal are the first $D$ eigenvalues of the Gram matrix. Since the first $D$ eigenvalue are non-zero (are related with the $D$-coefficients $b_i$ from eq. (\[e-values\]), which can be taken to be non-zero), the inverse in this basis is diagonal with elements $1/\lambda_i$. Therefore Eq. (\[SRM-states\]) becomes which are the DFT transformed $\ket{\gamma_k}$’s. We now obtain It is worth mentioning that the operator $B$ with the matrix elements $B_{i,j}$ is both circulant and symmetric.
The cost of making the SRM, for a cost matrix $C_{i,j}$, is given by We will see later that under certain circumstances this is also the minimum cost. For now, let us assume that the prior probabilities are equal, $\eta_i=1/N$, and that the cost matrix is circulant, i.e. that the matrix elements obey $C_{i,i+k}=C_{j,j+k}=\sum_k c_k \delta_{k,j-i}$. We then obtain The minimum-error probability, which is the cost for $C_{i,j}=1-\delta_{i,j}$, i.e. $c_k=1-\delta_{k,0}$, becomes
When is the SRM the minimum-cost measurement?
---------------------------------------------
In this section we will investigate under what conditions the minimum-cost measurement for $N$ symmetric states is the SRM, with a minimum cost given by Eq. (\[symmetric\_min\_cost\_circulant\]). In particular, we will examine the Helstrom conditions separately, and see what sufficient conditions we can impose on the cost matrix, such that the SRM is the optimal minimum-cost measurement. For circulant and symmetric cost matrices, the three first Helstrom conditions are satisfied by the SRM, as shown in supplementary material of [@Expr]. Here we will give an easier way to prove those conditions. We will then show that if the cost matrix obeys one more condition, then the last Helstrom condition, the inequality, also holds for the SRM, and thus the minimum-cost measurement for this type of cost matrices is the SRM.
\[1-3-conditions-circulant\] Let $\rho_i=\ket{\psi_i}\bra{\psi_i}$ be $N$ symmetric pure states, with equal prior probabilities $\eta_i=1/N$, and let $C_{i,j}$ be an $N \times N$ cost matrix which is circulant and symmetric. The three first Helstrom conditions for minimum-cost measurements can be re-written as the three first Helstrom conditions for a minimum-error measurement of the modified states $\rho'_i:=\sum_j C_{i,j}\rho_j$. That is, for all $i,j$. This condition holds for $\Pi_i=\ket{\phi_i}\bra{\phi_i}$ ,which is the SRM for the initially considered pure states. Eq. (\[helstrom\_1\]) becomes We can find, for every term in the first sum, a corresponding term in the second sum, so that these terms cancel. The elements of the cost matrix $C_{i,j}$ and the terms $\bra{\phi_i}\psi_j\rangle$ depend only on the difference $j-i$ of the two indices, and it also holds that Therefore, each term with a given $k$ in the first sum, will be exactly cancelled by the term with $l=i+j-k$ in the second sum (recall that addition in indices is modulo $N$). Therefore the whole sum vanishes. We can see this by first noting that What remains is to show that $C_{i,k}=C_{j,l}$ for $l=i+j-k$. This is the case because by assumption the cost matrix is both circulant and symmetric,\
We now proceed to investigate when the fourth Helstrom condition holds.
\[thm-inequality-circulant\] Consider a collection of $N$ equiprobable symmetric states $\ket{\psi_i}$. If the cost matrix $C$ is (1) symmetric, $C_{i,j}=C_{j,i}$, (2) circulant, $C_{i,i+k}=C_{j,j+k}=c_k$, (3) the coefficients $c_k$ are non-positive, $c_k\leq0\quad\forall\quad k$ and (4) the cost matrix is negative semidefinite (its eigenvalues are all non-positive), then the SRM satisfies the inequality Helstrom condition for the minimum-cost measurement for the above cost matrix. Therefore, since the first three conditions are satisfied by theorem \[1-3-conditions-circulant\], the SRM is the minimum-cost measurement. First, note that the eigenvalues of a circulant matrix are given by the discrete Fourier transform of the coefficients $c_k$. Thus the fourth condition of the above theorem reads The Helstrom inequality condition is where $\eta_i=1/N$. To prove that the operator in the LHS is positive definite, we need to prove that if we “sandwich” it with any general state $\ket{\chi}$, this always gives a positive number. We write where $\ket{\gamma_k}$ is the $D$-dimensional orthonormal basis that we used earlier, that is, the Fourier transform of the basis $\ket{\phi_i}$ of the SRM. The Helstrom inequality condition becomes We use the same definitions of $\Pi_i,\ket{\phi_i},\ket{\psi_i}$ as in the previous section. Moreover, note that since the cost matrix is circulant and symmetric, we have We call the first term of eq. (\[inequality\_circulant\]) $A$ and the second term $B$. By using the definitions we obtain where on the second line, we have used $l=j-i$ and $C_{i,i+l}=c_l$. We also obtain Writing $C_{m,j}=c_l$, where $l=m+l$, and using the fact that ${\sum_m \exp(2\pi I m(k_1-k_2) /N)=N\delta_{k_1,k_2}}$, we obtain We now take $A-B$, renaming $k_3$ as $k_2$, The above expressions followed since $C_{i,j}$ is symmetric, which implies that The fourth condition of the theorem states that $\bar c_n$, the eigenvalues of the cost matrix, are always negative. Therefore Eq. (\[inequality\_circulant\]) can further be written as (note that the remaining sums, in the following equations, take values from $k=0$ to $k=D-1$) $$\begin{aligned}
& &\sum_{k_1,k_2}|\bar c_{k_1-k_2}|\sqrt{\lambda_{k_1}\lambda_{k_2}}\left(|a_{k_1}|^2-a^*_{k_1}a_{k_2}\exp(2\pi Ii(k_1-k_2)/N)\right)\nonumber\\
& &
=\frac12\sum_{k_1,k_2}|\bar c_{k_1-k_2}|\sqrt{\lambda_{k_1}\lambda_{k_2}}
\times\nonumber\\ & &\times\left(|a_{k_1}|^2+|a_{k_2}|^2-a^*_{k_1}a_{k_2}\exp(2\pi Ii(k_1-k_2)/N)-a_{k_1}a^*_{k_2}\exp(-2\pi Ii(k_1-k_2)/N)\right)\nonumber\\& &
=\frac12\sum_{k_1,k_2}|\bar c_{k_1-k_2}|\sqrt{\lambda_{k_1}\lambda_{k_2}}
\left(|a_{k_1}|^2+|a_{k_2}|^2-2Re[a^*_{k_1}a_{k_2}\exp(2\pi Ii(k_1-k_2)/N)]\right)\nonumber\\
& &\geq\frac12\sum_{k_1,k_2}|\bar c_{k_1-k_2}|\sqrt{\lambda_{k_1}\lambda_{k_2}}
\left(|a_{k_1}|^2+|a_{k_2}|^2-2|a_{k_1}||a_{k_2}|\right)\nonumber\\
& &=\frac12\sum_{k_1,k_2}|\bar c_{k_1-k_2}|\sqrt{\lambda_{k_1}\lambda_{k_2}}(|a_{k_1}|-|a_{k_2}|)^2\geq0\end{aligned}$$ which completes the proof. Note that (a) we have multiplied the expressions with $N$, (b) in the second line we used the general property $\sum_{k_1,k_2}L_{k_1,k_2}=1/2\sum_{k_1,k_2}(L_{k_1,k_2}+L_{k_2,k_1})$, where $L$ was the full expression in the sum over $k_1,k_2$, and (c) the inequality from the third to te forth line comes from the property $Re[z_1z_2]\leq|z_1||z_2|$ of complex numbers.
To illustrate what conditions on $c_k$’s are imposed by the requirement that the eigenvalues of the cost matrix are all non-positive, we consider the case $N=4$: $$\begin{aligned}
\bar c_0&=&c_0+c_2+2c_1\nonumber\\
\bar c_1&=&c_0-c_2\nonumber\\
\bar c_2&=&c_0+c_2-2c_1\nonumber\\
\bar c_3&=&c_0-c_2=\bar c_1,\end{aligned}$$ where we have used that $c_1=c_3$. Given that $c_0,c_1,c_2,c_3\leq0$, the SRM will be the minimum-cost measurement for this cost matrix, if
Example: Bounding the minimum cost using SRM for coherent symmetric states
--------------------------------------------------------------------------
Here we will consider an example of four symmetric coherent states, given by $\{\ket{\alpha},\ket{i\alpha},\ket{-\alpha},\ket{-i\alpha}\}$, for amplitude $\alpha=2$. This symmetric set of states occurs in an implementation of quantum digital signatures [@QDS_Exp]. The choice of protocol parameters, such as signature length, in order to guarantee sufficient security, depends on the ability of a malevolent party to forge a message. This in turn depends on the minimum cost of the best measurement a malevolent party could make on all signature copies they can obtain. In finding a bound for how well signed messages can be forged, it is crucial to bound the minimum cost for a generic cost matrix (which in general comes from experimental parameters). We will give a method for how to obtain such bounds, using, as an example, a cost matrix that was actually obtained in an experiment on quantum digital signatures [^1] [@QDS_Exp]. This cost matrix is given by One can of course numerically compute the minimum cost using semi-definite programming. However, here we provide some analytical bounds using the properties we derived above, and the expressions for the SRM.
Before attempting to bound the minimum cost, we will first compute the SRM states $\ket{\phi_i}$ for this case and the corresponding minimum-error probability. The elements of the Gram matrix are given by Its eigenvalues are calculated as From this we can now write the states $\ket{\phi}$ using the Fourier orthonormal basis $\ket{\gamma_k}$, and the $B_{i,j}$ as The minimum error is then given by We now return to the minimum-cost measurement for the cost matrix in Eq. (\[cost-exp\]). In order to analytically bound the minimum cost using the methods given in the previous sections, we follow five steps.
1. We rewrite the cost matrix $C$ as sum of a constant-row matrix $C^h$ and the smallest possible non-negative remaining matrix $C'$. This is achieved by subtracting, from all elements of each row, the smallest element on that row. The cost for the constant-row matrix $C^h$ is the smallest cost one can possibly obtain, even if one knows what state is actually sent, and is given by $\bar C^h\sum_i\eta_i\min_j C_{i,j}$. For our example, the smallest cost in every row is on the diagonal. Thus the cost for $C^h$ is $\bar C^h=1/4\sum_i C_{i,i}=1.38\times10^{-4}$. We obtain the matrix
2. We further subtract the greatest fully constant matrix with $C^c_{i,j}=M$ for all $i,j$, so that the remaining cost matrix is strictly non-positive, i.e. $C'_{i,j}=M+C''_{i,j}$. This means subtracting, from all elements $C'_{i,j}$, the greatest element in that matrix. For our example, the greatest element is $1.55\times10^{-3}=M$, and this leads to (note the minus sign outside the matrix) The overall cost so far is $\bar C(\Pi)=\bar C^h+M+\bar C''(\Pi)$, where the cost of $C''$ is a function of the measurement made, and takes a negative value, since all the elements of the matrix are negative.
3. The cost of any cost matrix which is smaller, element by element, than $C''$, bounds the overall cost from below. To find the tightest bound, we look for such a matrix with the largest possible elements, which also satisfies the conditions of theorem \[thm-inequality-circulant\], so that the minimum cost is given by the SRM. For our example, the largest cost matrix which is smaller than $C''$ and is circulant, symmetric and has negative eigenvalues, is given by $C^l=\{c_0=-1.55\times10^{-3},c_1=-0.92\times10^{-3},c_2=-0.51\times10^{-3}\}$. Note that the condition for negative eigenvalues is satisfied, $c_2\geq c_0$ and $c_1\geq\frac{c_0+c_2}2$. It follows that the SRM gives the minimum cost for $C^l$, and this cost is $\bar C^l_{min}=-1.54989\times10^{-3}$. This gives a lower bound for the minimum cost of $C$,
4. Similarly, to find an upper bound, we seek a cost matrix which is larger than $C''$, element by element, which is the smallest possible matrix which also satisfies the conditions of theorem \[thm-inequality-circulant\]. This matrix is given by $C^u=\{c_0=-1.55\times10^{-3},c_1=-0.21\times10^{-3},c_2=0\}$. We can also confirm that its eigenvalues are negative, since the conditions for this are satisfied. Therefore the SRM is the minimum-cost measurement for $C^u$, with the cost $\bar C^u_{min}=-1.54978\times10^{-3}$. This leads to an upper bound for the minimum cost of $C$ as
We therefore obtain the bounds We see that these bounds are relatively tight. The minimum cost is of the order of $10^{-4}$, while the accuracy that the minimum cost is bounded by is of order $10^{-7}$. Another point to mention is that in the case the cost matrix after subtracting the constant-row $C^h$ is circulant, then it is likely that the two bounds coincide. In other words, in that case we obtain the exact minimum cost. A final point to stress here is that both the upper and lower bounds are important for different type of circumstances. If, for example, the minimum cost corresponds to the probability that some malevolent party correctly guesses the state, thereby undermining the security of some cryptographic protocol, then we are interested in the worst-case scenario, which is that he makes the best possible guess. We then use the lower bound of the minimum cost in order to make sure that our protocol is secure. If, on the other hand, some honest party is required to make the guess, then the worst case scenario corresponds to the upper bound for the minimum cost.
Minimum-error measurement and probabilities for mixed states of symmetric pure states
=====================================================================================
In the previous section we have seen that the minimum-cost measurement for a wide class of cost matrices for symmetric pure states is the SRM. More specifically, using the results of section \[Section-general-results-cost\], we see that if we can make a cost matrix circulant, with non-positive entries and negative semidefinite, by adding (subtracting) constant-row matrices, then the minimum-cost measurement is the SRM. Moreover, the cost can be easily analytically computed using the expressions for the SRM in terms of the eigenvalues of the Gram matrix of the symmetric states.
Here we will use the above result, and the equivalence between minimum-cost measurements for pure states and minimum-error measurements for mixed states, which we discussed in section \[Section-error-cost-general\], to obtain the minimum-error probability for a class of mixed states which are mixtures of pure symmetric states. We will similarly provide bounds on the minimum-error probability for a larger class of mixed states.
The first observation is that for any collection of mixed states of the form $\bar\rho_i=\sum_j a_{i,j}\ket{\psi_j}\bra{\psi_j}$, where $\ket{\psi_j}$ are symmetric states, we can rephrase any constraints (for a given measurement to be optimal) on the cost matrix with elements $C_{i,j}$ in terms of conditions on the $a_{i,j}$. In particular, assuming for simplicity that the prior probabilities $\eta_i$ of the different mixed states $\bar\rho_i$ are all equal to $1/N$, we obtain Requiring that the cost matrix $C$ is symmetric and circulant implies that the matrix with elements $a_{i,j}$ should also be symmetric and circulant, while requiring that the cost matrix $C$ is negative semidefinite, implies the requirement that $a_{i,j}$ define a positive semidefinite matrix. The results in the previous section imply that if the states $\bar\rho_i$ are such that the $a_{i,j}$ define a circulant, symmetric and positive definite matrix, then the SRM is the minimum-error measurement for the mixed states $\bar\rho_i$’s.
An interesting thing to point out is that mixed states generated by a circulant, symmetric matrix $a_{i,j}$, from pure symmetric states, are also symmetric states, induced by the same symmetry unitary. We can see that, since We have therefore shown that the SRM is the minimum-error measurement, even for mixed symmetric states defined as above, provided the eigenvalues of $a_{i,j}$ are non-negative. This is in agreement with the result of ref. [@Chou_Hsu].
An other interesting consequence concerns the case where the mixed states are arbitrary mixtures of symmetric states. In other words, when the matrix defined by $a_{i,j}$ is more general. As we have outlined in the example in the previous section, we are able to provide upper and lower bounds for the minimum error of those mixed states (given by the minimum cost for the corresponding pure states), using the explicit and easy form of the SRM for symmetric pure states. In particular, if those bounds are accurate, compared to other significant parameters that may interest us, then we can use the bounds provided by the SRM to estimate the minimum-error probability for the mixed states.
Finally, one should note that if the analytical form of the minimum-cost measurement for some class of cost matrices is known (as in our examples the SRM), then one can obtain bounds for the minimum error for a related class of mixed states using the methods we described.
\[Section-tensor-cost\]Minimum cost for sequences of states
===========================================================
In this section we will consider tensor products of states. In particular, we will focus on a special case, which is important for quantum cryptography. The Hilbert space is a tensor product of identical Hilbert spaces $\hil_{tot}=\otimes_{i=1}^L \hil_i$. We refer to the whole state as global, and the individual states as local. The set of possible states that we are going to consider consists of all (tensor product) combinations of the $N$ different local states, for the $L$ different subsystems that make the global state.
Such states occur frequently in quantum information science. The local states comprise an alphabet of possible quantum “letter" states, whereas the total tensor product state form a quantum “message". Such states occur in, for example, QKD, where the total system Alice sends to Bob is a sequence of $L$ local states. In BB84, the local states belong to two mutually unbiased bases. Appropriately ordered, the states form a set of symmetric states. Analogous situations occur in quantum digital signatures, universal blind quantum computing, and other protocols. Considering the entire global system, as opposed to individual components, which was the topic of previous sections, leads to collective (or coherent) measurement strategies which can be uses to gain information about the system.
To each individual local system one can assign a local minimum-cost problem, which is the situation we discussed previously. From the collection of local problems, one can derive a global minimum-cost problem, where the global cost is some function of local costs. A typical example of this is the scenario in which a party wishes to identify the message sent, in a way which minimizes the number of local states for which a misidentification occurred. In this paper we will consider the more general case of global cost matrices where the cost for each global state is some (general) function of the sum of the (local) costs of the subsystems. We further assume that the local cost matrices are all identical for the different subsystems. This type of systems and cost matrices are widely used. The question of whether the optimal measurement is a tensor product of local measurements, in scenarios where the possible states are tensor product states, was crucial in the development of QKD. The optimal measurement for obtaining the parity of a bit string, in the context of QKD, was examined by Fuchs and Graaf in [@FG] and by Bennett, Mor and Smolin in [@BMS]. It turns out that whether or not a sequence of local measurements is optimal depends on the global cost matrix, that is, on the specific global cost function. In particular, it was shown that the parity of a string of bits, encoded in qubits as in the BB84 protocol, can be best guessed by measuring in an entangled basis. The parity of a string is equal to addition modulo 2 of the bits, and the global cost becomes a function of the local costs. It may seem counter-intuitive that entangled measurements outperform local ones, since the possible states are all tensor products, and there are no correlations between individual bits or qubits in the example with the parity. However, for correctly determining the parity of the string of bits in this example, there is no optimum “local” measurement strategy. If we obtain the correct bit value after measuring the first qubit, then the best strategy is to guess the second bit correctly. But if we have guessed the first bit wrong, it is beneficial to make another mistake for the second one so that the parity is guessed correctly. The overall cost is a periodic function of the sum of the local costs. What is more surprising, however, is that even if the global cost is a monotone function of the local costs, then it is still not guaranteed that the global optimal measurement is non-entangled. In this section we will first prove that for a total cost matrix which is a linear function of the sum of the local costs, the minimum-cost measurement is a tensor product of local measurements. We will then provide bounds for total costs which are convex and concave functions of the sum of local costs. Finally we will give an example of a monotone function, a step function, for which the minimum-cost measurement is a measurement in an entangled basis. This example is interesting for various reasons. First, this type of cost matrix appears in protocols for QDS. Second, it is closely related to conclusive state elimination [@stevebook; @OppenUSE]. Third, this type of measurement is the one used to argue that an epistemic view of the wavefunction is impossible [@PBR]. We should introduce some notation here. The total number of local subsystems is $L$, and we call the global space of all subsystems $\Omega$. We label the global possible states as $\rho^{tot}_k=\otimes_i \rho_{k(i)}$. We will use the index $k$ for the global space, that is, it takes $N^L$ different values. To refer to different such global states, we will use subscripts (e.g. $k_1,k_2,\cdots$). When we want to refer to the state of a particular subsystem, e.g. the $i$’th, we will write $k(i)$. We assume that each subsystem is identical, and has $N$ different possible states. The states of the subsystems are independent of each other, so that the prior probabilities for the global states can be written as products $\eta^{tot}_k=\prod_i \eta_{k(i)}$. Note that $\sum_{k(i)}\eta_{k(i)}=1$ for all $i$, since the probabilities of each subsystem sum to one.
The cost matrices we are considering have entries of the form $C_{k_1,k_2}= f(\sum_i C^i_{k_1(i),k_2(i)})$. $C_{k_1, k_2}$ is the cost of choosing outcome $k_2$ if the global state was $\rho_{k_1}$. $C^i_{k_1(i),k_2(i)}$ are the entries of the local cost matrices. The cost of a global measurement corresponding to a POVM $\Pi$ with elements $\{\Pi_{k}\}$ is given by Indices in the above take values from one to $N^L$, as they will always do, unless the particular element is specified. For example, $k_1(i)$ is the index for the $i$’th subsystem, in the sequence $k_1$. The task here is to find under what conditions on $C_{k_1,k_2}$ (which is in our case is a function of the sum of the local costs) the minimum-cost measurement is to make optimal local minimum-cost measurements. For those cases, the value of the minimum cost can also be computed.
Cost matrix in the form of a linear function of the sum of local costs
----------------------------------------------------------------------
\[linear\_cost\] Assume a set of product states with independent prior probabilities for the subsystems. Assume that the global cost matrix $C_{k_1,k_2}$, is a linear function of the sum of some local cost matrices entries $C^i_{k_1(i),k_2(i)}$. In other words,
with $f(x)=ax+b$. Then (i) the minimum-cost measurement is the tensor product of the local minimum-cost measurements for the local costs $C^i$ and (ii) the minimum cost is given as $\bar C_{min}=a\sum_i \bar C^i_{min}+b$. In order to prove the above theorem, we first need few lemmas. \[indep\_subspace\_cost\] Consider a subset $A$ of $\Omega$ that consists of a collection of local subspaces $i\in A$ and call $\bar A=\Omega\setminus A$. The Hilbert space associated with $A$ is $\hil_A=\otimes_{i\in A}\hil_i$. Assume that the global cost matrix depends only on $i\in A$, i.e. $C_{k_1,k_2}=f(i\in A)$. Then for any global measurement $\Pi\in\hil_{tot}$, there exists another measurement of the form $\bar\Pi_{A}\otimes\one_{\bar A}$, with $\bar\Pi\in\hil_{A}$, that gives the same cost $\bar C(\Pi_\Omega)=\bar C(\bar\Pi_{A}\otimes\one_{\bar A})$. First we should note that the prior probabilities are of the form $\eta_k=\eta_{k(A)}\eta_{k(\bar A')}$, i.e. independent for $\hil_A$ and $\hil_{\bar A}$. We will prove the lemma by explicit construction. From eq. (\[tensor\_min\_cost\]) we obtain the following expression for the cost, where the subscripts for the POVMs indicate on which subsystems they act, An important thing to note is that the sums in Eq. (\[tensor\_min\_cost\]) run over all $k_1,k_2$, where we have decomposed these sums to summing over the different possibilities for the subsystems (summing over $k_1(A),k_1(\bar A),k_2(A),k_2(\bar A)$). The operator $\Pi_{k_2}$ has also been expressed as function of $k_2(A)$ and $k_2(\bar A)$, without implying that it has product structure. Finally, note that the cost matrix, by the assumptions in the lemma, depends only on the indices belonging to $A$.
By defining a POVM which acts on $\hil_A$ (note the partial trace) as it follows that the lemma holds since one can easily check that Therefore, for all possible costs, one can find a measurement acting non-trivially only on $\hil_A$, achieving that cost. In the cases described by this lemma, with no loss of generality, for any optimization we can restrict out attention to measurements acting on $\hil_A$.
\[product\_cost\_matrix\] If the cost matrix depends only on a subsystem $A$, i.e. $C_{k_1,k_2}=f(i\in A)$, and we have any measurement with a POVM of the form $\Pi_A\otimes\Pi_{\bar A}$, then the cost of the measurement is independent of the measurement on subsystem $\bar A$, that is, Since both the state and the elements of the POVM, are factorizable, the trace is simply the product of the trace of the subsystems $A,\bar A$, and Eq. (\[tensor\_min\_cost\]) becomes where we have used the fact that $C_{k_1,k_2}$ is independent of $k_1(\bar A)$ and $k_2(\bar A)$ to move the second sum in (\[eq:globalcost\]) inside the trace, and also that $\sum_{k_2(\bar A)}\Pi_{k_2(\bar A)}=\one_{\bar A}$, the trace of the density matrix is one and $\sum_{k_1(\bar A)}\eta_{k_1(\bar A)}=1$.
If the global cost is constant function ($f(x)=C$) and therefore $C_{k_1,k_2}=C$, then all measurements give same cost equal to that constant $C$.This follows from the definition of cost. \[lemma:product\] Consider a global cost matrix that is equal to the sum of the local cost matrices $C_{k_1,k_2}=\sum_i C^i_{k_1(i),k_2(i)}$ (corresponds to the case of a function $f(x)=x$ of the sum of the individual cost matrices). Then, the minimum-cost measurement is given by tensor product of local minimum-cost measurements. Moreover the minimum cost is given by $\bar C_{min}=\sum_i \bar C^i_{min}$.Eq. (\[tensor\_min\_cost\]) can be rewritten as where we defined Intuitively, each $C^i$ corresponds to a cost matrix that has no cost for any declaration for any subsystems except for subsystem $i$. The minimum cost of $C^i$ is denoted by $\bar C^i_{min}$. By noting that each $C^i$ depends only on the $i$’th element and using lemma \[indep\_subspace\_cost\], we have where $\Pi_{\Omega\setminus\{i\}}$ is any element of a POVM acting on that space, and the second equality follows from Lemma \[product\_cost\_matrix\]. Moreover, from Lemma \[sum of cost matrices\], it follows that the minimum total cost cannot be less than the sum of the minimum costs of each term in the sum. However, since for each term of the sum we have a measurement that has relevant support only on one subspace (the measurement on the remaining subsystems can be arbitrary), it is possible to have a measurement that achieves the minimum cost for all terms simultaneously, and thus the lower bound of lemma \[sum of cost matrices\] can actually be achieved. The measurement is given by the operators $\otimes_i \Pi^{min}_i$, and gives the cost $\bar C_{min}=\sum\bar C^i_{min}$.\
Note that we have shown that there exists a minimum cost measurement that is local. Since the optimal measurement is not unique, there may also be non-local measurement that achieves the same minimum cost.
Finally, it follows that Theorem \[linear\_cost\] holds from the last two lemmas and the definition of the cost matrix.
Convex, concave, monotonic and general functions
------------------------------------------------
Here we will consider bounds and statements which apply when the global cost matrix is a general function of the sum of some local costs.
Assume that we have a global cost matrix that is a convex function of the sum of some local costs $C_{k_1,k_2}=f(\sum_i C^i_{k_1(i),k_2(i)})$. Then the global minimum cost is upper bounded by the sum of local minimum costs, This follows by noting that $f((1/N)\sum_i C^i_{k_1(i),k_2(i)})\leq (1/N)\sum_i f(C^i_{k_1(i),k_2(i)})$, and by Lemma \[lemma:product\], which says that the minimum cost, for a global cost function which is a sum of local costs, is given by the sum of the local minimum costs for local cost functions $(1/N) f(C^i_{k_1(i),k_2(i)})$. The minimum cost obtained by making the local optimal measurements is therefore greater than or equal to the minimum possible cost for the cost function $C_{k_1,k_2}=f(\sum_i C^i_{k_1(i),k_2(i)})$, and thus provides an upper bound for the cost we are interested in. Assume that we have a global cost matrix which is a concave function of the sum of some local costs, $C_{k_1,k_2}=f(\sum_i C^i_{k_1(i),k_2(i)})$. Then the global minimum cost is lower bounded by the sum of local minimum costs, This again follows by noting that $f((1/N)\sum_i C^i_{k_1(i),k_2(i)})\geq (1/N)\sum_i f(C^i_{k_1(i,k_2(i))})$, and by Lemma \[lemma:product\]. The minimum cost obtained by optimal local measurements for the local costs $(1/N)f(C^i_{k_1(i),k_2(i)})$ is less or equal to the minimum cost in question, and thus provide a lower bound for this minimum cost.
Functions for which local measurements are sub-optimal, state elimination, and the PBR argument
-----------------------------------------------------------------------------------------------
In this section, we will give an example, which proves that even if the function of the local costs is monotonically increasing, the minimum cost measurement is not necessarily given by local minimum-cost measurements. We will consider a cost matrix which is a step function of the sum of the local costs. A step function is an important example, since in cryptographic protocols such as QDS [@QDS; @Expr; @QDS_VPE; @QDS_Exp], a party will accept a signed message as genuine if it contains fewer mismatches than a particular threshold. This means that achieving fewer mismatches than this threshold carries no cost, since the signed message is accepted as genuine, while exceeding the threshold has cost equal to one, since the message is rejected.
Consider a sequence of two qubits, each of them is either in the state $\ket{0}$ or in the state $\ket{+}=1/\sqrt{2}(\ket{0}+\ket{1})$. The global cost is given by a step function of the sum of the local costs, where the local cost matrices are the error probability $C_{i,j}=1-\delta_{i,j}$. In particular, we will consider the case where if both bits are wrong then the cost is one, while if only one or none of the bits are wrong, then there is no cost at all. In other words, we want to be sure that we do not make mistake for both elements, but either zero or one error is fine.
The best local measurement is clearly to perform a minimum-error measurement for each qubit. The cost for this measurement is given by which is the minimum probability of error for both (independent) elements. However, there exists a measurement in an entangled basis (we will call this the *PBR basis*), that gives a smaller cost. If we measure in the following basis, then we will *never* make two mistakes. The cost for this measurement, $\bar C(PRB)$, is therefore exactly zero.
We should make two comments here. First, this measurement basis was given by Pusey, Barrett and Rudolph (PBR) [@PBR] in an argument for proving that the nature of the wavefunction in quantum mechanics is not epistemic. Here we give a simplified version of this argument. The PBR argument started with the assumption that the wavefunction represents an epistemic distribution over some underlying different ontic states. Since the local states are non-orthogonal, they concluded that some ontic states are compatible with both $\ket{0}$ and $\ket{+}$ with some non-zero probability that is directly related to $p_{min}$. Having a pair of uncorrelated, non-interacting local states, would imply that there are some global ontic states, with probability $p_{min}^2$, that are compatible with all four possible wavefunctions $\{\ket{00},\ket{0+},\ket{+0},\ket{++}\}$. However, if one measures in the PBR basis, any outcome that is obtained is incompatible with (rules out) one of the four possible initial states. This manifest itself by the fact that the $\bar C(PBR)$ is zero. Therefore, the assumption that the wavefunction has purely epistemic character has to be rejected. It is very interesting that this deep philosophical insight is immediately connected to the security of cryptographic protocols.
The second comment is that this exact type of measurement can be understood as quantum state elimination or quantum state exclusion [@stevebook; @OppenUSE; @QDS_Exp]. Depending on which of the four possible outcomes is obtained, we can, with 100% probability, rule out one of the possible states. In particular, we can rule out the state for which both qubits are different compared with our result. This again is slightly counter-intuitive, since we started with four possible linearly independent non-orthogonal states. While it is well known that we cannot determine the state with certainty, we can rule out (eliminate) a state with certainty.
Finally, an interesting observation is that it is the inequality Helstrom condition that is expected to fail for the local measurements. In a sense, the local minimum-cost measurements corresonds to a “local minimum”, in the sense that it is optimal compared to other slight perturbations. However, there is an entangled basis which is globally optimal. In the appendix \[app\] we see that for sequences of symmetric states with a global cost matrix which is any function of the sum of the local costs, the three first conditions of Helstrom hold for local SRMs. It is the failure of inequality condition, however, that leads to an optimal measurement in an entangled basis for certain global cost functions[^2].
Summary and Conclusions
=======================
In this paper we examined minimum-cost measurements in order to obtain useful tools for quantum information and quantum communications. Knowledge of optimal measurements is important for example for bounding the ability of adversaries in cryptographic protocols to forge messages or learn about a secret key. We obtained a series of results concerning minimum-cost measurements. In particular, we showed (1) that the minimum-cost measurement remains the same if we add a constant-row cost matrix to the cost matrix, (2) one can bound the minimum cost from above (below) with an element-by-element greater (smaller) cost matrix, (3) one can bound the cost for a sum of cost matrices by the sum of the minimum costs for the individual cost matrices in the sum. We also (4) derived a formal mathematical equivalence between minimum-cost measurements for pure states and minimum-error measurements for mixtures of those pure states. Then we focused on the case of symmetric states, where we (5) derived an expression for the square-root measurement (SRM) and the minimum error for pure states in terms of the eigenvalues of the Gram matrix for the states which takes a surprisingly simple form (Eq. (\[min-error-sym\])), and (6) showed that when the cost matrix is circulant, symmetric, has negative elements and is negative semidefinite, then the SRM is the minimum-cost measurement. We (7) gave a particular example, where we obtained lower and upper bounds for the minimum cost of an arbitrary cost matrix. These results lead us to (8) obtain the minimum-error probability for mixed states which are a particular kind of mixtures of pure symmetric states, and a method to bound the minimum-error probability for a larger class of mixed states.
Finally we (9) considered sequences of (that is, tensor products of) individual systems, where the global cost is a function of the local costs. We (i) showed that if this function is linear, then a combination of local minimum-cost measurements is the global minimum-cost measurement, (ii) if the function is convex or concave we obtain bounds (upper/lower) from the local minimum cost measurements. We moreover (iii) showed that this is not the case for general functions of the local costs, even if the function is monotonic, and pointed out the connection between this, quantum state elimination measurements and the PBR argument regarding the nature of the wave function.
*Acknowledgments. Support by EPSRC grants EP/G009821/1, EP/K022717/1 and an EPSRC Doctoral Fellowship is gratefully acknowledged. PW is also partially supported by COST Action MP1006.*
\[app\]Minimum-cost measurements on tensor products of symmetric states
=======================================================================
Consider a minimum-cost measurement on a sequence of individual symmetric states. We will here show that the first three Helstrom conditions are satisfied by the local SRMs, if the local (individual) states are symmetric, for any global cost matrix that is a function of the sum of the local costs.
We consider tensor product states of symmetric local states, where the local costs are circulant and symmetric, and the global cost is some function of the sum of the local costs. We will prove that the tensor product of local SRMs satisfies the first three Helstrom conditions. However, as expected, the inequality conditions are not in general satisfied. Here we will use the same notation and terminology as in section \[Section-tensor-cost\].
Assume a global tensor product state of local pure symmetric states, and a global cost matrix that is (any) function of the sum of some local cost matrices. If the local cost matrices are circulant and symmetric, then the first three Helstrom conditions hold for the measurement with measurement which is a combination of local SRMs. We rewrite the minimum-cost measurement for the global system as a minimum-error measurement for newly defined states $\bar\rho_{k_i}=\sum_{k_j}C_{k_i,k_j}\rho_{k_j}$, with the same convention for indices as in section \[Section-tensor-cost\]. The Helstrom condition for the minimum-error measurement is then for all global states labelled by $k_1,k_2$. By assumption, the cost matrix is of the form We can view the sum of the local cost matrices as a distance of the string $k_1$ from the string $k_2$, and therefore the cost matrix is some function of the distance between the two states. The claim is that the tensor product of local SRMs satisfies Eq. (\[Helstrom\_1\_tensor\]). The global states corresponding to the SRM are of the form $\ket{\phi_k}=\otimes_i\ket{\phi_{k(i)}}$ and eq. (\[Helstrom\_1\_tensor\]) becomes To prove that this holds, it is sufficient to show that each term in the first sum cancels a term in the second sum, in a way so that the whole sum vanishes. We can explicitly show that this is the case. For any given $k_1,k_2,k_3$, choose $k_4$ so that for each element $k_4(i)=k_1(i)+k_2(i)-k_3(i)$, where the addition and subtraction is done for the labels of the local symmetric states, and is done modulo $N$. This gives a bijective map between terms in the two sums. Since the cost matrix is a function of the sum of the local cost matrices, and the local cost matrices are circulant, the total cost matrix is also circulant and therefore where the addition of global indices is understood as element by element addition modulo $N$. What remains for the proof is to show that for the choice of $k_4$ we made above. Note that i.e. these terms are circulant. The r.h.s. of Eq. (\[3.1\]) becomes using the fact that the local cost matrices are circulant. The last line is equal to the l.h.s. of Eq. (\[3.1\]) which then shows that eq. (\[Helstrom\_1\_tensor\]) holds and completes the proof.
The important thing to note is that we did not need to make any assumptions on the exact form of the global cost function. One can explicitly check that the inequality condition also holds for linear global cost functions, which is expected due to the results of section \[Section-tensor-cost\]. As we show, it turns out that this condition is often not satisfied, even for certain monotonic functions.
[99]{}
Helstrom C W 1976 , Academic Press, New York.
Clarke R B M, Kendon V M, Chefles A, Barnett S M, Riis E, and Sasaki M 2001 [*Phys. Rev. A*]{} [**64**]{} 012303
Waldherr G, Dada A C, Neumann P, Jelezko F, Andersson E and Wrachtrup J 2012 [*Phys. Rev. Lett.*]{} [**109**]{} 180501
Franke-Arnold S, Andersson E, Barnett S M, and Stenholm S 2001 [*Phys. Rev. A*]{} [**63**]{} 052301
Andersson E, Barnett S M, Gilson C R and Hunter K 2002 [*Phys. Rev. A*]{} **65** 052308
Barnett S 2001 [*Phys. Rev. A*]{} **64** 030303
K. Nakahira 2012 [*IEEE Transactions on Information Theory*]{} **58** 1215
Chou C-L and Hsu L-Y 2003 *Phys. Rev. A* [**68**]{} 042305
Hunter K 2004 [*AIP Conf. Proc.*]{} **734** 83; Hunter K 2004 [*Optimal Generalised Measurement Strategies*]{} PhD thesis, University of Strathclyde
Bae J 2013 [*New J. Phys.*]{} **15** 073037
Andersson E 2012 [*Phys. Rev. A*]{} [**86**]{} 012120
Bennett C and Brassard G 1984 *Proceedings of IEEE International Conference on Computers Systems and Signal Processing* pp. 175–179
Barbosa G A, Corndorf E, Kumar P and Yuen H P 2003 *Phys. Rev. Lett* [**90**]{} 227901
Sych D and Leuchs G 2010 [*New J. Phys.*]{} **12** 053019
Broadbent A, Fitzsimons J, and Kashefi E 2009 *Proceedings of the 50th Annual IEEE Symposium on Foundations of Computer Science* (IEEE Computer Society, Los Alamitos, USA, 2009) pp. 517–526
Gottesman D, and Chuang I 2001 preprint [*arXiv:quant-ph*]{}/0105032v2
Clarke P J, Collins R J, Dunjko V, Andersson E, Jeffers J and Buller G S 2012 *Nat. Commun.* [**3**]{} 1174
Dunjko V, Wallden P and Andersson E 2013 preprint \[arXiv:1309.1375\]
Collins R J, Donaldson R J, Dunjko V, Wallden P, Clarke P J, Andersson E, Jeffers J and Buller G S 2013 preprint \[arXiv:1311.5760\]
Pusey M F, Barrett J and Rudolph T 2012 *Nat. Phys.* [**8**]{}, 475
Barnett S 2009 [*Quantum Information*]{}, Oxford University Press, pp 103-104
Bandyopadhyay S, Jain R, Oppenheim J and Perry C 2013 preprint [*arXiv*]{}:1306.4683
Holevo A S 1973 [*J. Multivar. Anal.*]{} **3** 337
Yuen H P, Kennedy R S and Lax M 1975 [*IEEE Trans. Inform. Theory*]{} **IT-21** 125
Fuchs C A and van de Graaf J 1999 [*IEEE Trans. Inform. Theory*]{} **45** 1216
Bennett C H, Mor T and Smolin J A 1996 [*Phys. Rev. A*]{} **54** 2675
[^1]: The actual data that was used in that work was slightly different. The technique used to bound the forging probability was similar, but not identical, to the one presented here. We chose to use this data to better illustrate the use of the results presented in this paper.
[^2]: Note that even in the example with a step function, it is not always the case that global measurements outperform local ones. This depends on the particular value at which the step occurs. In the example we presented, if the step function was such that we accept only if both states are correct, then the optimum measurement would be a combination of local measurements.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A bounded operator $A$ on a real or complex separable infinite-dimensional Banach space $Z$ is *universal* in the sense of Glasner and Weiss if for every invertible [ergodic]{} [measure]{}-preserving transformation $T$ of a standard Lebesgue [probability]{} space $(X,{\mathcal{B}},\mu )$, there exists an $A$-[invariant]{} [probability]{} [measure]{} $\nu $ on $Z$ with full support such that the two dynamical systems $(X,{\mathcal{B}},\mu ;T)$ and $(Z,{\mathcal{B}}_{Z},\nu ;A)$ are isomorphic. We present a general and simple criterion for an operator to be universal, which allows us to characterize universal operators among unilateral or bilateral weighted shifts on $\ell_{p}$ or $c_{0}$, to show the existence of universal operators on a large class of Banach spaces, and to give a criterion for universality in terms of unimodular eigenvectors. We also obtain similar results for operators which are universal for all ergodic systems (not only for invertible ones), and study necessary conditions for an operator on a Hilbert space to be universal.'
address: 'CNRS, Laboratoire Amiénois de Mathématique Fondamentale et Appliquée, UMR 7352, Université de Picardie Jules Verne, 33 rue Saint-Leu, 80039 Amiens Cedex 1, France'
author:
- Sophie Grivaux
title: Some new examples of universal hypercyclic operators in the sense of Glasner and Weiss
---
[^1]
Introduction and main results {#Sec1}
=============================
Let $G$ be a topological group, and $Z$ a real or complex separable infinite-dimensional Banach space. We denote by $\mathcal{B}(Z)$ the set of bounded linear [operators]{} on $Z$. Let $S:G\rightarrow {Z}$, $g\mapsto S_{g}$ be a representation of the group $G$ by bounded [operators]{} on $Z$. If ${\mathcal{B}}{}_{Z}$ denotes the Borel $\sigma $-field of $Z$, and $\nu $ is a Borel [probability]{} [measure]{} on $Z$ which is $S$-[invariant]{} (i.e. $\nu $ is $S_{g}$-[invariant]{} for every $g$ in $G$), then $S$ naturally defines a [probability]{}-preserving action of the group $G$ on the [probability]{} space $(Z,{\mathcal{B}}_{Z},\nu )$. Recall that the measure $\nu $ is said to have full support if $\nu (U)>0$ for any non-empty open subset $U$ of $Z$.
Glasner and Weiss introduced in the paper [@GW] the following notion of a universal representation:
\[DefGW\][@GW] The representation $S=(S_{g})_{g\in G}$ of the group $G$ on the Banach space $Z$ is said to be *universal* if for every [ergodic]{} [probability]{}-preserving free action $T=(T_{g})_{g\in G}$ of $G$ on a standard Lebesgue [probability]{} space $(X,{\mathcal{B}},\mu )$, there exists a Borel [probability]{} [measure]{} $\nu
$ on $Z$ with full support which is $S$-[invariant]{} and such that the two actions of $T$ and $S$ of $G$ on $(X,{\mathcal{B}},\nu )$ and $(Z,{\mathcal{B}}_{Z},\nu )$ respectively are isomorphic.
Recall that $(T_{g})_{g\in G}$ is *free* if for any element $g\in G$ different from the identity, $\mu (\{x\in X;\, T_{g}x=x\})=0$, and *[ergodic]{}* if the following holds true: if $A\in \mathcal{B}$ is such that $T_{g}^{-1}(A)=A$ for every $g\in G$, then $\mu (A)(1-\mu (A))=0$.
A universal representation of $G$ thus simultaneously models every possible free [ergodic]{} action of $G$ on a [probability]{} space. The existence of a universal representation is shown in [@GW] for a large class of groups $G$, including all countable discrete groups and all locally compact, second countable, compactly generated groups.
When $G=\Z$, the main result of [@GW] states thus that there exists a bounded invertible [operator]{} $S$ on $H$ which is *universal* in the following sense: for every invertible [ergodic]{} [probability]{}-preserving transformation $T$ of a standard Lebesgue [probability]{} space $(X,{\mathcal{B}},\mu )$, there exists an $S$-[invariant]{} [probability]{} [measure]{} $\nu $ on $H$ with full support such that the two dynamical systems $(X,{\mathcal{B}},\mu ;T)$ and $(H,{\mathcal{B}}_{H},\nu ;S)$ are isomorphic. Observe that any invertible ergodic [probability]{}-preserving transformation $T$ of $(X, \mathcal{B},\mu)$ acts freely on $(X, \mathcal{B},\mu)$: for any $n\in\Z$, the set $\{x\in X \textrm;\ T^{n}x=x\}$ is $T$-invariant, and the ergodicity of $T$ implies that it is of $\mu$-measure zero. This definition of a universal operator is thus coherent with Definition \[DefGW\].
Any of the systems $(H,{\mathcal{B}}_{H},\nu ;S)$ is what is called a *linear dynamical system*, i.e. a system given by the action of a bounded linear [operator]{} $A$ on an infinite-dimensional separable Banach space $Z$. These systems can be studied from both the topological point of view and the ergodic point of view (when one endows the Banach space $Z$ with an $A$-[invariant]{} [probability]{} [measure]{}), and we refer the reader to the two books [@BM] and [@GP] for more on this particular class of dynamical systems.
The result of [@GW], when specialized to the case where $G=\Z$, thus says that *any* invertible ergodic [probability]{}-preserving dynamical system can be represented as a linear dynamical system where the underlying space is a Hilbert space, and, moreover, the same operator $S$ on $H$ can serve as a model for any such dynamical system. Given the apparent rigidity entailed by linearity, the universality result of [@GW] in the case where $G=\Z$ may seem rather surprising. It is worth pointing out here that a topological version of this result had been obtained previously by Feldman in [@F]: there exists a bounded operator $A$ on the Hilbert space ${\ell_{2}(\N)}$ which has the following property: whenever $\varphi $ is a continuous self-map of a compact metrizable space $K$, there exists a compact subset $L$ of ${\ell_{2}(\N)}$ which is $A$-invariant and an homeomorphism $\Phi :K\rightarrow L$ such that $\varphi =\Phi
^{-1}\circ A\circ\Phi $. The proof of this topological result is rather straightforward, but it already gives a hint at the richness of the class of linear dynamical systems.
A bounded operator $A$ acting on the Banach space $Z$ is said to be *[hypercyclic]{}* if it admits a vector $z\in Z$ whose orbit $\{A^{n}z;\ n\ge 0\}$ is dense in $Z$, and *[frequently hypercyclic]{}* if there exists a vector $z\in Z$ such that for every non-empty open subset $V$ of $Z$, the set $\{n\ge 0;\ A^{n}z\in V\}$ has positive lower density. If $A$ admits an invariant probability measure with full support with respect to which it is [ergodic]{}, then Birkoff’s [ergodic]{} theorem is easily seen to imply that almost all vectors of $Z$ are [frequently hypercyclic]{} for $A$. Thus any universal operator is [frequently hypercyclic]{}. Let us now say a few words about the construction of universal operators of [@GW].
The universal operators constructed in [@GW] are shift operators on certain weighted $\ell_{p}$-spaces of sequences on $\Z$ for $1<p<+\infty $, or, equivalently, weighted shift operators on ${\ell_{p}(\Z)}$. The proof uses in a crucial way an ergodic theorem for certain random walks of Jones, Rosenblatt, and Tempelman [@JRT]. This theorem states in particular that whenever $\eta $ is a symmetric strictly aperiodic probability measure on $\Z$, the following holds true: for any probability-preserving dynamical system $(X,{\mathcal{B}},\mu ;T)$ and any function $f\in L^{p}(X,{\mathcal{B}},\mu )$, $1<p<+\infty $, the powers $A_{\eta }^{n}f$ of the random walk operator on $\Z$ defined by $$A_{\eta }f(x)=\sum_{k\in\Z}f(T^{k}x)\,\eta (k)$$ converge for almost every $x\in X$ to the projection $P_{\mathcal{J}}f$ of $f$ onto the subspace $\mathcal{J}$ of $L^{p}(X,{\mathcal{B}},\mu )$ consisting of $T$-invariant functions. This ergodic theorem can be applied for instance starting from the measure $\eta =(\delta _{-1}+\delta _{0}+\delta_{1} )/3$ on $\Z$. If $(p_{n})_{n\ge 1}$ is a sequence of positive real numbers such that $\sum_{n\ge 1}p_{n}=1$ and $\sup\, (p_{n}/p_{n+1})<+\infty $, the weights considered in [@GW] are defined by setting $w_{k}:=\sum_{n\ge 1}p_{n}\,\eta ^{*n}(k)$ for every $k\in\Z$. If $S$ is the shift operator defined on $${\ell_{p}(\Z,w)}:=\{\xi =(\xi _{k})_{k\in\Z};\ \sum_{k\in \Z}|\xi
_{k}|^{p}\,w_{k}<+\infty
\}$$ by setting $S\xi =(\xi_{k+1} )_{k\in\Z}$ for each $\xi \in {\ell_{p}(\Z,w)}$, then $S$ is shown in [@GW] to be bounded, and the ergodic theorem of [@JRT] is then used to prove that for any function $f\in L^{2p}(X,{\mathcal{B}},\mu )$ the sequence $(f(T^{k}x))_{k\in\Z}$ belongs to ${\ell_{p}(\Z,w)}$ for $\mu $-almost every $x\in X$. Setting $$\begin{aligned}
\Phi _{f}:(X,{\mathcal{B}},\mu
)&\longrightarrow({\ell_{p}(\Z,w)},{\mathcal{B}}_{{\ell_{p}(\Z,w)}},\nu _{f})\\
x&\longmapsto (f(T^{k}x))_{k\in\Z}\end{aligned}$$ where $\nu _{f}$ is the measure on ${\ell_{p}(\Z,w)}$ defined by $\nu
_{f}(B)=\mu (\Phi _{f}^{-1}(B))$ for any Borel subset $B$ of ${\ell_{p}(\Z,w)}$, it is easy to check that $\Phi
_{f}$ intertwines the actions of $T$ on $(X,{\mathcal{B}},\mu )$ and of $S$ on ${\ell_{p}(\Z,w)}$. The last (and most difficult) step of the proof of [@GW] is then to construct a function $f$ such that $\Phi _{f}$ is an isomorphism of dynamical systems and $\nu _{f}$ has full support.
Our aim in this paper is to present an alternative construction of universal operators, which is elementary in the sense that it avoids the use of an ergodic theorem such as the one of [@JRT]. It is also more flexible than the construction of [@GW], yields some rather simple criteria for universality, and allows us to show the existence of universal operators on a large class of Banach spaces. Moreover, this construction makes it possible to exhibit operators which are universal for *all* ergodic dynamical systems, not only for invertible ones. As we will often need to make a distinction between these two notions, we introduce the following definition:
\[Def0\] Let $A$ be a bounded operator on a real or complex Banach space $Z$.
1. We say that $A$ is *universal for invertible ergodic systems* if for every invertible ergodic dynamical system $(X,{\mathcal{B}},\mu ;T)$ on a standard Lebesgue probability space there exists a probability measure $\nu $ on $Z$ with full support which is $A$-invariant and such that the dynamical systems $(X,{\mathcal{B}},\mu ;T)$ and $(Z,{\mathcal{B}}_{Z},\nu ;A)$ are isomorphic.
2. We say that $A$ is *universal for ergodic systems* if the same property holds true for all ergodic dynamical systems $(X,{\mathcal{B}},\mu ;T)$ on a standard Lebesgue probability space.
Universal operators in the sense of Glasner and Weiss are universal for invertible ergodic systems. When we use simply the term “universal operator” in the rest of the paper, we will mean an operator which is universal either for all ergodic systems or just for invertible ones. Before stating our main results, we introduce the following intuitive notation: suppose that $A$ is a bounded operator on a real or complex separable Banach space $Z$, and suppose that $(z_{n})_{n\in\Z}$ is a sequence of vectors of $Z$ such that, for every $n\in\Z$, $Az_{n}=z_{n+1}$. We then write $z_{n}=A^{n}z_{0}$ for every $n\in\Z$.
Our first result consists of a general and simple criterion for an operator to be universal for invertible ergodic systems.
\[Theo1\] Let $A$ be a bounded operator on a real or complex separable Banach space $Z$. Suppose that there exists a sequence $(z_{n})_{n\in\Z}$ of vectors of $Z$ such that, for every $n\in\Z$, $Az_{n}=z_{n+1}$, and such that the following three properties hold true:
1. the vector $z_{0}$ is bicyclic, i.e. $\overline{\vphantom{[}\operatorname{span}}\,[A^{-n}z_{0};\ n\in\Z\,]$=Z;
2. there exists a finite subset $F$ of $\Z$ such that $\overline{\vphantom{[}\operatorname{span}}\,[A^{-n}z_{0};\ n\in\Z\setminus F\,]\neq Z$;
3. the series $\sum_{n\in\Z}A^{-n}z_{0}$ is unconditionally convergent in $Z$.
Then $A$ is universal for invertible ergodic systems.
There is a very similar criterion which implies that an operator is universal for all ergodic systems:
\[Theo1bis\] Let $A$ be a bounded operator on a real or complex separable Banach space $Z$. If $A$ satisfies the assumptions of Theorem \[Theo1\], and if moreover the sequence $(z_{n})_{n\in \Z}$ is such that $A^{r}z_{0}=0$ for some $r\in\Z$ (or, equivalently, such that $z_{0}=0$), $A$ is universal for ergodic systems.
We have already mentioned that a universal operator is necessarily frequently hypercyclic. An operator satisfying the assumptions of either Theorem \[Theo1\] or Theorem \[Theo1bis\] is easily seen to satisfy the Frequent Hypercyclicity Criterion of [@BoGr] (see also [@BM] or [@GP]), and so is in particular frequently hypercyclic and chaotic.
The proofs of Theorems \[Theo1\] and \[Theo1bis\] largely rely on the ideas of [@GW], but some extra work is needed, in particular in order to cope with the condition (b) in both theorems. The proofs would be simpler if we assumed that $F=\{0\}$ (which is what happens in some of the examples, in particular in those of [@GW]), but the generality of assumption (b) is needed in several of the examples given in Section \[Sec4\].
The proofs of Theorems \[Theo1\] and \[Theo1bis\] are presented in Section \[Sec2\], as well as two generalizations of these results (Theorems \[Theo1-3\] and \[Theo1-4\]) in which assumption (a) is relaxed. The next two sections are devoted to applications and examples. In Section \[Sec3\], we characterize universal operators (both for ergodic systems and for invertible ergodic systems) among unilateral or bilateral weighted backward shifts on the spaces ${\ell_{p}(\N)}$, $1\le p<+\infty $ or ${c_{0}(\N)}$. Recall that if $(e_{n})_{n\ge 0}$ denotes the canonical basis of ${\ell_{p}(\N)}$, or ${c_{0}(\N)}$, and $(w_{n})_{n\ge 1}$ is a bounded sequence of non-zero complex numbers, the weighted backward shift ${B_{w}}$ is defined on ${\ell_{p}(\N)}$ or ${c_{0}(\N)}$ by setting ${B_{w}}e_{0}=0$ and ${B_{w}}e_{n}=w_{n}e_{n-1}$ for every $n\ge 1$. In the same way, if $(f_{n})
_{n\in \Z}$ is the canonical basis of ${\ell_{p}(\Z)}$ or ${c_{0}(\Z)}$, and $(w_{n})_{n\in\Z}$ is again a bounded sequence of non-zero complex numbers, the bilateral weighted shift $S_{w}$ on ${\ell_{p}(\Z)}$ or ${c_{0}(\Z)}$ is defined by setting $S_{w}e_{n}=w_{n}e_{n-1}$ for every $n\in\Z$. Here is the characterization of universal weighted shifts which can be obtained thanks to Theorems \[Theo1\] and \[Theo1bis\]:
\[Theo2\] With the notations above, the unilateral backward weighted shift ${B_{w}}$ is universal for (invertible) ergodic systems on ${\ell_{p}(\N)}$, $1\le p<+\infty $, if and only if the series $$\sum_{n\ge 1}\frac{1}{|w_{1}\dots w_{n}|^{p}}$$ is convergent. It is universal for (invertible) ergodic systems on ${c_{0}(\N)}$ if and only if $|w_{1}\dots w_{n}|\longrightarrow 0$ as $n\longrightarrow +\infty $.
In the same way, the bilateral backward weighted shift $S_{w}$ is universal for (invertible) ergodic systems on ${\ell_{p}(\Z)}$, $1\le p<+\infty $, if and only if the series $$\sum_{n\ge 1}\frac{1}{|w_{1}\dots w_{n}|^{p}}+\sum_{n\ge 1}{|w_{0}\dots w_{-(n-1)}|^{p}}$$ is convergent. It is universal for (invertible) ergodic systems on ${c_{0}(\Z)}$ if and only if $|w_{1}\dots w_{n}|\longrightarrow 0$ and $|w_{0}\dots w_{-(n-1)}|\longrightarrow +\infty $ as $n\longrightarrow +\infty $.
This result shows in particular the existence of universal operators for ergodic systems living on any of the spaces ${\ell_{p}(\N)}$, $1\le
p<+\infty $, or ${c_{0}(\N)}$. The existence of universal operators for invertible [ergodic]{} systems on ${\ell_{p}(\N)}$, $1< p<+\infty $, is already proved in [@GW]. A natural question, asked in [@GW], is to determine which Banach (or Fréchet) spaces support a universal operator. As a universal operator is necessarily frequently hypercyclic, and some Banach spaces (like the hereditarily indecomposable spaces, for instance), do not support frequently hypercyclic operators, it follows that not all Banach spaces support a universal operator. But, as a consequence of Theorem \[Theo1\], we obtain the existence of such operators on Banach spaces with a sufficiently rich structure.
\[Theo4\] Let $Z$ be a separable infinite dimensional Banach space containing a complemented copy of a space with a sub-symmetric basis. Then $Z$ supports an operator which is universal for all ergodic systems.
This result implies for example that any separable Banach space containing a complemented copy of one of the spaces ${\ell_{p}(\N)}$, $1\le p<+\infty $, or $c_{0}(\N)$, supports a universal operator. This is the case for all spaces ${L^{p}}(\Omega ,\mu )$, where $(\Omega ,\mu )$ is a $\sigma $-finite measured space.
If $A$ is a bounded operator on a complex infinite dimensional separable Hilbert space $H$, it is known (see [@BG1], or [@BM Ch.5]) that $A$ admits an invariant measure with respect to which it is ergodic, and which additionally has full support and admits a second order moment, if and only if its unimodular eigenvectors are *perfectly spanning*: this means that there exists a continuous probability measure $\sigma $ on the unit circle $\T=\{\lambda \in\C,\ |\lambda |=1\}$ such that for any Borel subset $B$ of $\T$ with $\sigma (B)=1$, $$\overline{\vphantom{(}\textrm{span}}\, \bigl[\,\ker(A-\lambda
),\ \lambda \in B \bigr]=H.$$ In this case, the unimodular eigenvectors of $A$ are said to be *$\sigma $-spanning*. An eigenvectorfield $E$ of $A$ is a map $E:\T\longrightarrow Z$ such that $A\,E(\lambda )=\lambda \,E(\lambda )$ for every $\lambda \in\T$. We will often be dealing in the rest of the paper with eigenvectorfields $E$ belonging to ${L^{2}}(\T,\sigma ;Z)$ where $\sigma $ is a certain probability measure on $\T$: this means that $E:\T\longrightarrow Z$ is $\sigma $-measurable, with $$\displaystyle\int_{\T}||E(\lambda
)||^{2}d\sigma (\lambda )<+\infty$$ and $A\,E(\lambda )=\lambda \,E(\lambda )$ $\sigma $-almost everywhere. When we write simply that $E$ belongs to ${L^{2}}(\T ;Z)$, this means that $\sigma $ is assumed to be the normalized Lebesgue measure $d\lambda $ on $\T$.
As a universal operator on $H$ is necessarily ergodic with respect to a certain invariant measure with full support (although this measure is not required to have a second-order moment), it is natural to look for conditions involving the unimodular eigenvectors of the operator $A$ which imply its universality. This is done in Section \[Sec4\], where we prove the two following general results:
\[Theo5\] Let $A$ be a bounded operator on a complex Banach space $Z$. Suppose that there exists an eigenvectorfield $E\in{L^{2}}(\T;Z)$ for $A$ such that
1. whenever $B$ is a Borel subset of $\T$ of full Lebesgue measure, $\overline{\vphantom{(}\textrm{span}}\,[E(\lambda ),
\ \lambda \in B]=Z$;
2. there exists a non-zero functional $z_{0}^{*}\in
Z^{*}$ and a trigonometric polynomial $p$ such that ${\ensuremath{{\langle z_{0}^{*},E(\lambda )\rangle}}}=p(\lambda )$ almost everywhere on $\T$;
3. if we set for every $n\in \Z$ $$\widehat{E}(n)=\displaystyle\int_{\T}\lambda ^{-n}E(\lambda )\,d\lambda
,$$ then the series $\sum_{n\in\Z}\widehat{E}(n)$ is unconditionally convergent.
Then the operator $A$ is universal for invertible ergodic systems. If moreover $\widehat{E}(-r)=0$ for some integer $r\in\Z$ (so that $\widehat{E}(-n)=0$ for every $n\ge
r$), then $A$ is universal for ergodic systems.
Remark that if $E$ is sufficiently smooth (of class $\mathcal{C}^{1}$ for instance), then the sequence of Fourier coefficients $(\widehat{E}(n))_{n\in\Z}$ goes to zero sufficiently rapidly for the series $\sum_{n\in\Z}||\widehat{E}(n)||$ to be convergent. Hence the assumption (iii) is automatically satisfied in this case. If $E$ is analytic in a neighborhood of $\T$, it can be renormalized in such a way that assumption (ii) is also satisfied, and this yields
\[Theo6\] Suppose that $A\in {\mathcal{B}}(Z)$ admits an eigenvectorfield $E$ which is analytic in a neighborhood of $\T$, and that $\overline{\vphantom{(}\emph{span}}\,\bigl[E(\lambda );\ \lambda \in\T
\bigr]=Z$. Then $A$ is universal for invertible ergodic systems. If $E$ is analytic in a neighborhood of the closed unit disk $\overline{\vphantom{(}\D}$ and $\overline{\vphantom{(}\emph{span}}\,\bigl[E(\lambda );\ \lambda \in\T
\bigr]=Z$, then $A$ is universal for ergodic systems.
Several applications of these two theorems are given in Section \[Sec4\], in particular to adjoints of multipliers on $H^{2}(\D)$ (Example \[Example2\]) and to the rather unexpected case of a Kalish-type operator on ${L^{2}}(\T)$ (Example \[Example4\]).
In Section \[Sec5\] we try to exhibit some necessary conditions for an operator to be universal. In this generality, and with the present definition of universality, this seems to be delicate. But if we restrict ourselves to operators acting on a Hilbert space, and if we additionally require in the definition of universality that the measure $\nu $ admits a moment of order $2$ (see Definition \[Definition10\]), then we obtain:
[\[Theo7\]]{} Suppose that $A\in {\mathcal{B}}(H)$ is universal for ergodic systems in the modified sense presented above. Then the unimodular eigenvalues of $A$ form a subset of $\T$ of Lebesgue measure 1.
It is a rather puzzling fact that we do not know whether Theorem \[Theo7\] can be extended to operators which are universal for invertible ergodic systems only. This point is discussed in Section \[Sec5\], as well as some open questions.
**Acknowledgement:** I am grateful to the referee for his/her careful reading of the manuscript and his/her suggestions which enabled me to clarify some points in the presentation of the text and to simplify some arguments.
A criterion for universality: proofs of Theorems \[Theo1\] and \[Theo1bis\] {#Sec2}
===========================================================================
The proofs of Theorems \[Theo1\] and \[Theo1bis\] being very similar, we concentrate on the proof of Theorem \[Theo1\], and will indicate briefly afterwards the modifications needed for proving Theorem \[Theo1bis\].
General pattern of the proof of Theorem \[Theo1\] {#SousSec2a}
-------------------------------------------------
Let $A$ be a bounded operator on the infinite-dimensional separable Banach space $Z$ satisfying the assumptions of Theorem \[Theo1\]. We will suppose in the rest of the proof that $Z$ is a complex Banach space, but the proof obviously holds true for real spaces as well. Let $(X,{\mathcal{B}},\mu ;T)$ be an invertible ergodic dynamical system on a standard probability space. For any complex-valued function $f\in{L^{\infty }}(X,{\mathcal{B}},\mu )$ let $\Phi _{f}$ be the map from $(X,{\mathcal{B}},\mu )$ into $Z$ defined by setting $$\Phi _{f}(x)=\sum_{{k\in\Z}}f({{T}^{k}}x)\,{{A}^{-k}}z_{0}.$$ Since $f$ is essentially bounded and the series $\sum_{{k\in\Z}}{{A}^{-k}}z_{0}$ is unconditionally convergent, $\Phi _{f}(x)$ is well-defined for $\mu $-almost every $x\in
X$. Also we have $$\Phi_{f}(Tx)=\sum_{{k\in\Z}}f({{T}^{k+1}}x)\,{{A}^{-k}}z_{0}
=\sum_{{k\in\Z}}f({{T}^{k}}x)\,{{
A}^{-k+1}}z_{0}=A\,\Phi _{f}(x)$$ since ${{A}^{-k+1}}z_{0}=z_{-k+1}=Az_{-k}=A.{{A}^{-k}}z_{0}$.
If we denote by ${\mathcal{B}}_{Z}$ the Borel $\sigma $-field of $Z$, and by $\nu _{f}$ the Borel probability measure on $Z$ which is the image of $\mu $ under the map $\Phi _{f}$ (i.e. $\nu _{f}(B)=\mu (\Phi
_{f}^{-1}(B))$ for every Borel subset $B$ of $Z$), it then follows that $\Phi _{f}:(X,{\mathcal{B}},\mu ;T)\longrightarrow(Z,{\mathcal{B}}_{Z},\nu _{f};A)$ is a factor map. This first argument is rather similar to the one employed in [@GW], but the map $\Phi _{f}$ is defined differently in [@GW], and for any $f\in{L^{4}}(X,{\mathcal{B}},\mu )$, thanks to the ergodic theorem of [@JRT]. The goal in [@GW] is then to construct a function $f\in{L^{4}}(X,{\mathcal{B}},\mu )$ as a limit of certain finitely-valued functions $f_{n}\in{L^{\infty}}(X,{\mathcal{B}},\mu )$, in such a way that $\Phi
_{f}$ becomes an isomorphism of dynamical systems and the measure $\nu
_{f}$ has full support. As $\Phi _{f}$ is not necessarily well-defined here when $f\in{L^{4}}(X,{\mathcal{B}},\mu )$ (or $f\in{L^{p}}(X,{\mathcal{B}},\mu )$, $1<p<+\infty$), we will construct, in the same spirit as in [@GW], a sequence of finitely-valued functions $f_{n}\in{L^{\infty }}(X,{\mathcal{B}},\mu )$ such that $\Phi _{f_{n}}$ converges in ${L^{2}}(X,{\mathcal{B}},\mu;Z )$ to a certain function $\Phi \in{L^{2}}(X,{\mathcal{B}},\mu ;Z)$ which will be an isomorphism between the two systems $(X,{\mathcal{B}},\mu ;T)$ and $(Z,{\mathcal{B}}_{Z},\nu ;A)$, where $\nu $ is the image of $\mu $ under the map $\Phi $.
Let $(Q_{j})_{j\ge 0}$ be a sequence of Borel subsets of $X$ which is dense in $({\mathcal{B}},\mu )$ (i.e. for every $\varepsilon>0$ and every $B\in{\mathcal{B}}$, there exists a $j\ge 0$ such that $\mu
(Q_{j}\vartriangle B )<\varepsilon $) with $Q_{0}=X$. Moreover, we suppose that for any $i\ge 0$, the set $J_{i}=\{j\ge i\; ;\; Q_{j}=Q_{i}\}$ is infinite. Since the span of the vectors ${{A}^{k}}z_{0}$, ${k\in\Z}$, is dense in $Z$, there exists a sequence $(u_{n})_{n\ge 1}$ of vectors of $Z$ of the form $$u_{n}=\sum_{|k|\le d_{n}}{{a}^{(n)}}_{k}\,{{A}^{-k}}z_{0},\quad
{{a}^{(n)}}_{k}\in\C,\quad \max_{|k|\le d_{n}}|{{a}^{(n)}}_{k}|>0$$ which is dense in $Z$. Let, for each $n\ge 1$, $r_{n}$ be a positive number such that the open balls $U_n=B(u_{n},r_{n})$ centered at $u_{n}$ and of radius $r_{n}$ form a basis of the topology of $Z$. We set $U_{0}=Z$. Lastly, by assumption (b) of Theorem \[Theo1\] there exists a finite subset $F$ of $\Z$ and a non-zero functional $z_{0}^{*}\in Z^{*}$ such that ${\ensuremath{{\langle z_{0}^{*},{{A}^{-n}}z_{0}\rangle}}}=0$ for all $n\in\Z\setminus F$ and ${\ensuremath{{\langle z_{0}^{*},{{A}^{-n}}z_{0}\rangle}}}\neq 0$ for all $n\in F$ (we may have to modify the initial set $F$ to obtain this property). If we replace the vector $z_{0}$ by the vector $z_{0}'=A^{-p}z_{0}$, then ${\ensuremath{{\langle z_{0}^{*},{{A}^{-n}}z_{0}'\rangle}}}\neq 0$ if and only if $n\in F'=F-p$. If we choose $p\in F$, ${\ensuremath{{\langle z_{0}^{*},z_{0}'\rangle}}}\neq 0$. So, replacing $z_{0}$ by $z_{0}'$ and $F$ by $F'$, we can suppose that $0\in F$ and that $z_{0}$ and $z_{0}^{*}$ are such that $c_{n}={\ensuremath{{\langle z_{0}^{*},{{A}^{-n}}z_{0}\rangle}}}$ is non-zero [if and only if]{} $n\in F$. We let $d=\max |F|$.
Construction of the functions $f_{n}$, $n\ge
0$ {#SousSec2b}
---------------------------------------------
We are now ready to start the construction of the functions $f_{n}$. This construction is very much inspired from that of [@GW], but many technical details need to be adjusted to the present situation. For any $z\in\C$ and $r>0$, $D(z,r)$ denotes the open disk centered at $z$ of radius $r$.
We construct by induction
1. a sequence $(f_{n})_{n\ge 0}$ of functions of ${L^{\infty }}(X,{\mathcal{B}},\mu ;\C )$;
2. sequences $(\alpha
_{n})_{n\ge 0}$, $(\beta _{n})_{n\ge 0}$, $(\gamma _{n})_{n\ge 0}$, $(\delta _{n})_{n\ge 0}$, and $(\eta _{n})_{n\ge 0}$ of positive real numbers, decreasing to zero extremely fast;
3. for each $n\ge 0$, families $(D_{i,\,0}^{(n)})_{0\le i\le n}$ and $(D_{i,\,1}^{(n)})_{0\le i\le n}$, $(E_{i,\,0}^{(n)})_{0\le i\le n}$ and $(E_{i,\,1}^{(n)})_{0\le i\le n}$, $(F_{i,\,0}^{(n)})_{0\le i\le n}$ and $(F_{i,\,1}^{(n)})_{0\le i\le n}$ of Borel subsets of $\C$;
4. for each $n\ge 0$, families $(G_{i}^{(n)})_{0\le i\le n}$ and $(H_{i}^{(n)})_{0\le i\le
n}$ of $\mu $-measurable subsets of $X$, and two measurable subsets $B_{n}$ and $C_{n}$ of $X$
such that
1. the sets $E_{i\,, 0}^{(n)}$ and $E_{i\,, 1}^{(n)}$, $0\le i\le n$, are finite, and the range of $f_{n}$ is equal to $$\left(\bigcup_{i=0}^{n}E_{i,\,0}^{(n)}\right)\cup\left(\bigcup_{i=0}^{n}E_{
i,\,1}^{(n)}
\right)\!;\label{Property1}$$ moreover, for every $i\in\{0,\dots, n\}$, $E_{i,\,0}^{(n)}\cap
E_{i,\,1}^{(n)}=\varnothing$;
2. we have \[Property2\]
1. \[Property2a\]$\mu (C_{n})>1-\eta _{n}$;
2. \[Property2b\] if we set, for every $i\in\{0,\dots, n\}$
$$\begin{aligned}
D_{i,\,0}^{(n)}&=\Bigl\{\sum_{p\in F}c_{p}\,f_{n}({{T}^{p}}x);\
x\in
C_{n}\ \textrm{and}\ f_{n}(x)\in E_{i,\,0}^{(n)}\Bigr\}\end{aligned}$$
and $$\begin{aligned}
D_{i,\,1}^{(n)}&=\Bigl\{\sum_{p\in F}c_{p}\,f_{n}({{T}^{p}}x);\
x\in
C_{n}\ \textrm{and}\ f_{n}(x)\in E_{i,\,1}^{(n)}\Bigr\},\end{aligned}$$
then $D_{i,\,0}^{(n)}\cap D_{i,\,1}^{(n)}=\varnothing$;
3. \[Property2c\] if we set, for every $i\in\{0,\dots, n\}$ $$F_{i,\,0}^{(n)}=D_{i,\, 0}^{(n)}+D(0,\beta _{n})\quad\textrm{and}\quad
F_{i,\,1}^{(n)}=D_{i,\, 1}^{(n)}+D(0,\beta _{n}),$$ then $F_{i,\,0}^{(n)}\cap F_{i,\,1}^{(n)}=\varnothing$;
3. \[Property3\] for every $i\in\{0,\dots, n\}$,
1. \[Property3a\]$\mu (H_{i}^{(n)})<\alpha _{i}(1-2^{-n})$;
2. \[Property3b\]$H_{i}^{(n-1)}\subseteq H_{i}^{(n)}$ for every $i\in\{0,\dots, n-1\}$;
3. \[Property3c\] for every $x\in Q_{i}\setminus H_{i}^{(n)}$, $f_{n}(x)\in E_{i,\,0}^{(n)}$,
and
for every $x\in (X\setminus Q_{i})\setminus H_{i}^{(n)} $, $f_{n}(x)\in
E_{i,\,1}^{(n)}$;
4. \[Property4\] we have
1. \[Property4a\]$\mu (B_{n})>1-\eta _{n}$;
2. \[Property4b\]for every $x\in B_{n}$, $|f_{n}(x)-f_{n-1}(x)|<\gamma
_{n}$;
3. \[Property4c\]for every $x\in B_{n}$, $||\Phi
_{f_{n}}(x)-\Phi
_{f_{n-1}}(x)||<\gamma _{n}$;
5. \[Property5\]we have
1. \[Property5a\] $ ||f_{n}-f_{n-1}||_{{L^{2}}(X,{\mathcal{B}},\mu
)}<2^{-n}$;
2. \[Property5b\]$||\Phi _{f_{n}}-\Phi
_{f_{n-1}}||_{{L^{2}}(X,{\mathcal{B}},\mu;Z
)}<2^{-n}$;
6. \[Property6\]for every $i\in\{0,\dots,n\}$,
1. \[Property6a\]$\mu (G_{i}^{(n)})\ge \delta _{i}(1+2^{-n})$;
2. \[Property6b\]$G_{i}^{(n)}\subseteq G_{i}^{(n-1)}$ for every $i\in\{0,\dots,n-1\}$;
3. \[Property6c\]$\Phi _{f_{n}}(x)\in U_{i}$ for every $x\in
G_{i}^{(n)}$.
We start the construction by setting (recall that $Q_{0}=X$ and $U_{0}=Z$): $E_{0,\,0}^{(0)}=\{0\}$, $E_{0,\, 1}^{(0)}=\varnothing$, $B_{0}=C_{0}=X$, $\alpha _{0}=\beta _{0}=\gamma _{0}=\delta _{0}=\eta
_{0}=1/8$, $G_{0}^{(0)}=X$, $H_{0}^{(0)}=\varnothing$ and $f_{0}=0$.
Suppose now that the construction has been carried out until step $n$. At step $n+1$, we start by introducing
1. an integer $N\ge 1$, which will be chosen very large at the end of the construction;
2. two positive numbers $\eta $ and $\gamma $, independent of each other, which will be chosen very small at the end of the construction.
As $T$ is invertible and ergodic, there exists a measurable subset $E$ of $X$ such that $\mu (E)>0$, $\mu \left(\bigcup_{|k|\le N}{{T}^{k}}E\right)<\eta $ and the sets ${{T}^{k}}E$, $|k|\le N$, are pairwise disjoint.
Recall that $$u_{n+1}=\sum_{|k|\le d_{n+1}}a_{k}^{{(n+1)}}\,{{A}^{-k}}z_{0}\quad\textrm{and}\quad
U_{n+1}=B(u_{n+1},r_{n+1}).$$ We suppose that $N\ge d_{n+1}$.
**Step 1:** We first define an auxiliary function $g_{n+1}$ on $X$ in the following way: $$g_{n+1}(x)=
\begin{cases}
a_{k}^{{(n+1)}}&\quad\textrm{if}\ x\in{{T}^{k}}E,\quad |k|\le d_{n+1}\\
0&\quad\textrm{if}\ x\in{{T}^{k}}E,\quad d_{n+1}<|k|\le N\\
f_{n}(x)&\quad\textrm{if}\ x\not\in \bigcup\limits_{|k|\le
N}{{T}^{k}}E.
\end{cases}$$ The function $g_{n+1}$ thus defined is finite-valued and it coincides with $f_{n}$ on the set $$B=X\setminus\bigcup\limits_{|k|\le N}{{T}^{k}}E,$$ which has $\mu $-measure larger than $1-\eta $. The range of $g_{n+1}$ is equal to $$\textrm{Ran}(f_{n}\restriction_{B})\cup\{0\}\cup\{a_{k}^{{(n+1)}};\ |k|\le d_{n+1}\},$$ and we write this finite set as $\{c_{l}^{{(n+1)}};\ 0\le l\le l_{n+1}\}$, with all numbers $c_{l}^{{(n+1)}}$ distinct.
By the Rokhlin Lemma, we can choose a subset $E'\in{\mathcal{B}}$ of $X$ and an integer $M\ge d$ such that the sets ${{T}^{k}}E'$, $|k|\le M$, are pairwise disjoint, and $$\mu \bigl(\bigcup_{|k|\le M-d}{{T}^{k}}E' \bigr)>1-\eta .$$
**Step 2:** We state and prove in this step a simple abstract lemma, which will be used in the forthcoming Steps 3 and 4 in order to approximate certain finite families of scalars (like the family $(c_{l}^{(n+1)})_{0\le l\le l_{n+1}}$) by other families of scalars with further additional properties.
\[Lemme0\] Let $r\ge 1$ and let $\mkern 1.5 mu\pmb{d}=(d_{1},\ldots, d_{r})$ be an $r$-tuple of positive integers. We denote by $E_{\mkern 1.5 mu\pmb{d}}$ the subset of $\Z^{r}$ defined by $$E_{\mkern 1.5 mu\pmb{d}}=\{\mkern 1.5 mu\pmb{u}=(u_{1},\ldots, u_{r});\ 0\le u_{i}\le d_{i}\; \textrm{for every } i=1,\ldots, r\}.$$ For every $i=1,\ldots, r$, let $\lambda _{i}$ be a map from $F$ into $\{0,\ldots, d_{i}\}$. We denote by $\mkern 1.5 mu\pmb{\lambda}$ the $r$-tuple of maps $\mkern 1.5 mu\pmb{\lambda}=(\lambda_1,\ldots, \lambda_{r})$ from $F$ into $\{0,\ldots, d_{1}\}\times\ldots\times \{0,\ldots, d_{r}\}$. For any such $\mkern 1.5 mu\pmb{\lambda }$, let $\sigma _{\mkern 1.5 mu\pmb{\lambda }}$ be the functional on the vector space of functions from $E_{\mkern 1.5 mu\pmb{d}}$ into $\C$, identified with $\C^{\#E_{\mkern 1.5 mu\pmb{d}}}$, defined by $$\begin{aligned}
\sigma _{\mkern 1.5 mu\pmb{\lambda } }\,:\qquad \C^{\#E_{\mkern 1.5 mu\pmb{d}}}&\longrightarrow\C\\ \ \ \bigl( \gamma _{\mkern 1.5 mu\pmb{u}}\bigr)_{\mkern 1.5 mu\pmb{u }\in E_{\mkern 1.5 mu\pmb{d }}}&\longmapsto \sum_{p\in F}c_{p}\, \gamma _{\mkern 1.5 mu\pmb{\lambda }(p)}.\end{aligned}$$ There exists a dense subset of $\C^{\#E_{\mkern 1.5 mu\pmb{d}}}$ consisting of elements $(\gamma _{\mkern 1.5 mu\pmb{u }})_{\mkern 1.5 mu\pmb{u }\in E_{\mkern 1.5 mu\pmb{d }}}$ with the following property: $$\sigma _{\mkern 1.5 mu\pmb{\lambda } }((\gamma _{\mkern 1.5 mu\pmb{u }})_{\mkern 1.5 mu\pmb{u }\in E_{\mkern 1.5 mu\pmb{d }}})\not = \sigma _{\mkern 1.5 mu\pmb{\lambda' } }((\gamma _{\mkern 1.5 mu\pmb{u }})_{\mkern 1.5 mu\pmb{u }\in E_{\mkern 1.5 mu\pmb{d }}})$$ for every maps $\mkern 1.5 mu\pmb{\lambda }$ and $\mkern 1.5 mu\pmb{\lambda' }$ such that $\mkern 1.5 mu\pmb{\lambda }(F)\not = \mkern 1.5 mu\pmb{\lambda' }(F)$.
Let us first observe that if $\mkern 1.5 mu\pmb{\lambda }(F)\not = \mkern 1.5 mu\pmb{\lambda' }(F)$, $\sigma _{\mkern 1.5 mu\pmb{\lambda }}\not = \sigma _{\mkern 1.5 mu\pmb{\lambda' }}$. This follows from the fact that all coefficients $c_{p}$, $p\in F$, are distinct. Let then $$\mkern 1.5 mu\pmb{\Sigma }=\{(\mkern 1.5 mu\pmb{\lambda },\mkern 1.5 mu\pmb{\lambda' });\ \mkern 1.5 mu\pmb{\lambda }(F)\not = \mkern 1.5 mu\pmb{\lambda' }(F)\}.$$ For each $(\mkern 1.5 mu\pmb{\lambda },\mkern 1.5 mu\pmb{\lambda' })\in\mkern 1.5 mu\pmb{\Sigma }$, the kernel $\ker(\sigma _{\mkern 1.5 mu\pmb{\lambda }}-\sigma _{\mkern 1.5 mu\pmb{\lambda' }})$ is different from the whole space $\C^{\#E_{\mkern 1.5 mu\pmb{d}}}$. The set $\mkern 1.5 mu\pmb{\Sigma }$ being finite, the Baire Category Theorem yields that $$\bigcup_{(\mkern 1.5 mu\pmb{\lambda },\mkern 1.5 mu\pmb{\lambda' })\in\mkern 1.5 mu\pmb{\Sigma }}(\sigma _{\mkern 1.5 mu\pmb{\lambda }}-\sigma _{\mkern 1.5 mu\pmb{\lambda' }})^{-1}(\C^{*})$$ is dense in $\C^{\#E_{\mkern 1.5 mu\pmb{d}}}$, which proves our claim.
**Step 3:** We define a second auxiliary function $h_{n+1}$ on $X$ by setting $$h_{n+1}(x)=
\begin{cases}
c_{l,\,k}^{{(n+1)}}&\quad\textrm{if}\ g_{n+1}(x)=c_{l}^{{(n+1)}}\ \textrm{and}\
x\in{{T}^{k}}E'\ \textrm{for some}\ |k|\le M,\\
c_{l}^{{(n+1)}}&\quad\textrm{if}\ g_{n+1}(x)=c_{l}^{{(n+1)}}\ \textrm{and}\
x\not\in\bigcup_{ |k|\le M}{{T}^{k}}E',
\end{cases}$$ where for every $l\in\{0,\dots,l_{n+1}\}$, $c_{l,\,k}^{(n)}$ is so close to $c_{l}^{{(n+1)}}$ for each $ |k|\le M$ that $$||h_{n+1}-g_{n+1}||_{\infty }\le \frac{\gamma }{2},$$ all the numbers $c_{l,\,k}^{{(n+1)}}$, $l\in\{0,\dots,l_{n+1}\}$, $
|k|\le
M$, and $ c_{l}^{{(n+1)}}$, $l\in\{0,\dots,l_{n+1}\}$, are distinct, and, moreover, the numbers $c_{l,\,k}^{(n+1)}$, $l\in\{0,\dots,l_{n+1}\}$, $|k|\le M$, have the following property:
whenever $\tau ,\,\tau '$ are two maps from $F$ into $\{0,\dots,l_{n+1}\}$, and $k,\, k'$ are two integers with $|k|,\,|k'|\le M-d$, we have $$\sum_{p\in F}c_{p}\,c_{\tau (p),\,k+p}^{(n+1)}\neq\sum_{p\in F}c_{p}\,c_{\tau '(p),\,k'+p}^{(n+1)}$$ as soon as there exists a $p\in F$ such that $(\tau (p),k+p)\neq(\tau '(p),k'+p)$.
Observe that since $|k|,\,|k'|\le M-d$ and $d=\max |F|$, $|k+p|,\,|k'+p|\le M$ for every $p\in F$, so that the quantities $c_{\tau (p),\,k+p}^{(n+1)}$ and $c_{\tau '(p),\,k'+p}^{(n+1)}$ in the expression above are well-defined.
That the scalars $c_{l,\,k}^{(n+1)} $ can indeed be chosen so as to satisfy these properties is a consequence of Lemma \[Lemme0\]. Denote by ${\Sigma }$ the set of all $4$-tuples $(\tau,\tau',k,k')$, where $\tau,\tau'$ are maps from $F$ into $\{0,\ldots, l_{n+1}\}$ and $k,k'$ are integers with $|k|,\,|k'|\le M-d$, such that there exists a $p\in F$ with $(\tau (p),k+p)\neq(\tau '(p),k'+p)$. For any map $\tau : F {\longrightarrow}\{0,\dots, l_{n+1}\}$ and any integer $k$ with $|k|\le M-d$, let $$\begin{aligned}
\mkern 1.5 mu\pmb{\lambda } _{\tau ,\,k}\,:\qquad F &\longrightarrow
\{0,\dots, l_{n+1}\}\times \{-M,\dots, M\}.\\ \ \ p&\longmapsto (\tau(p), p+k)\end{aligned}$$ Let us check that if $(\tau,\tau',k,k')$ belongs to ${\Sigma }$, $\mkern 1.5 mu\pmb{\lambda } _{\tau ,\,k}(F)\not =\mkern 1.5 mu\pmb{\lambda } _{\tau' ,\,k'}(F)$. If $\mkern 1.5 mu\pmb{\lambda } _{\tau ,\,k}(F) =\mkern 1.5 mu\pmb{\lambda } _{\tau' ,\,k'}(F)$, then $$\bigl\{(\tau (p),\,k+p);\ p\in F\bigr\}=\bigl\{(\tau '(p), k'+p);\ p\in F\bigr\},$$ so that $k+F=k'+F$. As the set $F$ is finite, $k=k'$, and thus for every $p\in F$ there exists a $p'\in F$ such that $(\tau (p), k+p)=(\tau' (p'), k+p')$. So $p=p'$ and $\tau (p)=\tau' (p)$. Thus $(\tau(p), k+p)=(\tau' (p), k+p)$ for every $p\in F$, which is contrary to our assumption. So $\mkern 1.5 mu\pmb{\lambda } _{\tau ,\,k}(F)\not =\mkern 1.5 mu\pmb{\lambda } _{\tau' ,\,k'}(F)$ as soon as $(\tau,\tau',k,k')$ belongs to ${\Sigma }$.
Applying Lemma \[Lemme0\], it follows from the observation above that we can choose a family of scalars $\bigl( c_{l,\,k}^{(n+1)}\bigr)_{0\le l\le l_{n+1},\,|k|\le M}$ such that $$\bigl|c_{l,\,k}^{(n+1)}-c_{l}^{(n+1)}\bigr|<\frac{\gamma }{2}\quad \textrm{ for every } l\in\{0,\ldots, l_{n+1}\} \textrm{ and }|k|\le M,$$ all the numbers $c_{l,\,k}^{n+1}$ and $c_{l}^{(n+1)}$ are distinct, and $$\bigl( \sigma _{\mkern 1.5 mu\pmb{\lambda } _{\tau ,\,k}}-\sigma _{\mkern 1.5 mu\pmb{\lambda } _{\tau' ,\,k'}}\bigr)\,\bigl(\bigl( c_{l,\,k}^{(n+1)}\bigr)_{0\le l\le l_{n+1},\,|k|\le M}\bigr)\neq 0\quad \textrm{for every}\ (\tau ,\,\tau',\,k,\,k')\in {{\Sigma }} ,$$ i.e. $$\displaystyle \sum_{p\in F}c_{p}\,c_{\tau (p),\,k+p}^{(n+1)}\neq\sum_{p\in F}c_{p}\,c_{\tau '(p),\,k'+p}^{(n+1)}
\quad \textrm{for every}\ (\tau ,\,\tau ',\,k,\,k')\in {{\Sigma }} .$$
Now the funtion $h_{n+1}$ has been defined, we observe that it is finite-valued, and we write its range as $$\{b_{j}^{{(n+1)}};\ 0\le j\le
j_{n+1}\}$$ where all the numbers $b_{j}^{{(n+1)}}$ are distinct. We also set $$C_{n+1}=\bigcup_{|k|\le M-d }{{T}^{k}}E'.$$ By our assumptions on $M$ and $E'$, $\mu
(C_{n+1})>1-\eta$.
**Step 4:** We construct in this step complex numbers $b_{j,\,0}^{(n+1)}$ and $b_{j,\,1}^{(n+1)}$, $0\le j\le j_{n+1}$, which are such that all the numbers $b_{j,\,0}^{(n+1)}$ and $b_{j,\,1}^{(n+1)}$ are distinct, and both $b_{j,\,0}^{(n+1)}$ and $b_{j,\,1}^{(n+1)}$ are so close to $b_{j}^{(n+1)}$ for each $j\in\{0,\dots, j_{n+1}\}$ that $$\sup_{j\in\{0,\dots, j_{n+1}\}}\Bigl(\, |b_{j,\,0}^{(n+1)}-b_{j}^{(n+1)}|+|b_{j,\,1}^{(n+1)}-b_{j}^{(n+1)}|\,\Bigr)<\dfrac{\gamma }{2}\cdot$$ Moreover, if $b_{j}^{(n+1)}=c_{l,\,k}^{(n+1)}$ for some $l\in\{0,\dots,l_{n+1}\}$ and $|k|\le M$, we write $$b_{j,\,0}^{(n+1)}=c_{l,\,k,\,0}^{(n+1)}\quad \textrm{ and }\quad b_{j,\,1}^{(n+1)}=c_{l,\,k,\,1}^{(n+1)},$$ and we require that the following holds true:
for any maps $\theta ,\,\theta '\,:\ F\longrightarrow \{0,1\}$, $\tau ,\,\tau '\,:\ F\longrightarrow \{0,\dots,l_{n+1}\}$ and any integers $k,\,k'$ with $|k|,|k'|\le M-d$, $$\sum_{p\in F}c_{p}\,c_{\tau (p),\,k+p,\,\theta (p)}^{(n+1)}\neq\sum_{p\in F}c_{p}\,c_{\tau'(p),\,k'+p,\,\theta '(p)}^{(n+1)}$$ as soon as there exists a $p\in F$ such that $(\tau (p),k+p,\theta (p))\neq (\tau' (p),k'+p,\theta '(p))$.
The proof of the existence of such numbers again relies on Lemma \[Lemme0\]. Denote by $\mathcal{F}$ the set of all $6$-tuples $(\tau,\tau',k,k',\theta
,\theta ')$, where $\tau,\tau'$ are maps from $F$ into $\{0,\ldots, l_{n+1}\}$, $\theta ,\theta '$ maps from $F$ into $\{0,1\}$, and $k,k'$ integers with $|k|,|k'|\le M-d$, such that there exists a $p\in F$ with $(\tau (p),k+p,\theta (p))\neq (\tau' (p),k'+p,\theta
'(p))$. For any maps $\tau \,:\ F\longrightarrow \{0,\ldots, l_{n+1}\}$, $\theta \,:\ F\longrightarrow \{0,1\}$, and any integer $k$ with $|k|\le M-d$, let $$\begin{aligned}
\mkern 1.5 mu\pmb{\lambda} _{\tau ,\,k,\,\theta }\,:\qquad F&\longrightarrow \{0,\ldots, l_{n+1}\}\times \{-M,\ldots, M\}\times\{0,1\}\\ \ \ p&\longmapsto (\tau(p), k+p,\theta (p)).\end{aligned}$$ We claim that if $(\tau ,\,\tau ',\,k,\,k',\,\theta ,\,\theta ')$ belongs to $\mathcal{F}$, then $\mkern 1.5 mu\pmb{\lambda} _{\tau ,\,k,\,\theta }(F)\not =
\mkern 1.5 mu\pmb{\lambda} _{\tau' ,\,k',\,\theta' }(F)$. Indeed, if these two sets were were equal, we would have $$\bigl\{\bigl( \tau (p),\,k+p,\,\theta (p)\bigr);\ p\in F\bigr\}=\bigl\{\bigl( \tau '(p),\,k'+p,\,\theta' (p)\bigr);\ p\in F\bigr\}\cdot$$ Hence $k+F=k'+F$, so that $k=k'$. Thus for every $p\in F$ there exists $p'\in F$ such that $\bigl( \tau (p),\,k+p,\,\theta (p)\bigr)=\bigl( \tau '(p'),\,k+p',\,\theta '(p')\bigr)$. Necessarily, $p=p'$, so that $\tau (p)=\tau '(p)$ and $\theta (p)=\theta '(p)$. Hence $\bigl( \tau (p),\,k+p,\,\theta (p)\bigr)=\bigl( \tau '(p),\,k+p,\,\theta '(p)\bigr)$ for every $p\in F$, and this contradicts our initial assumption. So $\mkern 1.5 mu\pmb{\lambda} _{\tau ,\,k,\,\theta }(F)\not = \mkern 1.5 mu\pmb{\lambda} _{\tau' ,\,k',\,\theta' }(F)$. It thus follows from Lemma \[Lemme0\] that numbers $c_{l,\,k,\,0}^{(n+1)}$ and $c_{l,\,k,\,1}^{(n+1)}$ can be chosen as close to $c_{l,\,k}^{(n+1)}$ as we wish, all distinct, and such that $$\sum_{p\in F}c_{p}\,c_{\tau (p),\,k+p,\,\theta (p)}^{(n+1)}\neq\sum_{p\in F}c_{p}\,c_{\tau '(p),\,k'+p,\,\theta '(p)}^{(n+1)}\quad \textrm{ for every } (\tau ,\,\tau ',\,k,\,k',\,\theta ,\,\theta ')\in\mathcal{F}.$$
This defines $b_{j,\,0}^{(n+1)}$ and $b_{j,\,1}^{(n+1)}$ when $b_{j}^{(n+1)}=c_{l,\,k}^{(n+1)}$ for some $l\in\{0,\dots,l_{n+1}\}$ and $|k|\le M$. It is then easy to define the numbers $b_{j,\,0}^{(n+1)}$ and $b_{j,\,1}^{(n+1)}$ for the remaining indices in such a way that they are sufficiently close to $b_{j}^{(n+1)}$, distinct, and distinct from all the numbers $c_{l,\,k,\,0}^{(n+1)}$ and $c_{l,\,k,\,1}^{(n+1)}$.
**Step 5:** We can now define the function $f_{n+1}$ on $X$ by setting $$f_{n+1}(x)=
\begin{cases}
b_{j,\,0}^{{(n+1)}}&\quad\textrm{if}\ h_{n+1}(x)=b_{j}^{{(n+1)}}\ \textrm{and}\
x\in
Q_{n+1}\\
b_{j,\,1}^{{(n+1)}}&\quad\textrm{if}\ h_{n+1}(x)=b_{j}^{{(n+1)}}\ \textrm{and}\
x\in
X\setminus Q_{n+1}.
\end{cases}$$ Obviously $$||h_{n+1}-f_{n+1}||_{\infty }
<\frac{\gamma }{2}\cdot$$ If $x$ belongs to $C_{n+1}$, then there exists an integer $k$ with $|k|\le M-d$ such that $x\in T^{k}E'$. Hence $T^{p}x\in T^{p+k}E'$ for every $p\in F$, and $|k+p|\le M$. It follows that there exists a map $\tau \,:\ F {\longrightarrow}\{0,\dots,l_{n+1}\}$ such that $$h_{n+1}(T^{p}x)=c_{\tau (p),\,k+p}^{(n+1)}\quad \textrm{ for every } p\in F.$$ By the definition of the function $f_{n+1}$, there exists a map $\theta :\ F\longrightarrow \{0,1\}$ such that $$f_{n+1}\bigl( T^{p}x\bigr)=c_{\tau (p),\,k+p,\,\theta (p)}^{(n+1)}\quad \textrm{ for every } p\in F.$$ This map $\theta $ satisfies $\theta (0)=0$ if $x\in Q_{n+1}$ and $\theta (0)=1$ if $x\in X\setminus Q_{n+1}$. We have $$\sum_{p\in F} c_{p} f_{n+1}\bigl( T^{p}x\bigr)=\sum_{p\in F}c_{p}\,c_{\tau (p),\,k+p,\,\theta (p)}^{(n+1)}.$$
**Step 6:** For every $i\in\{0,\dots,n\}$, let $$\begin{aligned}
J_{i,\,0}^{{(n+1)}}=\Bigl\{j\in\{0,\dots,j_{n+1}\}\ \textrm{ ; } &\textrm{there exists}\ l\in\{0,\dots,l_{n+1}\}\textrm{ such that} \ c_{l}^{{(n+1)}}\in E_{i,\,0}^{(n)}\ \\
&\textrm{with
either}\\
&b_{j}^{{(n+1)}}=c_{l}^{{(n+1)}}\ \textrm{or}\ b_{j}^{{(n+1)}}=c_{l,\,k}^{{(n+1)}}\
\textrm{for some}\ |k|\le M\Bigr\}\\[1ex]
J_{i,\,1}^{{(n+1)}}=\Bigl\{j\in\{0,\dots,j_{n+1}\}\ \textrm{ ; } &\textrm{there exists}\ l\in\{0,\dots,l_{n+1}\}\textrm{ such that}\ c_{l}^{{(n+1)}}\in E_{i,\,1}^{(n)}\\\
& \textrm{with
either}\\
&b_{j}^{{(n+1)}}=c_{l}^{{(n+1)}}\ \textrm{or}\ b_{j}^{{(n+1)}}=c_{l,\,k}^{{(n+1)}}\
\textrm{for some}\ |k|\le M\Bigr\} .\end{aligned}$$
Set for $i\in\{0,\dots,n\}$
$$\begin{aligned}
E_{i,\,0}^{{(n+1)}}&=\bigl\{b_{j,\,0}^{{(n+1)}};\ j\in
J_{i,0}^{{(n+1)}}\bigr\}\cup\bigl\{b_{j,\,1}^{{(n+1)}};\ j\in J_{i,0}^{{(n+1)}}\bigr\},\\
E_{i,\,1}^{{(n+1)}}&=\bigl\{b_{j,\,0}^{{(n+1)}};\ j\in
J_{i,1}^{{(n+1)}}\bigr\}\cup\bigl\{b_{j,\,1}^{{(n+1)}};\ j\in J_{i,1}^{{(n+1)}}\bigr\},\\
E_{n+1,\,0}^{{(n+1)}}&=\bigl\{b_{j,\,0}^{{(n+1)}};\ j\in
\{0,\dots,j_{n+1}\}\bigr\},\\
E_{n+1,\,1}^{{(n+1)}}&=\bigl\{b_{j,\,1}^{{(n+1)}};\ j\in
\{0,\dots,j_{n+1}\}\bigr\}.\end{aligned}$$
**Step 7:** With these definitions, let us check that property (1) holds true. Of course, all the sets $E_{i,\,0}^{{(n+1)}}$ and $E_{i,\,1}^{{(n+1)}}$ are finite and $$\textrm{ran}(f_{n+1})=\Bigl(
\bigcup_{i=0}
^{n+1}E_{i,\,0}^{{(n+1)}}\Bigr)\cup\Bigl( \bigcup_{i=0}^{n+1 }E_{i,\,1}^{{(n+1)}}\Bigr)=E_{n+1,\,0}^{{(n+1)}}\cup E_{n+1,\,1}^{{(n+1)}}.$$ All the numbers $b_{j,\,0}^{{(n+1)}}$ and $b_{j,\,1}^{{(n+1)}}$, $j\in\{0,\dots,j_{n+1}\}$ are distinct, and for every index $i\in{\{0,\dots,n\}}$, $J_{i,\,0}^{{(n+1)}}\cap
J_{i,\,1}^{{(n+1)}}=\varnothing$ (because $E_{i,\,0}^{(n)}\cap
E_{i,\,1}^{(n)}=\varnothing$). So $E_{i,\,0}^{{(n+1)}}\cap E_{i,\,1}^{{(n+1)}}=\varnothing$ for all $i\in\{0,\dots,n\}$. Also clearly $E_{n+1,\,0}^{{(n+1)}}\cap E_{n+1,\,1}^{{(n+1)}}=\varnothing$. So property (1) holds true.
**Step 8:** In order to check property (2), let us fix $i\in\{0,\dots,n+1\}$, and $x,\,y\in C_{n+1}$ such that $f_{n+1}(x)\in E_{i,\,0}^{{(n+1)}}$ and $f_{n+1}(y)\in E_{i,\,1}^{{(n+1)}}$. By Step 5 above, there exist maps $\tau ,\,\tau '\,:F\ {\longrightarrow}\{0,\dots, l_{n+1}\}$, integers $k,\,k'$ with $|k|,\,|k'|\le M-d$, and maps $\theta ,\,\theta '\,:\ F\longrightarrow\{0,1\}$ such that $$f_{n+1}\bigl( T^{p}x\bigr)=c_{\tau (p),\,k+p,\,\theta (p)}^{(n+1)}\quad\textrm{and}\quad
f_{n+1}\bigl( T^{p}y\bigr)=c_{\tau '(p),\,k'+p,\,\theta' (p)}^{(n+1)} \quad \textrm{for every } p\in F.$$ Recall that $0\in F$. Since $E_{i,\,0}^{(n+1)}\cap E_{i,\,1}^{(n+1)}=\varnothing$, $f_{n+1}(x)\neq f_{n+1}(y)$, so that $(\tau (0),\,k,\,\theta (0))\neq(\tau '(0),\,k',\,\theta '(0))$. Hence $(\tau ,\,\tau ',\,k,\,k',\, \theta ,\,\theta ')$ belongs to $\mathcal{F}$, and $$\begin{aligned}
\sum_{p\in F}c_{p}\,c_{\tau (p),\,k+p,\,\theta (p)}^{(n+1)}&\neq \sum_{p\in F}c_{p}\,c_{\tau '(p),\,k'+p,\,\theta' (p)}^{(n+1)}, \end{aligned}$$ i.e. $$\begin{aligned}
\sum_{p\in F}c_{p}\,f_{n+1}\bigl( T^{p}x\bigr)&\neq \sum_{p\in F} c_{p}\,f_{n+1}\bigl( T^{p}y\bigr).\end{aligned}$$
Thus $D_{i,\,0}^{{(n+1)}}\cap D_{i,\,1}^{{(n+1)}}=\varnothing$ for every $i\in\{0,\dots,n+1\}$. Once $\eta _{n+1}$ is fixed (and this will be done only later on in the construction), one can choose $\eta <\eta _{n+1}$, and then $\beta _{n+1}$ so small that properties (2a), (2b), and (2c) hold true.
**Step 9:** Our next step is to define the sets $H_{i}^{{(n+1)}}$ for $i\in{\{0,\dots,n+1\}}$ and to prove property (3). We set $$\begin{aligned}
H_{i}^{{(n+1)}}&=H_{i}^{(n)}\cup\Bigl(\bigcup_{|k|\le N}{{T}^{k}}E
\Bigr)\quad \textrm{ for every } i\in{\{0,\dots,n\}}\\
H_{n+1}^{{(n+1)}}&=\varnothing.\end{aligned}$$ Then for every $i\in{\{0,\dots,n\}}$, $\mu (H_{i}^{{(n+1)}})\le \mu (H_{i}^{(n)})+\eta
< \alpha
_{i}(1-2^{-n})+\eta $. So, if $\eta $ is chosen sufficiently small, $$\mu
(H_{i}^{{(n+1)}})<\alpha _{i}(1-2^{-{(n+1)}}).$$ Also, $\mu
(H_{n+1}^{{(n+1)}})=0<\alpha
_{n+1}
(1-2^{-{(n+1)}})$ whatever the value of $\alpha _{n+1}$. So (3a) holds. As (3b) is obvious, it remains to check (3c).
Let $i\in{\{0,\dots,n\}}$ and $x\in Q_{i}\setminus H_{i}^{{(n+1)}}$. Then $x\in X\setminus \bigl(\bigcup_{|k|\le N}{{T}^{k}}E
\bigr)$ so that $$g_{n+1}(x)=f_{n}(x)=c_{l}^{{(n+1)}} \textrm{ for some } l\in{\{0,\dots,l_{n+1}\}}.$$ Also $x\in Q_{i}\setminus H_{i}^{(n)}$, so $f_{n}(x)\in E_{i\,,0}^{(n)}$ by the induction assumption, that is $c_{l}^{{(n+1)}}
\in E_{i,\,0}^{(n)}$. We have either $$h_{n+1}(x)=c_{l}^{{(n+1)}} \quad \textrm{or}\quad h_{n+1}(x)=c_{l,\,k}^{{(n+1)}} \textrm{ for some } |k|\le M.$$ If we write $h_{n+1}(x)=b_{j}^{{(n+1)}}$ for some $j\in{\{0,\dots,j_{n+1}\}}$, then $j$ belongs to $J_{i,\,0}^{{(n+1)}}$. So $b_{j,\,0}^{{(n+1)}}$ and $b_{j,\,1}^{{(n+1)}}$ belong to $E_{i,\,0}^{{(n+1)}}$. Since $f_{n+1}(x)$ is equal to either $b_{j,\,0}^{{(n+1)}}$ or $b_{j,\,1}^{{(n+1)}}$, it follows that $f_{n+1}(x)$ belongs to $E_{i,\,0}^{{(n+1)}}$. In the same way, if $x\in\bigl(X\setminus Q_{i} \bigr)\setminus H_{i}^{{(n+1)}}$, then $f_{n+1}
(x)$ belongs to $E_{i,\,1}^{{(n+1)}}$.
Let now $i=n+1$. Let $x\in Q_{n+1}$, and let $j\in{\{0,\dots,j_{n+1}\}}$ be such that $h_{n+1}
(x)=b_{j}^{{(n+1)}}$. Then $f_{n+1}(x)=b_{j,\,0}^{{(n+1)}}$, and so by definition of the set $E_{n+1,\,0}^{{(n+1)}}$, $f_{n+1}(x)$ belongs to $E_{n+1,\,0}^{{(n+1)}}$. Similarly, if $x\in X\setminus Q_{n+1}$, then $f_{n+1}(x)$ belongs to $E_{n+1,\,1}^{{(n+1)}}$. This proves property (3c).
**Step 10:** We now have to check properties (4) and (5). We have $g_{n+1}(x)=f_{n}(x)$ for every $x\in B$ and $\mu (B)>1-\eta$ (the set $B$ has been defined in Step 1). If $\gamma
>0$ is an arbitrarily small positive number, the numbers $c_{l,\,k}^{{(n+1)}}$, $b_{j,\,0}^{{(n+1)}}$ and $b_{j,\,1}^{{(n+1)}}$ have been chosen so close to $c_{l}^{{(n+1)}}$ and $b_{j}^{{(n+1)}}$ respectively that $||f_{n+1}-g_{n+1}||_{\infty
}<\gamma $, so that in particular $$|f_{n+1}(x)-f_{n}(x)|<\gamma \quad \textrm{ for
every }
x\in B.$$ Moreover, $$\begin{aligned}
||f_{n+1}-f_{n}||_{\infty }&\le& ||f_{n+1}||_{\infty }+||f_{n}||_{\infty }\\
&\le& ||g_{n+1}||_{\infty }
+\gamma +||f_{n}||_{\infty }\\
&\le&
||f_{n}||_{\infty }+\max_{|k|\le
d_{n+1}}|a_{k}^{{(n+1)}}|+\gamma +||f_{n}||_{\infty }\\
&\le&
2\bigl(||f_{n}||_{\infty }+\max_{|k|\le
d_{n+1}}|a_{k}^{{(n+1)}}|\bigr)\end{aligned}$$ if $\gamma <\max_{|k|\le
d_{n+1}}|a_{k}^{{(n+1)}}|$. The quantity on the righthand side depends only on the construction until step $n$ and on the vector $u_{n+1}$, but not on the rest of the construction at step $n+1$. In particular, it does not depend on $\gamma $ nor on $\eta $. We have $$\int_{X}||f_{n+1}(x)-f_{n}(x)||^{2}d\mu (x)<\gamma ^{2}+\mu (X\setminus
B)\,||f_{n+1}-f_{n}||_{\infty }^{2}\le \gamma ^{2}+\eta
\,||f_{n+1}-f_{n}||_{\infty }^{2}.$$ If both $\gamma $ and $\eta $ are chosen sufficiently small, we can ensure that for instance $$||f_{n+1}-f_{n}||_{{L^{2}}(X,{\mathcal{B}},\mu )}<2^{-{(n+1)}}.$$ This proves (5a).
Let us now estimate, for $x\in X$, $$||\Phi _{f_{n+1}}(x)-\Phi _{f_{n}}(x)||=\Bigl|\!\Bigl|\sum_{k\in\Z}
\bigl(f_{n+1}({{T}^{k}}x)-f_{n}({{T}^{k}}x)
\bigr){{A}^{-k}}z_{0}\Bigr|\!\Bigr|.$$ The series $\sum_{k\in\Z}{{A}^{-k}}z_{0}$ being unconditionally convergent in $Z$, there exists for every $\rho >0$ a positive integer $k_{\rho }$ such that, for every bounded sequence $(a_{k})_{k\in\Z}$ of complex numbers, $$\Bigl|\!\Bigl|\sum_{|k|\ge k_{\rho}}a_{k}\,{{A}^{-k}}z_{0}\Bigr|\!\Bigr|\le
\rho
\,\sup_{|k|\ge k_{\rho}}|a_{k}|.$$ So $$\begin{aligned}
||\Phi _{f_{n+1}}(x)-\Phi _{f_{n}}(x)||&\le
\sup_{|k|<k_{\rho
}}\bigl|f_{n+1}({{T}^{k}}x)-f_{n}({{T}^{k}}x)\bigr|\,.\,
\sum_{|k|<k _{\rho }}||{{A}^{-k}}z_{0}||\\
&+\rho \,.\,\sup_{|k|\ge
k_{\rho}}\bigl|f_{n+1}({{T}^{k}}x)-f_{n}({{T}^{k}}x)\bigr|\\
&\le C_{\rho
}\sup_{|k|<k_{\rho}}\bigl|f_{n+1}({{T}^{k}}x)-f_{n}({{T}^{k}}x)
\bigr|+\rho \,||f_{n+1}-f_{n}||_{\infty},\end{aligned}$$ where $C_{\rho}=\sum_{|k|<k_{\rho }}||{{A}^{-k}}z_{0}||$. We have seen already that $||f_{n+1}-f_{n}||_{\infty}$ does not depend on the quantities introduced at step $n+1$ of the construction, so let us fix $\rho =
\rho _{n+1}>0$ so small that $\rho _{n+1}
||f_{n+1}-f_{n}||_{\infty}<\gamma $. Then $k_{n+1}=k_{\rho _{n+1}}$ depends on the construction until step $n$, but not on $\eta $. We set $$B_{n+1}=\bigcap_{|k|<k_{n+1}}{{T}^{-k}}B.$$ Then $\mu
(B_{n+1})>1-(2k_{n+1}-1)\eta $ which can be made as close to $1$ as we wish provided $\eta $ is small enough. We also have $$\sup_{|k|<k_{n+1}}
\bigl|f_{n+1}({{T}^{k}}x)-f_{n}({{T}^{k}}x)\bigr|<\gamma \quad \textrm{ for every }
x\in B_{n+1}.$$ It follows that for every $x\in
B_{n+1}$, $||\Phi _{f_{n+1}}(x)-\Phi _{f_{n}}(x)||\le C_{\rho
_{n+1}}\,.\,\gamma +\gamma $. So if $\gamma '$ is any positive number, we can ensure by taking $\gamma $ sufficiently small that $$||\Phi _{f_{n+1}}(x)-\Phi _{f_{n}}(x)||\le \gamma '\quad \textrm{ for
every } x\in B_{n+1}.$$ Thus if $\gamma _{n+1}$ and $\eta_{n+1}$ are any fixed positive numbers, taking $\gamma $ and $\eta$ sufficiently small yields that properties (4a), (4b) and (4c) are true.
Also, we have $$\begin{aligned}
\int_{X}||\Phi _{f_{n+1}}(x)-\Phi_{f_{n}}(x)||^{2}d\mu
(x)&<\gamma'{} ^{2}+\mu (X\setminus
B_{n+1})\,\bigl(C_{\rho_{n+1}}+\rho _{n+1} \bigr)^{2}
\,.\,||f_{n+1}-f_{n}||_{\infty}^{2}\\
&\le \gamma '{}^{2}+(2k_{n+1}-1)\,\eta \,\bigl(C_{\rho_{n+1}}+\rho
_{n+1} \bigr)\,.\,||f_{n+1}-f_{n}||_{\infty}^{2}.\end{aligned}$$ Since $\eta $ can be chosen as small as we wish compared to $k_{n+1}$, $\rho _{n+1}$, and $||f_{n+1}-f_{n}||_{\infty}$, we can make the bound above as small as we wish provided $\gamma
$ and $\eta $ are sufficiently small. So we can in particular ensure that $$||\Phi _{f_{n+1}}-\Phi_{f_{n}}||_{{L^{2}}(X,{\mathcal{B}},\mu
;Z )}<2^{-(n+1)},$$ which is (5b).
**Step 11:** It remains to construct the sets $G_{i}^{{(n+1)}}$, $i\in{\{0,\dots,n+1\}}$, in such a way that property (6) holds true.
We know that for every $i\in{\{0,\dots,n\}}$ and every $x\in G_{i}^{(n)}$, $\Phi
_{f_{n}}(x)$ belongs to $ U_{i} $, i.e. $||\Phi
_{f_{n}}(x)-u_{i}||<r_{i}$, and that $\mu (G_{i}^{(n)})\ge \delta _{i}(1+2^{-n})$. Let $0<\kappa <r_{i}$ be so small that for every $i\in{\{0,\dots,n\}}$, $$\mu \bigl(\bigl\{x\in G_{i}^{(n)} \textrm{ ; } ||\Phi
_{f_{n}}(x)-u_{i}||<r_{i}-\kappa \bigr\}\bigr)\ge\delta
_{i}\bigl(1+\frac{3}{4}2^{-n} \bigr) .$$ This number $\kappa $ only depends on the construction until step $n$. We set for $i\in{\{0,\dots,n\}}$ $$G_{i}^{{(n+1)}}=\bigl\{x\in G_{i}^{(n)} \textrm{ ; } ||\Phi
_{f_{n}}(x)-u_{i}||<r_{i}-\kappa\bigr\}\cap B_{n+1}$$ and $G_{n+1}^{{(n+1)}}=E$. Then obviously $G_{i}^{{(n+1)}}\subseteq
G_{i}^{(n)}$ for $i\in{\{0,\dots,n\}}$ and $$\mu (G_{i}^{{(n+1)}})\ge \delta
_{i}\,(1+\frac{3}{4} 2^{-n})-(2k_{n+1}-1)\,\eta \ge \delta
_{i}\,(1+2^{-{(n+1)}})$$ if $\eta $ is small enough. Also $\mu (G_{n+1}^{{(n+1)}})=\mu (E)\ge
\delta _{n+1}(1+2^{-{(n+1)}})$ if $\delta _{n+1}$ is small enough. So properties (6a) and (6b) are true.
If $i\in{\{0,\dots,n\}}$ and $x\in G_{i}^{{(n+1)}}$, then $||\Phi
_{f_{n}}(x)-u_{i}||<r_{i}-\kappa$. Also $x\in B_{n+1}$ so that $||\Phi
_{f_{n+1}}(x)-\Phi
_{f_{n}}(x)||<\gamma '$. If $\gamma '$ is chosen less than $\kappa $ (which is possible since $\kappa$ depends only on the construction until step $n$), we have $||\Phi
_{f_{n+1}}(x)-u_{i}||<r_{i}$, i.e. $\Phi _{f_{n+1}}(x)\in U_{i}$. If $i=n+1$, and $x\in G_{n+1}^{{(n+1)}}=E$, then ${{T}^{k}}x\in{{T}^{k}}E$ for every $k\in\Z$. Hence $$\Phi _{g_{n+1}}(x)=\sum_{|k|\le d_{n+1}}a_{k}^{{(n+1)}}\,{{A}^{-k}}z_{0}
+\sum_{|k|> N}g_{n+1}({{T}^{k}}x)\,{{A}^{-k}}z_{0}$$ by the definition of the function $g_{n+1}$. Thus $$||\Phi _{g_{n+1}}(x)-u_{n+1}||\le \bigl|\bigl|\sum_{|k|>
N}g_{n+1}({{T}^{k}}x)\,{{A}^{-k}}z_{0}\bigr|\bigr|.$$ Let now be $\rho
>0$ such that $\rho \,||g_{n+1}||_{\infty }<r_{n+1}/2$. Since $$||g_{n+1}||_{\infty}\le ||f_{n}||_{\infty}+\max_{|k|\le d_{n+1}}|a_{k}^{(n)}|,$$ the number $\rho $ only depends on the construction until step $n$, and we can choose $N$ so large that $N>k_{\rho }$. Then we have for every $x\in E$ $$||\Phi _{g_{n+1}}(x)-u_{n+1}||\le\rho \,\sup_{|k|>N}|g_{n+1}({{T}^{k}}x)|\le \rho \,||g_{n+1}||_{\infty }<\frac{r_{n+1}}{2}\cdot$$ Recall that there exists a positive constant $C$ such that $$\bigl|\bigl|\sum_{k\in
\Z}a_{k}\,{{A}^{-k}}z_{0}\bigr|\bigr|\le C\,\sup_{k\in\Z}|a_{k}|$$ for all bounded sequences $(a_{k})_{k\in\Z}$ of complex numbers. Since $||f_{n+1}-g_{n+1}||_{\infty}<\gamma $, we can assume by taking $\gamma $ sufficiently small that $||f_{n+1}-g_{n+1}||_{\infty
}<r_{n+1}/(2C)$. Then we have, for every $x\in X$, $$||\Phi _{f_{n+1}}(x)-\Phi _{g_{n+1}}(x)||\le C\,||f_{n+1}-g_{n+1}||_{\infty
}<\frac{r_{n+1}}{2}\cdot$$ Hence $||\Phi _{f_{n+1}}(x)-u_{n+1}||<r_{n+1}$ for every $x\in E$, i.e. $\Phi _{n+1}(x)$ belongs to $ U_{n+1}$ for every $x\in E=G_{n+1}
^{{(n+1)}}$. Thus property (6c) is satisfied.
This finishes the construction by induction of the functions $f_{n}$.
Construction of the isomorphism $\Phi $ and proof of Theorem \[Theo1\] {#SousSec2c}
----------------------------------------------------------------------
By property (5b), the sequence $(\Phi _{f_{n}} )_{n\ge 0}$ converges in ${L^{2}}(X,{\mathcal{B}},\mu ;Z)$ to a function $\Phi$ which belongs to ${L^{2}}(X,{\mathcal{B}},\mu ;Z)$. Our aim is now to prove that the probability measure $\nu $ on $Z$ defined by $\nu (B)=\mu (\Phi ^{-1}(B))$ for any Borel subset $B$ of $Z$ has full support, and that $\Phi $ is an isomorphism between the two dynamical systems $(X,{\mathcal{B}},\mu ;T)$ and $(Z,{\mathcal{B}}_{Z},\nu
;A)$.
Observe that the property (5a) of the sequence of functions $(f_{n})_{n\ge 0 }$ implies that $(f_{n})_{n\ge 0}$ converges in ${L^{2}}(X,{\mathcal{B}},\mu )$ to a certain function $f\in{L^{2}}(X,{\mathcal{B}},\mu )$. Obviously, $f$ does not belong to ${L^{\infty
}}(X,{\mathcal{B}},\mu )$, and thus it makes no sense to speak of the map $\Phi
_{f}$ (this is a difference with what happens in the construction of [@GW]). But a link between $\Phi $ and $f$ can be obtained thanks to the assumption (b) of Theorem \[Theo1\]: for $\mu$-almost every $x\in X$ and every $n\ge 0$ we have ${\ensuremath{{\langle z_{0}^{*},\Phi _{f_{n}}(x)\rangle}}}=\sum_{p\in
F}c_{p}\,f_{n}({{T}^{p}}x)$, where the set $F$ is finite. Since there exists a strictly increasing sequence $(n_{k})_{k\ge 0}$ of integers such that $\Phi _{f_{n_{k}}}\longrightarrow\Phi $ and $f_{n_{k}}\longrightarrow
f$ $\mu $-almost everywhere, it follows that $${\ensuremath{{\langle z_{0}^{*},\Phi(x)\rangle}}}=\sum_{p\in
F}c_{p}\,f({{T}^{p}}x)\quad \mu \textrm{-almost everywhere}.$$
$\bullet$ **Let us first show that $\nu $ has full support.** By property (4c) we have that for every $n\ge 0 $ and every $x\in \widetilde{B}_{n}=\bigcap_{k\ge n}B_{k}$, $||\Phi _{f_{k}}(x)-\Phi _{f_{k-1}}(x)||<\gamma _{k}$ for every $k\ge n$. Observe that $$\mu (\widetilde{B}_{n})\ge 1-\sum_{k\ge n}\eta _{k}>1-2\eta_{n}$$ if the sequence $(\eta_{n} )_{n\ge 0}$ decreases sufficiently fast, so that $\mu (\widetilde{B}_{n})\rightarrow 1$ as $n\rightarrow +\infty$. For every $n\ge 0$ and every $x\in \widetilde{B}_{n}$ we have in particular $$\bigl|\bigl|\Phi _{f_{k}}(x)-\Phi _{f_{n}}(x)\bigr|\bigr|\le
\sum_{j=n+1}^{k}\bigl|\bigl|\Phi _{f_{j}}(x)-\Phi _{f_{j-1}}(x)\bigr|\bigr|
\le \sum_{j\ge n+1}\gamma _{j}<2\gamma _{n+1}$$ if the sequence $(\gamma _{k})_{k\ge 0}$ is sufficiently rapidly decreasing. As $(\Phi _{f_{k}})_{k\ge 0}$ converges to $\Phi $ in ${L^{2}}(X,{\mathcal{B}},\mu ;Z)$, there exists a subsequence of $(\Phi
_{f_{k}})_{k\ge 0}$ which converges to $\Phi $ $\mu $-almost everywhere on $X$ and thus we have: $$\textrm{for } \mu \textrm{-almost every } x\in \widetilde{B}_{n},\quad
||\Phi (x)-\Phi _{f_{n}}(x)||\le 2\gamma _{n+1}.$$ Let us now fix $i\ge
0$. By property (6c), we have that, for every $n\ge i$ and every $x\in G_{i}^{(n)}$, $||\Phi_{f_{n}}(x)-u_{i}
||<r_{i}$. Let $$G_{i}=\bigcap_{n\ge i}G_{i}^{(n)}.$$ As the sequence of sets $\bigl( G_{i}^{(n)}\bigr)_{n\ge i}$ is decreasing by (6b), we have $\mu (G_{i})=\lim_{n\to+\infty }\mu
(G_{i}^{(n)})$ so that $\mu (G_{i})\ge\delta _{i}>0$ by property (6a). Let now $n\ge i$ be so large that $\mu (G_{i}\cap\widetilde{B}_{n})>0$. If $x\in G_{i}\cap
\widetilde{B}_{n}$, then $$||\Phi (x)-u_{i}||\le||\Phi (x)-\Phi _{f_{n}}(x)||+||\Phi
_{f_{n}}(x)-u_{i}||<2\gamma _{n+1}+r_{i}.$$ Hence $\nu \bigl(B(u_{i},r_{i}+2\gamma _{n+1}) \bigr)>0$ for all $n$ sufficiently large, so that it follows in particular that $\nu \bigl(B(u_{i},2r_{i})
\bigr)>0$. This being true for all $i\ge 0$, the measure $\nu$ has full support.
Let us observe at this point that the measure $\nu $ admits a moment of order $2$. Indeed $$\int_{Z}||z||^{2}\,d\nu (z)=\int_{X}||\Phi (x)||^{2}\, d\mu (x)<+\infty$$ since $\Phi \in{L^{2}}(X,{\mathcal{B}},\mu ;Z)$.
$\bullet$ **It remains to show that $\Phi $ is an isomorphism between $(X,{\mathcal{B}},\mu ;T) $ and $(Z,{\mathcal{B}}_{Z},\nu ;A)$.** For this it suffices to prove (see for instance [@Wa p.59–60]) that the two transformations $T$ and $A$ are conjugated via the map $\Phi $. First of all, we need to check that $\Phi (Tx)=A\Phi (x)$ for $\mu $-almost every $x\in X$. Since $\Phi
_{f_{n}}$ tends to $\Phi $ in ${L^{2}}(X,{\mathcal{B}},\mu ;Z)$ as $n$ tends to infinity, there exists a subsequence $\bigl(\Phi _{f_{n_{k}}} \bigr)
_{k\ge 0 }$ of $\bigl(\Phi _{f_{n}}\bigr)$ which tends to $\Phi $ $\mu $-almost everywhere. As $\Phi _{f_{n}}(Tx)=A\Phi _{f_{n}}(x)$ $\mu $-almost everywhere for each $n\ge 0$, it follows that $\Phi (Tx)=A\Phi (x)$ $\mu $-almost everywhere.
The second point is to check that for every subset $Q$ of $X$, $Q\in {\mathcal{B}}$, there exists a subset $B$ of $Z$, $B\in
{\mathcal{B}}_{Z}$, such that $\mu \bigl(Q\vartriangle \Phi ^{-1}(B) \bigr)=0$. So let $Q\in{\mathcal{B}}$. Suppose that we are able to exhibit a Borel subset $C$ of $\C$ such that $Q=\bigl\{x\in X;\ \sum_{p\in
F}c_{p}\,f({{T}^{p}}x)\in C\bigr\}$ up to a set of measure zero. Setting $B=\bigl\{z\in Z;\ {\ensuremath{{\langle z_{0}^{*},z\rangle}}}\in C\bigr\}$, and remembering that ${\ensuremath{{\langle z_{0}^{*},\Phi(x)\rangle}}}=\sum_{p\in
F}c_{p}\,f({{T}^{p}}x)$ $\mu $-almost everywhere, we obtain that $B$ is a Borel subset of $Z$ such that $\Phi ^{-1}(B)=Q$ (up to a set of $\mu $-measure zero). So it suffices to find $C$ with the property above.
Let us first introduce some notation: for each $n\ge 0$ we define the function $F_{n}$ on $X$ by setting $F_{n}(x)=\sum_{p\in F}c_{p}\,
f_{n}({{T}^{p}}x)$ and the function $F$ by setting $F(x)=\sum_{p\in
F}c_{p}\,f({{T}^{p}}x)$. For every $i\ge 0$, let $$C_{i,\,0}=\bigcap_{n\ge i}F_{i,\,0}^{(n)}\quad \textrm{and} \quad
C_{i,\,1}=\bigcap_{n\ge i}F_{i,\,1}^{(n)}.$$ These are Borel subsets of $\C$. Recall that $J_{i}=\{j\ge i;\ Q_{j}=Q_{i}\}$ is supposed to be infinite. Let $$\Gamma _{i,\,0}=\bigcup_{j\ge 0}\bigcap_{\genfrac{}{}{0pt}{}{k\ge
j}{k\in J_{i}}}C_{k,\,0}\quad \textrm{and}\quad
\Gamma _{i,\,1}=\bigcup_{j\ge 0}\bigcap_{\genfrac{}{}{0pt}{}{k\ge
j}{k\in J_{i}}}C_{k,\,1}.$$ These two sets are Borel in $\C$. Moreover $\Gamma _{i,\,0}\cap \Gamma
_{i,\,1}=\varnothing$. Indeed, if it is not the case, we have $$\Bigl( \bigcap_{\genfrac{}{}{0pt}{}{k\ge j_{1}}{k\in
J_{i}}}C_{k,\,0}\Bigr)\cap \bigl( \bigcap_{\genfrac{}{}{0pt}{}{k\ge
j_{2}}{k\in J_{i}}}C_{k,\,1}\bigr)\neq\varnothing$$ for some integers $j_{1}$ and $j_{2}$. In particular, if $k\in J_{i}$ is such that $k\ge \max(j_{1},j_{2})$ (and such a $k$ does exist because $J_{i}$ is infinite), $C_{k,\,0}\cap C_{k,1}\neq \varnothing$. So for all $n\ge k$, $F_{k,\,0}^{(n)}\cap F_{k,\,1}^{(n)}\neq\varnothing$, which is a contradiction with (2c).
For every $i\ge 0$, we consider the subsets of $X$ $$\begin{aligned}
D_{i}&=\bigcap_{p\in F}\Bigl(\bigcap_{k\ge i}{{T}^{-p}}B_{k}
\Bigr)\cap\Bigl( \bigcap_{k\ge i}C_{k}\Bigr)\\
H_{i}&=\bigcup_{n\ge i}H_{i}^{(n)}\\
\Omega _{i}&=\bigcap_{k\ge i}\bigl(D_{k}\setminus H_{k} \bigr)\\
\Omega &=\bigcup_{i\ge 0}\Omega _{i}.\end{aligned}$$ We have $\mu (D_{i})\ge 1-(2d+2)\sum_{k\ge i}\eta _{k}\ge 1-4(d+1)\eta
_{i}$ if the sequence $(\eta _{i})_{i\ge 0}$ decreases sufficiently fast. Also, (3a) and (3b) imply that $\mu
(H_{i})\le \alpha _{i}$ so that $$\mu (\Omega _{i})\ge 1-\sum_{k\ge i}\bigl(\mu (X\setminus D_{k})
+\mu (H_{k}) \bigr)\ge 1-4(d+1)\sum_{k\ge i}\eta _{k}-\sum_{k\ge i}\alpha
_{k}\ge 1-8(d+1)(\alpha _{i}+\eta _{i})$$ if the sequences $(\eta _{i})_{i\ge 0}$ and $(\alpha _{i})_{i\ge 0}$ decrease to zero sufficiently fast. Thus $(\Omega _{i})_{i\ge 0}$ is an increasing sequence of sets such that $\mu (\Omega _{i})\longrightarrow 1$ as $i\longrightarrow +\infty $. So $\mu (\Omega )=1$.
For every $n\ge 0$, every $x\in D_{n}$, every $p\in F$, and every $k\ge
n$, by property (4b), $|f_{k+1}({{T}^{p}}x)-f_{k}({{T}^{p}}x)|<\gamma _{k+1}$. Hence $|F_{k+1}(x)-F_{k}(x)|\le(\sup_{p\in F}|c_{p}|)\,\gamma _{k+1}$. Since $(F_{k})_{k\ge 0}$ converges to $F$ in $L^{2}(X,\mathcal{B}, \mu )$, there exists a subsequence $(F_{k_{j}})_{j\ge 0}$ of $(F_{k})_{k\ge 0}$ which converges to $F$ $\mu $-almost everywhere. Hence for ($\mu $-almost) every $x\in D_{n}$ we have $$|F(x)-F_{n}(x)|\le\sup_{p\in F}|c_{p}|\sum_{k\ge n}\gamma _{k+1}<\beta _{n}$$ if $\gamma _{n+1},\gamma _{n+2},\dots$ are chosen at steps $n+1,\,n+2,\dots$ sufficiently small with respect to $\beta _{n}$. Thus $F(x)\in D(F_{n}
(x),\beta _{n})$ for every $x\in D_{n}$.
After these preliminaries, our aim is now to show that for every $i\ge 0$, $F^{-1}(\Gamma _{i,\,0})=Q_{i}$ and $F^{-1}(\Gamma
_{i,\,1})=X\setminus Q_{i}$. So, let us fix $i\ge 0$ and $k\in J_{i}$. Suppose that $x\in Q_{i}\cap\, \Omega _{k}$. Then for every $n\ge k$, $x\in Q_{i}\cap (D_{k}\setminus H_{k})\cap(D_{n}\setminus H_{n})$ so that $
x\in Q_{i}\cap\bigl(D_{n}\setminus H_{k}^{(n)} \bigr)=Q_{k}\cap
(D_{n}\setminus H_{k}^{(n)} \bigr)
$. By property (3c), $f_{n}(x)\in
E_{k,\,0}^{(n)}$. Since $x\in D_{n}\subseteq C_{n}$, this implies that $F_{n}(x)\in
D_{k,\,0}^{(n)}$. It follows that for every $n\ge k$, $F(x)\in
F_{k,\,0}^{(n)}$, so that $F(x)\in\bigcap_{n\ge k}F_{k,\,0}^{(n)}
=C_{k,\,0}$. We have thus proved that if $k\in J_{i}$ and $x\in Q_{i}\cap \Omega _{k}$, then $F(x)\in C_{k,\,0}$.
Let now $j\ge 0$. If $k\ge j$ and $k\in J_{i}$, then, since $\Omega _{j}\subseteq \Omega
_{k}$, we have $Q_{i}\cap \Omega _{j}\subseteq Q_{i}\cap \Omega _{k}$. It follows that if $x\in Q_{i}\cap \Omega _{j}$, $F(x)\in C_{k,\,0}$ for every $k\in J_{i}$, $k\ge j$, and so $$F(x)\in\bigcap _{\genfrac{}{}{0pt}{}{k\ge j}{k\in
J_{i}}}C_{k,\,0}.$$ Suppose now that $x\in Q_{i}\cap\Omega $: there exists a $j\ge 0$ such that $x\in Q_{i}\cap \Omega _{j}$. Hence $$F(x)\in\bigcup_{j\ge 0 }\bigcap _{\genfrac{}{}{0pt}{}{k\ge j}{k\in
J_{i}}}C_{k,\,0}= \Gamma
_{i,\,0 }.$$ In exactly the same way, if $x\in(X\setminus Q_{i})\cap\Omega $ then $F(x)\in \Gamma_{i,\,1} $. Since $\Gamma_{i,\,0}\cap \Gamma_{i,\,1}=
\varnothing$ and $\mu (\Omega )=1$, we have proved that $$F^{-1}(\Gamma _{i,\,0})=Q_{i},\quad\textrm{and}\quad F^{-1}(\Gamma
_{i,\,1})=X\setminus Q_{i}$$ up to sets of $\mu $-measure zero. The proof is now nearly finished. Let $$\begin{aligned}
\mathcal{Q}=\bigl\{Q\in{\mathcal{B}};\ \textrm{there exists a Borel subset}\
C\ \textrm{of}\ \C\textrm{ such that } \mu(Q\vartriangle F^{-1}(C))=0\bigr\}\cdot\end{aligned}$$ Then $\mathcal{Q}$ is a $\sigma $-algebra which contains all the sets $Q_{i}$, $i\ge 0$, and these sets generate ${\mathcal{B}}$. Thus $\mathcal{Q}={\mathcal{B}}$. This finishes the proof that $\Phi $ is an isomorphism of dynamical systems, and Theorem \[Theo1\] is proved.
Proof of Theorem \[Theo1bis\] {#SousSec2d}
-----------------------------
The proof of Theorem \[Theo1bis\] is exactly similar in spirit, but some technical points need to be adjusted. We briefly list the most important ones. If $(X,
{\mathcal{B}},\mu ;T)$ is an ergodic dynamical system, which is not necessarily invertible, the maps $\Phi _{f}$, $f\in{L^{\infty }}(X,{\mathcal{B}},\mu)$ are defined as $\Phi _{f}(x)=\sum_{k\ge 1}f({{T}^{k}}x)\,{{A}^{-k+r}}z_{0}$ for $\mu $-almost every $x\in X$, where $r\in\Z$ is such that $A^{r}z_{0}=0$. Then $$\Phi _{f}(Tx)=\sum_{k\in\Z}f({{T}^{k}}x)\,{{A}^{-k+1+r}}z_{0}=\sum_{k\ge
1}f({{T}^{k}}x)\,{{A}^{-k+1+r}}z_{0}$$ since $A^{r}z_{0}=0$. Hence $\Phi _{f}(Tx)=A\,\Phi _{f}(x)$ $\mu $-almost everywhere. As in the proof of Theorem \[Theo1\], we can suppose by shifting the sequence $(z_{n})_{n\in\Z}$ that $0\in F$ and that there exists a non-zero functional $z_{0}^{*}\in Z^{*}$ such that $c_{n}={\ensuremath{{\langle z_{0}^{*},A^{-n+r}z_{0}\rangle}}}$ is non-zero [if and only if]{} $n\in F$. The vectors $u_{n}$ are defined as $$u_{n}=\sum_{k=0}^{d_{n}}
a_{k}^{(n)}\,{{A}^{-k+r}}z_{0}, \quad a_{k}^{(n)}\in\C.$$ The set $E$ and the integer $N$ are chosen in such a way that $\mu (E)>0$, $\mu \bigl(\bigcup_{k=0}^{N}T^{-k}E \bigr)<\eta $, and the sets ${{T}^{-k}}E$ are pairwise disjoint. This entails a modification of the definition of the function $g_{n+1}$: $$g_{n+1}(x)=
\begin{cases}
a_{k}^{{(n+1)}}&\quad\textrm{if}\ x\in{{T}^{-N+k}}E,\quad 0\le k\le d_{n+1}
\\
0&\quad\textrm{if}\ x\in{{T}^{-N+k}}E,\quad d_{n+1}<k\le N\\
f_{n}(x)&\quad\textrm{if}\ x\in X\setminus\bigcup_{k=0}^{N}{{T}^{-k}}E.
\end{cases}$$
The set $E'$ and the integer $M$ are chosen so that $\mu \bigl(
\bigcup_{k=0}^{M-d}{{T}^{-k}}E'\bigr)<1-\eta $ and the sets ${{T}^{-k}}E'$, $0\le k\le M$, are pairwise disjoint (where $d=\max |F|$). The set $G_{n+1}^{{(n+1)}}$ needs to be defined as $G_{n+1}^{{(n+1)}}={{T}^{-N}}E$. Up to these modifications, the proof is the same, and we leave the details to the reader.
Some remarks {#SousSec2e}
------------
Now we have carried out the proofs of these two results, a few remarks are in order.
\[Remarque10\] We have seen during the proofs of Theorem \[Theo1\] and \[Theo1bis\] that the measures $\nu $ constructed on $Z$ admit a second order moment. Thus operators satisfying the assumptions of these theorems are $2$-universal for (invertible) ergodic systems in the sense of Definition \[Definition10\] (see Section \[Sec5\]). It is not difficult to check that one can ensure that all the measures $\nu $ actually admit moments of all orders.
\[Remarque11\] If $A$ is an invertible operator on $Z$, then it cannot be universal for all ergodic systems: indeed, suppose that $(X,{\mathcal{B}},\mu ;T)$ is a dynamical system which is isomorphic to $(Z,{\mathcal{B}}_{Z},\nu ;A)$ for some $A$-invariant measure $\nu $ with full support. Then $\nu $ is $A^{-1}$-invariant too. If $\Phi :X\longrightarrow Z$ is an isomorphism between $(X,{\mathcal{B}},\mu ;T)$ and $(Z,{\mathcal{B}}_{Z},\nu ;A)$ then set $S=\Phi ^{-1}A^{-1}\Phi $. The measure $\mu $ is $S$-invariant, and since $T=\Phi ^{-1}A\Phi $, we have $TS=ST=\textrm{id}_{X}$, so that $T$ is invertible. Thus an invertible operator may only be universal for invertible ergodic systems.
If $A$ is a universal operator for ergodic systems and $(X,{\mathcal{B}},\mu ;T)$ is an invertible ergodic system, there exists a probability measure $\nu $ on $Z$ such that $(X,{\mathcal{B}},\mu ;T)$ and $(Z,{\mathcal{B}}_{Z},\nu ;A)$ are isomorphic. Hence $A$ is invertible as a [measure]{}-preserving transformation of $(Z,{\mathcal{B}}_{Z},\nu)$, although, as observed above, it is not invertible as a topological map from $Z$ into itself. There is no contradiction in this: there indeed exists a Borel subset $M$ of $Z$ with $A(M)= M$ and $\nu (M)=1$ such that $A\,:\, M{\longrightarrow}M$ is an invertible [measure]{}-preserving transformation, i.e. there exists a map $S\,:\, M{\longrightarrow}M$ which is measurable and [measure]{}-preserving such that $AS(x)=SA(x)=x$ for every $x\in M$. But $M$ being distinct from the whole space $X$, and $S$ not necessarily uniformly continuous on $M$, there is no reason why $S$ should extend into a topological inverse of $A$ on $Z$.
\[Remarque12\] The ergodicity of the system $(X,
{\mathcal{B}},\mu ;T)$ has been used in the proofs of Theorems \[Theo1\] and \[Theo1bis\] only via the Rokhlin lemma. So we could have supposed instead that the system was aperiodic, and the operators $A$ of Theorems \[Theo1\] and \[Theo1bis\] are in fact universal for (invertible) aperiodic systems.
We finish this section with a generalization of Theorems \[Theo1\] and \[Theo1bis\], in which assumption (a) is relaxed. This generalization is useful for proving the universality of some operators (see Section \[Sec4\]) for which the (bi)cyclicity assumption (a) is not easily seen to hold true.
\[Theo1-3\] Let $A\in{\mathcal{B}}(Z)$ be a bounded operator on $Z$. Suppose that there exist sequences $\bigl(z_{n}^{(\iota)} \bigr)_{n\in\Z}$ of vectors of $Z$, $\iota\in I$, where either $I={\{0,\dots,N\}}$ for some integer $N\ge 0$, or $I=\Z^{+}$, such that $Az_{n}^{(\iota)}=z_{n+1}^{(\iota)}$ for every $\iota\in I$ and $n\in\Z$. Suppose that the following three properties hold true:
1. $\overline{\vphantom{(}\emph{span}}\,\bigl[
A^{-n}z_{0}^{(\iota)};\ n\in\Z,\ \iota\in I\bigr]=Z$;
2. there exists a finite subset $F$ of $\Z$ such that $$\overline{\vphantom{(}\emph{span}}\,\bigl[
\{A^{-n}z_{0}^{(0)};\ n\in \Z\setminus
F\}\cup\{A^{-n}z_{0}^{(\iota)};\ n\in\Z,\ \iota\in I\setminus\{0\}\}\bigr]\neq
Z;$$
3. the series $\sum_{n\in\Z}A^{-n}z_{0}^{(\iota)}$, $\iota\in I$, are unconditionally convergent in $Z$.
Then $A$ is universal for invertible ergodic systems.
The analogue statement for universality for ergodic systems is
\[Theo1-4\] Let $A\in{\mathcal{B}}(Z)$ be a bounded operator on $Z$. With the notations of Theorem \[Theo1-3\], suppose that there exist sequences $\bigl(z_{n}^{(\iota)} \bigr)_{n\in \Z}$, $\iota\in I$, of vectors of $Z$ and integers $r^{(\iota)}$, $\iota \in I$, where either $I={\{0,\dots,N\}}$ for some integer $N\ge 0$, or $I=\Z^{+}$, such that $Az_{n}^{(\iota)}=z_{n+1}^{(\iota)}$ for every $n\in \Z$ and $A^{r^{(\iota )}}z_{0}^{(\iota)}=0$, for every $\iota \in I$.
If assumptions (a’), (b’) and (c’) of Theorem \[Theo1-3\] hold true, $A$ is universal for ergodic systems.
The proofs are almost the same as those of Theorems \[Theo1\] and \[Theo1bis\], with the additional complication that several functions $f^{(\iota)}\in{L^{2}}(X,{\mathcal{B}},\mu )$, $\iota\in I$, have to be considered, and several sequences $(f_{n}^{(\iota)})_{n\ge \iota}$ of functions of ${L^{\infty }}
(X,{\mathcal{B}},\mu )$ introduced. We sketch here the proof of Theorem \[Theo1-3\], and leave the proof of Theorem \[Theo1-4\] to the reader.
Let $(u_{n})_{n\ge 0}$ be a sequence of vectors of $Z$ which is dense in $Z$ and has the following property: there exist for each $n\ge 0$ an integer $d_{n}$ and complex coefficients $a_{k}^{(\iota ,\,n)}$, $k\in{\{0,\dots,n+1\}}$, $\iota \in{\{0,\dots,n\}}$, such that $$u_{n}=\sum_{\iota =0}^{n}\sum_{k=0}^{d_{n}}a_{k}^{(\iota ,\,n)}{{A}^{-n}}z_{0}
^{(\iota)}.$$ Let $(r_{n})_{n\ge 0}$ be a decreasing sequence of positive radii such that the balls $U_{n}=B(u_{n},r_{n})$, $n\ge 0$, form a basis of the topology of $Z$. Let also $z_{0}^{*}$ be a non-zero functional and $F$ a finite subset of $\Z$ such that ${\ensuremath{{\langle z_{0}^{*},{{A}^{-n}}z_{0}^{(\iota)}\rangle}}}
=0$ for every $\iota \in I\setminus \{0\}$ and every $n\in\Z$, $c_{n}={\ensuremath{{\langle z_{0}^{*},{{A}^{-n}}z_{0}^{(0)}\rangle}}}
=0$ for every $n\in\Z\setminus F$, and $c_{n}={\ensuremath{{\langle z_{0}^{*},{{A}^{-n}}z_{0}^{(0)}\rangle}}}$ is non-zero for every $n\in F$ (the set $F$ can be assumed to contain $0$).
Then we construct sequences $\bigl(f_{n}^{(\iota)} \bigr)_{n\ge \iota }$, $\iota \in I$, as in the proof of Theorem \[Theo1\]: at each step $n$ we construct the functions $f_{n}^{(\iota)}$ for $\iota \in{\{0,\dots,n\}}$: for each such $\iota$, the open set used in the construction is $U_{n}^{(\iota)}=
B(u_{n}^{(\iota)},{{2}^{-(\iota +1)}}r_{n})$, where $$u_{n}^{(\iota)}=\sum_{k=0}^{d_{n}}a_{k}^{(\iota ,\,n)}{{A}^{-n}}z_{0}^{(\iota)}$$ if $\iota \in{\{0,\dots,n\}}$, $u_{n}^{(\iota)}=0$ otherwise (if $I$ is finite, $\iota \in{\{0,\dots,n\}}$ means that $\iota \in{\{0,\dots,\min(|I|,n)\}}$). Hence $u_{n}=\sum_{\iota \in I}u_{n}^{(\iota)}$ for every $n\ge 0$.
We carry out the construction in such a way that properties (1)–(6) hold true for $f_{n}^{(0)}$, and properties (4)–(6) hold true for $f_{n}^{(\iota)}$, $\iota \in\{1,\dots,n\}$. The important points in this construction, when compared to the proof of Theorem \[Theo1\] are the following:
– the initial functions $f_{\iota }^{(\iota)}$, $\iota \in I$, are such that $||f_{\iota }^{(\iota)}||_{{L^{\infty}}(X,{\mathcal{B}},\mu )}<2^{-(\iota +1)}\frac{r^{(\iota)}}{C_{\iota }}$, so that $$||\Phi_{\iota }^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu ;Z)}\le C_{\iota }
||f_{\iota }^{(\iota)}||_{{L^{\infty}}(X,{\mathcal{B}},\mu )}
<2^{-(\iota +1)}{r^{(\iota )}},$$ where $C_{\iota }$ is a positive constant such that $$\left|\left|\sum_{k\in\Z} a_{k}^{(\iota ) } A^{-k} z_{0}^{(\iota )} \right|\right|\le C_{\iota }\sup _{k\in\Z}|a_{k}^{(\iota) }|$$ for every $(a_{k}^{\iota })_{k\in\Z}\in\ell_{\infty}(\Z)$.
– at a given step of the construction, the sets $E$ and $E'$ and the parameters $\gamma $ and $\eta $ can be chosen independently of the index $\iota \in{\{0,\dots,n\}}$. It follows that the sets $B_{n}$ and $G_{i}^{(n)}$, which may a priori depend on the index $\iota $, can be constructed so as not to depend on it, as well as the parameters $\gamma
_{n}$, $\eta _{n}$, and $\delta _{n}$ (the other parameters in the construction, as well as the sets $C_{n}$, $D_{i,\,0}^{(n)}$, $D_{i,\,1}^{(n)}$, $F_{i,\,0}^{(n)}$, $F_{i,\,1}^{(n)}$, and $H_{i}^{(n)}$ are involved only for the index $\iota =0$). More precisely, Step 11 in the proof of the construction becomes: we know that for every $i\in{\{0,\dots,n\}}$ and every $x\in G_{i}^{(n)}$, we have $$||\Phi_{f_{n}^{(\iota )}}-u_{i}^{(\iota )}||<2^{-(\iota +1)}r_{i}\quad \textrm{for every }\iota \in {\{0,\dots,n\}}.$$ Let $\kappa $ be so small that for every $i\in{\{0,\dots,n\}}$, $$\mu \bigl(\bigl\{x\in G_{i}^{(n)} \textrm{ ; } \textrm{for every }\iota \in {\{0,\dots,n\}},\, ||\Phi
_{f_{n}^{(\iota )}}(x)-u_{i}^{(\iota )}||<2^{-(\iota +1)}(r_{i}-\kappa) \bigr\}\bigr)\ge\delta
_{i}\bigl(1+\frac{3}{4} 2^{-n} \bigr) .$$ We set for $i\in{\{0,\dots,n\}}$ $$G_{i}^{{(n+1)}}=\bigl\{x\in G_{i}^{(n)} \textrm{ ; }\textrm{for every }\iota \in {\{0,\dots,n\}},\, ||\Phi
_{f_{n}^{(\iota )}}(x)-u_{i}^{(\iota )}||<2^{-(\iota +1)}(r_{i}-\kappa)\bigr\}\cap B_{n+1}$$ and $G_{n+1}^{{(n+1)}}=E$. Properties (6a) and (6b) are clearly true. It remains to prove property (6c). It is not difficult to see that for every $i\in{\{0,\dots,n\}}$ and every $\iota \in{\{0,\dots,n\}}$, $\Phi
_{f_{n+1}^{(\iota )}}(x)\in U_{i}^{(\iota )}$ for every $x\in G_{i}^{(n+1)}$. We then have to check that for every $i \in{\{0,\dots,n\}}$ and every $x\in G_{i}^{(n+1)}$, $\Phi
_{f_{n+1}^{(n+1 )}}(x)\in U_{i}^{(n+1)}$, i.e. that $$||\Phi
_{f_{n+1}^{(n+1)}}(x)-u_{i}^{(n+1)}||<2^{-(n+2)}r_{i}.$$ But $u_{i}^{(n+1)}=0$ and $$||\Phi
_{f_{n+1}^{(n+1)}}||_{{L^{\infty}}(X,{\mathcal{B}},\mu )}<2^{-(n +2)}r_{n+1}<2^{-(n+2)}r_{i},$$ so we do have that $\Phi
_{f_{n+1}^{(n+1 )}}(x)\in U_{i}^{(n+1)}$. The last item in the proof of property (6c) is to show that $\Phi
_{f_{n+1}^{(\iota )}}(x)\in U_{n+1}^{(\iota )}$ for every $x\in G_{n+1}^{(n+1)}=E$ and every $\iota \in {\{0,\dots,n\}}$, and here the proof is again exactly the same.
We thus have $$||f_{n}^{(\iota)}-f_{n-1}^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu )}<2^{-n}
\quad
\textrm{and} \quad ||\Phi _{f_{n}}^{(\iota)}-\Phi _{f_{n-1}}^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu ;Z
)}<2^{-n}$$ for every $n\ge 1$ and $\iota\in{\{0,\dots,n\}}$ and $$||\Phi
_{f_{n}^{(\iota)}}(x)-u_{n}^{(\iota)}||<2^{-(\iota+1)}r_{n}$$ for every $x\in
G_{i}^{(n)}$, $i\in{\{0,\dots,n\}}$, $\iota \in{\{0,\dots,n\}}$. It follows that the sequence $\bigl(f_{n}^{(\iota)} \bigr)_{n\ge \iota }$ converges in ${L^{2}}(X,{\mathcal{B}},\mu )$ to a function $f^{(\iota)}$ which satisfies $$\begin{aligned}
||f^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu )}&\le ||f_{\iota}^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu )}+\sum_{n\ge \iota}
||f_{n+1}^{(\iota)}-f_{n}^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu )}\\
&< 2^{-(\iota+1)}+\sum_{n\ge \iota}2^{-(n+1)}\le
2^{-(\iota-1)}.\end{aligned}$$ In the same way, $\bigl(\Phi _{f_{n}}^{(\iota)} \bigr)_{n\ge \iota}$ converges in ${L^{2}}(X,{\mathcal{B}},\mu ;Z)$ to a function $\Phi ^{(\iota)}$ which satisfies $||\Phi ^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu ;Z)}<2^{-(\iota-1)}$. In particular, the series $\sum_{\iota\in I}||\Phi ^{(\iota)}||_{{L^{2}}(X,{\mathcal{B}},\mu;Z )}$ is convergent.
We now consider the map $\Phi :X\longrightarrow Z$ defined as $$\Phi =\sum_{\iota\in I}\Phi ^{(\iota)}.$$ Since $$\int_{X}\sum_{\iota\in I}||\Phi ^{(\iota)}(x)||\,d\mu (x)\le\sum_{\iota\in I}\Bigl(
\int_{X}||\Phi ^{(\iota)}(x)||^{2}d\mu (x)\Bigr)^{1/2}\le
\sum_{\iota\in I}2^{-(\iota -1)}
<+\infty,$$ $\Phi (x)$ is defined $\mu $-almost everywhere.
If $\nu $ is the measure defined on $Z$ by setting $\nu (B)=\mu (\Phi
^{-1}(B))$ for every $B\in{\mathcal{B}}_{Z}$, then it is clear that $\Phi :(X,{\mathcal{B}},\mu ;T)\longrightarrow (Z,{\mathcal{B}}_{Z}, \nu ;A)$ is a factor map. It is an isomorphism of dynamical systems for the same reason as in the proof of Theorem \[Theo1\]: if $Q$ is any Borel subset of $X$, we know that there exists a Borel subset $C$ of $\C$ such that $$\bigl\{x\in X;\ \sum_{p\in F}c_{p}f^{(0)}({{T}^{p}}x)\in C\bigr\}=Q.$$ Let now $B=\bigl\{z\in Z;\ {\ensuremath{{\langle z_{0}^{*},z\rangle}}}\in C\bigr\}$. We have $$\Phi ^{-1}(B)=\bigl\{x\in X;\ {\ensuremath{{\langle z_{0}^{*},\Phi (x)\rangle}}}=
\sum_{p\in F}c_{p}\,f^{(0)}({{T}^{p}}x)\in C\bigr\}=Q.$$ So $\Phi $ is an isomorphism between $(X,{\mathcal{B}},\mu ;T)$ and $
(Z,{\mathcal{B}}_{Z}, \nu ;A)$. It remains to prove that the measure $\nu $ has full support: for every $i\ge 0$, every $n\ge i$, every $\iota \in{\{0,\dots,n\}}$, and every $x\in G_{i}=\bigcap_{n\ge i}G_{i}^{(n)}$, $||\Phi _{f_{n}^{(\iota)}}-u_{n}^{(\iota)}||<2^{-(\iota +1)}r_{i}$. We also know that for every $k\ge 0$, every $\iota \in{\{0,\dots,k\}}$ and every $x\in B_{k}$, $$||\Phi _{f_{k}^{(\iota)}}(x)-\Phi _{f_{k-1}^{(\iota)}}(x)||<\gamma _{k}.$$ So, if $x\in G_{i}\cap\bigl(\bigcap_{k\ge n }B_{k} \bigr)$ and $\iota \in{\{0,\dots,n\}}$, $$\begin{aligned}
||\Phi _{f^{(\iota)}}(x)-u_{i}^{(\iota)}||&\le\sum_{k\ge n+1}
||\Phi _{f_{k}^{(\iota)}}(x)-\Phi _{f_{k-1}^{(\iota)}}(x)||+
||\Phi _{f_{n}^{(\iota)}}(x)-u_{i}^{(\iota)}||\\
&\le\sum_{k\ge n+1}\gamma _{k}+2^{-(\iota +1)}r_{i}<\gamma
_{n}+2^{-(\iota +1)}r_{i}.\end{aligned}$$ Let us now fix $i\ge 0$. Let $(n_{\iota })_{\iota \in I}$ be an increasing (finite or infinite) sequence of integers such that $\gamma _{n_{\iota }}<2^{-(\iota +1)}r_{i}$ and $\mu \bigl(G_{i}\cap\bigl(\bigcap_{k\ge n_{0}}B_{k} \bigr) \bigr)>0$. If $x\in G_{i}\cap\bigl(\bigcap_{k\ge n_{0}}B_{k} \bigr)$, then $x\in G_{i}\cap\bigl(\bigcap_{k\ge n_{\iota }}B_{k} \bigr)$ for every $\iota \in I$, so that $$||\Phi _{f^{(\iota)}}(x)-u_{i}^{(\iota)}
||<2^{-\iota }r_{i}\quad \textrm{ for every } \iota \in I.$$ Hence $$||\sum_{\iota \in I}\Phi _{f^{(\iota)}}(x)-\sum_{\iota \in I}u_{i}^{(\iota)}||<2r_{i}\quad \textrm{ for every }x\in G_{i}\cap\bigl(\bigcap_{k\ge n_{0 }}B_{k} \bigr),$$ i.e. $\Phi (x)\in B(u_{i},2r_{i})$. This being true for every $i\ge 0 $, it follows that the measure $\nu $ has full support.
Universal unilateral and bilateral weighted shifts {#Sec3}
==================================================
The easiest class of operators to which the criteria of Theorems \[Theo1\] and \[Theo1bis\] can be applied is the class of weighted shifts on ${\ell_{p}(\N)}$ or ${\ell_{p}(\Z)}$, $1\le p<\infty $, or on ${c_{0}(\N)}$ or ${c_{0}(\Z)}$. We first give the proof of Theorem \[Theo2\].
Proof of Theorem \[Theo2\] {#SousSec3a}
--------------------------
Suppose that ${B_{w}}$ is a unilateral weighted shift with weights $(w_{n})_{n\ge 1}$ on one of the spaces ${\ell_{p}(\N)}$, $1\le p<\infty $, or ${c_{0}(\N)}$, with ${B_{w}}e_{0}=0$ and ${B_{w}}e_{n}=w_{n}e_{n-1}$ for every $n\ge
1$. Let us set $z_{0}=e_{0}$ and $z_{-n}=1/({w_{1}\dots w_{n}})\,e_{n}$ for every $n\ge 1$. Then $${B_{w}}z_{-n}=\dfrac{1}{{w_{1}\dots w_{n}}}\,w_{n}e_{n-1}=z_{-(n-1)}\quad\textrm{for every}\ n\ge 1,
\ \textrm{and}\ {B_{w}}z_{0}=0.$$ If we write $z_{-n}={B_{w}}^{-n}e_{0}$ for every $n\ge 1$, then the vectors ${B_{w}}^{-n}z_{0}$, $n\ge 0$, span a dense subset of ${\ell_{p}(\N)}$ (or ${c_{0}(\N)}$). Also, the linear span of the vectors ${B_{w}}^{-n}z_{0}$, $n\ge 1$, is not dense, so that assumption (b) of Theorem \[Theo1bis\] always holds true with $F
=\{0\}$.
The series $\sum_{n\ge 0}{B_{w}}^{-n}e_{0}$ is unconditionally convergent in ${\ell_{p}(\N)}$ (resp. ${c_{0}(\N)}$) if and only if the series $\sum_{n\ge 1}
1/|{w_{1}\dots w_{n}}|^{p}$ is convergent (resp. if and only if $|{w_{1}\dots w_{n}}|\longrightarrow +\infty $ as $n\longrightarrow +\infty $). So if this last condition is satisfied, ${B_{w}}$ is universal on ${\ell_{p}(\N)}$ (resp. ${c_{0}(\N)}$) for ergodic systems.
This condition is also necessary for ${B_{w}}$ to be universal for either ergodic or invertible ergodic systems, but the arguments are different depending on whether ${B_{w}}$ acts on ${\ell_{p}(\N)}$ for some $1\le p<+\infty$ or on ${c_{0}(\N)}$.
– if ${B_{w}}$ acts on ${\ell_{p}(\N)}$, $1\le p< +\infty $, and is universal for ergodic systems, then ${B_{w}}$ is necessarily frequently hypercyclic, and by the characterization of frequently hypercyclic weighted shifts on $\ell_{p}(\N)$ of [@BR], the series $\sum_{n\ge 1}1/|{w_{1}\dots w_{n}}|^{p}$ is convergent;
– if ${B_{w}}$ acts on ${c_{0}(\N)}$, then the same argument does not apply since the condition $|{w_{1}\dots w_{n}}|\longrightarrow +\infty $ as $n\longrightarrow
+\infty $ does not characterize frequently hypercyclic backward weighted shifts on ${c_{0}(\N)}$ (see [@BG2] and [@BR] for details). But if ${B_{w}}$ is universal for (invertible) ergodic systems, then it is necessarily strongly mixing with respect to some invariant measure with full support, hence in particular topologically mixing. So, for every $\varepsilon >0$, there exists an integer $n_{\varepsilon }$ such that, for every $n\ge
n_{\varepsilon }$, there exists a vector $x^{(n,\,\varepsilon )}\in c_{0}(\N)$ with $||x^{(n,\,\varepsilon )}||_{\infty}<\varepsilon $ such that $||{B_{w}}^{n}\,x^{(n,\,\varepsilon )}-e_{0}||_{\infty}<1/2$. In particular, $|x_{n}^{(n,\,\varepsilon )}\,{w_{1}\dots w_{n}}-1|<1/2$, so that $|x_{n}^{(n,\,\varepsilon )}\,{w_{1}\dots w_{n}}|>1/2$. Thus $\varepsilon |{w_{1}\dots w_{n}}|>1/2$ for every $n\ge n_{\varepsilon }$. It follows that $|{w_{1}\dots w_{n}}|\longrightarrow +\infty $ as $n\longrightarrow +\infty $.
The arguments for bilateral weighted shifts are exactly the same and we leave them to the reader.
Proof of Theorem \[Theo4\] {#SousSec3b}
--------------------------
The proof of Theorem \[Theo4\] relies on the following simple idea: suppose that $Z$ is a Banach space admitting a biorthogonal system $(u_{n},u_{n}^{*})_{n\ge 0}$ having the following property:
*there exists a bounded sequence $(\omega _{n})_{n\ge 1}$ of non-zero weights such that $B_{\omega}$ defined by $B_{\omega}u_{n}=\omega _{n}u_{n-1}$ for every $n\ge 1$ and $B_{\omega} u_{0}=0$ is a bounded operator on $Z$ and $$\sum_{n\ge 1}\frac{||u_{n}||}{|\omega _{1}\ldots\omega _{n}|}<+\infty.$$*
Then $B_{\omega}$ is a universal operator for ergodic systems on $Z$. The proof of this statement is exactly similar to that of Theorem \[Theo2\]: we set $z_{0}=u_{0}$ and $z_{-n}=(1/|\omega _{1}\ldots\omega _{n}|)\,u_{n}$. Then $\overline{\vphantom{(}\textrm{span}}\,\bigl[z_{-n};\ n\ge 0 \bigr]=Z$, ${\ensuremath{{\langle u_{0}^{*},z_{-n}\rangle}}}=0$ for every $n\ge 1$ so that $\overline{\vphantom{(}\textrm{span}}\,\bigl[z_{-n};\ n\ge 1 \bigr]\neq Z$, and lastly the series $\sum_{n\ge 0}z_{-n}$ is unconditionally convergent since $\sum_{n\ge 0}||z_{-n}||<+\infty $. So Theorem \[Theo1bis\] applies.
Suppose that $Z$ can be decomposed as a topological sum $Z=E\oplus Y$ where $E$ has a sub-symmetric basis $(e_{n})_{n\ge 0}$ (i.e. the basis $(e_{n})_{n\ge 0}$ is unconditional and equivalent to each of its subsequences). Let $(e_{n}^{*})_{n\ge 0}$ denote the family of biorthogonal functionals on $E$, which we extend to $Z$ by setting ${\ensuremath{{\langle e_{n}^{*},y\rangle}}}=0$ for every $y\in Y$. Let also $(y_{n},y_{n}^{*})_{n\ge 0}$ be a bounded biorthogonal system for $Y$, where each $y_{n}^{*}$ is extended to $Z$ by setting ${\ensuremath{{\langle y_{n}^{*},e\rangle}}}=0$ for every $e\in E$. We denote by $P_{E}$ and $P_{Y}$ the projections of $Z$ onto $E$ and $Y$ respectively, associated to the decomposition $Z=E\oplus Y$, and we let $M=\max(||P_{E}||,||P_{Y}||)$. Since the basis $(e_{n})_{n\ge 0}$ is sub-symmetric, $B_{\alpha }$ defined by $B_{\alpha } e_{n}=\alpha_{n}e_{n-1}$ for every $n\ge 1$ and $B_{\alpha } e_{0}=0$ is a bounded operator on $E$ for any bounded sequence of weights $(\alpha _{n})_{n\ge 1}$. Let $(n_{k})_{k\ge 0}$ be a strictly increasing sequence of integers with $n_{0}=0$. Consider the biorthogonal system $(u_{n},u_{n}^{*})$ of $Z$ defined by setting $$\begin{aligned}
u_{n}&=
\begin{cases}
y_{k}&\quad\textrm{if}\ n=n_{k}\ \textrm{for some}\ k\ge 0\\
e_{n-k-1}&\quad\textrm{if}\ n\in\{n_{k}+1,\dots,n_{k+1}-1\}\ \textrm{for
some}\ k\ge 0
\end{cases}\\
u_{n}^{*}&=
\begin{cases}
y_{k}^{*}&\quad\textrm{if}\ n=n_{k}\ \textrm{for some}\ k\ge 0\\
e_{n-k-1}^{*}&\quad\textrm{if}\ n\in\{n_{k}+1,\dots,n_{k+1}-1\}\
\textrm{for
some}\ k\ge 0.
\end{cases}\end{aligned}$$ With this definition, $\{u_{n} ;\ n_{k}<n<n_{k+1}\}=\{e_{n};\ n_{k}-k
\le n<n_{k+1}-(k+1)\}$, and $\{u_{n};\ n\in\{n_{k};\ k\ge 0\}\}=\{y_{n};\ n\ge 0\}$, so that $(u_{n},u_{n}^{*})_{n\ge 0}$ is indeed a bounded biorthogonal system of $Z$.
We define an operator $A$ on $Z$ by setting for every $z\in Z$ $$Az=\sum_{k\ge 1}u_{n_{k}}^{*}(z)\,w_{k}u_{n_{k}-1}+2\sum_{k\ge
0}\,\,\Bigl(
\sum_{n=n_{k}+2}^{n_{k+1}-1}u_{n}^{*}(z)\,u_{n-1}\Bigr)+\sum_{k\ge 0}
u_{n_{k}+1}^{*}(z)\,w'_{k}u_{n_{k}}$$ where the weights $w_{k}$ and $w'_{k}$ are defined by $$w_{k}=\frac{2^{-k}}{||y_{k}^{*}||}\quad \textrm{and} \quad w'_{k}=\frac{2^{-k}}{||y_{k}||},\quad k\ge 0.$$ Observe that these weights do not depend on the sequence $(n_{k})_{k\ge
0}$.
This operator is a backward weighted shift with respect to the biorthogonal system $(u_{n},u_{n}^{*})_{n\ge 0}$: $Au_{n}=\omega
_{n}u_{n-1}$ where $$\omega _{n}=
\begin{cases}
w_{k}&\quad\textrm{if}\ n=n_{k}\ \textrm{for some}\ k\ge 1\\
2&\quad\textrm{if}\ n\in\{n_{k}+2,\dots,n_{k+1}-1\}\ \textrm{for some}\
k\ge 0\\
w'_{k}&\quad\textrm{if}\ n=n_{k}+1\ \textrm{for some}\ k\ge 1.\\
\end{cases}$$ Let us first check that $A$ is bounded. We have for every $z\in Z$ $$Az=\sum_{k\ge 1}y_{k}^{*}(z)\,w_{k}e_{n_{k-1}-k}+\sum_{k\ge
0}\,\,\left(
\sum_{n=n_{k}+2}^{n_{k+1}-1}e_{n-k-1}^{*}(z)\,2\,e_{n-k-2}\right)+\sum_{
k\ge 0}
e^{*}_{n_{k}-k}(z)\,w'_{k}y_{k}.$$ Since the basis $(e_{n})_{n\ge 0}$ is sub-symmetric, there exists a positive constant $C$ such that for every $z\in Z$, $$\left|\left|\sum_{k\ge 0}\left(\sum_{n=n_{k}+2}^{n_{k+1}-1} e^{*}_{n-k-1}(z)\,e_{n-k-2}\right)\right|\right|\le C||z||.$$ Hence $$||Az||\le||z||\sum_{k\ge 1}||y_{k}^{*}||\,w_{k}||e_{n_{k-1}-k}||+2C||z||+
||z||\sum_{k\ge 0}||e^{*}_{n_{k}-k}||\,w'_{k}||y_{k}||.$$ Since $\sup_{n\ge 0}||e_{n}||$ and $\sup_{n\ge 0}||e_{n}^{*}||$ are finite, the conditions on the weights $w_{k}$ and $w'_{k}$ imply that $A$ is bounded.
In order to show that $A$ is universal for ergodic systems, it remains to choose the sequence $(n_{k})_{k\ge 0}$ in such a way that the series $$\sum_{n\ge 1}\frac{||u_{n}||}{\omega _{1}\dots\omega _{n} }$$ is convergent. We have $$\begin{aligned}
\sum_{n\ge 1}\dfrac{||u_{n}||}{\omega _{1}\dots\omega _{n} }&=
\sum_{k\ge 0}
\sum_{n=n_{k}+1}^{n_{k+1}}\dfrac{||u_{n}||}{\omega
_{1}\dots\omega _{n} }\\
&\!\!\!\!\!\!=\sum_{k\ge 0}\dfrac{1}{w_{1}\dots w_{k}\,w'_{1}\dots
w'_{k}\,2^{n_{k}-2k}}
\left(\sum_{n=n_{k}+1}^{n_{k+1}-1}\dfrac{||u_{n}||}{2^{n-(n_{k}+1)}}
+\dfrac{||u_{n_{k+1}}||}{2^{(n_{k+1}-1)-(n_{k}+1)}w_{k+1}}\right).\end{aligned}$$ If we write $C_{k}=\max\{||u_{n}||;\ n_{k}+1\le n\le n_{k+1}\}, $ this yields $$\sum_{n\ge 1}\dfrac{||u_{n}||}{\omega _{1}\dots\omega _{n} }\le
\sum_{k\ge 0}\dfrac{C_{k}}{w_{1}\dots w_{k}\,w'_{1}\dots
w'_{k}\,2^{n_{k}-2k}}\Biggl(
2+\dfrac{1}{2^{n_{k+1}-n_{k}-2}\,w_{k+1}}\Biggr).$$ Since the weights $w_{k}$ and $w_{k}'$ are defined independently of the sequence $(n_{k})_{k\ge 0}$, we can choose this sequence growing so fast that $$2^{n_{k}}>\max\Biggl(\dfrac{2^{n_{k-1}+2}}{w_{k}},\,
\dfrac{2^{3k}\,C_{k}}{w_{1}\dots w_{k}\,w'_{1}\dots
w'_{k}}\Biggr)\quad\textrm{for every}\ k\ge 1.$$ Then $$\sum_{n\ge 1}\displaystyle \dfrac{||u_{n}||}{\omega _{1}\dots\omega _{n}
}\le \sum_{k\ge 0} 3\,.\, 2^{-k},$$ from which it follows that the series $$\sum_{n\ge 1}\frac{||u_{n}||}{\omega
_{1}\dots\omega _{n} }$$ is convergent. This terminates the proof of Theorem \[Theo4\].
Unimodular eigenvectors and universality {#Sec4}
========================================
We begin this section with the proof of Theorem \[Theo5\], which gives a straightforward criterion in terms of unimodular eigenvectors for an operator to be universal for (invertible) ergodic systems.
Proof of Theorem \[Theo5\] {#SousSec4a}
--------------------------
Let $A$ be a bounded operator on the complex separable infinite-dimensional Banach space $Z$, admitting a unimodular eigenvectorfield $E$ satisfying assumptions (i), (ii), and (iii) of Theorem \[Theo5\]. Set, for each $n\in\Z$, $z_{n}=\widehat{E}(-n)$. Then $A{z_{n}}=z_{n+1}$ for every $n\in
\Z$. The vectors $z_{n}$, $n\in\Z$, span a dense subspace of $Z$. Indeed, suppose that $z^{*}\in Z^{*}$ is such that ${\ensuremath{{\langle z^{*},z_{n}\rangle}}}=0$ for every $n\in\Z$. Then the function ${\ensuremath{{\langle z^{*},E(\,.\,)\rangle}}}$ is zero almost everywhere, and by (i) it follows that $z^{*}=0$. So $z_{0}$ is bicyclic for $A$. Let now $F$ be the spectrum of the polynomial $q$ defined by $q(e^{i\theta })=\sum_{n\in\Z} \widehat{p}(-n)e^{in\theta }$. For every $n\in\Z\setminus
F$, we have $${\ensuremath{{\langle z_{0}^{*},z_{n}\rangle}}}=\int_{\T}\lambda ^{n}\,{\ensuremath{{\langle z_{0}^{*},E
(\lambda )\rangle}}}\,d\lambda =\widehat{q}(n)=0.$$ Since $z_{0}^{*}$ is non-zero, it follows that the linear span of the vectors $z_{n}$, $n\in\Z\setminus F$, is not dense in $Z$. Lastly, assumption (iii) of Theorem \[Theo5\] states that the series $\sum_{n\in\Z }z_{n}$ is unconditionally convergent. So the hypotheses of Theorem \[Theo1\] are satisfied, and $A$ is universal for invertible ergodic systems. If $\widehat{E}(-r)=0$ for some integer $r\in\Z$, then $A^{r}z_{0}=0$, and Theorem \[Theo1bis\] applies.
Theorem \[Theo6\] can now be obtained as a consequence of Theorem \[Theo5\].
Proof of Theorem \[Theo6\] {#SousSec4b}
--------------------------
Suppose that $A\in{\mathcal{B}}(Z)$ admits a unimodular eigenvectorfield $E$ which is analytic in a neighborhood $\Omega $ of $\T$, and such that the vectors $E(\lambda )$, $\lambda \in\T$, span a dense subspace of $Z$. Then the restriction of $E$ to $\T$ is $\sigma $-spanning for any measure $\sigma $ with infinite support. In particular, it is $d\lambda
$-spanning, where $d\lambda $ is the normalized Lebesgue measure on $\T$.
Let $z_{0}^{*}$ be any non-zero element of $Z^{*}$. The function $\varphi
$ defined on $\Omega $ by setting $$\varphi
(\lambda)={\ensuremath{{\langle z_{0}^{*},E(\lambda )\rangle}}},\quad \lambda \in\T,$$ is analytic on $\Omega $, and not identically zero since the span of the vectors $E(\lambda )$, $\lambda \in\T$, is dense in $Z$. Thus $\varphi $ admits only finitely many zeroes $z_{1},\dots,z_{r}$ on $\T$, with respective multiplicities $d_{1},\dots,d_{r}$. Let $p(z)=\prod_{j=1}^{r}(z-z_{j})^{d_{j}}$. There exists a function $\psi $, which is analytic on $\Omega $ and does not vanish on a neighborhood $\Omega '$ of $\T$ such that $\varphi (z)=p(z)\psi (z)$ for every $z\in\Omega '$.
Consider the eigenvectorfield $F:\Omega '\longrightarrow Z$ defined by $F(\lambda )=E(\lambda )/\psi (\lambda )$. The span of the vectors $F(\lambda )$, $\lambda \in\T$, is dense in $Z$. Moreover, for every $\lambda \in \T$, $${\ensuremath{{\langle z_{0}^{*},F(\lambda )\rangle}}}=\dfrac{1}{\psi (\lambda )}\,{\ensuremath{{\langle z_{0}^{*},E(\lambda )\rangle}}}=p(\lambda ).$$ Lastly, since $F$ is analytic on a neighborhood of $\T$, there exists $a\in(0,1)$ such that $$||\widehat{F}(n)||=\textrm{O}(a^{|n|})\quad \textrm{as }|n|
\longrightarrow +\infty .$$ Hence the series $\sum_{n\in\Z}\widehat{F}(n)$ is unconditionally convergent. The assumptions of Theorem \[Theo5\] are thus satisfied, and $A$ is universal for invertible ergodic systems.
If $E$ is analytic in a neighborhood of $\overline{\D}$, the same kind of reasoning applies: if $z_{1},\dots,z_{r}$ are the zeroes of $\varphi (\lambda )={\ensuremath{{\langle z_{0}^{*},E(\lambda )\rangle}}}$ on $\overline{\D}$ with multiplicities $d_{1},\dots,d_{r}$, and if $p(z)=\prod_{j=1}^{r}(z-z_{j})^{d_{j}}$, then $\varphi (z)=p(z)\psi (z)$ on a neighborhood of $\overline{\D}$, where $\psi $ is an analytic function on this neighborhood which does not vanish. If we consider again the eigenvectorfield $F$ defined by $F(\lambda )=
E(\lambda )/\psi (\lambda )$, then $F$ is analytic in a neighborhood of $\overline{\D}$ so that $\widehat{F}(n)=0$ for every $n<0$. So Theorem \[Theo5\] applies again and $A$ is universal for ergodic systems.
Let us mention here that, using Theorems \[Theo1-3\] and \[Theo1-4\], Theorems \[Theo5\] and \[Theo6\] can be generalized to the case where $A$ admits several eigenvectorfields.
Examples and applications {#SousSec4c}
-------------------------
We present in this section several examples of operators which can be shown to be universal thanks to Theorems \[Theo5\] and \[Theo6\]. First, one can easily retrieve the universality of some of the weighted shifts considered in Theorem \[Theo2\] above.
\[Example1\] Let $\alpha B$, $|\alpha |>1$, be a multiple of the unweighted backward shift on ${\ell_{p}(\N)}$, $1\le p<+\infty $, or ${c_{0}(\N)}$. Then $\alpha B$ admits an eigenvectorfield $E$ defined on the disk of radius $|\alpha |$ by $$E(\lambda )=\sum_{n\ge 0}\left(\frac{\lambda }{\alpha }\right)^{n-1}e_{n}.$$ So $E$ is analytic on $\D(0,|\alpha |)$, and it is easy to check that the eigenvectors $E(\lambda )$, $\lambda \in\T$, span a dense subspace of ${\ell_{p}(\N)}$ or ${c_{0}(\N)}$. Thus Theorem \[Theo6\] applies, and $\alpha B$ is universal for ergodic systems. The same argument applies for instance to the weighted shift $S_{w}$ on ${\ell_{p}(\Z)}$ or ${c_{0}(\Z)}$, where the weight $w$ is given by $w_{n}=2$ if $n\ge 1$ and $w_{n}=1/2$ if $n\le 0$: $S_{w}$ is universal for invertible ergodic systems.
Our second class of examples is given by adjoints of multipliers $M_{\varphi }^{*}$ on $H^{2}(\D)$. This is a natural class of operators to consider here, since their dynamical properties (hypercyclicity, frequent hypercyclicity, ergodicity) are rather well understood. See for instance [@BM] for details.
\[Example2\] Denoting by $\D$ the open unit disk, let $\varphi :\D\longrightarrow\C$ be an analytic map belonging to $H^{\infty }(\D)$, and consider the adjoint $M_{\varphi }^{*}$ of the multiplier $M_{\varphi }$ defined on the Hardy space $H^{2}(\D)$ by setting $M_{\varphi }\,f=\varphi f$ for every $f\in H^{2}(\D)$. If $\T\subseteq\varphi (\D)$, then $M^{*}_{\varphi }$ is universal for invertible ergodic systems.
Suppose that the analytic map $\varphi :\D{\longrightarrow}\C$ is such that $\varphi (\D)$ contains the unit circle. Since $\varphi $ is open, $K=\varphi ^{-1}(\T)$ is a compact subset of $\D$, and there exists $\rho \in(0,1)$ such that $K\subseteq D(0,\rho )$. The derivative $\varphi '$ of $\varphi $ can vanish only finitely many times on $K$, and we denote by $z_{1},\dots,z_{r}$ the distinct zeroes of $\varphi '$ on $K$, with respective multiplicities $m_{1},\dots,m_{r}$. There exists for every $j\in\{1,\dots,r\}$ a disk $D(z_{j},\varepsilon _{j})$ with $\varepsilon _{j}>0$ and two holomorphic functions $\psi _{j}$ and $\sigma _{j}$ on $D(z_{j},\varepsilon _{j})$ which do not vanish here such that for every $z\in D(z_{j},\varepsilon _{j})$, $$\varphi (z)=\lambda _{j}+(z-z_{j})^{d_{j}}\psi _{j}(z)\quad \textrm{and}\quad\varphi '(z)=(z-z_{j})^{m_{j}}\sigma _{j}(z)$$ where $d_{j}=m_{j}+1$ and $\lambda _{j}=\varphi (z_{j})$. Also, there exists a holomorphic function $\beta _{j}$ on $D(z_{j},\varepsilon _{j})$ such that $\psi _{j}=\beta _{j}^{d_{j}}$. If we set $\alpha _{j}(z)=(z-z_{j})\beta _{j}(z)$ for $z\in D(z_{j},\varepsilon _{j})$, we can assume that $\alpha _{j}$ is a biholomorphism from a certain open neighborhood $V_{j}$ of $z_{j}$ contained in $D(z_{j},\varepsilon _{j})$ onto an open disk $D(0,\delta _{j})$. Let $$\Omega _{j,\,0} =D(0,\delta _{j}^{d_{j}})\setminus [0,\delta _{j}^{d_{j}})$$ and let $\gamma _{j,\,0} $ be an holomorphic determination of the $d_{j}$-th root of $z$ on $\Omega _{j,\,0} $: $\gamma _{j,\,0} (z)^{d_{j}}=z$ for every $z\in\Omega _{j,\,0} $. We also set $V_{j,\,0} =\bigl\{z\in V_{j};\ \alpha _{j}(z)^{d_{j}}\in\Omega _{j,\,0} \bigr\}$, $U_{j,\,0} =\lambda _{j}+\Omega _{j,\,0}$ and $U_{j}=D(\lambda_{j} ,\delta_{j} ^{d_{j}})$. Observe that $U_{j,\,0}\cap\T$ contains a set of the form $\Gamma_j\setminus\{\lambda_j\}$, where $\Gamma_j$ is an open sub-arc of $\T$ containing the point $\lambda_j$.
We now claim that $\varphi $ is a biholomorphism from $V_{j,\,0} $ onto $U_{j,\,0} $. Let us first check that $\varphi (V_{j,\,0} )=U_{j,\,0} $: $z\in V_{j}$ belongs to $V_{j,\,0} $ if and only if $\alpha _{j}(z)^{d_{j}}=\varphi (z)-\lambda _{j}$ belongs to $\Omega _{j,\,0}$, i.e. if and only if $\varphi (z)$ belongs to $U_{j,\,0}$. Let us set, for $z\in U_{j,\,0}$, $$\varphi ^{-1}_{j,\,0}(z)=\alpha ^{-1}_{j}\circ\gamma _{j,\,0}\,(z-\lambda _{j}).$$ This definition makes sense: if $z\in U_{j,\,0}$, $z-\lambda _{j}\in\Omega _{j,\,0}$, so that $\gamma _{j,\,0}(z-\lambda _{j})\in D(0,\delta _{j})$, and $\alpha _{j}$ is a biholomorphism from $V_{j}$ onto $D(0,\delta _{j})$. This function $\varphi^{-1}_{j,\,0}$ is thus well-defined and holomorphic on $U_{j,\,0}$, and we have for every $z\in U_{j,\,0}$ $$\varphi \bigl( \varphi ^{-1}_{j,\,0}(z)\bigr)=\lambda _{j}+\alpha _{j}\bigl( \varphi ^{-1}_{j,\,0}(z)\bigr)^{d_{j}}=\lambda _{j}+\bigl( \gamma _{j,\,0}(z-\lambda _{j})\bigr)^{d_{j}}=\lambda _{j}+z-\lambda _{j}=z$$ since $z-\lambda _{j}$ belongs to $\Omega _{j,\,0}$ and $\gamma _{j,\,0}(z)^{d_{j}}=z$ for every $z\in\Omega _{j,\,0}$. It follows from this that $\varphi :V_{j,\,0}{\longrightarrow}U_{j,\,0}$ is a biholomorphism, the inverse of which is $\varphi _{j,\,0}^{-1}$. Restricting the sets $V_{j,\,0}$ and $U_{j,\,0}$, we can and do assume that, for every $j\in\{1,\ldots,r\}$, $U_{j,\,0}\cap\T=\Gamma_j\setminus\{\lambda_j\}$, where $\Gamma_j$ is an open sub-arc of $\T$ containing the point $\lambda_j$, and that the arcs $\Gamma_j$, $j\in\{1,\ldots, r\}$, do not intersect.
If $z\in K$ is such that $\varphi (z)\not \in\{\lambda_{1},\dots,\lambda_{r}\}$ and $\varphi '(z)\neq 0$, $\varphi $ is a biholomorphism in a neighborhood of $z$. For every such $z$, let $V_{z}$ be an open neighborhood of $z$, and $U_{z}$ a disk centered at $\varphi (z)$ of radius $\rho _{z}>0$ such that $\varphi :V_{z}{\longrightarrow}U_{z}$ is a biholomorphism, $\varphi'$ does not vanish on $V_{z}$, and the closure of the set $U_{z}$ contains none of the points $\lambda_{j}$, $j\in\{1,\ldots,r\}$. The disks $U_{z}$, $z\in \varphi^{-1}(\T\setminus \{\lambda_{1},\dots,\lambda_{r}\})$ and $U_{j}$, $j\in\{1,\ldots,r\}$, form an open covering of $\T$ (remember our assumption that $\T\subseteq \varphi(\D)$), so one can extract from it a finite covering of the form $U_{\xi _{1}},\dots,U_{\xi _{s}},U_{1},\dots,U_{r}$, $s\ge 1$. Denote by $\bigl( \Omega_{l}\bigr)_{l\in\Lambda } $ the finite family of open subsets of $\D$ consisting of the sets $U_{\xi _{i}}$, $1\le i\le s$, and $U_{j,\,0}$, $1\le j\le r$. For every $l\in \Lambda $, $\Omega'_l=\Omega _{l}\cap \T$ is either an open subarc of $\T$ whose closure is contained in $\T\setminus\{\lambda _{1},\dots,\lambda _{r}\}$ (when $\Omega_{l}=U_{\xi _{i}}$ for some $i\in\{1,\ldots, s\}$) or an open subarc minus one point (when $\Omega_{l}=U_{j,\,0}$ for some $j\in\{1,\ldots, r\}$, in which case $\Omega'_{l}=\Gamma_j\setminus\{\lambda_j\}$). We have $$\bigcup_{l\in\Lambda }\Omega' _{l}\cap\T=\T\setminus\{\lambda _{1},\dots,\lambda _{r}\}.$$ Observe that for every $j\in\{1,\ldots, r\}$, there is a unique index $l\in \Lambda $ such that $\Omega'_{l}$ contains the point $\lambda_j$ in its closure. Writing this index as $l_{j}$, we have $\Omega _{l_{j}}=U_{j,\,0}$ and $\Omega'_{l_{j}}=\Omega _{l_{j}}\cap\T=\Gamma_j\setminus\{\lambda_j\}$. For every $l\in\Lambda $, we denote by $\varphi ^{-1}_{l}$ the inverse of $\varphi $ defined on the set $\Omega _{l}$.
Let $\bigl( v_{l}\bigr)_{l\in\Lambda }$ be a finite $\mathcal{C}^{1}$-partition of the unity associated to the finite covering $\bigl( \Omega' _{l}\bigr)_{l\in\Lambda }$ of $\T\setminus\{\lambda _{1},\dots,\lambda _{r}\}$, with $v_{l}$ supported on $\Omega'_l$ for every $l\in\Lambda$. If $\Omega'_l=U_{\xi_i}\cap\T$ for some $i\in\{1,\ldots, s\}$, the function $v_{l}$, which is supported on $\Omega '_{l}$, obsviously extend into a $\mathcal{C}^{1}$ function on the whole circle $\T$ by setting $v_{l}(\lambda _{j})=0$ for every $j\in\{1,\ldots,r\}$. If $\Omega'_l=\Omega '_{l_{j}}$ for some $j\in\{1,\ldots, r\}$, the observation above shows that there exists an open sub-arc $\Gamma '_{j}$ of $\Gamma _{j}$, containing $\lambda _{j}$, such that the only index $l\in \Lambda $ for which $v_{l}$ does not identically vanish on $\Gamma '_{j}\setminus\{\lambda _{j}\}$ is $l_{j}$. Hence $v_{l_{j}}$ is equal to $1$ on $\Gamma '_{j}\setminus\{\lambda_j\}$, and it follows that $v_{l_{j}}$ can be extended into a $\mathcal{C}^{1}$ function on the whole circle $\T$ by setting $v_{l_{j}}(\lambda _{j})=1$. We can thus assume that all the functions $v_{l}$, $l\in\Lambda$, are defined and of class $\mathcal{C}^{1}$ on $\T$.
For every $z\in\D$, let $k_{z}$ be the reproducing kernel of the space $H^{2}(\D)$ at the point $z$: $$k_{z}(\xi)= \sum_{n\ge 0}\overline{\vphantom{t}z}^{\,n}\xi
^{n}=\dfrac{1}{1-\overline{\vphantom{t}z}\,\xi },\quad \xi \in\D,$$ and $k_{z}$ is characterized by the property that ${\ensuremath{{\langle f,k_{z}\rangle}}}=f(z)$ for every $f\in H^{2}(\D)$. We have $M_{\varphi}^{*}k_{z}=\overline{\varphi(z)}k_{z}$ for every $z\in\D$. We then introduce the polynomial $$p(z)=\prod_{j=1}^{r}(z-\lambda _{j})^{2}$$ and the maps $F:\T\setminus\{\lambda _{1},\dots,\lambda _{r}\}{\longrightarrow}H^{2}(\D)$ and $E:\T\setminus\{{\overline{\lambda }}_{1},\dots,{\overline{\lambda }}_{r}\}{\longrightarrow}H^{2}(\D)$ defined by $$\begin{aligned}
\forall\,\lambda \in\T\setminus\{\lambda _{1},\dots,\lambda _{r}\},\qquad &F(\lambda )={\overline{p(\lambda )}}\,\sum_{l\in\Lambda }v_{l}(\lambda )\,k_{\varphi ^{-1}_{l}(\lambda )}
\intertext{and}
\forall\,\lambda \in\T\setminus\{{\overline{\lambda}} _{1},\dots,{\overline{\lambda}} _{r}\},\qquad &E(\lambda )=F({\overline{\lambda}} ).\end{aligned}$$ Observe that the quantity $k_{\varphi ^{-1}_{l}(\lambda )}$ in the expression of $F(\lambda)$ above only makes sense when $\lambda$ belongs to $\Omega'_l$. But since $v_l$ is supported on $\Omega'_l$, the function $\lambda\mapsto v_l(\lambda)k_{\varphi ^{-1}_{l}(\lambda )} $ extends into a $\mathcal{C}^{\infty}$ function on $\T\setminus\{{\lambda} _{1},\dots,{\lambda} _{r}\}$ by defining it to be zero outside the set $\Omega'_l$, so that $F$ is a well-defined $\mathcal{C}^{\infty}$ function on $\T\setminus\{{\lambda} _{1},\dots,{\lambda} _{r}\}$.
For every $\lambda \in\T\setminus\{\lambda _{1},\dots,\lambda _{r}\}$ we have $$M^{*}_{\varphi} F(\lambda )={\overline{p(\lambda )}}\,\displaystyle\sum_{l\in\Lambda }v_{l}(\lambda ){\overline{\varphi \bigl( \varphi ^{-1}_{l}(\lambda )\bigr)}}\,k_{\varphi ^{-1}_{l}(\lambda )}$$ so that $M^{*}_{\varphi }F(\lambda )={\overline{\lambda }}\,F(\lambda )$. It follows that $E:\T\setminus\{{\overline{\lambda }}_{1},\dots,{\overline{\lambda }}_{r}\}{\longrightarrow}H^{2}(\D)$ is a unimodular eigenvectorfield for $M^{*}_{\varphi }$.
Our aim is to show that the assumptions of Theorem \[Theo5\] are satisfied. Assumptions (i) and (ii) are easy to check: suppose that $f\in H^{2}(\D)$ is such that ${\ensuremath{{\langle f,E(\lambda )\rangle}}}=0$ for every $\lambda \in B$ where $B\subseteq\T\setminus\{{\overline{\lambda}} _{1},\dots,{\overline{\lambda}} _{r}\}$ is a Borel subset of $\T$ of full Lebesgue measure. Then by continuity $${\overline{p(\lambda )}}\,\displaystyle\sum_{l\in\Lambda }v_{l}(\lambda )\,f\bigl( \varphi ^{-1}_{l}(\lambda )\bigr)=0 \quad \textrm{ for every } \lambda \in\T\setminus\{\lambda_{1},\dots,\lambda_{r}\},$$ so that $$\displaystyle\sum_{l\in\Lambda }v_{l}(\lambda )\,f\bigl( \varphi ^{-1}_{l}(\lambda )\bigr)=0 \quad \textrm{ for every } \lambda \in\T\setminus\{\lambda_{1},\dots,\lambda_{r}\}.$$ It follows that for every $j\in\{1,\ldots, r\}$ and every $\lambda \in\Gamma'_{j}\setminus\{\lambda _{j}\} $, $f\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)=0$. Since $\varphi ^{-1}_{l_{j}}(\Gamma'_{j}\setminus\{\lambda _{j}\} )$ has accumulation points in $\D$, $f=0$, and $\textrm{span}\,\{ E(\lambda );\ \lambda \in B\}$ is dense in $H^{2}(\D)$. Assumption (i) is thus satisfied. Assumption (ii) clearly holds true: if $f\equiv 1$, then ${\ensuremath{{\langle f,E(\lambda )\rangle}}}={\overline{p({\overline{\lambda }})}}=q(\lambda )$ for every $\lambda \in\T\setminus\{{\overline{\lambda}} _{1},\dots,{\overline{\lambda}} _{r}\}$, where $q({e^{i\theta }})=\sum_{n\in\Z}{\overline{\hat{p}(n)}}e^{in\theta }$. The main difficulty is to check that the series $\sum_{n\in\Z}\widehat{E}(n)$, or equivalently the series $\sum_{n\in\Z}\widehat{F}(n)$, is unconditionally convergent. Since $H^{2}(\D)$ does not contain a copy of $c_{0}$, it suffices to prove that for every $f\in H^{2}(\D)$, the series $\sum_{n\in\Z}|{\ensuremath{{\langle f,\widehat{F}(n)\rangle}}}|$ is convergent.
We are going to show that for every $f\in H^{2}(\D)$, the function $\phi _{f}$ defined on $\T\setminus\{\lambda _{1},\dots,\lambda _{r}\}$ by $$\phi _{f}(\lambda )={\ensuremath{{\langle f,F(\lambda )\rangle}}}={\overline{p(\lambda )}}\,\displaystyle\sum_{l\in\Lambda }v_{l}(\lambda )\,f\bigl( \varphi ^{-1}_{l}(\lambda )\bigr)$$ extends into a function of class $\mathcal{C}^{1}$ on $\T$. Bernstein’s Theorem will then imply that the series $\sum_{n\in\Z}|{\ensuremath{{\langle f,\widehat{F}(n)\rangle}}}|$ is convergent.
We have seen that the function $F$ is of class $\mathcal{C}^{\infty }$ on $\T\setminus\{\lambda _{1},\dots,\lambda _{r}\}$. Since, for every $l\in\Lambda $, $\sup\bigl\{|\varphi ^{-1}_{l}(\lambda )|;\ \lambda \in\Omega'_{l} \bigr\}\le \rho<1$, the quantity $\sup\bigl\{\bigl| \bigl| k_{\varphi ^{-1}_{l}(\lambda )}\bigr|\bigr|;\ \lambda \in\Omega' _{l}\bigr\}$ is finite for every $l\in\Lambda $. Now $p(\lambda _{j})=0$ for every $j\in\{1,\dots,r\}$, and since each function $v_{l}$, $l\in\Lambda$, is uniformly bounded on $\T$ it follows that $F$ can be extended into a continuous map on $\T$ by setting $F(\lambda _{j})=0$ for every $j\in\{1,\dots,r\}$. So $\phi _{f}$ is actually continuous on $\T$ for every $f\in H^{2}(\D)$. Let us now compute the derivative of $\phi _{f}$. Writing, for $0\le \theta < 2\pi $, $\lambda =e^{i\theta }$ and $$\begin{aligned}
\phi _{f}( e^{i\theta })=&{\overline{\vphantom{\bigl( }p( e^{i\theta })}}\,\sum_{l\in\Lambda }v_{l}( e^{i\theta })\,f\bigl(\varphi ^{-1}_{l}( e^{i\theta }) \bigr)
\intertext{we have}
\dfrac{d\phi _{f}}{d\theta }(e^{i\theta })=&-i\,e^{-i\theta }\,{\overline{\vphantom{\bigl( }p'( e^{i\theta })}}\,\sum_{l\in\Lambda }v_{l}( e^{i\theta })\,f\bigl(\varphi ^{-1}_{l}( e^{i\theta }) \bigr)\\&+
{\overline{\vphantom{\bigl( }p( e^{i\theta })}}\,\sum_{l\in\Lambda }ie^{i\theta }\,\dfrac{dv_{l}}{d\theta }(e^{i\theta })\,f\bigl(\varphi ^{-1}_{l}( e^{i\theta }) \bigr)\\
&+{\overline{\vphantom{\bigl( }p( e^{i\theta })}}\,\sum_{l\in\Lambda }v_{l}( e^{i\theta })\,(i\,e^{i\theta })^{2}\dfrac{1}{\varphi '\bigl( \varphi ^{-1}_{l}(e^{i\theta })\bigr)}\,f'\bigl( \varphi ^{-1}_{l}(e^{i\theta })\bigr)\end{aligned}$$ (the notation $\dfrac{d}{d\theta }$ is used for the derivative of a function on $\T$ with respect to the real variable $\theta $, while the sign ’ is used for the complex derivative of a holomorphic function).
The same argument as above, using the facts that $\Lambda $ is finite, that the functions $v_{l}$ are of class $\mathcal{C}^{1}$ on $\T$ (and hence have uniformly bounded derivatives on $\T$), and that $p(\lambda _{j})=p'(\lambda _{j})=0$ for every $j\in\{1,\dots,r\}$, shows that the first two terms in this expression tend to $0$ as $\lambda =e^{i\theta }\in\T\setminus\{\lambda _{1},\ldots, \lambda _{r}\}$ tends to $\lambda _{j}$, $j\in\{1,\dots,r\}$. So it remains to deal with the last term. If $\lambda \in\T\setminus\{\lambda _{1},\ldots, \lambda _{r}\}$ tends to $\lambda _{j}$ for some $j\in\{\lambda _{1},\ldots, \lambda _{r}\}$, we can suppose without loss of generality that $\lambda $ belongs to $\Omega '_{l_{j}}$. The third term in the expression above is then equal to $${\overline{p(\lambda )}}\dfrac{1}{\varphi '\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)}f'(\varphi ^{-1}_{l_{j}}(\lambda )), \quad \textrm{where } \lambda =e^{i\theta }.$$ As $\sup\bigl\{\bigl|f'\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)\bigr|; \ \lambda\in\Omega' _{l_{j}} \bigr\}$ is finite, it suffices to show that $$\dfrac{{\overline{p(\lambda )}}}{\varphi '\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)}\to 0\quad \textrm{ as } \lambda \to\lambda _{j},\; \lambda \in\Omega '_{l_{j}}.$$ We have seen that for every $z\in D(z_{j}\varepsilon _{j})$, $\varphi '(z)=(z-z_{j})^{m_{j}}\sigma _{j}(z)$, where $\sigma _{j}$ is an holomorphic function which does not vanish on $D(z_{j},\varepsilon _{j})$. Hence the quantity $$\sup\Bigl\{\dfrac{|z-z_{j}|^{m_{j}}}{|\varphi '(z)|};\ z\in D(z_{j},\varepsilon _{j})\setminus\{z_{j}\}\Bigr\}$$ is finite. So there exists a positive constant $C$ such that $$\sup\Bigl\{\dfrac{\bigl|\varphi ^{-1}_{l_{j}}(\lambda )-z_{j}\bigr|^{m_{j}}}{\bigl|\varphi '\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)\bigr|};\ \lambda \in\Omega _{l_{j}}\Bigr\}\le C.$$ We have $$p(\lambda )=p\bigl( \varphi \bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)\bigr)=p\bigl( \lambda _{j}+\bigl( \varphi ^{-1}_{l_{j}}(\lambda )-z_{j}\bigr)^{d_{j}}\psi _{j}\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)\bigr)$$ for every $\lambda \in\Omega _{l_{j}}$, and thus $p(\lambda )=\bigl( \varphi ^{-1}_{l_{j}}(\lambda )-z_{j}\bigr)^{2d_{j}}q_{j}\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)$ for every $\lambda \in\Omega _{l_{j}}$, where $q_{j}$ is an holomorphic function which is bounded on $\varphi ^{-1}_{l_{j}}(\Omega _{l_{j}})$. There exists hence a positive constant $C'$ such that $$\begin{aligned}
\biggl|\dfrac{p(\lambda )}{\varphi '\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)}\biggr|&=
\biggl|\dfrac{\bigl( \varphi ^{-1}_{l_{j}}(\lambda )-z_{j}\bigr)^{2d_{j}}}{\bigl( \varphi ^{-1}_{l_{j}}(\lambda )-z_{j}\bigr)^{m_{j}}}\biggr|\,.\, \bigl|q_{j}\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)\bigr|\,.\,
\biggl|\dfrac{\bigl( \varphi ^{-1}_{l_{j}}(\lambda )-z_{j}\bigr)^{m_{j}}}{\varphi '\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)}\biggr|\\
&\le C'\,.\,\bigl|\varphi ^{-1}_{l_{j}}(\lambda )-z_{j}\bigr|^{2d_{j}-m_{j}}\qquad \textrm{for every } \lambda \in\Omega _{l_{j}}.\end{aligned}$$ The righthand bound tends to $0$ as $\varphi ^{-1}_{l_{j}}(\lambda )$ tends to $z_{j}$ since $2d_{j}=2(m_{j}+1)>m_{j}$. We now claim that if $\lambda\in \Omega _{l_{j}} $ tends to $\lambda _{j}$, then $\varphi ^{-1}_{l_{j}}(\lambda )$ tends to $z_{j}$. Indeed, since $\Omega _{l_{j}}=U_{j,\,0}$, $\varphi ^{-1}_{l_{j}}(\lambda )=\varphi _{j,\,0}^{-1}(\lambda )=
\alpha _{j}^{-1}\bigl( \gamma _{j,\,0}(\lambda -\lambda _{j})\bigr)$. If $\lambda \to \lambda _{j}$, then $\gamma_{j,\,0}(\lambda -\lambda _{j})\to 0$. Now the map $\alpha _{j}:D(z_{j},\varepsilon _{j}){\longrightarrow}D(0,\delta _{j})$ is a biholomorphism such that $\alpha _{j}(z_{j})=0$. Thus $\alpha _{j}^{-1}\bigl( \gamma_{j,\,0}(\lambda -\lambda _{j})\bigr)\to z_{j}$ as $\lambda \to\lambda _{j}$, $\lambda \in\Omega _{l_{j}}$, and this proves our claim.
So we have proved that if $\lambda \in\Omega _{l_{j}}$ tends to $\lambda _{j}$, then $$\biggl|\dfrac{{\overline{\vphantom{\bigl( }p(\lambda )}}} {\varphi '\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)}\biggr|=\biggl|\dfrac{p(\lambda )} {\varphi '\bigl( \varphi ^{-1}_{l_{j}}(\lambda )\bigr)}\biggr|\quad \textrm{tends to}\ 0.$$ As explained above, this shows that the third term in the expression of $\dfrac{d\phi _{f}}{d \theta }(e^{i\theta })$ given above tends to $0$ as $d\bigl( e^{i\theta } ,\{\lambda _{1},\dots,\lambda _{r}\}\bigr)$ tends to $ 0$ with $e^{i\theta }\in\T\setminus\{\lambda _{1},\ldots,\lambda _{r}\}$. So $\phi _{f}$ is a map of class $\mathcal{C}^{1}$ on $\T$, and this finishes the proof of Example \[Example2\].
\[Example3\] Using the notation of Example \[Example1\], the operator $\alpha B^{2}$ is universal on ${\ell_{p}(\N)}$ or ${c_{0}(\N)}$ for any $\alpha $ with $|\alpha |>1$.
This can be proved in several ways. One of these is to observe that $\alpha B^{2}$ admits two eigenvectorfields $E_{1}$ and $E_{2}$ which are analytic on $D(0,|\alpha |)$: $$E_{1}(\lambda )=\sum_{n\ge 0}\biggl(\dfrac{\lambda }{\alpha }
\biggr)^{\!\!n}
e_{2n+1}\quad\textrm{and}\quad
E_{2}(\lambda )=\sum_{n\ge 0}\biggl(\dfrac{\lambda }{\alpha }
\biggr)^{\!\!n}
e_{2n}\cdot$$ We have $\overline{\vphantom{'}\textrm{span}}\,\bigl[E_{1}(\lambda ),
E_{2}(\lambda );\ \lambda \in\T\bigr]=Z$. Using the generalization of Theorem \[Theo6\] following from Theorem \[Theo1-4\], we obtain that $\alpha B^{2}$ is universal for ergodic systems.
If one is interested only in the universality of $\alpha B^{2}$ for invertible ergodic systems, one can use simply Theorem \[Theo5\] and the following argument: let $\varphi $ be a function of class $\mathcal{C}^{\infty }$ on $\T$ such that $0\le \varphi \le 1$ and the support of $\varphi
$ is a non-trivial closed sub-arc $\Gamma $ of $\T$. Consider the $\mathcal{C}^{\infty }$ eigenvectorfield of $\alpha B^{2}$ defined by $$E(\lambda )=\varphi (\lambda )\,E_{1}(\lambda )+(1-\varphi (\lambda ))\,
E_{2}(\lambda ),\quad \lambda \in\T.$$ Let us show that $\overline{\vphantom{'}\textrm{span}}\,\bigl[E(\lambda );\ \lambda
\in\T\bigr]=Z$: if $x^{*}$ is a functional such that ${\ensuremath{{\langle x^{*},E(\lambda
)\rangle}}}=0$ for every $\lambda \in\T$, then $(1-\varphi (\lambda ))\,{\ensuremath{{\langle x^{*},E_{2}(\lambda )\rangle}}}={\ensuremath{{\langle x^{*},E_{2}(\lambda )\rangle}}}=0$ for every $\lambda
\in\T\setminus\Gamma $. Hence by analyticity of $E_{2}$, ${\ensuremath{{\langle x^{*},E_{2}(\lambda )\rangle}}}=0$ for every $\lambda \in\T$. It follows that $\varphi (\lambda )\,{\ensuremath{{\langle x^{*},E_{1}(\lambda )\rangle}}}=0$ for every $\lambda
\in\Gamma $, and the same argument shows that ${\ensuremath{{\langle x^{*},E_{1}(\lambda
)\rangle}}}=0$ for every $\lambda \in\T$. Since $\overline{\vphantom{'}\textrm{span}}\,\bigl[E_{1}(\lambda ),E_{2}(\lambda
);\ \lambda
\in\T\bigr]=Z$, $x^{*}=0$, and thus $\overline{\vphantom{'}\textrm{span}}\,\bigl[E(\lambda );\ \lambda
\in\T\bigr]=Z$. Moreover, ${\ensuremath{{\langle e_{0}^{*}+e_{1}^{*},E(\lambda )\rangle}}}=1$ for every $\lambda \in\T$ (where $e_{i}^{*}$ is the functional on $\ell_{p}(\N)$ or $c_{0}(\N)$ mapping a vector $x$ of the space on its $i^{th}$ coordinate). Lastly, $E$ being of class $\mathcal{C}^{\infty }$ on $\T$, the series $$\sum_{n\in\Z}||\widehat{E}(n)||$$ is convergent. So Theorem \[Theo5\] applies. If we consider $\alpha B^{2}$ as acting on ${\ell_{2}(\N)}$, Example \[Example2\] applies directly since $\alpha B^{2}$ is unitarily similar to $M_{\varphi }^{*}$ where $\varphi (z)=\overline{\alpha}
z^{2}$, $z\in\D$.
Many other examples can be obtained along these lines (such as $\bigoplus
_{\ell_{p}}\alpha B$ on the infinite direct sum $\bigoplus_{\ell_{p}}\ell_{p}$, $1\le
p<+\infty $, $|\alpha| >1$,…).
One can observe that all the universal operators presented until now admit eigenvectorfields which are analytic in a neighborhood of some points of $\T$, and in particular have a rather large spectrum. So one may naturally wonder whether this condition is necessary for $A$ to be universal. Our last example, which is rather unexpected, shows that it is not the case. It is to be found within the class of Kalish-type operator on ${L^{2}}(\T)$ (see [@BM] for details).
\[Example4\] Let $A$ be the operator defined on ${L^{2}}(\T)$ by setting for every $f\in{L^{2}}(\T)$ and every $0\le\theta <2\pi $, $$Af(e^{i\theta })=e^{i\theta }f(e^{i\theta })-\int_{0}^{\theta
}ie^{it}f(e^{it})dt.$$ Then $A$ is universal for invertible ergodic systems.
It is not difficult to check that for every $\lambda =e^{i\theta }\in\T$, $0\le\theta<2\pi $, $\ker(A-\lambda )=\textrm{span}\bigl[E(\lambda )
\bigr]$ where $E(\lambda )=\chi _{(\lambda ,1)}$. Here $\chi _{(\lambda ,1)}$ denotes the indicator function of the arc $\Gamma _{\lambda }=\bigl\{e^{i\tau };\ \theta <\tau <2\pi \bigr\}$. The eigenvectorfield $E$ is $1/2$-Hölderian on $\T$. Since the spectrum of $A$ coincides with $\T$, $A$ does not admit any eigenvectorfield which is analytic in a neighborhood of some point of $\T$.
The most obvious idea is to try to apply Theorem \[Theo5\]. Setting $$y_{n}(e^{it})=\int_{0}^{2\pi }e^{-in\theta
}E(e^{i\theta })(e^{it})\dfrac{d\theta }{2\pi }$$ for every $n\in\Z$, we have $$y_{0}(e^{it})=\dfrac{t}{2\pi }\quad\textrm{and}\quad y_{n}(e^{it})
=\dfrac{1-e^{-int}}{2i\pi n}\quad \textrm{for every}\
n\in\Z\setminus\{0\}\cdot$$ The series $\sum_{n\in\Z} y_{n}$ is obviously not unconditionally convergent in ${L^{2}}(\T)$, and Theorem \[Theo5\] cannot be applied this way. So we consider instead of $E$ the unimodular eigenvectorfield $F$ defined by $F(\lambda )=(1-\lambda )E(\lambda )$, and we set $$z_{n}(e^{i\theta })=\int_{0}^{2\pi }e^{-in\theta }F(e^{i\theta })(e^{it })
\dfrac{d\theta }{2\pi }=
\int_{0}^{2\pi }\bigl(e^{-in\theta }-e^{-i(n-1)\theta }
\bigr)E(e^{i\theta })(e^{it })\dfrac{d\theta }{2\pi }\cdot$$ Thus $z_{n}=y_{n}-y_{n-1}$ for every $n\in\Z$. We have for every $n\in\Z\setminus\{0,1\}$ $$z_{n}(e^{it})=\dfrac{1}{2i\pi
}\biggl(\dfrac{e^{-i(n-1)t}}{n-1}-\dfrac{e^{-int}}{n}-\dfrac{1}{n(n-1)}
\biggr)\cdot$$ Since the series $\sum_{|n|\ge 2}1/(n(n-1))$ is convergent, and the series $\sum_{|n|\ge 2}e^{-int}/n$ is unconditionally convergent in ${L^{2}}(\T)$, it follows that the series $\sum_{|n|\ge 2}z_{n}$ (and hence the series $\sum_{n\in\Z} z_{n}$) is unconditionally convergent in ${L^{2}}(\T)$.
Let us now check that the functions $z_{n}$, $n\in\Z$, span a dense subspace of ${L^{2}}(\T)$. Suppose that $f\in{L^{2}}(\T)$ is such that ${\ensuremath{{\langle f,z_{n}\rangle}}}=0$ for every $n\in\Z$. Then $$\begin{aligned}
\dfrac{\widehat{f}(n)}{n}&=\dfrac{\widehat{f}(n-1)}{n-1}+\widehat{f}
(0)\biggl( \dfrac{1}{n}-\dfrac{1}{n-1}\biggr)\quad \textrm{for every }n\not\in\{0,1\}.
\intertext{Summing these equalities for $n\ge 2$, we obtain}
\sum_{n\ge 2}\dfrac{\widehat{f}(n)}{n}&=\widehat{f}(1)+\sum_{n\ge 2}
\dfrac{\widehat{f}(n)}{n}-\widehat{f}(0), \textrm{ so that }
\widehat{f}(1)=\widehat{f}(0).
\intertext{As}
\widehat{f}(n)&=\dfrac{n}{n-1}\widehat{f}(n-1)-\dfrac{1}{n-1}\widehat{f}(0)\end{aligned}$$ for every $n\ge 2 $, $\widehat{f}(2)=2\widehat{f}(1)-\widehat{f}(0)=\widehat{f}(0)$. By induction $\widehat{f}(n)=\widehat{f}(0)$ for every $n\ge 2$, so that $\widehat{f}(n)=0$ for every $n\ge 0$. So $$\widehat{f}(n-1)=\frac{n-1}{n}
\widehat{f}(n)\quad \textrm{ for every }
n\ge -1,$$ which implies that $\widehat{f}(-n)=n\widehat{f}(-1)$ for every $n\ge
1$. Hence $f=0$. So $\overline{\vphantom{'}\textrm{span}}\,\bigl[z_{n};\
n\in\Z
\bigr]={L^{2}}(\T)$ (we have actually proved that $\overline{\vphantom{'}\textrm{span}}\,\bigl[z_{n};\
n\in\Z\setminus\{0,1\}
\bigr]={L^{2}}(\T)$ ).
Lastly, observe that if we set $f_{0}(e^{i\theta })=e^{i\theta }$, then ${\ensuremath{{\langle f_{0},z_{n}\rangle}}}=0$ for every $n$ with $|n|\ge 2$, ${\ensuremath{{\langle f_{0},z_{1}\rangle}}}=-{\ensuremath{{\langle f_{0},z_{-1}\rangle}}}=-1/2i\pi $ and a simple computation shows that ${\ensuremath{{\langle f_{0},z_{0}\rangle}}}=0$. So ${\ensuremath{{\langle f_{0},F(e^{i\theta })\rangle}}}=\bigl(e^{i\theta }-e^{-i\theta } \bigr)/2i\pi $, and all the assumptions of Theorem \[Theo5\] (or, directly, Theorem \[Theo1\]) are satisfied. So $A$ is universal for invertible ergodic systems.
Miscellaneous results and comments {#Sec5}
==================================
We begin this section by investigating necessary conditions for an operator to be universal. We have already seen some such necessary conditions: a universal operator must be frequently hypercyclic, and topologically mixing. If it is universal for all ergodic systems, it cannot be invertible. Without any additional assumption, it seems difficult to say more. But looking at the examples of universal operators presented in Sections \[Sec3\] and \[Sec4\], we observe that all of them admit continuous unimodular eigenvectorfields, and that every $\lambda
\in\T$ is an eigenvalue. This is not completely a coincidence: if $A\in
{\mathcal{B}}(Z)$ satisfies the assumptions of Theorem \[Theo1\], then $A$ admits a continuous unimodular eigenvectorfield defined as $$E(\lambda )=\sum_{n\in\Z}\lambda ^{n}A^{-n}z_{0}.$$ If $A$ satisfies the assumptions of Theorem \[Theo1bis\] with $r=1$ for instance, then $A$ admits a continuous eigenvectorfield on $\overline{\D}$ defined as $E(\lambda
)=\sum_{n\in\N}\lambda ^{n}A^{-n}z_{0}$. In both cases assumption (b) implies that all $\lambda \in\T$ except possibly finitely many are eigenvectors of $A$. Indeed, there exists a finite subset $F$ of $\Z$ and a [non-zero]{} functional $z_{0}^{*}\in Z^{*}$ such that $${\ensuremath{{\langle z_{0}^{*},E(\lambda )\rangle}}}=\sum_{n\in F}\lambda ^{n}{\ensuremath{{\langle z_{0}^{*},A^{-n}z_{0}\rangle}}}.$$ The function ${\ensuremath{{\langle z_{0}^{*},E(\,.\,)\rangle}}}$ is a [non-zero]{} trigonometric polynomial, so it has only finitely many zeroes on $\T$. This implies that $E(\lambda )$ is [non-zero]{} for all $\lambda \in\T$ except possibly finitely many.
So it comes as a natural question to ask whether a universal operator necessarily admits some (or many) unimodular eigenvectors. It is possible to answer this question in the affirmative under two additional assumptions: first that the operator lives on a Hilbert space, and, second, that we add in the definition of the universality the requirement that all the measures $\nu $ on $Z$ involved in the definition have a moment of order $2$.
\[Definition10\] A bounded operator $A$ on $Z$ is said to be $2$-universal for (invertible) ergodic systems if for every (invertible) ergodic dynamical system $(X,{\mathcal{B}},\mu ;T)$ there exists a Borel probability measure $\nu $ on $Z$ which is $A$-invariant, has full support, has a moment of order $2$ (i.e. $\int_{Z}||z||^{2}d\nu (z)<+\infty$), and is such that the two dynamical systems $(X,{\mathcal{B}},\mu ;T)$ and $(Z,{\mathcal{B}}_{Z},\nu ;A)$ are isomorphic.
We have seen in Section \[Sec2\] that operators satisfying the assumptions of Theorems \[Theo1\], \[Theo1bis\], \[Theo1-3\] or \[Theo1-4\] are $2$-universal. Thus all the universal operators presented in Sections \[Sec3\] and \[Sec4\] above are $2$-universal. A first necessary condition for an operator on a Hilbert space to be $2$-universal is
\[Proposition11\] Let $H$ be a complex separable infinite-dimensional Hilbert space, and let $A$ be a $2$-universal operator on $H$. Then $A$ admits a perfectly spanning (and even a $\mathcal{U}_{0}$-perfectly spanning) set of unimodular eigenvectors.
The proof is an easy consequence of some results of [@BG2] and [@BM2] concerning the ergodic theory of linear dynamical systems (see also [@BM]). Since $A$ is $2$-universal, it is weakly mixing with respect to some probability measure $\nu $ with full support which has a moment of order $2$. Consider the centered Gaussian probability measure $m$ on $H$ whose covariance operator $S$ is given by $${\ensuremath{{\langle Sx,y\rangle}}}=\int_{H}{\ensuremath{{\langle x,z\rangle}}}\,\overline{{\ensuremath{{\langle y,z\rangle}}}}\,d\nu
(z)\quad\textrm{for every}\ x,y\in H.$$ Then $$\int_{H}{\ensuremath{{\langle x,z\rangle}}}\,\overline{{\ensuremath{{\langle y,z\rangle}}}}\,dm(z)={\ensuremath{{\langle Sx,y\rangle}}}=\int_{H}{\ensuremath{{\langle x
,z\rangle}}}\,\overline{{\ensuremath{{\langle y,z\rangle}}}}\,d\nu(z)$$ for every $x,y\in H$. Hence $m$ is $A$-invariant and has full support. Moreover, since $A$ is weakly mixing with respect to $\nu $, $$\dfrac{1}{N}\sum_{n=0}^{N-1}\,\,\biggl|\int_{H}{\ensuremath{{\langle x,A^{n}z\rangle}}}\,\overline{
{\ensuremath{{\langle y ,z \rangle}}} }\,d\nu (z)\biggr|^{2}=\dfrac{1}{N}\sum_{n=0}^{N-1}\,\,\biggl|\int_{H}{\ensuremath{{\langle x,A^{n}z\rangle}}}\,\overline{
{\ensuremath{{\langle y ,z \rangle}}} }\,dm (z)\biggr|^{2}$$ tends to $0$ as $N$ tends to infinity for every $x,y\in H$, and this implies (see [@BG1] or [@BM] for details) that $A$ is weakly mixing with respect to $m$. Since it is proved in [@BG2] that any operator on a space of cotype $2$ which is weakly mixing with respect to some Gaussian measure with full support has perfectly spanning unimodular eigenvectors, the first part of the result follows. The second part is proved in exactly the same way: $A$ is necessarily strongly mixing [with respect to]{} some probability measure on $Z$ with full support, and is hence strongly mixing [with respect to]{} some Gaussian probability measure on $Z$ with full support. It is proved in [@BM2] that this implies that the unimodular eigenvectors of $A$ are $\mathcal{U}_{0}$-perfectly spanning (i.e. for any Borel set $B\subseteq \T$ which is a set of extended uniqueneness, $\overline{\vphantom{'}\textrm{span}}\,\bigl[\ker(T-\lambda );\, \lambda \in\T\setminus B\bigr]=H$. See [@BM2] for more about these questions).
If $A$ is supposed to be $2$-universal for ergodic systems (and not only for invertible ones), we can moreover prove that the unimodular point spectrum of $A$ is a subset of $\T$ of full Lebesgue measure. These are the contents of Theorem \[Theo7\], which we now prove.
Let $A$ be a $2$-universal operator on $H$ for ergodic systems. Consider the dynamical system $T$ defined on $\bigl([0,1],{\mathcal{B}}_{[0,1]},dx \bigr)$, where $dx$ is the Lebesgue measure on $[0,1]$, by $Tx=2x\mod 1$. Then $T$ is strongly mixing, and has the following property of decay of correlations: for any $f,g\in{L^{2}}\bigl([0,1] \bigr)$ and $n\ge 0$, define the $n$-th correlation between $f$ and $g$ as $$\mathcal{C}_{n}(f,g)=\int_{0}^{1}f(T^{n}x)\,\overline{\vphantom{'}g(x)}\,dx-
\biggl(\int_{0}^{1}f \biggr)\,.\,\overline{\biggl(\int_{0}^{1}g \biggr)}.$$ Then we have for every $f\in{L^{2}}\bigl([0,1] \bigr)$, $g\in\mathcal{C}^{1}\bigl([0,1] \bigr)$, and $n\ge 0$ $$|\,\mathcal{C}_{n}(f,g)|\le 2^{-n}\,\dfrac{||f||_{2}\,.\,||g'||_{\infty
}}{\sqrt{3}}\cdot$$ Thus the correlations decay exponentially fast provided one of the two functions $f$ and $g$ is sufficiently smooth.
Since $A$ is $2$-universal for ergodic systems, there exists an $A$-invariant measure $\nu $ on $H$ with full support and with a moment of order $2$, for which there exists an isomorphism $\Phi $ between the two dynamical systems $\bigl([0,1],{\mathcal{B}}_{[0,1]},dx;T \bigr)$ and $\bigl(H,
{\mathcal{B}}_{H},\nu ;A\bigr)$. We denote by ${L^{2}}_{0}([0,1])$ (resp. ${L^{2}}_{0}(H,{\mathcal{B}}_{H},\nu )$) the set of functions $f\in
{L^{2}}([0,1])$ (resp. $F\in {L^{2}}(H,{\mathcal{B}}_{H},\nu )$) such that $\int_{0}^{1}f(x)\,dx=0$ (resp. $\int_{H}F(z)\,d\nu (z)=0$). For every functions $F,G\in{L^{2}}_{0}(H,{\mathcal{B}}_{H},\nu )$ such that $G=g\circ\Phi
^{-1}$ for some function $g\in\mathcal{C}^{1}\bigl([0,1] \bigr)
\cap{L^{2}}_{0}\bigl([0,1] \bigr)$, there exists a positive constant $c(F,G)$ such that if we denote for every $n\ge 0$ by $C_{n}(F,G)$ the correlation $$C_{n}(F,G)=\int_{Z}F(A^{n}z)\,\overline{\vphantom{'}G(z)}\,d\nu (z),$$ then $\bigl|C_{n}(F,G) \bigr|\le c(F,G)\,2^{-n}$. Since the measure $\nu $ has a moment of order $2$, one can consider its covariance operator $S$ on $H$ defined as: $${\ensuremath{{\langle Sx,y\rangle}}}=\int_{Z}{\ensuremath{{\langle x,z\rangle}}}\overline{\vphantom{'}{\ensuremath{{\langle y,z\rangle}}}}\,
d\nu (z)$$ for every $x,y\in H$. The operator $S$ is self-adjoint, positive, and of trace class. Since $\nu $ has full support, $S$ has dense range. Hence there exist an orthonormal basis $(e_{l})_{l\ge 1}$ of $H$ and a sequence $(\sigma ^{2}_{l})_{l\ge 1}$ of positive numbers with $\sum_{l\ge 1} \sigma _{l}^{2}<+\infty$ such that $Se_{l}
=2\sigma ^{2}_{l}e_{l}$ for every $l\ge 1$. Also it follows from the orthogonality of the vectors $e_{l}$ in $H$ that the functions ${\ensuremath{{\langle e_{l},\,.\,\rangle}}}$ are orthogonal in ${L^{2}}(H,\mathcal{B}_{H},\nu )$. Let $\mathcal{E}=\overline{\vphantom{'}\textrm{span}}\,\bigl[
{\ensuremath{{\langle e_{l},\,.\,\rangle}}};\ l\ge 1\bigr]$, where the closed linear span is taken in ${L^{2}}(H,\mathcal{B}_{H},\nu )$: $\mathcal{E}$ is a closed subspace of ${L^{2}}(H,\mathcal{B}_{H},\nu )$ which consists of all functions $F\in\mathcal{E}$ which can be written as a convergent series in ${L^{2}}(H,\mathcal{B}_{H},\nu )$ of the form $$F=\sum_{l\ge 1} a_{l}\,{\ensuremath{{\langle e_{l},\,.\,\rangle}}}, \quad \textrm{where}\quad
\sum_{l\ge 1}
|a_{l}|^{2}\,\sigma _{l}^{2}<+\infty .$$ Remark that the function ${\ensuremath{{\langle x,\,.\,\rangle}}}$ belongs to $\mathcal{E}$ for every $x\in H$. We denote by $\iota $ the injection operator $\iota :H{\longrightarrow}\mathcal{E}$ defined by $\iota (x)={\ensuremath{{\langle x,\,.\,\rangle}}}$. If $U_{A}$ denotes the Koopman operator associated to $(H,{\mathcal{B}}_{H},\nu ;A)$, then $U_{A}(\mathcal{E})\subseteq \mathcal{E}$. Proceeding as in the proof of [@BG2 Th. 4.1], we apply the spectral decomposition theorem to $U_{A}$, which is an isometry on ${L^{2}}
(H,{\mathcal{B}}_{H},\nu )$: there exists a finite or countable family $(H_{i})_{i\in I}$ of Hilbert spaces, with either $H_{i}=H^{2}(\T)$ or $H_{i}={L^{2}}(\T,\sigma _{i})$ for some probability measure on $\T$, and an invertible isometry $J:\bigoplus_{i\in I}H_{i}\longrightarrow {L^{2}}
(H, \mathcal{B}_{H},\nu )$ such that $U_{A}J=JM$. Here $M$ acts on $\bigoplus_{i\in
I}H_{i}$ as $M=\bigoplus_{i\in I}M_{i}$, where $M_{i}$ is the multiplication operator by $\lambda $ on $H_{i}$: $(M_{i}f_{i})(\lambda )=
\lambda f_{i}(\lambda )$ for every $f_{i}\in H_{i}$.
Let now $K:\bigoplus_{i\in I}H_{i}\longrightarrow H$ be the operator defined as $K=\iota ^{*}J$. For every $x\in H$, we have $${\ensuremath{{\langle Sx,x\rangle}}}=\int_{H}|{\ensuremath{{\langle x,z\rangle}}}|^{2}\,d\nu (z)=||\iota (x)||^{2}=||K^{*}x||^{2}.$$ It follows that $K^{*}$ is a Hilbert-Schmidt operator, and so there exists for every $i\in I $ a unimodular eigenvectorfield $E_{i}\in {L^{2}}(\T,\sigma _{i};H)$ such that $K^{*}x=\bigoplus_{i\in I}
{\ensuremath{{\langle x,E_{i}(\,.\,)\rangle}}}$ (see [@VTC], [@BG2] or [@BM] for more details). We have thus for every $x,y\in H$, $${\ensuremath{{\langle Sx,y\rangle}}}={\ensuremath{{\langle K^{*}x,K^{*}y\rangle}}}=\sum_{i\in
I}\int_{\T}{\ensuremath{{\langle x,E_{i}(\lambda )\rangle}}}\,\overline{\vphantom{'}{\ensuremath{{\langle y,E_{i}
(\lambda )\rangle}}}}\, d\sigma _{i}(\lambda ).$$ Let now $G\in{L^{2}}_{0}(H,\mathcal{B}_{H},\nu )$ be a function of the form $G=g\circ \Phi
^{-1}$ with $g\in {L^{2}}_{0}\bigl([0,1] \bigr)\cap \mathcal{C}^{1}
\bigl([0,1] \bigr)$, which is such that its orthogonal projection on $\mathcal{E}$, which we denote by $F$, is non-zero. Such a function $G$ does exist because $\bigl\{g\circ\Phi ^{-1};\ g\in {L^{2}}_{0}\bigl([0,1]
\bigr)\cap\mathcal{C}^{1}
\bigl([0,1] \bigr)\bigr\}$ is dense in ${L^{2}}(H,\mathcal{B}_{H},\nu )$. For every $n\ge 0$ we have $$\mathcal{C}_{n}(F,G)=\int_{Z}F(A^{n}z)\,\overline{\vphantom{'}G(z)}\,d\nu
(z)={\ensuremath{{\langle U_{A}^{n}F,G\rangle}}}
={\ensuremath{{\langle U_{A}^{n}F,F\rangle}}}+{\ensuremath{{\langle U_{A}^{n}F,G-F\rangle}}}.$$ Since $U_{A}(\mathcal{E})\subseteq \mathcal{E}$ and $G-F$ is orthogonal to $\mathcal{E}$, the second term vanishes and $$\mathcal{C}_{n}(F,G)=\mathcal{C}_{n}(F,F)=\int_{Z}F(A^{n}z)\,\overline{
\vphantom{'}F(z)}\,d\nu
(z)\quad \textrm{for every}\ n\ge 0.$$ Hence there exists a positive constant $C$ such that for every $n\ge 0$, $|\mathcal{C}_{n}(F,F)|\le C.2^{-n}$. Writing $F$ as $F=\sum_{l\ge
1}a_{l}\,{\ensuremath{{\langle e_{l},\,.\,\rangle}}}$, where $\sum_{l\ge 1}|a_{l}|\,\sigma
_{l}^{2}<+\infty $ and one at least of the coefficients $a_{l}$ is non-zero, we can write $\mathcal{C}_{n}(F,F)$ as $$\begin{aligned}
\mathcal{C}_{n}(F,F)&=\sum_{k,l\ge 1}a_{k}\overline{\vphantom{t}a_{l}}\,
{\ensuremath{{\langle SA^{*n}\,e_{k},e_{l}\rangle}}}\\&=\sum_{k,l\ge
1}a_{k}\overline{\vphantom{t}a_{l}}\,\sum_{i\in I}\int_{\T}
{\ensuremath{{\langle A^{*n}e_{k},E_{i}(\lambda )\rangle}}}\,\overline{\vphantom{'}{\ensuremath{{\langle e_{l},E_{i}(\lambda )\rangle}}}}\,d\sigma _{i}(\lambda )\\
&=\int_{\T}\,\lambda ^{n}\,\sum_{i\in I}\Bigl(
\sum_{k,l\ge
1}a_{k}\overline{\vphantom{t}a_{l}}\,{\ensuremath{{\langle e_{k},E_{i}(\lambda )\rangle}}}\,
\overline{\vphantom{'}{\ensuremath{{\langle e_{l},E_{i}(\lambda )\rangle}}}}\Bigr)\,d\sigma
_{i}(\lambda )\\
&=\int_{\T}\,\,\lambda ^{n}\,\sum_{i\in I}\,\,\,\Bigl|\,
\sum_{k\ge 1}
a_{k}\,{\ensuremath{{\langle e_{k},E_{i}(\lambda )\rangle}}}\,
\Bigr|^{2}\,d\sigma _{i}(\lambda ).\end{aligned}$$ All these computations make sense because $$\int_{\T}\,\sum_{i\in I}\,\Bigl|\lambda ^{n}\,
\sum_{k\ge 1}
a_{k}\,{\ensuremath{{\langle e_{k},E_{i}(\lambda )\rangle}}}
\Bigr|^{2}\,d\sigma _{i}(\lambda )=\int_{Z}|F(z)|^{2}d\nu (z)<+\infty.$$ Let us denote by $\sigma $ the positive finite measure $$\sigma =\sum_{i\in I}\,\,\,\Bigl|\,
\sum_{k\ge 1}
a_{k}\,{\ensuremath{{\langle e_{k},E_{i}(\,.\,)\rangle}}}\,
\Bigr|^{2}\sigma _{i}.$$ Then $|\widehat{\sigma }(n)|\le C.2^{-n}$ for every $n\ge 0$, and since $\sigma $ is a positive measure, $\widehat{\sigma }(-n)={\overline{\widehat{\sigma }(n)}}$ for every $n\ge 0$, so that $|\widehat{\sigma }(n)|\le C.2^{-|n|}$ for every $n\in\Z$. Hence there exists a function $\varphi $ which is analytic in $\mathbb{A}_{1/2}=
\bigl\{\lambda \in\C;\ 1/2<|\lambda |<2\bigr\}$ such that $d\sigma =
\varphi \,d\lambda $, where $d\lambda $ is the normalized Lebesgue measure on $\T$. Since $\mathcal{C}_{0}(F,F)=\widehat{\sigma }(0)>0$ (recall that the function $F$ is [non-zero]{}), the function $\varphi $ cannot be identically zero. If $E$ is a subset of $\T$ of positive Lebesgue measure, it is thus impossible that $E_{i}(\lambda )=0$ for every $\lambda \in E$ and every $i\in I$. So the unimodular point spectrum of $A$ has full Lebesgue measure in $\T$. This proves Theorem \[Theo7\].
The proof of Theorem \[Theo7\] does not extend to operators which are $2$-universal for invertible ergodic systems: the proof uses in a crucial way that the correlations $\mathcal{C}_{n}(f,g)$ of the system $x\mapsto 2x \mod 1$ on $[0,1]$ decay exponentially fast for *all* $f\in{L^{2}}
\bigl([0,1] \bigr)$ and sufficiently smooth $g\in{L^{2}}\bigl([0,1]
\bigr)$. This system is not invertible, and this seems to be in the nature of things that for an invertible system, the correlations decay exponentially fast only for sufficiently smooth functions $f$ *and* $g$ (see [@B] for more on these questions). So the following question remains open:
\[Question12\] If $A$ is a $2$-universal operator for invertible ergodic systems on a Hilbert space, is it true that the unimodular point spectrum of $A$ has full Lebesgue measure?
We do not know any example of a $2$-universal operator whose unimodular point spectrum is not the whole unit circle:
\[Question12bis\] If $A$ is a $2$-universal operator for (invertible) ergodic systems on a Hilbert space, is it true that the unimodular point spectrum of $A$ is equal to $\T$?
If an affirmative answer to this question could be obtained, it would be a first step towards a characterization of symbols $\varphi \in H^{\infty }(\D)$ such that $M^{*}_{\varphi} $ acting on $H^{2}(\D)$ is $2$-universal for invertible ergodic systems.
\[Question13\] Let $\varphi \in H^{\infty }(\D)$. Is it true that $M^{*}_{\varphi} \in
{\mathcal{B}}(H^{2}(\D))$ is $2$-universal for invertible ergodic systems if and only if $\T\subseteq\varphi (\D)$?
It would also be interesting to obtain a characterization of adjoints of multipliers on $H^{2}(\D)$ which are universal for ergodic systems.
\[Question14\] Let $\varphi \in H^{\infty }(\D)$. If $\overline{\D}\subseteq \varphi
(\D)$ (where $\overline{\D}$ denotes the closure of the unit disk $\D$), is $M^{*}_{\varphi} $ universal for ergodic systems? Is it true that $M^{*}_{\varphi}$ is $2$-universal for ergodic systems if and only if $\overline{\D}\subseteq\varphi (\D)$?
Of course things would be simpler if we knew that a universal operator is necessarily $2$-universal, but this does not seem easy to prove. It is not even known whether a frequently hypercyclic operator on a reflexive space admits an invariant measure with full support having a moment of order $2$, although it is known that it admits invariant measures with full support (see [@GM]).
\[Question15\] Does there exist a universal operator which is not $2$-universal?
This brings us back to questions about the existence of unimodular eigenvectors for universal operators.
\[Question16\] Does there exist universal (or $2$-universal) operators admitting no unimodular eigenvalue? What about universal operators on a Hilbert space?
The second half of this question seems hard, again because we do not know whether a frequently hypercyclic operator on a Hilbert space necessarily has some unimodular eigenvalue. The first half of Question \[Question16\] may be more tractable, and a potential example would be the Kalish-type operator $A$ of Example \[Example4\] acting on the space $\mathcal{C}_{0}([0,2\pi])$ of continuous functions on $[0,2\pi]$ vanishing at $0$. It is proved in [@BG2] that although this operator has no unimodular eigenvalue, it admits a Gaussian invariant measure with full support with respect to which it is strongly mixing. We have seen that $A$ acting on ${L^{2}}(\T)$ is universal for invertible ergodic systems, so one may naturally wonder about the following:
\[Question17\] Let $A$ be the bounded operator on $\mathcal{C}_{0}([0,2\pi])$ defined by setting, for every $f\in\mathcal{C}_{0}([0,2\pi])$ and every ${\theta }\in [0,2\pi]$, $$Af({\theta })=e^{i\theta }f({\theta
})-\int_{0}^{\theta}ie^{it}f(e^{it})\,dt.$$ Is $A$ a universal (or $2$-universal) operator for invertible ergodic systems on $\mathcal{C}_{0}([0,2\pi])$?
A positive answer to Question \[Question17\] cannot be obtained via an application of Theorem \[Theo1\], or any variant of it, since we have seen that assumption (c) of Theorem \[Theo1\] for instance implies that $A$ admits a continuous unimodular eigenvectorfield $E$ which is such that $\overline{\vphantom{'}\textrm{span}}\bigl[E(\lambda );\ \lambda \in\T
\bigr]=Z$. So we finish the paper with this last question:
Does there exist any universal operator for invertible ergodic systems which does not satisfy the assumptions of Theorem \[Theo1\]?
[99999]{}
<span style="font-variant:small-caps;">V. Baladi,</span> Positive transfer operators and decay of correlations, *Advanced Series in Nonlinear Dynamics*, **16**, World Scientific Publishing Co., 2000.
<span style="font-variant:small-caps;">F. Bayart, S. Grivaux,</span> Frequently hypercyclic operators, *Trans. Amer. Math. Soc.* [**358**]{} (2006), no. 11, 5083–5117.
<span style="font-variant:small-caps;">F. Bayart, S. Grivaux,</span> Invariant Gaussian measures for operators on Banach spaces and linear dynamics, *Proc. Lond. Math. Soc.* [**94**]{} (2007), 181–210.
<span style="font-variant:small-caps;">F. Bayart, É. Matheron,</span> Dynamics of linear operators, Cambridge Tracts in Mathematics [**179**]{}, Cambridge University Press (2009).
<span style="font-variant:small-caps;">F. Bayart, É. Matheron,</span> Mixing operators and small subsets of the circle, *J. Reine Angew. Math.* (to appear), available at http://arxiv.org/abs/1112.1289. <span style="font-variant:small-caps;">F. Bayart, I. Z. Rusza,</span> Difference sets and frequently hypercyclic weighted shifts, *Erg. Th. Dynam. Systems* **35** (2015), 691–709.
<span style="font-variant:small-caps;">A. Bonilla, K.-G. Grosse-Erdmann,</span> Frequently hypercyclic operators and vectors, *Erg. Th. Dynam. Systems* **27** (2007), 383–404.
<span style="font-variant:small-caps;">N. Feldman,</span> Linear chaos?, preprint 2001, available at http://home.wlu.edu/ feldmann/research.html.
<span style="font-variant:small-caps;">S. Grivaux, É. Matheron,</span> Invariant measures for frequently hypercyclic operators, *Adv. Math.* **265** (2014), 371–427.
<span style="font-variant:small-caps;">E. Glasner, B. Weiss,</span> A universal hypercyclic representation, *J. Funct. Anal.* **268** (2015), 3478–3491.
<span style="font-variant:small-caps;">K. -G. Grosse-Erdmann, A. Peris,</span> Linear Chaos, Universitext, Springer (2011).
<span style="font-variant:small-caps;">R. Jones, J. Rosenblatt, A. Tempelman,</span> Ergodic theorems for convolutions of a measure on a group, *Illinois J. Math.* [**38**]{} (1994), 521–553.
<span style="font-variant:small-caps;">P. Walters,</span> An introduction to ergodic theory, *Graduate Texts in Mathematics* [**79**]{} (1982), Springer, New-York.
<span style="font-variant:small-caps;">N. Vakhania, V. Tarieladze, S. Chobanyan,</span> Probability distributions on Banach spaces, *Mathematics and its Applications (Soviet Series)* **14** (1987), D. Reidel, Dordrecht.
[^1]: This work was supported in part by the Labex CEMPI (ANR-11-LABX-0007-01)
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'M. Guainazzi, P.Grandi, A.Comastri,'
- 'G.Matt'
date: 'Received ; accepted '
title: |
The hard X-ray view of the low-luminosity blazar\
in the radio galaxy NGC 6251
---
Introduction
============
NGC 6251 is an E2 nearby ($z=0.02488$; if $H_{\rm 0} = 70$ km s$^{-1}$ Mpc$^{-1}$, as assumed in this paper, 1$\arcsec$ corresponds to about 500 pc) radio galaxy with a Faranoff-Riley I morphology, known to host a giant radio jet ([@wagget77]), a Seyfert 2 nucleus ([@werner00]), and almost edge-on dust lanes ([@nieto83]). The nucleus is likely to contain a supermassive black hole with mass $M \sim 4$–$8 \times 10^8 M_{{\rm \odot}}$ ([@ferrarese99]), as suggested by the presence of nuclear gas and a dust disc on scales of a few hundreds parsecs. NGC 6251 belongs to the outskirts of the cluster Zw 1609.0+8212 ([@young79]), whose influence on the properties of the galaxy should be, however, marginal ([@prestage88]).
The NGC 6251 jet is one of the most spectacular radio objects of the whole sky. It is a 4$\arcmin$.5 long, highly collimated (opening angle 7.4$^{\circ}$) structure ([@perley84]). Regions of enhanced radio emission along the jet were labeled by Birkinshaw & Worrall (1993) in the 330 MHz radio map as A (10–40$\arcsec$ from the nucleus), B (40–126$\arcsec$; actually structured in smaller sub-structures) and D (178-264$\arcsec$, where the jet bends towards the North; it is as well highly structured). Between B and D one finds a region of low radio brightness (C). None of these region was observed to be a significant source of X-ray emission by ROSAT ([@birkinshaw93]). A sub-pc counter-jet was discovered only recently with Very Long Baseline Interferomer observations ([@sudou01]). U-band [*Hubble Space Telescope*]{} images unveiled a region of extended emission, lying nearly perpendicular to the radio jet axis and the dust ring ([@crane97]), which is likely to originate from scattering of a nuclear continuum source.
Discovered in the X-rays by the [*Einstein*]{} Imaging Proportional Counter ([@jones86]), NGC 6251 showed in the ROSAT PSPC observation ([@birkinshaw93]) an unresolved core ($FWHM < 4\arcsec$), with a possible extended halo on scales $\approxlt 100$ kpc ([@mack97]). The first observation of NGC 6251 covering the intermediate X-rays (i.e.: 2–10 keV) was performed by ASCA. The presence of a bright K$_{\alpha}$ fluorescence iron from ionized iron (centroid energy, $E_c \simeq 6.68$ keV; Equivalent Width, $EW \simeq 600$ eV; [@turner97]; [@sambruna99]), and the fact that the continuum could be best fit with a combination of a standard AGN power-law component (photon index, $\Gamma$, of 2.11) plus a thermal soft excess, suggested that ionized gas may significantly contribute in this energy band. Electron scattering of an otherwise invisible nuclear continuum was an interesting possibility, in light of the HST discovery of extended ionized gas, which could potentially act as a nuclear mirror (“warm mirror" hereinafter). Recently, an association has been proposed between NGC 6251 and the EGRET source 1EGJ1621+8203 ([@mukherjee02]). If this is confirmed, NGC 6251 would be one of the three radio galaxies - alongside with Cen A ([@sreekumar99]) and 3EG J1735-1500 ([@combi03]) - detected in high-energy $\gamma$-rays. No detection by the [*Extreme Ultraviolet Explorer*]{} is reported in the literature ([@marshall95]).
The deep BeppoSAX observation, described in this paper, aimed at verifying the interpretation scenario emerging from the ASCA outcomes. Thanks to its unsurpassed sensitivity in hard X-rays (i.e.: $>$10 keV; [@boella97a]), BeppoSAX would be easily able to detect a transmitted nuclear component piercing through a Compton-thick absorber ([@matt00]). The results of this observation are presented in this paper, together with an analysis of archival ASCA and [*Chandra*]{} data of the same source. The log of the observations discussed in this paper is presented in Table \[tab4\].
---------------------- ------------- ---------------------- --------------------------------------------------------
Satellite Start time Exposure time Count rate
(ks) ($10^{-2}$ s$^{-1}$)
ASCA 28-Oct-1994 36$^a$/44$^b$ $7.2 \pm 0.2$$^a$/$3.63 \pm 0.13$$^b$
[*Chandra*]{}/ACIS-I 11-Sep-2000 25 $8.1 \pm 0.2$
BeppoSAX 19-Jul-2001 28$^c$/80$^d$/72$^e$ $2.78 \pm 0.12$$^c$/$5.55 \pm 0.09$$^d$/$13 \pm 4$$^e$
---------------------- ------------- ---------------------- --------------------------------------------------------
$^a$SIS0, $^b$GIS2, $^c$LECS, $^d$MECS, $^e$PDS.
While this paper was in the process of being accepted, we became aware of a recently accepted paper by Chiaberge et al. (2003), who discuss the same X-ray observations, together with literature and HST data. Their X-ray results are largely coincident with ours.
BeppoSAX results
================
Data reduction
--------------
BeppoSAX data were reduced according to standard procedures as in, e.g., Guainazzi et al. (1999). Scientific products for the imaging Low Energy Concentrator Spectrometer (LECS, [@parmar97], 0.5–4 keV) and Medium Energy Concentrator Spectrometer (MECS, [@boella97b], 1.8–10.5 keV) were extracted from circular regions of 8$\arcmin$ and 4$\arcmin$, respectively. Background spectra were extracted from blank sky field event lists provided by the BeppoSAX Science Data Center, and appropriate for the date of the observation. Products for the Phoswitch Detector System (PDS, [@frontera97], 13–200 keV) were extracted by plain subtraction of those corresponding to intervals when the instrument was observing NGC 6251, and a region $\pm$3.5$^{\circ}$ aside, according to the standard 96 s cycle.
All the spectra employed in this paper have been rebinned in order to oversample the intrinsic energy resolution of the instruments by a factor not larger than 3, and to have a number of counts in each spectral channel higher than 30, in order to ensure the applicability of the $\chi^2$ test. In this paper: energies are quoted in the source rest frame; uncertainties on the spectral parameters are quoted at the 90% confidence level for one interesting parameter; uncertainties on the count rates are at 1$\sigma$ level, unless otherwise specified.
Spectral results
----------------
In Fig. \[fig1\] the results of the fit of the NGC 6251
BeppoSAX spectrum with a simple absorbed power-law is shown. The fit is statistically acceptable ($\chi^2 = 96.7/92$ degrees of freedom, dof). The best-fit parameters and results are reported in Table \[tab1\].
[lcccccccc]{} Model & $N_H$ & $\Gamma$ & $kT_1$/$\Gamma_{{\rm soft}}$ & $kT_2$/$E_{{\rm break}}$ & $E_c$ & $I_c$ & $EW$ & $\chi^2/$ dof\
Mission & ($10^{21}$ cm$^{-2}$) & & (keV)/ & (keV) & (keV) & ($10^{-6}$ ph cm$^{-2}$ s$^{-1}$) & (eV) &\
\
PL & $0.8 \pm^{0.6}_{0.3}$ & $1.79 \pm^{0.06}_{0.07}$ & ... & ... & ... & ... & ... & 96.7/92\
PL+TH & $1.0 \pm^{0.12}_{0.05}$ & $1.70 \pm^{0.12}_{0.16}$ & $1.4 \pm^{0.9}_{0.7}$ & ... & 6.4$^{\dag}$ & $<8.0$ & $< 160$ & 92.3/90\
& & & & ... & 6.59$^{\dag}$ & $<4.5$ & $<80$ &\
& & & & ... & 6.7$^{\dag}$ & $<3.0$ & $<60$ &\
& & & & ... & 6.96$^{\dag}$ & $<2.1$ & $<50$ &\
\
PL & $0.9 \pm^{0.4}_{0.3}$ & $2.06 \pm^{0.08}_{0.07}$ & ... & ... & ... & ... & ... & 259.7/225\
PL+TH & $1.6 \pm^{0.7}_{0.8}$ & $2.00 \pm^{0.08}_{0.07}$ & $0.85 \pm^{0.34}_{0.15}$ & ... & ... & ... & ... & 240.6/223\
PL+2$\times$TH & $1.6 \pm^{2.1}_{0.8}$ & $2.3 \pm 0.4$ & $0.8 \pm^{0.3}_{0.2}$ & $6 \pm^6_2$ & ... & ... & ... & 231.8/221\
PL+TH+GA & $1.7 \pm^{0.5}_{0.4}$ & $2.05 \pm^{0.11}_{0.10}$ & $0.84 \pm^{0.27}_{0.18}$ & & $6.59 \pm 0.16$ & $8 \pm 4$ & $500 \pm 200$ & 228.1/221\
BKPL+GA & $4.0 \pm^{1.8}_{1.5}$ & $2.28 \pm^{0.17}_{0.14}$ & $3.7 \pm^{1.2}_{0.9}$ & $1.38 \pm^{0.15}_{0.09}$ & $6.62 \pm^{0.12}_{0.18}$ & $10 \pm4 $ & $800 \pm 300$ & 245.0/226\
$^{\dag}$Gaussian emission profiles added to the best-fit continuum. Centroid energies $E_c$ are fixed. Upper limits are at the 90% confidence level
The PDS data points lay actually above the best-fit model, which is mainly determined by the statistics of the imaging instruments. The excess is, however, not highly significant. The ratio between the PDS flux and the extrapolation of the best fit model in the LECS/MECS energy band is $1.4 \pm 0.6$, therefore only marginally exceeding the typical values of the cross-normalization factors between the PDS and the MECS (0.80–0.85; [@fiore98]). Not surprisingly, several models provide comparably good descriptions of this hard excess, but none of them is required from the statistical point of view. A broken power-law, for instance, yields a $\Delta \chi^2=3.4$ for a reduction of the number of degrees of freedom ($\Delta \nu$) by 2 (this quantity will be indicated as $\Delta \chi^2/\Delta \nu$ hereinafter), corresponding to a confidence level $\simeq$83.0%
The BeppoSAX large source spectrum extraction region encompasses the putative ROSAT extended emission. We therefore tried to add a soft X-ray component, modeled by a collisionally ionized plasma (we used the [mekal]{} implementation in [Xspec v11.0]{} throughout this paper; [@mewe85], [@liedahl95], [@arnaud92]). Again, the improvement in the quality of the fit is not significant ($\Delta \chi^2 / \Delta \nu = 4.4/2$). Consequently, the temperature is rather poorly determined ($kT = 1.4 \pm^{0.9}_{0.7}$). The 0.5–2.4 keV flux of this component would be $(1.3 \pm 0.7) \times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$, corresponding to about 15% of the non-thermal component in the same energy band. This model requires an absorbing column density slightly in excess with respect to the Galactic contribution along the line-of-sight to NGC 6251 ($5.7 \times 10^{20}$ cm$^{-2}$, [@murphy96]).
No systematic residual feature is present at the energies, where K$_{\alpha}$ fluorescence transitions of iron are expected. Negligible improvements of the $\chi^2$ are yielded by the addition of a narrow (i.e., intrinsic width, $\sigma = 0$) Gaussian profile to the power-law model. 90% upper limits on the intensity of a neutral (6.4 keV) or He-like (6.7 keV) iron lines are 8.5 and $3.0 \times 10^{-6}$ photons cm$^{-2}$ s$^{-1}$, respectively, corresponding to EWs of 160 and 60 eV, respectively. These upper limits are largely inconsistent with the ASCA detection ([@turner97]; cf. also Sect. 3.1 later). Simulations show that a 600 eV EW line would have been detected at the 6-7$\sigma$ confidence level in the MECS spectrum.
The 0.5–10 keV flux during the BeppoSAX observation was $7.5 \times 10^{-12}$ erg cm$^{-2}$ s$^{-1}$, corresponding to un unabsorbed rest-frame luminosity of $1.13 \times 10^{43}$ erg s$^{-1}$ in the same energy band.
Comparison with previous X-ray observations
===========================================
ASCA
----
We have retrieved the data of the ASCA observation of NGC 6251 from the public archive as screened event lists. Spectra were extracted from regions of radii 4$\arcmin$, 3$\arcmin$.1 and 3$\arcmin$.75 in the SIS0, SIS1 (grade 0, 2, 3, 4) and GIS detectors, respectively (they therefore encompass as well the ROSAT extended emission region). Background spectra were extracted from regions in the field of view of the detectors, free from contaminating sources. Response matrices were generated with the packages included in the [Lheasoft]{} v5.0 software. The spectral analysis was performed in the 0.5–10 keV and 0.7–10 keV energy bands, for the SIS and GIS instruments, respectively. The results of our analysis substantially coincide with those presented by Turner et al. (1997) and Sambruna et al. (1999), and we summarize them in this paper for the sake of clarity only.
A simple power-law is a fairly good representation of the ASCA spectra ($\chi^2 = 259.7/225$ dof). However, the addition of a [mekal]{} component improves significantly the quality of the fit ($\Delta \chi^2/\Delta \nu = 19.1 / 2$, corresponding to an F-test confidence level $\simeq$$99.97\%$). The addition of a multitemperature blackbody (model [diskbb]{} in [Xspec]{}) yields, on the contrary, a negligible improvement to the quality of the fit. Modeling the continuum in terms of a broken power-law yields a comparatively worse fit as well. A systematic excess around the energy where fluorescent K$_{\alpha}$ transitions from iron are expected is observed (see Fig. \[fig2\]).
It can be cured with an unresolved Gaussian profile ($\Delta \chi^2/ \Delta \nu = 12.5/2$, corresponding to a confidence level $\simeq$$99.7\%$), or with a hotter ($kT \simeq 6$ keV) thermal component ($\Delta \chi^2 / \Delta \nu = 8.8/2$, corresponding to a confidence level $\simeq$$98.6\%$).
In the best-fit model (cf. Table \[tab1\]), the power-law index ($\Gamma \simeq 2.06$) is significantly steeper than observed by BeppoSAX. The temperature of the colder thermal component is $kT \simeq 0.8$ keV, and its 0.5–2.4 keV flux \[$(1.6 \pm 0.4) \times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$\] is in good agreement with that of the corresponding component in the BeppoSAX spectrum. The iron line centroid energy is consistent with K$_{\alpha}$ fluorescence from mildly ionized to He-like iron. Its $EW$ is $500 \pm 200$ eV, significantly larger than the upper limits measured by BeppoSAX. The difference is most likely due to a change in the underlying continuum, as the intensity of the ASCA line \[$ (8 \pm 4) \times 10^{-6}$ photons cm$^{-2}$ s$^{-1}$\] is marginally consistent with the BeppoSAX upper limits. However, an intrinsic variability of the iron line cannot be ruled out.
The observed 0.5–10 keV flux during the ASCA observation was $2.7 \times 10^{-12}$ erg cm$^{-2}$ s$^{-1}$, corresponding to an unabsorbed total luminosity of $4.4 \times 10^{42}$ erg s$^{-1}$ in the same energy band.
[*Chandra*]{}
-------------
Data of an ACIS-I NGC 6251 [*Chandra*]{} observation were retrieved from the public archive. A bright source \[count rate $(8.1 \pm 0.2) \times 10^{-2}$ s$^{-1}$\] is detected with coordinates: $\alpha_{{\rm 2000}}=16^h32^m31^s.8$; $\delta_{{\rm 2000}}=+82^h32^m16^s$, i.e. 0.6$\arcsec$ distant from the optical nucleus of NGC 6251, well within the accuracy of the [*Chandra*]{} attitude reconstruction. At this count rate level, a [*Chandra*]{} source is likely to be substantially affected by pile-up, given the instrumental mode employed (Time Exposure Mode with a 3.2 s frame time). We have tried to fit the 0.5–8 keV spectrum extracted from the innermost 5$\arcsec$ with a photoelectrically absorbed power-law, corrected for pile-up according to the [Xspec]{} implementation of J.Davis’ algorithm (model [pileup]{} in [Xspec]{}; [@arnaud02]). The fit is acceptable ($\chi^2 = 57.1/78$ dof), and yields best-fit parameters which are consistent with those measured during the BeppoSAX observation, save a 60% lower flux: $N_H = (1.6 \pm 0.5) \times
10^{21}$ cm$^{-2}$; $\Gamma = 1.76 \pm 0.16$; 0.5–10 keV flux of $4.5 \times 10^{-12}$ erg cm$^{-2}$ s$^{-1}$.
Thanks to the unprecedented spatial resolution of the ACIS-S, the issue of the spatial extension can be better addressed. No evidence for extended emission along the arcminutes scale jet is detected, as the ROSAT observation had already shown in the soft X-rays only ([@birkinshaw93]). The upper limits on the X-ray fluxes of the regions A to D are reported in Table \[tab2\].
-------- ---------- -------------- ------------- ------ -----------
Region $d$ Area $CR$ $F$ $S_{\nu}$
(arcsec) (arcsec$^2$) ($10^{-4}$) $^a$ (nJy)
A 10-40 46.8 2.8 2.3 0.34
B 40-126 140.6 2.2 1.8 0.26
C 126-178 78.1 2.9 2.4 0.36
D 178-264 124.9 1.9 1.6 0.24
-------- ---------- -------------- ------------- ------ -----------
: 90% upper limits on the ACIS-I count rates ($CR$), 0.5–10 fluxes ($F$), and 1 keV flux density ($S_{\nu}$) across the NGC 6251 jet regions (following their definition in Birkinshaw & Worrall 1993). The distance $d$ is defined from the inner-outer border of each region to the core. Fluxes are calculated assuming a power-law spectrum, with $\Gamma=2$ and photoelectric absorption column density $N_H = 5.7 \times 10^{20}$ cm$^{-2}$.[]{data-label="tab2"}
$^a$in the 0.5–10 keV band, in units of $10^{-15}$ erg cm$^{-2}$ s$^{-1}$
They are by a factor 5 to 10 tighter than those determined by Birkinshaw & Worrall (1993). However, the upper limit on the knot “D" flux density is inconsistent with the detection ($S_{\nu} = 13 \pm 2$ nJy) obtained by Mack et al. (1997) from a reanalysis of the same ROSAT/PSPC observation discussed by Birkinshaw & Worrall (1993).
In the innermost 4$\arcmin$ around the NGC 6251 core only two sources, alongside with the nucleus itself, are detected at a signal-to-noise ratio higher than 3. Their positions and count rates are reported in Table \[tab3\].
---- ----------------------- ----------------------- ----------------------
\# $\alpha_{{\rm 2000}}$ $\delta_{{\rm 2000}}$ count rate
($10^{-4}$ s$^{-1}$)
1 $16^h32^m31^s.8$ $+82^h32^m16^s$ $810 \pm 20$$^a$
2 $16^h33^m00^s.5$ $+82^h31^m12^s$ $4.8 \pm 1.8$
3 $16^h33^m21^s.3$ $+82^h31^m56^s$ $5.6 \pm 1.8$
---- ----------------------- ----------------------- ----------------------
: Sources detected in the ACIS-I observation of the NGC 6251 field, within 4$\arcmin$ from the core. []{data-label="tab3"}
$^a$NGC 6251 nucleus
None of them exhibits a clear association with any radio structures.
Discussion
==========
The nature of the nuclear X-ray emission in NGC 6251
----------------------------------------------------
The main goal of the BeppoSAX observation described in this paper was to test whether the strong ($EW \sim 600$ eV), ionized K$_{\alpha}$ fluorescent iron line observed in the ASCA spectrum ([@turner97]) was due to a “warm mirror” reflection-dominated, Compton-thick Seyfert 2 spectrum. In the light of the BeppoSAX observation outcomes, this possibility is rather unlikely. The PDS detection ($0.13 \pm 0.04$ s$^{-1}$ in the 13–200 keV energy band) is statistically consistent with the extrapolation of the 2–10 keV spectrum. This constraints the column density of any Compton-thick absorber to the nucleus to be $\approxgt 3 \times 10^{24}$ ($6 \times 10^{24}$) cm$^{-2}$ for a 10% (1%) scattering fraction (see Fig. \[fig9\]).
![PDS count rates in the 13–200 keV energy band against the column density of a Compton-thick absorber covering the NGC 6251 nucleus, for different values of the scattering fraction $f$. The [*shaded area*]{} indicates the PDS detection yielded by the BeppoSAX observation of NGC 6251. $f$ is defined by the ratio between the normalizations of the transmitted and the warm scattered components, assuming an optically thin scatterer []{data-label="fig9"}](h4334F3.ps){width="8cm"}
Moreover, the 2–10 keV BeppoSAX spectrum lacks any trace of fluorescence K$_{\alpha}$ ionized iron lines, which are almost always observed in reflection-dominated Seyfert 2 galaxies ([@turner97b], [@matt00]). The upper limits on their EWs are rather strict, ranging between 60 and 80 eV for He- and H-like iron. The standard AGN 2–10 keV spectral index ($\Gamma \simeq 1.7$), and the lack of K$_{\alpha}$ neutral iron line (again, with a rather tight EW upper limit: 160 eV) in the BeppoSAX spectrum rule out a significant contribution by Compton-reflection from the inner side of the molecular torus ([@krolik94], [@ghisellini94], [@matt96]) In principle, the line flux could be suppressed by resonant trapping at the ionized surface of an accretion disk ([@matt93]). However, further pieces of evidence rule out that the X-ray spectrum is dominated by reflection of an otherwise invisible Seyfert nucleus.
Some hints come from the study of the multiwavelength Spectral Energy Distribution (SED). Recently, Fossati et al. (1998) have proposed a unified scheme to explain the multivawelenght SED of blazars, whose properties (peak frequency of the synchrotron and inverse-Compton components, luminosity ratio between them) are mainly governed by a single parameter related to the overall luminosity. This scheme can be applied to help identifying the nature of the bulk of the X-ray emission from the NGC 6251 nucleus. In Fig. \[fig6\] we show a
![Radio (6 cm), X-ray (0.5–4.5 keV) and IR (25$\mu$) color-color diagram. The [*dotted line*]{} represents the blazar track according to the blazar unified scheme of Fossati et al. (1998). The [ *filled circles*]{} represent the data points corresponding to the NGC 6251 ASCA and BeppoSAX observations. The [*cross*]{} corresponds to the Seyfert 2 sample of Mass-Hesse et al. (1995). The [*empty square*]{} represents the SED of radio-quiet quasars ([@elvis94]) []{data-label="fig6"}](h4334F4.ps){width="8cm"}
color-color diagram between the radio ($\nu = 6$ cm), soft X-ray (0.5–4.5 keV), and infrared ($\lambda = 25 \mu$) luminosities. The dotted line in this diagram represents the trace followed by blazars according to the Fossati scheme. In the same plot the cross represents the position of the Seyfert 2 galaxies of the Mass-Hesse et al. sample (1995). The 6 cm radio luminosity of NGC 6251 is $(1.0 \pm 0.1) \times 10^{40}$ erg s$^{-1}$ ([@jones86]). The data points corresponding to the ASCA and BeppoSAX observations lay intriguingly well on the blazar track. This supports the idea that the bulk of the nuclear emission in NGC 6251 is due to a jet.
Chiaberge et al. (1999) discovered a clear correlation between the optical ($F_O$) and the radio core luminosity ($F_R$) in a sample of FRI galaxies extracted from the 3C catalogue and observed with the HST WFC2. Assuming the NGC 6251 SED published by Ho (1999), $\log (F_R/F_O) \simeq 3.4$, in perfect agreement with the value derived from the 3C correlation ($3.7 \pm 0.4$; the uncertainties represent the r.m.s. scattering of the data points in the correlation), and largely inconsistent with values typically measured in radio-loud quasars ($\simeq$-1; [@elvis94]). This supports a common origin for the radio and optical emission as synchrotron radiation. HST observations in the U-band measured a rather high degree of polarization (close to 50%) in clumps close to the nucleus along the radio axis. Crane & Vernet (1997) suggest that the UV emission of these clumps is due to scattering. However, such an evidence is consistent as well with the possibility that the synchrotron-dominance extends well within the UV range.
We tried to fit the overall NGC 6251 SED from radio to $\gamma$-rays (X-rays represented by the BeppoSAX spectrum) with an homogeneous Synchrotron Self-Compton (SSC) model ([@tavecchio98]). This model assumes that synchrotron radiation is produced by relativistic electrons with density $n_e$ continuously injected in a spherical region of radius $r$ with a magnetic field $B$ and moving with bulk factor $\Gamma_{{\rm bulk}}$ at an angle $\theta$ with respect to the line of sight (the Doppler factor is therefore $\delta = [ \Gamma_{{\rm bulk}} (1- \beta \cos \theta)]^{-1}$). These photons are subsequently upscattered by the same electrons. Following Tavecchio et al. (1998), the electron energy distribution is modeled with a broken power-law with indices $n_1 <3$ and $n_2 > 3$ below and above a break energy $\gamma_b$, respectively. The IRAS data points, although in principle available, have not been included in the fit due to the unknown contamination from the host galaxy. The best fit from radio to $\gamma$ frequencies ([*dashed line*]{} in Fig. \[fig10\])
![NGC 6251 SED. Data are from not-simultaneous observations compiled by Ho (1999), save the and X-ray data (BeppoSAX, [*dots*]{}; this paper), and the putative EGRET detection. The [ *dashed line*]{} represents the best radio-$\gamma$ fit SSC model; the [*dotted line*]{} represents the fit with an SSC model, forced to reproduce the BeppoSAX spectral shape. The ASCA spectrum ([*empty circles*]{}) is shown as well for reference []{data-label="fig10"}](h4334F5.ps){width="8cm"}
fails to reproduce the BeppoSAX spectral shape and slightly deviates below $10^{10}$ Hz. If one forces the model to match the BeppoSAX data point ([*dotted line*]{} in Fig. \[fig10\]), the radio emission is more overproduced. This may indicate that self-absorption (which is not explicitly included in SSC models) plays an important role redwards the synchrotron peak. Given the fact that the SED data points correspond to non-simultaneous observations, one cannot push too far the comparison between the model and the observation. Nonetheless, the overall properties of the best-fit SSC models are not strongly dependent on the details of the fit. The synchrotron peak is below $10^{14}$ Hz, whereas the Compton peak remains constrained between $10^{21.5}$ and $10^{22.5}$ Hz. As a reference the best-fit parameters for the fits are in the range: $B = 0.01$–0.15 G, $\delta = 2$–4, $n_1 = 1.75$–2.6, $n_2 = 4.1$–5.4, $\gamma_b = 6.3 \times 10^3$–$1.6 \times 10^4$ $n_e=10^5$–10$^6$ cm$^{-3}$, and $r = 1$–8$\times10^{16}$ cm. The reader is referred to Chiaberge et al. (2003) for a more detailed comparison between the NGC 6251 SED and SSC models.
It is hard to interpret the spectral variability between the ASCA ([*empty circles*]{} in Fig. \[fig10\]) and the BeppoSAX spectra in terms of pure SSC model. In principle, the ASCA spectrum may represent the trailing edge of the synchrotron component. However, this would imply a shift of the synchrotron peak by more than 4 orders of magnitude, which is rather unlikely. Alternatively, the steeper ASCA spectrum may imply that a different spectral component may be emerging during phases of low X-ray flux. This component may be completely outshined during BeppoSAX-like, X-ray brighter states. We will further discuss this possibility in Sect.4.2, in connection with the strong fluorescent iron line observed in the ASCA spectra only.
Recently, it has been suggested that the bulk of the X-ray emission in radio galaxies may be due to a hot inner accretion flow, following a line of thought suggested more than 20 years ago by Rees et al. (1982). Ho (1999) estimates the bolometric luminosity of NGC 6251 as $\sim$$10^{-4}$ L$_{{\rm Edd}}$. This may indicate that an Advection Dominated Accretion Flow (ADAF; [@narayan95]) is responsible for the bulk of the emission in the NGC 6251 core. In this scenario, X-rays are mainly produced via bremsstrahlung by a distribution of thermal electrons with typical temperatures $kT \sim 100$ keV ([@dimatteo00]). In Fig. \[fig8\] we compare
![Radio between the 1 keV and the V-band flux density in: a) NGC 6251 ([ *filled circle*]{}); b) the blazar radio luminosity classes in the unified scenario after Fossati et al. (1998); c) ADAF models applied to a sample of nearby elliptical galaxies ([@dimatteo00]; details in text). In the last case, the error bars on the y-axis represent the r.m.s. of the sample values in the corresponding radio luminosity interval. []{data-label="fig8"}](h4334F6.ps){width="8cm"}
the ratio between the V and the 1 keV flux density (BeppoSAX measurement) in NGC6251 with the values expected by the blazar unification scenario ([@fossati98]) and by models of ADAF applied to the multiwavelength SED of nearby elliptical galaxies suspected to host low radiative efficiency accretion flows ([@dimatteo00]; we consider hereby models with wind, and outermost hot accretion radius of 300 gravitational radii). The NGC 6251 data point is in principle consistent with both scenarios. Applying a bremsstrahlung model to the BeppoSAX spectrum, one gets indeed a total X-ray luminosity $L_X/L_{{\rm Edd}} = (2 \pm 1) \times
10^{-3}$. However, the fit is significantly worse than with a simple power-law ($\chi^2 = 110.0/92$ dof). A composite bremsstrahlung and power-law fit (the latter component taking into account possible Comptonization effects) yields an implausibly low electron temperature ($kT \simeq 3$ keV). These pieces of evidence are in agreement with the conclusions drawn by Ferrarese & Ford (1999), who remark that the [*nuclear*]{} non-thermal bolometric optical luminosity in NGC 6251 is larger than predicted for accretion at the Bondi rate with 10% efficiency, making the case for an ADAF far less compelling.
Additional components in the nuclear X-ray spectrum of NGC 6251
---------------------------------------------------------------
As already noticed by Turner et al. (1997), a soft X-ray excess is observed in the NGC 6251 ASCA spectra. A multitemperature blackbody is not able to fit the excess, arguing against the possibility that this component originates in an accretion disk. The best-fit temperature of the soft excess, if modeled with an optically thin, collisionally excited plasma, is $kT \simeq 0.8$ keV, with an unabsorbed luminosity of $\simeq 3.6
\times 10^{41}$ erg s$^{-1}$. Such a temperature is typical of gaseous halos in elliptical galaxies ([@matsumoto97]). The parameters of this component are consistent with those of a similar component in the BeppoSAX spectrum, which is, however, not required from the statistical point of view. The ASCA temperature is in turn consistent, within the rather large statistical uncertainties, with the temperature of the extended halos measured by the [*Einstein*]{}/IPC and the ROSAT/PSPC in NGC 6251 (for the latter, $kT = 0.3$–0.8 keV; [@birkinshaw93] and references therein). Similarly, the 0.5–2.4 keV unabsorbed fluxes of this component measured by ASCA and BeppoSAX are consistent with the ROSAT/PSPC soft X-ray “halo” flux, integrated across its whole extension \[$(2.4 \pm ^{4.2}_{2.0}) \times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$; [@mack97]\]. It is therefore straightforward to identify the soft excess in the large ASCA aperture with the diffuse emission in ROSAT. Mack et al. (1997) discuss (and rule out) the possibility that such a plasma could be the confining medium for the jet, along the whole structure up to knot “D". A $\simeq$$10^7$ K thermal emission responsible for the confinement of the jet in NGC 6251 should not extend beyond scales larger than 60 pc. On such small scales, it might be marginally resolvable by [*Chandra*]{}. Unfortunately, the pile-up affecting the [*Chandra*]{} observation prevented us from deriving precious constraints on the X-ray extension around the NGC 6251 core.
The ASCA spectrum unveils the possible presence of a third spectral component, whose main signature is a bright ($EW \simeq 600$) eV and significantly ionized K$_{\alpha}$ fluorescence iron line. This component, even if present with comparable flux in the BeppoSAX observation, would have easily missed detection, due to the fact that the BeppoSAX non-thermal continuum 6 keV flux density was $\ge$6 times larger than in ASCA. This “third” component cannot be produced by the superposition of unresolved discrete sources, integrated in the large ASCA aperture. The [*Chandra*]{} image shows only two sources in the innermost 4$\arcmin$ around the NGC 6251 core, whose total count rate \[($1.1 \pm 0.3) \times 10^{-3}$\], corresponds to a 0.5–9 keV flux $\simeq$$1.0 \times 10^{-14}$ erg cm$^{-2}$ s$^{-1}$ (assuming a thermal plasma with $kT = 6$ keV). This is more than one order of magnitude less than the flux of the ASCA component in the same energy band ($2 \times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$). This component may therefore represent the “tip of the iceberg” of an underlying Seyfert nuclear “warm scattering” (Matt et al. 1996, 2001), whose relative contribution becomes not-negligible during phases of lower blazar activity. Alternatively, this component may represent a hotter and weaker phase of the thermal emission. It has already been pointed out that the kpc-scale jet structure in NGC 6251 requires an ambient gas with temperatures of 2–5 keV to confine it. So far, the lack of detection of a gas component with such an high-temperature had led to the problem of an “over-pressurized“ jet in NGC 6251. It would be tempting to speculate that the high-temperature thermal component measured by ASCA is the required ”missing link“ to ensure the thermodynamical stability of the kpc-scale jet. Unfortunately, only rather strict upper limits can be set on any X-ray emission associated with the radio jet (see Table \[tab2\]). Alternatively, the line could be produced in the interaction between the ”bloated" base of a jet with a stratified velocity structure and the circumnuclear matter, as proposed by Chiaberge et al. (2000) in the context of a possible unification model between BL Lac objects and FR I radio galaxies. The possibility that the iron line is associated with the highly ionized “skin" of an accretion disk cannot be in principle ruled out, although evidence for significantly ionized disk lines is rather scanty so far.
All the best-fit X-ray models discussed in this paper require an amount of cold photoelectric absorption in excess to the Galactic contribution. The weighted between the BeppoSAX and ASCA measurements is $N_H = (1.06 \pm 0.11) \times
10^{21}$ cm$^{-2}$. This is in excellent agreement with the column density through the dusty disk, as derived from its visual extinction ($A_V = 0.61 \pm 0.12$; [@ferrarese99]), if standard gas-to-dust Galactic ratios are assumed. This is in agreement with the idea that the standard pc-scale optically and geometrically thick torus - even if present in FRI low-luminosity radio galaxies - does not intercept our line-of-sight to the nucleus ([@chiaberge02]).
This paper benefitted of the stimulating scientific environment at the Workshop “AGN spectroscopy with [*Chandra*]{} and XMM-Newton”, held at the Max Planck Institut für Extraterrestrische Physik in Garching. Support by Maria Teresa Fiocchi in using the SSC models fitting facility at the A.S.I. Science Data Center is gratefully acknowledged. Last, but not least, comments by the referee (Dr. M. Boettcher) greatly helped us to sharpen the focus of the discussion.
Arnaud, K., & Dorman, B., “XSPEC User Manual” v11.0
Arnaud, M., & Raymond, J., 1992, ApJ, 398, 394
Birkinshaw, M., & Worrall, D.M., 1993, ApJ 412, 568
Boella, G., Butler, R.C., Perola, G.C., et al., 1997, A&AS 122, 299
Boella, G., Chiappetti, L., Conti, G., et al., 1997, A&AS 122, 327
Chiaberge, M., Capetti, A., & Celotti, A., 1999, A&A, 349, 77
Chiaberge, M., Celotti, A., Capetti, A., & Ghisellini, G., 2000, A&A, 358, 104
Chiaberge, M., Duccio Macchetto, F., Sparks, W.B., et al., 2002, ApJ, 571, 247
Chiaberge, M., Gilli, R., Capetti, A., & Duccio Macchetto, F., 2003, ApJ, in press (astroph/0307120)
Combi, J.A., Romero, C.E., Paredes, J.M., Torres, D.F., & Ribó, M., 2003, ApJ, 588, 731
Crane, P., & Vernet, J., 1997, ApJ 486, L91
Di Matteo, T., Quataert, E., Allen, S.W., Narayan, R., & Fabian, A.C., 2000, MNRAS, 311, 507
Elvis, M., Wilkes, B.J., McDowell, J.C., et al., 1994, ApJS, 95, 1
Ferrarese, L., & Ford, H.C., 1999, ApJ 515, 583
Fiore, F., Guainazzi, M., & Grandi, P., 1998, “Cookbook of BeppoSAX data analysis”, (BeppoSAX SDC:Roma)
Fossati, G., Maraschi, L., Celotti, A., Comastri, A., & Ghisellini, G., 1998, MNRAS, 299, 433
Frontera, F., Costa, E., dal Fiume, D., et al., 1997, A&AS 122, 347
Ghisellini, G., Haardt, F., & Matt, G., 1994, MNRAS 267, 743
Ho, L.C., 1999, ApJ, 516, 672
Krolik, J.H., Madau, P., & ${\rm \dot{Z}}$ycki P.T., 1994, ApJ 420, L57
Jones, D.L., Unwin, S.C., Readhead, A.C.S., et al., 1986, ApJ 305, 684
Liedahl, D.A., Osterheld, A.L., & Goldstein, W.H., 1985, ApJL, 438, 115
Mack, K.H., Kerp, J., & Klein, U., 1997, A&A, 324, 870
Marshall, H.L., Fruscione, A., & Carone, T.E., 1995, ApJ, 439, 90
Matsumoto, H., Koyama, K., Awaki, H., et al., 1997, ApJ, 482, 133
Matt, G., Fabian, A.C., & Ross, R.R., 1993, MNRAS, 261, 346
Matt, G., Brandt, W.N., & Fabian, A.C., 1996, MNRAS, 280, 823
Matt, G., Fabian, A.C., Guainazzi, M., et al., 2000, MNRAS, 318, 173
Matt, G., Guainazzi, M., Perola, G.C., et al., 2001, A&A, 377, L31
Mewe, R., Gronenschild, E.H.B.M., & van der Oord, G.H.J., 1985, A&AS, 62, 197
Mukherjee, R., Halpern, J., Mirabal, N., & Gotthelf, E.V., 2002, ApJ, 574, 693
Murphy, E.M., Lockman, F.J., Laor, A., & Elvis, M., 1996, ApJS, 105, 369
Narayan, R., & Yi, I., 1995, ApJ, 444, 231
Nieto, J.C., Coupinot, G., LeLievre, G., & Madsen, C., 1983, MNRAS, 203, 39P
Parmar, A.N., Martin, D.D.E., Bavdaz, M., et al., 1997, A&AS, 122, 309
Perley, R.A., Bridle, A.H., & Willis, A.G., 1984, ApJS, 54, 291
Prestage, R.M., & Peacock, J.A., 1988, MNRAS, 230, 131
Rees, M.J., Begelman, M.C., Blandford, R.D., & Phinney, E.S., 1982, Nat, 295, 17
Sambruna, R., Eracleous, M., & Mushotzky, R., 1999, ApJ, 526, 60
Sreekumar, P., Bertsch, D.L., Hartman, R.C., Nolan, P.L., & Thompson, D.J., 1999, Astropar. Physics, 11, 221
Sudou, H., Taniguchi, Y., Ohyama, Y., et al., 2001, PASJ, 52, 989
Tavecchio, F., Maraschi, L., & Ghisellini, G., 1998, ApJ, 509, 608
Turner, T.J., George, I.M., Nandra, K., & Mushotzky, R.F., 1997, ApJS, 113, 23
Turner, T.J., George, I.M., Nandra, K., & Mushotzky, R.F., ApJ, 488, 164
Wagget, P.C., Warner, P.J., & Baldwin, J.E., 1977, MNRAS, 181, 465
Werner, P.M., Worrall, D.M., & Birkinshaw, M., 2000. MNRAS, 317, 105
White, R.L., & Becker, R.H., 1992, ApJS, 79, 331
Young P.J., Sargent, W.L.W., Kristian, J., & Westphal, J.A., 1979, ApJ, 234, 76
|
{
"pile_set_name": "ArXiv"
}
|
hep-th/0009083\
September 2000\
[ Supersymmetric Brane World Scenarios from Off-Shell Supergravity]{}\
[Max Zucker]{}[^1]
[*[ Physikalisches Institut\
Universität Bonn\
Nussallee 12\
D-53115 Bonn, Germany]{}*]{}\
[Abstract]{}
Using $N=2$ off-shell supergravity in five dimensions, we supersymmetrize the brane world scenario of Randall and Sundrum. We extend their construction to include supersymmetric matter at the fixpoints.
Introduction
============
During the last months, an idea due to Randall and Sundrum [@Randall:1999ee] has gained a lot of attention. These authors consider five dimensional gravity with a cosmological constant on the orbifold $S^1/\mathbb{Z}_2$. In addition, there are cosmological constants located at the fixpoints of the $\mathbb{Z}_2$. The equations of motion of this theory are solved using a warped product ansatz which preserves four dimensional Poincaré invariance. The exponential dependence of the warp factor on the fifth direction then leads to an elegant solution of the hierarchy problem.
Clearly, it would be of great interest to supersymmetrize the Randall-Sundrum scenario. To be more precise, one may ask for five dimensional gauged supergravity [@D'Auria:1981kq] on $S^1/\mathbb{Z}_2$ with additional cosmological constants on the branes and supersymmetrizations thereof.
Some work has already been devoted to the topic of supersymmetric brane world scenarios. Let us mention the paper of Falkowski et al. [@Falkowski:2000er], which discusses models where the supergravity in the bulk is slightly modified; namely the bulk mass term for the gravitino contains a step function. Another approach is the work of Altendorfer et al. [@Altendorfer:2000rr], who keep the bulk theory as it stands and find the terms on the boundaries by requiring the warped product metric of Randall-Sundrum being a supersymmetric vacuum. However, their method has the disadvantage that it cannot be extended straightforwardly to include additional matter on the boundaries. Finally, we should mention a more recent preprint, which treats the subject in a more general framework [@Bergshoeff:2000zn].
In this work we present an alternative and as we think quite powerful technique, suited for the derivation of theories of the Randall-Sundrum type. This technique rests mainly on two ingredients. First, Mirabelli and Peskin [@Mirabelli:1998aj] presented a very elegant method to couple theories, which live in the bulk of the orbifold $S^1/\mathbb{Z}_2$ to fields located at the fixpoints in a supersymmetric way. However, they apply their method to rigidly supersymmetric theories only, which is in the light of Randall-Sundrum to restrictive. We stress that the method of [@Mirabelli:1998aj] rests upon the use of off-shell formulations. So the second ingredient is clear: An off-shell formulation of gauged supergravity in five dimensions is required in order to generalize the idea of Mirabelli and Peskin to local supersymmetry.
In two recent publications, we have worked out the $N=2$ off-shell multiplet calculus in five dimensions. Besides the minimal multiplet and the nonlinear multiplet [@Zucker:2000ej], we discussed in [@Zucker:1999fn] the linear and the super Yang-Mills multiplet. The multiplet calculus is completed with the hypermultiplet which may be found in [@diss] and the tensor multiplet, to be discussed in the following section. In a recent publication [@Kugo:2000hn] these topics have been reconsidered and extended using a different approach.
The purpose of this paper is to use our off-shell calculus in order to generalize the idea of Mirabelli and Peskin to local supersymmetry and derive in that way lagrangians of the Randall-Sundrum type.
The present paper is structured as follows. In section \[gos\] we construct the bulk theory of our orbifold $S^1/\mathbb{Z}_2$, namely gauged $N=2$ off-shell supergravity in five dimensions. In section \[sgs\] we develop a four dimensional $N=1$ tensor calculus at the fixpoints. Using these results we finally study various supersymmetric theories on $S^1/\mathbb{Z}_2$ in section \[chap4\]. Short conclusions and an outlook are presented in section \[conc\].
Gauged off-shell supergravity\[gos\]
====================================
In this section we construct the gauged supergravity which will be the bulk theory for our orbifold construction.
The fundamental building block is the minimal multiplet. It has been discussed in our earlier works, [@Zucker:2000ej; @Zucker:1999fn], so that we refer to these references for detailed treatments. In order to find a consistent supergravity theory, there is a compensator for the local $SU(2)_{\cal R}$ required. In principle there are three possibilities, leading to three different versions of off-shell supergravity: the nonlinear multiplet [@Zucker:2000ej] (version I), the hypermultiplet [@diss] (version II) and the tensor multiplet (version III) which is discussed in this work and will be used for the orbifolding procedure. These are all (minimal) off-shell versions of $N=2$ supergravity which exist in five dimensions. These results are easily extended to gauged supergravity.
A detailed presentation of the three versions of off-shell supergravity, including explicit expressions for the lagrangians, may be found in [@diss]. An analysis of the behavior of the gauged theories under $\mathbb{Z}_2$ orbifolding is included in that work.
The tensor multiplet {#tenm}
--------------------
We start by reminding the reader of the linear multiplet [@Zucker:1999fn]. It contains a Lorentz scalar isotriplet $\vec{Y}$, a spinor $\rho$, a scalar $N$ and a vector $W^A$. The vector $W^A$ is constrained (for detailed formulas see [@Zucker:1999fn] and for the centrally charged multiplet [@diss])[^2] $$\partial_A W^A+\ldots=0.\label{cons}$$ As in four dimensional conformal supergravity [@deWit:1983na], one may solve this constraint explicitly for vanishing gauge group and vanishing central charge. We can achieve that by the introduction of a 3-form tensor potential through $$W^A = \frac{1}{12}e_M^A\varepsilon^{MNPQR}\widehat{\cal D}_N
B_{PQR}\label{dual}.$$ This field forms together with the remaining fields of the linear multiplet the tensor multiplet. Schematically, the tensor multiplet is then given by $$(~\vec{Y},~\rho,~ B_{MNP},~ N).$$ The introduction of the field $B_{MNP}$ implies a new symmetry, namely tensor gauge transformations $$\delta_\Lambda(\lambda)B_{MNP}=3\partial_{[M}\lambda_{NP]}.\label{tenseich}$$ From the constraint (\[cons\]) it is possible to deduce the supersymmetry transformation law of the tensor field. One finds $$\delta B_{MNP} =
-i\bar{\varepsilon}\gamma_{MNP}\rho-3\bar{\varepsilon}\vec{\tau}\gamma_{[NP}\psi_{M]}\vec{Y}.$$ It then follows that the supercovariant derivative in (\[dual\]) is given by $$\widehat{\cal D}_{[M} B_{NPQ]} = \partial_{[M}
B_{NPQ]}+i\bar{\rho}\gamma_{[NPQ}\psi_{M]}-\frac{3}{2}\bar{\psi_{[N}}\vec{\tau}\gamma_{PQ}\psi_{M]}\vec{Y}.$$ The tensor gauge transformations (\[tenseich\]) appear in the supersymmetry algebra. The additional terms are $$[\delta_Q(\eta),\delta_Q(\varepsilon)]=\ldots +
\delta_\Lambda(i\bar{\varepsilon}\gamma^P\eta
B_{MNP}-\bar{\varepsilon}\vec{\tau}\gamma_{MN}\eta\vec{Y})$$ and the dots denote field dependent transformations which may be found in [@Zucker:2000ej].
The action for the tensor multiplet {#tenact}
-----------------------------------
An action for the tensor multiplet is found as follows [@deWit:1983na; @Bergshoeff:1986mz]: one starts with the coupling of a Maxwell multiplet to the tensor multiplet. The tensor multiplet is inert under the gauge group. One easily finds that $$\begin{split}
{\cal L} & = 2
\vec{X}\vec{Y}-2i\bar{\Omega}\rho-2MN-\frac{1}{12}\varepsilon^{ABCDE}G_{AB}B_{CDE}-16\vec{Y}\vec{t}M\\ &
-i\bar{\rho}\gamma^M\psi_M M
+\frac{1}{2}\bar{\psi}_A\vec{\tau}\gamma^{AB}\psi_B\vec{Y}M-\bar{\Omega}\vec{\tau}\gamma^M\psi_M\vec{Y}\label{tensoraction}
\end{split}$$ is invariant. Next, one forms a vector multiplet out of the fields of the tensor multiplet. Here one faces the first problem. The canonical embedding should start with $M\sim N+\ldots$ as it does in the case of rigid supersymmetry. However, this cannot be generalized straightforwardly to the local case [@deWit:1983na]. Instead, one has to move on to the improved tensor multiplet [@deWit:1983na; @deWit:1982fh].
Since the correspondence is rather complicated, we give only the lowest component $M$ of this vector multiplet (the full formulas may be found in [@diss]): $$M =
Y^{-1}N+\frac{1}{4}Y^{-3}\bar{\rho}\vec{\tau}\rho\vec{Y}+6Y^{-1}\vec{Y}\vec{t}.\label{emb1}$$ Here we have defined $$Y=(\vec{Y}\vec{Y})^{1/2}$$ and let us mention that similar cohomology problems as in four dimensions [@deWit:1983na] appear for the gauge field strength but are unimportant for our purposes. Acting repeatedly with the supersymmetry transformations on expression (\[emb1\]) gives the complete embedding.
Using this embedding of the tensor multiplet in the vector multiplet in the action formula (\[tensoraction\]) gives the desired action for the tensor multiplet: $$\begin{split}
{\cal L}_{tensor} = &
-\frac{1}{4}YR(\widehat{\omega})^{AB}{}_{AB}+4YC-\frac{1}{6}Y\widehat{F}_{AB}\widehat{F}^{AB}+Yv_{AB}v^{AB}+20
Y\vec{t}^2 - Y^{-1}N^2\\
& - 36Y^{-1}(\vec{t}\vec{Y})^2 -\frac{1}{4}Y^{-1}\widehat{\cal
D}_A\vec{Y}\widehat{\cal D}^A\vec{Y}-
\frac{1}{12}Y^{-1}\varepsilon^{MNPQR}\vec{Y}\vec{V}_M\partial_N B_{PQR}\\
& + Y^{-1}W_A W^A- 4Y^{-1}\bar{\lambda}\vec{\tau}\rho\vec{Y}
-
2iY\bar{\psi}_A\gamma^A\lambda-\frac{i}{4\sqrt{3}}Y^{-1}\bar{\rho}\gamma^{AB}\rho\widehat{F}_{AB}\\
& - \frac{i}{2}Y^{-1}\bar{\rho}\gamma^A {\cal
D}_A\rho-3Y^{-3}\bar{\rho}\vec{\tau}\rho\vec{Y}\tilde{t}\tilde{Y} -
\frac{1}{2}Y^{-3}\bar{\rho}\vec{\tau}\rho\vec{Y}N-\frac{i}{4}Y^{-1}\bar{\rho}\gamma^{AB}\rho
v_{AB}\\
& +\frac{1}{4}Y^{-3}\bar{\rho}\vec{\tau}\gamma^A\rho(\vec{Y}\times
\widehat{\cal D}_A\vec{Y})
+\frac{1}{2}Y^{-1}\vec{Y}\bar{\psi}_A\vec{\tau}\gamma^{AB}{\cal D}_B\rho+ iY^{-1}\bar{\rho}\psi_A W^A\\
& + \frac{1}{2}Y^{-3}\bar{\rho}\vec{\tau}\gamma^A\rho \vec{Y}W_A -
\frac{1}{2}Y^{-1}\bar{\rho}\vec{\tau}\gamma^{MN}{\cal D}_M\psi_N
\vec{Y}-\frac{i}{2}Y\bar{\psi}_P\gamma^{PMN}{\cal D}_M\psi_N\\
& -12 Y^{-1}N\vec{t}\vec{Y}-Y\bar{\psi}_A\vec{\tau}\gamma^{AB}\psi_B
\vec{t}-\frac{i}{2}Y\bar{\psi}_A\psi_B
v^{AB}+2Y^{-1}\bar{\psi}_A\vec{\tau}\gamma^A\rho(\vec{t}\times\vec{Y})\\
& +
\frac{1}{24}Y^{-3}\varepsilon^{MNPQR}\vec{Y}(\partial_M\vec{Y}\times\partial_N\vec{Y})B_{PQR}
-
\frac{i}{4\sqrt{3}}Y\bar{\psi}_A\gamma^{ABCD}\psi_B\widehat{F}_{CD}\\
& + Y^{-1}\bar{\rho}\vec{\tau}\gamma_B\psi_A v^{AB}\vec{Y}
+2iY^{-1}\bar{\rho}\gamma^A\psi_A\vec{t}\vec{Y}-
\frac{1}{2\sqrt{3}}Y^{-1}\bar{\rho}\vec{\tau}\gamma^{ABC}\psi_A\widehat{F}_{BC}\vec{Y}\\
& -
\frac{1}{4}Y^{-1}\bar{\psi}_A\vec{\tau}\gamma^{ABC}\psi_B(\vec{Y}\times\widehat{\cal D}_C\vec{Y}) -
\frac{1}{2}Y^{-3}\bar{\psi}_A\vec{\tau}\gamma^{AB}\rho\vec{Y}\tilde{Y}\widehat{\cal
D}_B\tilde{Y}+\frac{1}{8}Y^{-3}(\bar{\rho}\vec{\tau}\rho)^2\\
&
-\frac{i}{2}Y^{-3}\bar{\psi}_A\gamma^A\rho\bar{\rho}\vec{\tau}\rho\vec{Y}+\frac{1}{2}Y^{-1}\bar{\rho}\psi_B\bar{\rho}\gamma^{AB}\psi_A-\frac{i}{4}Y^{-1}\bar{\rho}\psi_B\bar{\psi}_A\vec{\tau}\gamma^{ABC}\psi_C\vec{Y}\\
& -\frac{3}{8}Y^{-5}(\bar{\rho}\vec{\tau}\rho\vec{Y})^2 +
\frac{1}{8}Y^{-3}\bar{\psi}_A\vec{\tau}\gamma^{AB}\psi_B\bar{\rho}\tilde{\tau}\rho\vec{Y}\tilde{Y}
+
\frac{i}{4}Y^{-3}\bar{\psi}_A\vec{\tau}\tilde{\tau}\gamma^A\rho\bar{\rho}\tilde{\tau}\rho\vec{Y}\\
&
-\frac{i}{4}Y^{-3}\bar{\rho}\vec{\tau}\gamma_B\rho\bar{\rho}\gamma^{AB}\psi_A\vec{Y} -
\frac{1}{8}Y^{-3}\bar{\rho}\vec{\tau}\gamma_B\rho\bar{\psi}_A\tilde{\tau}\gamma^{ABC}\psi_C\vec{Y}\tilde{Y}.\label{tensact}
\end{split}$$
An action for gauged supergravity {#sugact}
---------------------------------
Using the results of the preceding section we now construct an action for gauged off-shell supergravity following [@deWit:1983na]. As compensator for the $SU(2)_{\cal R}$ ${\cal R}$-symmetry we use the $\vec{Y}$-field. However, before dealing with the gauge fixing, we have to consider an old problem: The appearance of a term linear in $C$ in eq. (\[tensact\]) leads to inconsistent equations of motion. We can solve this problem in an elegant way, using the lagrangian ${\cal L}_{min}$ for the minimal multiplet, eq. (2.10) in [@Zucker:2000ej]. The lagrangian ${\cal L}={\cal L}_{tensor}+{\cal L}_{min}$ leads to consistent equations of motion. It is a well defined lagrangian for off-shell supergravity.
The gauged variant of this theory is also easily found: Using the lagrangian ${\cal L}_{lin}$ for the linear multiplet given in [@Zucker:1999fn], eq. (3.2), the desired lagrangian is $${\cal L}_{gauged}={\cal L}_{tensor}+{\cal L}_{min}-\frac{\sqrt{3}}{4}
g'{\cal L}_{lin}.\label{orbi1}$$ Here we have introduced the cosmological constant $g'$. Of course, wherever $W^A$ appears it has to be understood as supercovariant field strength for $B_{MNP}$, defined by (\[dual\]).
Before we proceed, we should discuss the gauge fixing. In our formulation of off-shell supergravity, we have an $SU(2)_{\cal R}$ ${\cal
R}$-symmetry which is gauged by an auxiliary field $\vec{V}_M$. This symmetry has to be fixed. We do that by setting [@deWit:1983na] $$\vec{Y}=e^u(0,1,0)^T\label{eichfixx}$$ where we have introduced a new scalar $u$. This breaks the original $SU(2)_{\cal R}$ but leaves a residual $SO(2)$, gauged by $V_M^2$, intact. In addition, there is still the $U(1)$ under which only the graviphoton $A_M$ transforms at this stage. The equations of motion for the auxiliary fields imply $V_M^2=2g'A_M$ so that after elimination of the auxiliary fields, we end up with the lagrangian of gauged supergravity given in [@Zucker:1999fn].
Unfortunately, the lagrangian (\[orbi1\]) after the gauge fixing (\[eichfixx\]) is quite complicated. Nevertheless it is the starting point for our orbifold construction so that we should present explicit formulas: $$\begin{split}
{\cal L}_{gauged} = &
e^u(-\frac{1}{4}R(\widehat{\omega})_{AB}{}^{AB}+4C-\frac{1}{6}\widehat{F}_{AB}\widehat{F}^{AB}+v_{AB}v^{AB}+20
\vec{t}\vec{t}-36{(t^2)}^2\\
& -\frac{1}{4}\partial^Au\partial_Au - \frac{1}{4}V_A^1V^{A1} -
\frac{1}{4}V_A^3V^{A3}+8\sqrt{3}g't^2-\frac{i}{2}\bar{\psi}_P\gamma^{PMN}{\cal
D}_M\psi_N\\
&-2i\bar{\psi}_A\gamma^A\lambda-\frac{\sqrt{3}g'}{4}\bar{\psi}_A\tau^2\gamma^{AB}\psi_B
- \frac{i}{2}\bar{\psi}_A\psi_Bv^{AB}) -12Nt^2+\sqrt{3}g'N\\
& -\frac{1}{\sqrt{3}}F_{AB}v^{AB}
-\frac{1}{6\sqrt{3}}\varepsilon^{ABCDE} A_AF_{BC}F_{DE}-
4\bar{\lambda}\tau^2\rho-2i\bar{\lambda}\gamma^A\psi_A\\
& +
\frac{1}{2}\bar{\rho}\tau^2\psi_A\partial^Au-\frac{1}{2}\bar{\rho}\tau^1\psi^MV_M^3+\frac{1}{2}\bar{\rho}\tau^3\psi^MV_M^1+\frac{1}{2}\bar{\psi}_A\tau^2\gamma^{AB}{\cal
D}_B\rho\\
& + 2i\bar{\rho}\gamma^A\psi_A t^2 - 2\bar{\psi}_A\tau^1\gamma^A\rho
t^3 + 2\bar{\psi}_A\tau^3\gamma^A\rho
t^1-\frac{1}{2}\bar{\psi}_A\tau^2\gamma^{AB}\rho \partial_B u\\
& - \frac{1}{12}\varepsilon^{MNPQR}(V_M^2-2g'A_M)\partial_NB_{PQR}-32
\vec{t}\vec{t}-\frac{\sqrt{3}ig'}{2}\bar{\psi}_A\gamma^A\rho \\
& - \frac{1}{2}\bar{\rho}\tau^2\gamma^{MN}{\cal D}_M\psi_N +
\bar{\rho}\tau^2\gamma_B\psi_Av^{AB} -
\frac{1}{2\sqrt{3}}\bar{\rho}\tau^2\gamma^{ABC}\psi_A\widehat{F}_{BC}\\
& -
\frac{i}{4\sqrt{3}}\bar{\psi}_A\gamma^{ABCD}\psi_B(e^u\widehat{F}_{CD} + \frac{1}{2}F_{CD}) +
(1-e^u)\bar{\psi}_A\vec{\tau}\gamma^{AB}\psi_B\vec{t}\\
& +e^{-u}\big( -
\frac{i}{4}\bar{\rho}\gamma^{AB}\rho(v_{AB}+\frac{1}{\sqrt{3}}\widehat{F}_{AB}) - 3\bar{\rho}\tau^2\rho t^2+W_AW^A-N^2\\
& -\frac{i}{2}\bar{\rho}\gamma^A{\cal D}_A\rho+i\bar{\rho}\psi_A
W^A\big)-4C -\frac{1}{2}e^{-2u}(\bar{\rho}\tau^2\rho N
-\bar{\rho}\tau^2\gamma^A\rho W_A) +{\cal L}_{4F}.
\end{split}$$ ${\cal L}_{4F}$ contains four fermion terms which play no rôle for us. They may be found in [@diss]. The covariant derivatives which appear here are covariant w.r.t. local Lorentz and local $SO(2)$ transformations, e.g. $${\cal
D}_M\rho=\partial_M\rho+\frac{1}{4}\widehat{\omega}_{MAB}\gamma^{AB}\rho-\frac{i}{2}\tau^2\rho V_M^2.$$ Let us stress, that we work in what follows till the end with the gauged $SU(2)_{\cal R}$. The very last step is to impose the condition (\[eichfixx\]).
Supergravity on $S^1/\mathbb{Z}_2$ {#sgs}
==================================
We now consider gauged supergravity on $S^1/\mathbb{Z}_2$. The $\mathbb{Z}_2$ acts on the fifth coordinate, $x^5\to -x^5$. A generic bosonic field transforms like $$\varphi(x^m,x^5)\to{\cal P}\varphi(x^m,-x^5),\qquad \mbox{with}\qquad
{\cal P}^2=1.$$ Fermionic fields transform like $$\psi(x^5)\to {\cal P}i\tau^3\gamma^{\dot{5}}\psi(-x^5).\label{proj}$$ An extended discussion of $\mathbb{Z}_2$ assignments to symplectic Majorana spinors can be found in [@Bergshoeff:2000zn]. We use the notation that a dot on an index $5$ denotes a Lorentz index.
The minimal multiplet {#sugzs}
---------------------
Starting from ${\cal P}(e_m^a)=+1$, the parity assignments of all fields belonging to the minimal multiplet are dictated by supersymmetry. We have collected these assignments in table \[table3022\].
Field $e_m^a$ $e_m^{\dot{5}}$ $e_5^a$ $e_5^{\dot{5}}$ $\psi_m$ $\psi_5$ $A_m$ $A_5$ $V_m^1$ $V_m^2$
----------------- --------- ----------------- --------- ----------------- ---------- ---------- ------- ------- --------- ---------
Parity $\cal P$ $+1$ $-1$ $-1$ $+1$ $+1$ $-1$ $-1$ $+1$ $-1$ $-1$
: Parities of the minimal multiplet.[]{data-label="table3022"}
------------------------------------------------------------------------------------------------------------------------------
Field $V_m^3$ $V_5^1$ $V_5^2$ $V_5^3$ $v_{ab}$ $v_{a\dot{5}}$ $\lambda$ $C$ $t^1$ $t^2$ $t^3$
--------------- --------- --------- --------- --------- ---------- ---------------- ----------- ------ ------- ------- -------
Parity $\cal $+1$ $+1$ $+1$ $-1$ $-1$ $+1$ $+1$ $+1$ $+1$ $+1$ $-1$
P$
------------------------------------------------------------------------------------------------------------------------------
: Parities of the minimal multiplet.[]{data-label="table3022"}
Note that the orbifold condition breaks the $SU(2)_{\cal R}$ at the fixpoints to a residual $U(1)$.
Let us define
$$\begin{aligned}
b_a & = & v_{a\dot{5}}\\
a_m & = & -\frac{1}{2}( V_m^3 - \frac{2}{\sqrt{3}}
\widehat{F}_{m5}e^5_{\dot{5}}+4e_m^av_{a\dot{5}})\\
S & = & C-\frac{1}{2}e_{\dot{5}}^5(\partial_5
t^3-\bar{\lambda}\tau^3\psi_5+V_5^1 t^2-V_5^2t^1)\end{aligned}$$
\[wichtig1\]
at the fixpoints. The important observation is then that the fields $$\big( e_m^a,~ \psi_m,~ b_a,~ a_m,~ \lambda,~ S, ~t^1,~t^2\big),$$ taken at the boundaries, form a non-minimal $N=1$ supergravity multiplet in four dimensions with $(16+16)$ components. This multiplet, sometimes called the intermediate multiplet has been studied in detail by Sohnius and West [@Sohnius:1983xs]. These authors also constructed the corresponding multiplet calculus.
The $U(1)$ which survives the orbifold projection is to be identified with the chiral $U(1)$ gauged by the auxiliary field $a_m$. The fields transform as follows under this symmetry (with parameter $\alpha$): $$\begin{gathered}
\delta\psi_m=\gamma^{\dot{5}}\psi_m\alpha,\qquad
\delta\lambda=\gamma^{\dot{5}}\lambda\alpha,\\
\delta a_m=\partial_m\alpha,\qquad \delta t^1=-2t^2\alpha,\qquad \delta
t^2=2t^1\alpha.\end{gathered}$$
The supersymmetry transformation laws on the boundary are in our notations $$\begin{split}
\delta e_m^a & = -i\bar{\varepsilon}\gamma^a\psi_m\\
\delta \psi_m & = {\cal D}_m\varepsilon -3\gamma^{\dot{5}}\varepsilon
b_m
+\gamma_m(\gamma^a\gamma^{\dot{5}}b_a+2i\tau^1t^1+2i\tau^2t^2)\varepsilon\\
\delta b_a & =
-\frac{i}{4}\bar{\varepsilon}\gamma_a{}^{bc}\gamma^{\dot{5}}\widehat{\cal
R}_{bc}-2i\bar{\varepsilon}\gamma_a\gamma^{\dot{5}}\lambda\\
\delta a_m & =
-\frac{i}{2}\bar{\varepsilon}\gamma^{\dot{5}}\gamma_m\lambda'\\
\delta \lambda & = -\frac{1}{4}\gamma^{\dot{5}}\varepsilon
\widehat{{\cal D}}_ab^a+\varepsilon S
-\frac{i}{2}\tau^1\gamma^a\varepsilon\widehat{{\cal D}}_a t^1-\frac{i}{2}\tau^2\gamma^a\varepsilon\widehat{{\cal
D}}_a t^2\\
\delta S & = -\frac{i}{2}\bar{\varepsilon}(\gamma^m\widehat{{\cal
D}}_m +\gamma^a\gamma^{\dot{5}} b_a)\lambda
-\bar{\varepsilon}(\tau^1t^1+\tau^2t^2)(\frac{1}{2}\gamma^{ab}\widehat{\cal R}_{ab}+8 \lambda)\\
\delta t^1 & = \bar{\varepsilon}\tau^1\lambda\\
\delta t^2 & = \bar{\varepsilon}\tau^2\lambda.
\end{split}\label{mino}$$ Note that fields like $\psi_5$ and $A_5$ have dropped out. The covariant derivatives which appear here are covariant w.r.t. four dimensional local Lorentz transformations and chiral $U(1)$ transformations. Further, we have introduced the useful definition [@Sohnius:1983xs] $$\lambda'=12\lambda+\gamma^{ab}\widehat{\cal R}_{ab}.$$ $\lambda'$ transforms quite simple under supersymmetry: $$\delta \lambda'=12\varepsilon S-\frac{1}{4}\varepsilon
\widehat{R}(\widehat{\omega})^{ab}{}_{ab}+48 \varepsilon ((t^1)^2+(t^2)^2)-6\varepsilon
b_ab^a-\frac{1}{2}\gamma^{ab}\gamma^{\dot{5}}\varepsilon\widehat{f}_{ab}$$ and $\widehat{f}_{mn}$ is the supercovariant field strength of $a_m$. We remind the reader that in these equations, the fermions $\psi_m$ and $\lambda$ satisfy the chirality constraints $$\lambda=i\tau^3\gamma^{\dot{5}}\lambda\qquad\mbox{and}\qquad
\psi_m=i\tau^3\gamma^{\dot{5}}\psi_m.$$
We can compute the gauge algebra for the multiplet (\[mino\]). It is the projection of the five dimensional gauge algebra: $$\begin{split}
[\delta_Q(\varepsilon),\delta _Q (\eta)] = &
\delta_{g.c.}(i\bar{\eta}\gamma^m\varepsilon)+\delta_Q(i\bar{\varepsilon}\gamma^m\eta\psi_m)+\delta_{ch}(i\bar{\varepsilon}\gamma^m\eta
a_m)\\
+ &
\delta_{Lt}(i\bar{\eta}\gamma^m\varepsilon\widehat{\omega}_{mab}-2i\bar{\varepsilon}\gamma_{abc}\gamma^{\dot{5}}\eta
b^c+4\bar{\varepsilon}\tau^1\gamma_{ab}\eta t^1+4\bar{\varepsilon}\tau^2\gamma_{ab}\eta
t^2).\label{algorbi}
\end{split}$$ $\delta_{ch}(\alpha)$ denotes a chiral transformation with parameter $\alpha$ and all transformation have to be understood in the four dimensional sense.
We are now in a position to elaborate on the matter multiplets which may be coupled to the supergravity multiplet (\[mino\]). This has been done in great detail in [@Sohnius:1983xs]. We restrict ourselves to the presentation of formulas which are required for the remaining sections. A more extended discussion may be found in the original work, for details using our notations, [@diss] should be consulted.
The chiral multiplet {#chm}
--------------------
This multiplet has also been given in [@Sohnius:1983xs]. Since it is very well known we restrict ourselves to the transformation laws using our conventions. The field content of the chiral multiplet is $$\mathbb{A}=(A,~B,~\psi,~F,~G).$$ It exists for arbitrary chiral weight $w$. The transformation laws are: $$\begin{aligned}
\delta A & = & \bar{\varepsilon}\tau^2 \psi+wB\alpha\\
\delta B & = & \bar{\varepsilon}\tau^2
\gamma^{\dot{5}}\psi-wA\alpha\\
\delta \psi & = &-\frac{i}{2}\gamma^a\tau^2 \varepsilon\widehat{\cal
D}_a A-\frac{1}{2}\varepsilon F-\frac{i}{2}\gamma^a\gamma^{\dot{5}}\tau^2
\varepsilon \widehat{\cal D}_a B
-\frac{1}{2}\gamma^{\dot{5}}\varepsilon G+(w-1)\gamma^{\dot{5}}\psi \alpha\\
\delta F & = & i\bar{\varepsilon}\gamma^a(\widehat{\cal
D}_a\psi+\gamma^{\dot{5}}\psi b_a) +4 \bar{\varepsilon}(\tau^1\psi
t^1-\tau^1\gamma^{\dot{5}}\psi t^2)\\
& + & \frac{w}{2}\bar{\varepsilon}\tau^2 ( A
+\gamma^{\dot{5}}B)\lambda'+(2-w)G\alpha\\
\delta G & = & i\bar{\varepsilon}\gamma^{\dot{5}}\gamma^a(\widehat{\cal
D}_a\psi+\gamma^{\dot{5}}\psi b_a)
-4\bar{\varepsilon}\gamma^{\dot{5}}(\tau^1\psi t^1-\tau^1\gamma^{\dot{5}}\psi t^2)\\
& + & \frac{w}{2}\bar{\varepsilon}\tau^2 \gamma^{\dot{5}}( A +
\gamma^{\dot{5}}B)\lambda'+(w-2)F\alpha.\end{aligned}$$ We have also indicated the transformation properties under chiral rotations.
For a chiral multiplet with weight $w=2$ we can write down an invariant $F$-term density: $$[\mathbb{A}]_F =
F+i\bar{\psi}_m\gamma^m\psi+\frac{1}{2}\bar{\psi}_m\tau^2 \gamma^{mn}(A+\gamma^{\dot{5}}B)\psi_n-12 t^2A-12
t^1B\label{chiral}.$$
The vector multiplet {#vec}
--------------------
The field content of the super Yang-Mills multiplet is $$\mathbb{Y}=(~u_m,~\chi,~ D).$$ The abelian vector multiplet can be derived straightforwardly from the general multiplet. We skip this multiplet, refer the interested reader for technical details to [@diss] and confine ourselves to the presentation of the transformation laws of the vector multiplet: $$\begin{aligned}
\delta u_m & = & i\bar{\varepsilon}\gamma_m\chi\\
\delta \chi & = & -\gamma^{\dot{5}}\varepsilon D +
\frac{1}{4}\gamma^{ab}\varepsilon \widehat{u}_{ab}\\
\delta D & =
&\frac{i}{2}\bar{\varepsilon}\gamma^{\dot{5}}\gamma^m\widehat{\cal D}_m\chi-\frac{3i}{2}\bar{\varepsilon}\gamma^a\chi b_a,\end{aligned}$$ where the fields take their values in the Lie algebra of the gauge group. We have defined the supercovariant field strength $\widehat{u}_{mn}$ of $u_m$. Of course, the gauge algebra (\[algorbi\]) is modified by the appearance of a field dependent gauge transformation.
An action can be derived by forming a chiral multiplet from fields of the Yang-Mills multiplet. One finds [@Sohnius:1983xs; @diss] $$\begin{split}
{\cal L}_{sym} = Tr~ [ & D D
-\frac{1}{4}\widehat{u}_{ab}\widehat{u}^{ab}-3i\bar{\chi}\gamma^a\gamma^{\dot{5}}\chi b_a\\
+ & i\bar{\chi}\gamma^a\widehat{\cal D}_a\chi
- \bar{\psi}_m\gamma^m(i\gamma^{\dot{5}}\chi
D+\frac{i}{4}\gamma^{ab}\chi\widehat{u}_{ab})\\
+ &
\frac{1}{4}\bar{\psi}_m\tau^2\gamma^{mn}\psi_n\bar{\chi}\tau^2\chi+\frac{1}{4}\bar{\psi}_m\tau^2\gamma^{mn}\gamma^{\dot{5}}\psi_n\bar{\chi}\tau^2\gamma^{\dot{5}}\chi].\label{lqed}
\end{split}$$
Lagrangians at the fixpoints {#chap4}
============================
In this section we construct lagrangians of the Randall-Sundrum type, i.e. theories which live on an orbifold $S^1/\mathbb{Z}_2$. In the bulk there is only gauged supergravity, but this can be extended easily to more complicated configurations using the results in [@Zucker:1999fn; @diss]. On the boundaries, the original Randall-Sundrum scenario [@Randall:1999ee] requires a cosmological constant. We present some generalizations of this model.
A cosmological constant – The Randall-Sundrum scenario {#cosm}
------------------------------------------------------
To start with the supersymmetrization of the Randall-Sundrum scenario we analyze the parity assignments of the tensor multiplet from section \[gos\]. The results are shown in table \[table22\].
Field $Y^1$ $Y^2$ $Y^3$ $\rho$ $N$ $B_{mnp}$ $B_{mn5}$
----------------- ------- ------- ------- -------- ----- ----------- -----------
Parity $\cal P$ $+1$ $+1$ $-1$ $+1$ +1 $+1$ $-1$
: Parities of the tensor multiplet.[]{data-label="table22"}
Inspection of the transformation laws shows, that on the boundaries, they form chiral multiplets with chiral weight $w=2$. The precise correspondence is $$\big(A,~B,~\psi,~F,~G\big)= \big(~Y^2,~Y^1,~\rho,~-2N+\widehat{\cal
D}_{\dot{5}}Y^3, ~+2 W^{\dot{5}}+12(t^1Y^2-Y^1t^2)\big),\label{emb}$$ on the boundary.
Using the result for the $F$-term density of a chiral multiplet, eq. (\[chiral\]), we are then in a position to write down the desired theory. Before doing that, however, we have to fix the $\cal R$-symmetry. This is achieved by imposing (\[eichfixx\]) which breaks the $SU(2)_{\cal R}$ in the bulk to a residual $U(1)$, gauged by $V_M^2$, and breaks the chiral $U(1)$ on the boundaries completely. The minimal multiplet on the boundary is then extended to an $N=1$, $D=4$ supergravity multiplet with $(20+20)$ components. One then uses the expressions (\[emb\]) in (\[chiral\]) to obtain $${\cal L}_{cc}=-2N + e^u V_{\dot 5}^1 - \bar{\rho}\tau^3\psi_{\dot
5}+i\bar{\psi}_m\gamma^m\rho+\frac{1}{2}e^u
\bar{\psi}_a\tau^2\gamma^{ab}\psi_b - 12 e^u t^2.\label{dM}$$ Our complete action is $$S=\int d^5x~e\left( {\cal L}_{gauged} + \Lambda_1\delta(x^5){\cal
L}_{cc}+\Lambda_2\delta(x^5-\ell){\cal L}_{cc}\right)\label{rs2}$$ where $\Lambda_1$ and $\Lambda_2$ are real constants. We use the unusual convention that $\delta(x^5)$ is a scalar, not a density. If one eliminates the auxiliary fields, terms of the form $\delta(x^5)^2$ appear but drop out at the end. One finds a result similar to the one of Altendorfer et al. [@Altendorfer:2000rr], $${\cal L}=\widetilde{\cal L}_{gauged}-
\big(\Lambda_1\delta(x^5)+\Lambda_2\delta(x^5-\ell)\big)\big(
\sqrt{3}g'-\frac{1}{2}\bar{\psi}_a\tau^2\gamma^{ab}\psi_b\big),\label{rslag}$$ with the on-shell lagrangian $\widetilde{\cal L}_{gauged}$ of gauged supergravity given in [@Zucker:1999fn]. $g'$ is the cosmological constant. The on-shell transformation laws are $$\begin{split}
\delta e_M^A = & - i\bar{\varepsilon}\gamma^A\psi_M\\
\delta \psi_M = & {\cal D}_M \varepsilon
+\frac{1}{4\sqrt{3}}\gamma_{MAB}\varepsilon\widehat{F}^{AB}-\frac{1}{\sqrt{3}}\gamma^N\varepsilon\widehat{F}_{MN}+\frac{ig'}{2\sqrt{3}}\gamma_M\tau^2\varepsilon\\
& + i\Lambda_1\delta_M^5\delta(x^5)\tau^2\gamma^{\dot{5}}\varepsilon
e_5^{\dot{5}}+i\Lambda_2\delta_M^5\delta(x^5-\ell)\tau^2\gamma^{\dot{5}}\varepsilon
e_5^{\dot{5}}\\
\delta A_M = & - \frac{\sqrt{3}i}{2}\bar{\varepsilon}\psi_M.
\end{split}\label{traflaw}$$ As a consistency check, we have verified the invariance of the lagrangian (\[rslag\]) under the on-shell transformations (\[traflaw\]).
One of the motivations for the present work where the hope to remove the fine tuning of the bulk and boundary cosmological constants, inherent to the original work of Randall and Sundrum, by supersymmetry. It should be clear that the method of Altendorfer et al. [@Altendorfer:2000rr] is not suited to answer the question whether this fine tuning is removed by supersymmetry. The alternative work [@Falkowski:2000er] implies that this fine tuning is removed by supersymmetry. This is not true for our result (\[rslag\]), at least on the lagrangian level, due to the appearance of the parameters $\Lambda_1$ and $\Lambda_2$.
Before proceeding, let us shortly comment on the solutions to the equation of motion and the Killing spinor equations. The Einstein equations are solved by the Randall-Sundrum metric $$ds^2 = e^{-2\sigma}\eta_{mn}dx^mdx^n-(d{x^5})^2\qquad \mbox{with}
\qquad\sigma\equiv \frac{g'}{\sqrt{3}}|x^5|,\label{rsmetric}$$ provided that $$\Lambda_1=-\Lambda_2=1.\label{val2}$$ Note that we have used according to common practice, that[^3] $\theta(x^5)^2=1$, where $\theta$ is the standard step function.
By decomposing the spinors in two-component objects $$\varepsilon^i=\binom{\varepsilon_+^i}{\varepsilon_-^i}.\label{zerl}$$ and using the representation of gamma matrices given in [@Mirabelli:1998aj], one may try to solve the Killing spinor equations $\delta\psi_M=0$ using the Ansatz of Altendorfer et al. [@Altendorfer:2000rr] $$\varepsilon_+^1=e^{-\frac{1}{2}\sigma}\eta,\qquad
\varepsilon_+^2=-i\theta(x^5)e^{-\frac{1}{2}\sigma}\eta,\label{damned}$$ with a constant spinor $\eta$. However, in this case one finds the constraint $$\Lambda_1 = -\Lambda_2 =2.\label{val1}$$ Note that (\[val1\]) is in contradiction with (\[val2\]) and thus the spinors (\[damned\]) represent no Killing spinors of the Randall-Sundrum solution.[^4] However, since the Randall-Sundrum metric (\[rsmetric\]) is a solution to the Einstein equations we have to expect that there exist consistent solutions to the Killing spinor equations. Otherwise we would have constructed a theory where the geometry of spacetime breaks supersymmetry spontaneously.
Matter at the fixpoints {#matt}
-----------------------
Our next models contain as bulk theory gauged supergravity and cosmological constants on the boundaries. In addition, we allow for chiral multiplets or super Yang-Mills multiplets on the boundaries. Let us start with the chiral multiplet.
For notational convenience we consider only one brane located at $x^5=0$. It is completely trivial to add the second one.
The systematic construction of a kinetic term for the chiral multiplet as it is needed here requires the general multiplet which we do not discuss. For technical details we refer to [@Sohnius:1983xs; @diss]. The bosonic part of the lagrangian for a chiral multiplet with chiral weight $w$ is $$\begin{split}
{\cal L}_{kin} = & -{\cal D}_a B {\cal D}^a B - {\cal D}_a A {\cal D}^a
A
-8t^2 GB- F^2-G^2\\
+ & 8t^2 FA+8t^1FB + 8t^1GA-4{\cal D}_a BA b^a+4{\cal D}_a AB b^a\\
+ &
(A^2+B^2)(12wS-\frac{w}{4}R_{ab}{}^{ab}+48(w-1)((t^1)^2+(t^2)^2)-6wb_ab^a-8S).
\end{split}\label{kinch}$$ Note the appearance of the curvature scalar. After inserting the definitions (\[wichtig1\]) and fixing the gauge via (\[eichfixx\]), we take our action to be $$S=\int d^5x~e\left({\cal L}_{gauged}+\Lambda_1\delta(x^5){\cal
L}_{cc}+\tilde{\Lambda}_1\delta(x^5){\cal L}_{kin}\right),$$ with $\tilde{\Lambda}_1$ a constant. After gauge fixing, the final step would be the elimination of the auxiliary fields. This turns out to be a remarkably complicated exercise which we have not completed.
The other possibility is to consider a super Yang-Mills multiplet on the boundary. This is conceptually similar to the model of Hořava and Witten [@Horava:1996ma; @Horava:1996qa].
Our model is the following. In the bulk, we have gauged supergravity. On the boundary there is a cosmological constant plus a super Yang-Mills multiplet. Our action is $$S=\int d^5x~e\left({\cal L}_{gauged}+\Lambda_1\delta(x^5){\cal
L}_{cc}+\tilde{\Lambda}_1\delta(x^5){\cal L}_{sym}\right)$$ with $\tilde{\Lambda}_1$ a constant and ${\cal L}_{sym}$ is given in (\[lqed\]). We next use the definitions (\[wichtig1\]) and fix the gauge using (\[eichfixx\]). In this case it is easy to eliminate the auxiliary fields from the lagrangian. The result is $$\begin{split}
{\cal L} & = \widetilde{\cal L}_{gauged}-\Lambda_1
\delta(x^5)(\sqrt{3}g'-\frac{1}{2}\bar{\psi}_a\tau^2\gamma^{ab}\psi_b)\\
+ &
\tilde{\Lambda}_1\delta(x^5)\big(-\frac{1}{4}\widehat{u}_{ab}\widehat{u}^{ab}-\frac{i\sqrt{3}}{4}\bar{\chi}\gamma^a\gamma^{\dot{5}}\chi\widehat{F}_{a\dot{5}}+i\bar{\chi}\gamma^m\widehat{D}_m\chi-\frac{i}{2}\bar{\chi}\gamma^{mab}\psi_m\widehat{u}_{ab}\\
+ & \frac{1}{4}\bar{\psi}_m\tau^2\gamma^{mn}\psi_n\bar{\chi}\tau^2\chi
+
\frac{1}{4}\bar{\psi}_m\tau^2\gamma^{mn}\gamma^{\dot{5}}\psi_n\bar{\chi}\tau^2\gamma^{\dot{5}}\chi\big).
\end{split}\label{sqedrs}$$
Conclusions and Outlook {#conc}
=======================
Using off-shell supergravity, we have presented a technique which allows the systematic construction of models of the Randall-Sundrum type. We assign consistently $\mathbb{Z}_2$ transformation laws to the fields belonging to the supergravity multiplet in the bulk. The fields with positive parity form a non-minimal $N=1$, $D=4$ supergravity multiplet at the fixpoints. Using this multiplet, we develop parts of an $N=1$ tensor calculus located at the branes. We apply this calculus to the explicit construction of lagrangians of the Randall-Sundrum type.
The configurations presented here are easily extended to more complicated ones, using the results in [@Zucker:2000ej; @Zucker:1999fn; @diss]. That is, one could introduce additional matter to the bulk which would then extend the possible couplings to matter at the fixpoints of the orbifold.
Of particular interest would be the study of supersymmetry breaking, since the brane world scenario represents a geometrization of hidden sector supergravity models (for a review see [@Nilles:1984ge]).
First steps in this direction were undertaken in [@Mirabelli:1998aj] and in [@gabi]. These authors considered rigidly supersymmetric theories on $S^1/\mathbb{Z}_2$. On one brane a supersymmetry breaking $D$- or $F$-term was placed, respectively, and the transmission of the supersymmetry breaking to the other wall studied. Using the techniques developed in the present work, it should be quite simple to extend these investigations to local supersymmetry. In this context we should also mention [@Falkowski:2000er], where gaugino condensation as supersymmetry breaking mechanism has been considered.
##### Acknowledgments:
We would like to thank Jonathan Bagger, Jan Conrad, Stefan Förste, Zygmunt Lalak and Hans-Peter Nilles for helpful discussions. This work was partially supported by the European Commission programs ERBFMRX-CT96-0045 and CT96-0090.
[99]{}
L. Randall and R. Sundrum, “A large mass hierarchy from a small extra dimension,” Phys. Rev. Lett. [**83**]{} (1999) 3370 \[hep-ph/9905221\].
R. D’Auria, E. Maina, T. Regge and P. Fre, “Geometrical First Order Supergravity In Five Space-Time Dimensions,” Annals Phys. [**135**]{} (1981) 237.
A. Falkowski, Z. Lalak and S. Pokorski, “Supersymmetrizing branes with bulk in five-dimensional supergravity,” hep-th/0004093.
E. Bergshoeff, R. Kallosh and A. V. Proeyen, “Supersymmetry in Singular Spaces,” hep-th/0007044.
R. Altendorfer, J. Bagger and D. Nemeschansky, “Supersymmetric Randall-Sundrum scenario,” hep-th/0003117v2.
E. A. Mirabelli and M. E. Peskin, “Transmission of supersymmetry breaking from a 4-dimensional boundary,” Phys. Rev. [**D58**]{} (1998) 065002 \[hep-th/9712214\].
M. Zucker, “Minimal off-shell supergravity in five dimensions,” Nucl. Phys. [**B570**]{} (2000) 267 \[hep-th/9907082\].
M. Zucker, “Gauged N = 2 off-shell supergravity in five dimensions,” JHEP [**08**]{} (2000) 016 \[hep-th/9909144\].
M. Zucker, “Off-shell supergravity in five dimensions and supersymmetric brane world scenarios,” thesis, Physikalisches Institut, Universität Bonn, August 2000, BONN-IR-2000-10.
T. Kugo and K. Ohashi, “Supergravity Tensor Calculus in 5D from 6D,” hep-ph/0006231.
B. de Wit, R. Philippe and A. Van Proeyen, “The Improved Tensor Multiplet In N=2 Supergravity,” Nucl. Phys. [**B219**]{} (1983) 143.
E. Bergshoeff, E. Sezgin and A. Van Proeyen, “Superconformal Tensor Calculus And Matter Couplings In Six-Dimensions,” Nucl. Phys. [**B264**]{} (1986) 653.
B. de Wit and M. Rocek, “Improved Tensor Multiplets,” Phys. Lett. [**B109**]{}, 439 (1982).
M. F. Sohnius and P. C. West, “Supergravity With One Auxiliary Spinor,” Nucl. Phys. [**B216**]{} (1983) 100.
J. O. Conrad, “Brane tensions and coupling constants from within M-theory,” Phys. Lett. [**B421**]{} (1998) 119 \[hep-th/9708031\].
P. Horava and E. Witten, “Eleven-Dimensional Supergravity on a Manifold with Boundary,” Nucl. Phys. [**B475**]{} (1996) 94 \[hep-th/9603142\].
P. Horava and E. Witten, “Heterotic and type I string dynamics from eleven dimensions,” Nucl. Phys. [**B460**]{} (1996) 506 \[hep-th/9510209\].
H. P. Nilles, “Supersymmetry, Supergravity And Particle Physics,” Phys. Rept. [**110**]{} (1984) 1.
G. Honecker, “Übertragung der Supersymmetrie-Brechung durch ein fünfdimensionales Hypermultiplett,” Diploma Thesis; Bonn, January 1999, (in german)
[^1]: e-mail: [email protected]
[^2]: We use the same conventions as in [@Zucker:2000ej], with the exception that for five dimensional indices we use capital letters and lower case letters for four dimensional indices. Letters from the middle of the alphabet are curved, letters from the beginning of the alphabet are flat.
[^3]: A counterexample is found as follows: Taking the derivative of $\theta(x^5)^2=1$ leads to $\delta(x^5)\theta(x^5)=0$ and multiplying this equation with $\theta(x^5)$ leads, using $\theta(x^5)^2=1$ to $\delta(x^5)=0$. More compactly stated: $\big(\theta(x^5)\theta(x^5)\big)\delta(x^5)\neq\theta(x^5)\big(\theta(x^5)\delta(x^5)\big)$. See also [@Conrad:1998ww]
[^4]: In equation (13) of [@Altendorfer:2000rr], the transformation law for the gravitino $\psi_m^+$ contains a term $\sim i\Lambda e_m{}^a\sigma_a\bar{\eta}_+$ which is missing in eq. (20). If this term is included, eqs. (21) and (22) have to be changed. After performing these corrections, one recovers the $\Lambda_1 =
-\Lambda_2 =2$ case and our calculation is in agreement with [@Altendorfer:2000rr]. We would like to thank Jan Conrad for extended discussions on this point.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'An adaptive moving mesh finite element method is studied for the numerical solution of the porous medium equation with and without variable exponents and absorption. The method is based on the moving mesh partial differential equation approach and employs its newly developed implementation. The implementation has several improvements over the traditional one, including its explicit, compact form of the mesh velocities, ease to program, and less likelihood of producing singular meshes. Three types of metric tensor that correspond to uniform and arclength-based and Hessian-based adaptive meshes are considered. The method shows first-order convergence for uniform and arclength-based adaptive meshes, and second-order convergence for Hessian-based adaptive meshes. It is also shown that the method can be used for situations with complex free boundaries, emerging and splitting of free boundaries, and the porous medium equation with variable exponents and absorption. Two-dimensional numerical results are presented.'
author:
- 'Cuong Ngo [^1]'
- 'Weizhang Huang [^2]'
title: |
A Study on Moving Mesh Finite Element Solution of\
the Porous Medium Equation
---
porous medium equation, adaptive moving mesh method, MMPDE method, finite element method, Hessian-based adaptivity, immersed boundary, free boundary
65M60, 65M50, 35Q35
Introduction {#SEC:intro}
============
We consider the numerical solution of the initial-boundary value problem (IBVP) of the porous medium equation (PME) in two dimensions, $$\begin{cases}
u_t = \div(|u|^m \grad u) , \quad & \text{in} \quad \Omega \times (t_0,T] \\
u({\boldsymbol{x}},t_0) = u_0({\boldsymbol{x}}) , \quad & \text{on} \quad \Omega \\
u({\boldsymbol{x}},t) = 0 , \quad & \text{on} \quad \p \Omega \times (t_0,T]
\end{cases}
\label{PME-1}$$ where $\Omega$ is a bounded polygonal domain, $u_0({\boldsymbol{x}})$ is a given function, and $m\ge 1$ is a physical parameter. PME is a nontrivial generalization of the heat equation. It is found in many areas of the physical sciences, including gas flow in porous medium, incompressible fluid dynamics, nonlinear heat transfer, and image processing; e.g., see [@Vazquez2007] and references therein. In the case of gas flow in porous medium, $u$ represents the density of the gas, $u^m$ the pressure, $u \grad (u^m)$ the flux, $\grad (u^m)$ the velocity, and $m$ is the isentropic coefficient. In the case of radiation diffusion in plasmas, $m$ stands for the power of temperature appearing in the nonlinear diffusion coefficient and can take values up to 5.5 [@Vazquez2007 Page 23]. PME itself represents a nonlinear diffusion process. One of its many interesting features is its degeneracy which induces a property of the finite propagation: if $u_0({\boldsymbol{x}})$ has a compact support, the solution will have a compact support for any time $t > t_0$. This in effect creates a free boundary that stays between the regions where $u$ is nonzero and where $u$ vanishes and propagates at a finite speed for all time. Contrary to the heat equation which smooths out its initial solutions, PME solution can become nonsmooth even in cases where it has a smooth initial solution. Moreover, for a certain type of initial solutions, the solution can exhibit a waiting-time phenomenon for which the free boundary will not move until a finite amount of time has elapsed. PME has been studied extensively in theory and there is a vast literature, including the earlier work by Ole[ĭ]{}nik et al. [@Oleinik1958], Kala[š]{}nikov [@Kalashnikov1967], Aronson [@Aronson1969], the more recent work by Shmarev [@Shmarev2005; @Shmarev2003], and the monograph by V[á]{}zquez [@Vazquez2007] and references therein.
The numerical solution of PME has also received considerable attention from researchers. Particularly, error estimates have been obtained for various finite element approximations. For example, using a regularization approach (to avoid degenerate or negative diffusion, for instance, with the diffusion coefficient $|u|^m$ being replaced by $\max (|u|, \epsilon/2)^m$ for some regularization parameter $\epsilon > 0$) and taking $\epsilon = \mathcal{O}(h^{\frac{2m+4}{m^2+4m+2}})$, Rose [@Rose-1983] shows that the error for a P1 finite element (for space) – backward Euler (for time) approximation of PME is bounded by $$\left (\sum_{n} \Delta t \| u_h^n - u \|_{L^{m+2}(\Omega)}^{m+2} \right )^{\frac{1}{m+2}}
\le C \left (\Delta t^{\frac{1}{m+1}} + \left (\ln \left (\frac{1}{h}\right )\right )^{\frac{1}{(m+1)(m+2)}} h^{\frac{2}{m+1}}\right ),
\label{rose-1}$$ where $h$ is the maximum element diameter and $u_h^n$ is the numerical approximation of $u$ at $t= t_n$. Nochetto and Verdi [@Nochetto-1988] consider a class of degenerate PDEs which includes PME as a special example and improve the result of [@Rose-1983]. They show that the error for a P1 finite element – 1st-order semi-implicit approximation of PME is bounded by $$\begin{aligned}
&\| u_h - u \|_{L^{\infty}(0,T; H^{-1}(\Omega))}
+ \| (u_{h})^{m+1} - u^{m+1} \|_{L^2(0,T; L^2(\Omega))}
+ \left \| \int_0^t ((u_{h})^{m+1} - u^{m+1})\right \|_{L^{\infty}(0,T; H^{1}(\Omega))}
\notag
\\
& \qquad \qquad
\le C \left (\frac{h^2}{\epsilon} + \frac{h^4}{\epsilon^2 \Delta t} + \Delta t\right )^{\frac 1 2}
\label{nochetto-1}
\\
& \qquad \qquad = \mathcal{O} (h^{\frac{m+2}{2(m+1)}}),\quad \text{if}\quad
\Delta t = \mathcal{O}(h^{\frac{m+2}{m+1}}),\quad \epsilon = \mathcal{O}(h^{\frac{m}{m+1}}).
\notag\end{aligned}$$ For the P1 finite element – backward Euler approximation, error bounds in various norm are obtained, for instance, $$\| u_h - u \|_{L^\infty(0,T; H^{-1}(\Omega))} \le C \left (\Delta t + \left (\ln \left (\frac{1}{h}\right )\right )^{\frac{2m+3}{2m+2}} h\right )
\label{rulla-1}$$ by Rulla and Walkington [@Rulla1996], $$\| u_h - u \|_{L^2(0,T; L^2(\Omega))} \le C h^{\frac{m^2+6m+8}{6m^2+14m+8}}
\quad \text{when}\quad \Delta t = \mathcal{O}(h^{\frac{5 m+4}{2m}})
\label{ebmeyer-1}$$ by Ebmeyer [@Ebmeyer-1998], and $$\| u_h - u \|_{L^{m+2}(0,T; L^{m+2}(\Omega))} \le
C \left ( \Delta t^{\frac 1 2} + h + h^{\frac{1}{m+1} \left (\frac{d m}{2m+4}+1\right )} \right )^{\frac{1}{m+2}}
\label{wei-1}$$ by Wei and Lefton [@Wei1999], where $d$ is the space dimension. It is remarked that these estimates are obtained for quasi-uniform meshes. The convergence rate is first order at best and decreases with $m$. Some of these estimates are shown to be optimal in the corresponding norm in lieu of the known regularity of the solution of PME. Moreover, it is worth mentioning that Ebmeyer and Liu [@Ebmeyer-2008] obtain error estimates in quasi-norm and Emmrich and [Š]{}i[š]{}ka [@Emmrich2012] prove that a Galerkin finite element – backward Euler approximation converges to the weak solution of PME using the theory of monotone operators. More recently, Duque et al. [@Duque2013] establish $L^{1+\max (\gamma/2)}$ error bounds for the approximation of a general order continuous Galerkin in space and a general order discontinuous Galerkin in time for PME with a variable exponent $m = \gamma({\boldsymbol{x}})$. Zhang and Wu [@Zhang-2009] consider the numerical simulation of the one-dimensional PME on a uniform mesh using a high-order local discontinuous Galerkin finite element method. The method can effectively eliminate unwanted, nonphysical oscillations in the computed solution near the free boundary and lead to a high-order convergence rate within the solution support and away from the free boundary.
The low regularity (steep gradient and corner shape near the moving free boundary) and evolving nature of the solution makes adaptive moving mesh methods an attractive tool to improve accuracy in the numerical solution of PME. A number of works exist in this direction. For example, Budd et al. [@BCHR99] investigate the numerical simulation of self-similarity solutions of one-dimensional PME using the Moving Mesh PDE (MMPDE) moving mesh method [@HRR94a; @HR11] and a specially designed monitor function to preserve the scaling invariance of PME. In a series of papers [@BHJ05; @BHJ05a; @BHJJ06] (also see the review paper [@Baines-2011]), Baines and his co-workers study the numerical solution of PME in one and two dimensions using a moving mesh finite element method that is based upon conserving a local proportion of the total mass that is present in the projected initial data. Numerical results show that their method gives a second-order convergence for $m=1$ but only a first-order convergence for $m=3$ when a uniform initial mesh is used [@BHJ05a; @Baines-2011]. For $m=3$, the second-order convergence can be recovered in one dimension if an optimal initial mesh is used. Unfortunately, such an optimal mesh is significantly more expensive to compute in two dimensions than in one dimension. Recently, Duque et al. [@Duque2014; @Duque2015] present a moving mesh finite element method based on an MMPDE for PME with variable exponents and with/without absorption. The method shows a first-order convergence when tested for the Barenblatt-Pattle solution of PME.
The objective of this paper is to study an adaptive moving mesh finite element method for the numerical solution of PME. The method is also based on an MMPDE but significantly different from the method of [@Duque2014; @Duque2015]. The MMPDE we use is formulated by minimizing an energy (cf. (\[Ih\])) based on the equidistribution and alignment conditions and the mesh adaptation is controlled through a matrix-valued function (i.e., a metric tensor) instead of a scalar function. The advantage of using a metric tensor is that it provides information not only to control the size of mesh elements but also their shape and orientation. Generally speaking, a so generated mesh has better alignment with the geometry of the physical solution than that with a scalar mesh adaptation function. Moreover, a newly developed compact formulation of the method (cf. (\[mmpde-2\]) and (\[mmpde-3\])) makes its implementation much easier and more efficient. Mesh adaptation based on the gradient and Hessian of the solution will be considered. The arclength metric tensor (a gradient-based metric tensor) has been widely used in the context of moving mesh methods because it results in more stable mesh movement than a Hessian-based metric tensor and works well for many problems. On the other hand, there is no theoretical guarantee that the arclength metric tensor will lead to the optimal convergence order for piecewise linear interpolation or finite element approximation since the error is determined by the Hessian of the solution in these situations. As a matter of fact, there are problems, although not very common, for which we have to use a Hessian-based metric tensor in order to achieve the optimal convergence order for piecewise linear interpolation or finite element approximation. Interestingly, PME is one of those problems. We shall show that the linear finite element approximation of the Barenblatt-Pattle solution of PME shows a first-order convergence for arclength-based adaptive meshes and a second-order convergence for Hessian-based adaptive meshes. Another feature of the method that is different from those of Duque et al. [@Duque2014; @Duque2015] and Baines et al. [@BHJ05; @BHJ05a; @Baines-2011; @BHJJ06] is that PME is solved on a large domain that contains the free boundary for the whole time period under consideration. In this way, there is no need to explicitly trace the movement of the free boundary and thus the method can readily deal with more complicated structures in the solution and in the differential equation. Numerical examples with simple free boundary (such as the Barenblatt-Pattle solution) and more complex and even merging free boundaries will be presented as well as those for PME with variable exponents and absorption.
In addition to the above mentioned MMPDE method, a number of other moving mesh methods have been developed in the past; e.g., see Hirt et al. [@HAC74] (ALE – Arbitrary Lagrangian-Eulerian) Miller and Miller [@MM81] (MFE – moving finite element), Liao and Anderson [@LA92] (deformation map), Li et al. [@LTZ01] (mesh rezoning), Cao et al. [@CHR02] (geometric conservation law), Baines et al. [@BHJ05a] (conservation of fraction mass), and Budd and Williams [@BW06] (parabolic [M]{}onge-[A]{}mpère equation). The interested reader is also referred to the books/review articles [@Bai94a; @Baines-2011; @Shashkov-2015; @BHR09; @HR11; @Tan05] and references therein. The outline of the paper is as follows. Some properties of PME that are relevant to the numerical simulation are described in Section \[SEC:PME-theory\]. Section \[SEC:mmfem\] is devoted to the description of the moving mesh finite element method, including the linear finite element discretion of PME on a moving mesh and the generation of an adaptive moving mesh using the MMPDE approach. Numerical examples are presented in Sections \[SEC:PME-numerics\] and \[SEC:PME-numerics-2\] for PME and PME with variable exponents and absorption, respectively. Finally, Section \[SEC:conclusion\] contains the conclusions and further comments.
Properties of the porous medium equation {#SEC:PME-theory}
========================================
Before we describe the finite element approximation of IBVP (\[PME-1\]), it is instructive to recall some of its properties that are relevant to numerical simulation. First of all, it is known (e.g., see V[á]{}zquez [@Vazquez2007]) that when $u_0^{m+2} \in L^{1}(\Omega)$, IBVP has a weak solution $u$ satisfying $u(\cdot, t)^{m+2} \in L^{1}(\Omega)$ for any $t \in (t_0, T]$ and $u^{m+1} \in L^{2}(0,T; H^1_0(\Omega))$. Moreover, a nonnegative weak solution exists if $u_0^{m+2} \in L^{1}(\Omega)$, $u_0 \in L^1(\Omega)$, and $u_0 \ge 0$. The uniqueness of the weak solution is guaranteed if it is further assumed that $u \in L^2(\Omega \times (0,T))$.
PME is degenerate whenever $u = 0$. Due to this degeneracy, PDE has the property of the finite speed of propagation: if compact-supported initially, its solution remains compact-supported at any finite time, with the support monotonically expanding as time evolves. The boundary of the support forms a moving interface $\Gamma(t)$ which is commonly referred to as a free boundary. The velocity of the free boundary is given by Darcy’s law (e.g., see Shmarev [@Shmarev2005]), i.e., $$\Gamma'(t) = - \lim_{{\boldsymbol{x}} \to \Gamma(t)^-} \grad \left( \frac{u^m({\boldsymbol{x}},t)}{m} \right),
\label{Darcy-law}$$ where the limit is taken from the interior of the support. In addition, PME exhibits a waiting-time phenomenon: for a certain type of initial solutions the free boundary does not move until a finite amount of time has elapsed. Loosely speaking, from (\[Darcy-law\]) we may expect to see this phenomenon for initial solutions having vanishing $\grad (u^m)$ at the initial free boundary.
A few classes of special solutions to IBVP have been known, among which is the Barenblatt-Pattle solution, viz., $$u(r,t) =
\begin{cases}
\frac{1}{\lambda^d(t)}
\left(
1 - \left( \frac{r}{r_0 \lambda(t)} \right)^2
\right)^{\frac{1}{m}} , \quad & \text{for} \quad |r| \leq r_0 \lambda(t) \\
0 , \quad & \text{for} \quad |r| > r_0 \lambda(t)
\end{cases}
\label{BP-soln}$$ where $$r = | {\boldsymbol{x}} | , \quad
\lambda(t) = \left(\frac{t}{t_0} \right)^{\frac{1}{2+d m}} ,
\quad
t_0 = \frac{r_0^2 m}{2 (2 + d m)} ,$$ and $r_0>0$ is a given parameter. It is radially symmetric, self-similar, and compact-supported for any finite time. Moreover, $u^m$ is Lipschitz continuous in ${\boldsymbol{x}}$ and $t$ and $\nabla (u^m)$ is bounded in the support of $u(\cdot, t)$, $\text{supp}(u(\cdot, t))$. Furthermore, the solution is Hölder continuous. The slope of the solution at the free boundary is finite for $m = 1$ and becomes infinite when $m>1$, which causes challenges for the numerical solution of PME. These regularity properties also hold for general compactly supported solutions of (\[PME-1\]), and their free boundaries can be shown to be at least Lipschitz continuous in both space and time; e.g., see [@Aronson1969; @Caffarelli1980; @Daskalopoulos1998a; @Shmarev2005].
The moving mesh finite element method {#SEC:mmfem}
=====================================
In this section we describe the adaptive moving mesh finite element approximation of IBVP . To begin with, we note that there are roughly two approaches for solving the IBVP. The first, as used in [@Rose-1983; @Zhang-2009], is to solve PME in a large domain containing the free boundary for the whole time period of the simulation. With this approach, there is no need to explicitly treat the free boundary, which makes the approach more amenable to problems with complex solution supports. The main disadvantage of this approach is that the solution has a corner shape between the regions of zero and nonzero solution values and thus its regularity on the whole domain is at most $H^1$. An $H^1$ regularity often means at best a first-order convergence in the numerical solution as the mesh is refined.
The second approach is to solve the problem only in the region of compact support; e.g., see [@BHJ05a; @Duque2014; @Duque2015]. One of the advantages of this approach is that a smaller spatial domain is used and thus fewer mesh points can be used to achieve the same computational accuracy. Moreover, the regularity of the solution is better on the support than on a larger domain since it does not have a corner shape. As a result, the numerical solution can have a higher convergence order than that with the first approach. The main disadvantage is that the boundary movement has to be treated explicitly using Darcy’s law (\[Darcy-law\]). We use the first approach in this work. We choose this approach due to its advantage of no need to explicitly treat the free boundary and its potential to deal with problems having complex solution supports. To better resolve the corner shape in the solution and improve the computational accuracy, we employ an MMPDE-based moving mesh method [@HR11] to adaptively and dynamically concentrate the mesh points around the free boundary. Linear finite elements and the fifth-order Radau IIA method (e.g., see Hairer and Wanner [@HW96]) are used for the spatial and temporal discretization, respectively. As we will see in Section \[SEC:PME-numerics\], a second-order convergence of the finite element approximation in space can be achieved when a properly adapted mesh is used.
Finite element discretization {#SEC:fem}
-----------------------------
We now describe the finite element discretization. Denote the time instants by $$t_0 = 0 < t_1 < \ldots < t_{n_f} \equiv T .
\label{time-grid}$$ For the moment, we assume that the simplicial meshes $\cTh^n$, $n = 0,\ldots,n_f$ for the physical domain $\Omega$ at these time instants are known and have the same connectivity and the same numbers of vertices and elements. (Their generation will be discussed in the next subsection.) Denote the coordinates of the vertices of $\cTh^n$ by ${\boldsymbol{x}}_j^n$, $j = 1, ..., N_v$, where $N_v$ is the number of all vertices. The mesh $\cTh(t)$ between any two time instants $t_n$ and $t_{n+1}$ is defined through linear interpolation, i.e., $$\begin{aligned}
& {\boldsymbol{x}}_j(t) =
\frac{t-t_n}{t_{n+1}-t_n} {\boldsymbol{x}}_j^{n+1} +
\frac{t_{n+1}-t}{t_{n+1}-t_n} {\boldsymbol{x}}_j^{n} , \quad \forall j = 1, ..., N_v
\\
&\dot{{\boldsymbol{x}}}_j (t) = \frac{{\boldsymbol{x}}_j^{n+1}-{\boldsymbol{x}}_j^{n}}{t_{n+1}-t_n},\quad
j = 1, ..., N_v.\end{aligned}$$ Denote by $\phi_j({\boldsymbol{x}}, t)$ the linear basis function associated with vertex ${\boldsymbol{x}}_j(t)$. For convenience, we assume that the vertices are arranged such that the first $N_{vi}$ vertices are the interior vertices. Let $$V_h(t) = \text{span}\{\phi_1(\cdot, t), ..., \phi_{N_{vi}}(\cdot, t)\} .$$ Then, the linear finite element approximation to the solution of IBVP is defined as $u_h(\cdot, t) \in V_h(t)$, $ t \in (t_0, T]$ such that $$\begin{cases}
& \int_{\Omega} \frac{\p u_h}{\p t} v~d{\boldsymbol{x}} = - \int_{\Omega} |u_h|^m \grad u_h
\cdot \grad v~d{\boldsymbol{x}} ,
\quad \forall v \in V_h(t),\quad t_0 < t \leq T \\
& \int_{\Omega} (u_h({\boldsymbol{x}},0) - u^0({\boldsymbol{x}})) v~d{\boldsymbol{x}} = 0, \quad \forall v \in V_h(t) .
\end{cases}
\label{fem0}$$ The above equation can be cast in matrix form. Indeed, expressing $$u_h({\boldsymbol{x}},t) = \sum_{j=1}^{N_{v_i}} u_j(t) \phi_j({\boldsymbol{x}},t)$$ and differentiating it with respect to $t$, we have $$\frac{\p u_h}{\p t} = \sum_{j=1}^{N_{v_i}} \frac{\p u_j}{\p t} \phi_j({\boldsymbol{x}},t) +
\sum_{j=1}^{N_{v_i}} u_j(t) \frac{\p \phi_j}{\p t} .$$ It can be shown (e.g. see Jimack and Wathen [@Jimack-1991 Lemma 2.3]) that $$\frac{\p \phi_j}{\p t} = - \nabla \phi_j \cdot \dot{{\boldsymbol{X}}},\quad a.e. \text{ in } \Omega$$ where the mesh velocity $\dot{{\boldsymbol{X}}}$ is defined as $$\dot{{\boldsymbol{X}}} ({\boldsymbol{x}},t) = \sum_{j=1}^{N_v} \dot{{\boldsymbol{x}}}_j(t) \phi_j ({\boldsymbol{x}}, t) .$$ Then, we get $$\frac{\p u_h}{\p t} = \sum_{j=1}^{N_{v_i}} \frac{\p u_j}{\p t} \phi_j({\boldsymbol{x}},t) - \nabla u_h\cdot \dot{{\boldsymbol{X}}} .$$ From this and taking $v = \phi_i$ ($i = 1, ..., N_{vi}$) in (\[fem0\]) successively, we obtain $$\sum_{j=1}^{N_{vi}} \left(\int_{\Omega}
\phi_j \phi_i~d{\boldsymbol{x}}\right) \frac{d u_j}{dt}
= \int_{\Omega} \grad u_h \cdot \left( \dot{{\boldsymbol{X}}} \phi_i
- u_h^m \grad \phi_i \right) ~d{\boldsymbol{x}}, \quad i = 1, ..., N_{vi}, \quad t_0 < t \leq T$$ which can be cast in the matrix form as $$B({\boldsymbol{X}}) \dot{{\boldsymbol{U}}} = F({\boldsymbol{U}}, {\boldsymbol{X}}, \dot{{\boldsymbol{X}}}),
\label{fem-2}$$ where $B$ is the mass matrix and ${\boldsymbol{X}}$ and ${\boldsymbol{U}}$ are the vectors representing the mesh and solution, respectively. This ODE system is integrated from $t_n$ to $t_{n+1}= t_n + \Delta t_n$ using the fifth-order Radau IIA method, with $\Delta t_n$ being determined by a standard time step size selection procedure (e.g., see Hairer et al. [@HNW93 Section II.4]) and using a two-step error estimator of Gonzalez-Pinto et al. [@Montijano2004]. The relative and absolute tolerances $rtol = 10^{-6}$ and $atol = 10^{-8}$ are taken in the computation.
The whole computation alternates between the integration of PME and the generation of the mesh. Starting with the current mesh $\cTh^n$ and a solution $u_h^n({\boldsymbol{x}})\approx u({\boldsymbol{x}}, t_n)$ defined thereon, a new mesh $\cTh^{n+1}$ is generated using the moving mesh strategy to be described in the next subsection. Then, the discrete PME (\[fem-2\]) is integrated from $t_n$ to $t_{n+1}$ (as described above) to obtain the solution approximation $u_h^{n+1}({\boldsymbol{x}})$.
An MMPDE-based moving mesh strategy {#SEC:mmpde}
-----------------------------------
We now describe the generation of $\cTh^n$, $ n = 1, ..., n_f$. We assume that the mesh $\cTh^n$ and a computed solution $u_h^n({\boldsymbol{x}})$ are known at $t = t_n$. We also assume that a reference computational mesh $\cTco = \{ \hat{{\boldsymbol{\xi}}}_j, j = 1, ..., N_v\}$ having the same connectivity and the same numbers of vertices and elements as $\cTh^n$ has been chosen. In our computation, it is taken as a uniform mesh (in the Euclidean metric) defined on $\Omega$. $\cTco$ stays fixed for the whole computation.
The generation of $\cTh^{n+1}$ is through the computational mesh $\cTc = \{ {\boldsymbol{\xi}}_j, j = 1, ..., N_v\}$ which serves as an intermediate variable. (A sketch of the relations among the meshes $\cTco$, $\cTc^{n+1}$, $\cTh^n$, and $\cTh^{n+1}$ is shown in Fig. \[fig:mesh-relation\].) First, an MMPDE-based mesh equation (to be described below) for the velocities of the computational vertices is employed. It takes the form (cf. (\[mmpde-2\])) $$\begin{cases}
\frac{d {\boldsymbol{\xi}}_j}{d t} = {\boldsymbol{v}}_j(\bM, \cTh^n; {\boldsymbol{\xi}}_1, ..., {\boldsymbol{\xi}}_{N_v}),&\quad j = 1, ..., N_v, \quad t \in (t_n, t_{n+1}]
\\
{\boldsymbol{\xi}}_j(t_n) = \hat{{\boldsymbol{\xi}}}_j,& \quad j = 1, ..., N_v
\end{cases}$$ where ${\boldsymbol{v}}_j$ denotes the mesh velocity for the $j$-th node which depends on $\cTh^n$, the metric tensor $\bM$ defined thereon, and $\cTc$. Here, the initial mesh is taken to be the reference computational mesh $\cTco$. The system is integrated and the mesh $\cTc^{n+1} \approx \cTc(t_{n+1})$ is obtained. During the integration, both $\bM$ and $\cTh^n$ are kept fixed. Notice that $\cTc^{n+1}$ and $\cTh^n$ form a correspondence relation, say, ${\boldsymbol{x}}_j^n = \Phi_h( {\boldsymbol{\xi}}_j^{n+1}),\, j = 1, ..., N_v$ or ${\boldsymbol{x}} = \Phi_h( {\boldsymbol{\xi}} )$. Then, the vertices of the new physical mesh $\cTh^{n+1}$ are defined as ${\boldsymbol{x}}_j^{n+1} = \Phi_h(\hat{{\boldsymbol{\xi}}}_j), \, j = 1, ..., N_v$. Since $\Phi_h$ is defined only at the vertices of $\cTc^{n+1}$, we need to compute $\Phi_h(\hat{{\boldsymbol{\xi}}}_j)$ using interpolation. Linear interpolation is used since it is important to keep the nonsingularity of the mesh while it is unnecessary to compute the mesh to high accuracy. The metric tensor is assumed to be symmetric and uniformly positive definite on $\Omega$. It is used to control the size, shape, and orientation of the elements of the mesh to be generated. We consider three types of mesh, one is uniform and the other two are arclength- and Hessian-based adaptive meshes. The metric tensors associated with the adaptive meshes are defined as $$\begin{aligned}
\bM & = (\mathbb{I} + \grad u_h^n (\grad u_h^n)^T)^{\frac{1}{2}} ,
\label{M-arclength}
\\
\bM & = \left[ \det \left( \mathbb{I} + |H(u_h^n)| \right)\right]^{-\frac{1}{6}}
\left( \mathbb{I} + |H(u_h^n)|\right) ,
\label{M-hessian}
\end{aligned}$$ where $\mathbb{I}$ is the $d\times d$ identity matrix, $H(u_h^n)$ is a recovered Hessian for the piecewise linear finite element solution $u_h^n$, and $|H(u_h^n)| = Q \text{diag} (|\lambda_1|, ..., |\lambda_d|) Q^T$ with $Q \text{diag} (\lambda_1, ..., \lambda_d) Q^T$ being the eigen-decomposition of $H(u_h^n)$. The tensor (\[M-arclength\]) is the frequently used arclength monitor function which, loosely speaking, places mesh points according to the uniformity in arclength. The tensor (\[M-hessian\]) is optimal [@HS03] for the $L^2$ norm of linear interpolation error. In our computation, we use a least squares fitting strategy for Hessian recovery (e.g., see [@Kamens09PhD; @KaHu2013]).
We now describe the formulation of the MMPDE-based mesh equation. A key component of the formulation is the $\bM$-uniform mesh concept with which any (nonuniform) adaptive mesh is viewed as a uniform one in some metric. It is known [@Hua06; @HR11] that such an $\bM$-uniform mesh $\cTh$ approximately satisfies the equidistribution and alignment conditions $$\begin{aligned}
& |K| \sqrt{\det(\bM_K)} = \frac{\sigma_h |K_c|}{|\Omega_c|} , \quad \forall K \in \cTh
\label{eq-1}
\\
&\frac{1}{d} {\text{trace}}\left(
\left( F'_K \right)^{-1}
\bM_K^{-1} \left (F'_K \right )^{-T}
\right)
=
\det \left(
\left( F'_K \right)^{-1}
\bM_K^{-1} \left (F'_K \right )^{-T}
\right)^{\frac{1}{d}} , \quad \forall K \in \cTh
\label{ali-1}\end{aligned}$$ where $|K|$ is the volume of $K$, $\bM_K$ is the average of $\bM$ over $K$, $\det(\cdot)$ and ${\text{trace}}(\cdot)$ denote the determinant and trace of a matrix, respectively, $|K_c|$ is the volume of the element $K_c \in \cTc$ corresponding to $K$, $F'_K$ is the Jacobian matrix of the affine mapping $F_K: K_c \to K$, and $$\sigma_h = \sum_{K \in \cTh} |K| \sqrt{\det(\bM_K)}, \quad
|\Omega_c | = \sum_{K_c \in \cTc} |K_c| .$$ The equidistribution condition (\[eq-1\]) requires that the volume of $K$ in the metric $\bM$ be proportional to $|K_c|$ with constant proportionality while the alignment condition (\[ali-1\]) requires that $K$ be similar to $K_c$. The meshes that closely satisfy these conditions can be obtained by minimizing the energy function $$\begin{aligned}
I_h & = \theta \sum_{K \in \cTh} |K| \sqrt{\det(\bM_K)}
\left(
{\text{trace}}({(F_K')}^{-1} {\bM}_{K}^{-1} {(F_K')}^{-T})
\right)^{\frac{d p}{2}} \nonumber \\
& \quad \quad + (1-2\theta) d^{\frac{dp}{2}}
\sum_{K \in \cTh} |K| \sqrt{\det (\bM_K)}
\left(
\frac{|K_c|}{|K| \sqrt{\det(\bM_K)}}
\right)^{p},
\label{Ih}
\end{aligned}$$ which is a Riemann sum of a continuous functional developed in [@Hua01b] based on equidistribution and alignment for variational mesh adaptation. Here, $\theta \in (0, \frac{1}{2}]$ and $p > 1$ are non-dimensional parameters. We choose $\theta = 1/3$ and $p = 2$ in our computation.
Notice that $I_h$ is a function of the position of the computational vertices $\{ {\boldsymbol{\xi}}_j\}_{j=1}^{N_v}$ and the physical coordinates $\{ {\boldsymbol{x}}_j\}_{j=1}^{N_v}$. For the current situation, we choose $\cTh$ to be $\cTh^n$ (the current physical mesh). Then, $I_h$ is the function of ${\boldsymbol{\xi}}_j$, $j = 1, ..., N_v$ only. Instead of taking direct minimization of $I_h$ with respect to these coordinates, we follow the MMPDE approach [@HRR94a] and define the moving mesh equation as a gradient system of $I_h$, $$\frac{d {\boldsymbol{\xi}}_j}{d t} = - \frac{P_j}{\tau} \left [ \frac{\partial I_h} {\partial {\boldsymbol{\xi}}_j} \right ]^T,
\quad j = 1, ..., N_v
\label{mmpde-1}$$ where the derivative of $I_h$ with respect to ${\boldsymbol{\xi}}_j$, ${\partial I_h}/{\partial {\boldsymbol{\xi}}_j}$, is considered as a row vector, $\tau > 0$ is a parameter used to control the response time of the mesh movement to the change in the metric tensor, and $P_j = \det(\bM({\boldsymbol{x}}_j))^{\frac{p-1}{2}} $ is chosen such that (\[mmpde-1\]) is invariant under the scaling transformation of $\bM$: $\bM \to c \bM$ for any positive constant $c$. The derivative of $I_h$ with respect to ${\boldsymbol{\xi}}_j$ can be found analytically using the notion of scalar-by-matrix differentiation; see [@HK2014]. With these analytical formulas, we can rewrite (\[mmpde-1\]) into $$\label{mmpde-2}
\frac{d {\boldsymbol{\xi}}_j} {d t}= \frac{P_j}{\tau} \sum_{K \in \omega_j} |K| {\boldsymbol{v}}_{j_K}^K , \quad j = 1, \dotsc, N_v$$ where $\omega_j$ is the element patch associated with the $j$-th vertex, $j_K$ is its local index of the vertex on $K$, and ${\boldsymbol{v}}_{j_K}^K$ is the velocity contributed by the element $K$ to the vertex $j_K$. The velocities contributed by $K$ to its vertices are given by $$\label{mmpde-3}
\begin{bmatrix} {({\boldsymbol{v}}_1^K)}^T \\ \vdots \\ {({\boldsymbol{v}}_d^K)}^T \end{bmatrix}
= - E_K^{-1} \frac{\partial G}{\partial \mathbb{J}} - \frac{\partial G}{\partial \det(\mathbb{J})}
\frac{\det(\hat{E}_K)}{\det(E_K)} \hat{E}_K^{-1},
\quad
{\boldsymbol{v}}_0^K = - \sum_{i=1}^d {\boldsymbol{v}}_d^K ,$$ where $E_K = [{\boldsymbol{x}}_1^K-{\boldsymbol{x}}_0^K, ..., {\boldsymbol{x}}_d^K-{\boldsymbol{x}}_0^K]$ is the edge matrix of $K$, $\hat{E}_K$ is the edge matrix for $K_c$ which is defined similarly, the function $G$ is associated with the energy (\[Ih\]) and defined as $$G(\mathbb{J}, \det(\mathbb{J}), \bM)
= \theta \sqrt{\det(\bM)} \left ( {\text{trace}}(\mathbb{J} \bM^{-1} \mathbb{J}^T) \right )^{\frac{dp}{2}}
+ (1-2\theta) d^{\frac{d p}{2}} \sqrt{\det(\bM)} \left (\frac{\det(\mathbb{J})}{\sqrt{\det(\bM)}} \right )^p,$$ and its derivatives (evaluated at $(\mathbb{J}, \det(\mathbb{J}), \bM) = ((F_K')^{-1}, \det(F_K')^{-1}, \bM_K)$) with respect to the first (the Jacobian matrix) and second arguments are given by $$\begin{aligned}
& \frac{\partial G}{\partial \mathbb{J}} = d p \theta \sqrt{\det(\bM)} \left ( {\text{trace}}(\mathbb{J} \bM^{-1} \mathbb{J}^T)
\right )^{\frac{dp}{2}-1}\bM^{-1} \mathbb{J}^T,
\\
& \frac{\partial G}{\partial \det(\mathbb{J})} = p (1-2\theta) d^{\frac{d p}{2}} \det(\bM)^{\frac{1-p}{2}}
\det (\mathbb{J})^{p-1} .\end{aligned}$$ Notice that ${\partial G}/{\partial \mathbb{J}}$ is a $d$-by-$d$ matrix.
In practical computation, we can first compute the local velocities ${\boldsymbol{v}}_j^K$, $j = 0, ..., d$ for all elements using (\[mmpde-3\]) and then obtain the velocity for any mesh point by summing the volume weighted contributions from its neighboring elements (cf. (\[mmpde-2\])). The mesh equation should be modified for boundary mesh points. For fixed points (such as corners), we can set the velocity to be zero. For those on a boundary edge or surface, the mesh velocities should be modified such that they do not move out of the domain.
The mesh equation (\[mmpde-2\]) (with proper modifications for boundary mesh points) is integrated from $t=t_n$ to $t_{n+1}$ starting with $\cTco$ as the initial mesh. In our computation, the Matlab ODE solver [*ode15s*]{} (an implicit scheme) is used to integrate (\[mmpde-2\]).
Equation (\[mmpde-2\]) is called the ${\boldsymbol{\xi}}$-formulation of the MMPDE moving mesh method since it has been formulated in terms of the derivatives of $I_h$ with respect to ${\boldsymbol{\xi}}_j$ and the velocities for the computational coordinates. We can obtain an ${\boldsymbol{x}}$-formulation by directly differentiating $I_h$ with respect to ${\boldsymbol{x}}_j$ (with $\cTc$ being taken as $\cTco$ and fixed) and the new physical mesh $\cTh^{n+1}$ by directly integrating this formulation. The main disadvantage of this formulation is that its formula is more complicated than that of the ${\boldsymbol{\xi}}$-formulation and the metric tensor, which is defined on $\cTh^{n}$, needs to be updated every time the physical mesh is changed during the time integration of the mesh equation for $\cTh^{n+1}$. It is analytically shown in [@HK2015] that the mesh governed by the ${\boldsymbol{x}}$-formulation will stay nonsingular when it is nonsingular initially. Although such a theoretical result is not available for the ${\boldsymbol{\xi}}$-formulation, our limited numerical experience shows that the ${\boldsymbol{\xi}}$-formulation also produces nonsingular meshes.
Numerical Results for PME {#SEC:PME-numerics}
=========================
In this section we present numerical results obtained with the moving mesh finite element method described in the previous section for a number of PME examples. They include the Barenblatt-Pattle solution and the generalizations of several one-dimensional examples studied by Zhang and Wu [@Zhang-2009]. These examples are selected to demonstrate the accuracy of our method as well as its ability to deal with solutions having complex support and the waiting-time phenomenon. For the cases having an exact solution, the error in the computed solution will be measured in the (global) $L^2$ norm, i.e., $$\|e_h\|_{L^2(t_0,T; L^2(\Omega))} = \left ( \int_{t_0}^{T} \int_{\Omega} e_h^2({\boldsymbol{x}}, t) d {\boldsymbol{x}} d t \right )^{\frac{1}{2}}.$$ We choose this norm because various error estimates are obtained in this norm, e.g., see (\[ebmeyer-1\]). (An exception is Fig. \[fig:pme4-adaptivity-L1\] where the convergence history in $L^1$ norm is plotted for comparison purpose.) In our computation, we use $\tau = 10^{-4}$ (for the mesh movement), the maximal allowed time step size $\Delta t_{max} = 10^{-3}$ (for integrating PME), and the Hessian-based metric tensor (\[M-hessian\]), unless stated otherwise.
\[exam4.1\]
We first consider the Barenblatt-Pattle solution (\[BP-soln\]) with $r_0 = 0.5$ and $T = (t_0+0.1)/2$. We use it to verify the accuracy of the numerical method and the effects of the mesh adaptivity and the physical parameter $m$ on the computational accuracy.
Typical meshes and computed solutions at the final time obtained with the uniform mesh and two adaptive mesh strategies are shown in Fig. \[fig:mesh-soln-compare\] and the convergence history is shown in Fig. \[fig:pme4-adaptivity\] for the cases $m = 1$ and $2$. We can see that for both uniform and the arclength-based adaptive meshes, the convergence order is about 1.5 (i.e., $\mathcal{O}(N^{-\frac{1.5}{2}})$) for $m=1$ and 1 for the case $m=2$, with an arclength-based adaptive mesh producing slightly more accurate solutions for both cases. We notice that the exact solution (\[BP-soln\]) is in $H^1(\Omega)$ for $m = 1$ and $W^{1,\frac{m}{m-1}-\epsilon}(\Omega)$ for $m> 1$, where $\epsilon$ is a small positive number. The observed convergence order is higher than what we can expect from the solution regularity. (For example, the theoretical estimate (\[ebmeyer-1\]) shows a convergence order of $15/28$ for $m=1$ and $6/15$ for $m=2$.) Even more surprisingly, Hessian-based adaptive meshes lead to a second-order convergence rate for both the $m=1$ and $2$ cases. We do not have a rigorous explanation for this but would like to point out two relevant observations. The first is that the mesh is denser near the free boundary with the Hessian-based metric tensor than with the arclength metric tensor (e.g., see Fig. \[fig:mesh-soln-compare\]). The other is that the exact solution has higher regularity in its support than on the whole domain $\Omega$. Indeed, it can be directly verified that $$\sqrt{\det(|H(u(\cdot, t))|)} \in L^{\frac{2m}{3m-2}-\epsilon}(\text{supp}(u(\cdot, t))),$$ where $\epsilon$ is a small positive number and $H(u(\cdot, t))$ denotes the Hessian of $u$. It is known [@HS03] that for an $\mathbb{M}$-uniform mesh associated with the metric tensor (\[M-hessian\]), the linear interpolation error on a polygonal domain $D$ is bounded by $$\| u - \Pi_1 u \|_{L^2(D)} \leq C N^{-1} \| \sqrt{\det(|H(u)|)}\|_{L^{\frac{2}{3}}(D)} + h.o.t. ,$$ where $h.o.t.$ stands for higher-order terms. From this we can expect a second-order convergence if we consider linear interpolation only in the support of the solution with a Hessian-based adaptive mesh. Although this analysis does not apply directly to our current situation with a larger domain than the support, it may shed some light on why the scheme with Hessian-based adaptive meshes shows a second-order convergence.
For comparison purpose, we plot the convergence history in the $L^1$ norm in Fig. \[fig:pme4-adaptivity-L1\]. It can be seen that the the $L^1$ norm of the error behaves similarly as the $L^2$ norm.
We have seen from Figs. \[fig:pme4-adaptivity\] and \[fig:pme4-adaptivity-L1\] that mesh adaptation, especially the Hessan-based one, can significantly improve the accuracy. But this comes with additional cost. To show if the mesh adaptation can also improve the efficiency, we plot the solution error against the required CPU time (in seconds) in Fig. \[fig:pme4-cpu-m2\] for the computation corresponding to Fig. \[subfig:pme4-adaptivity-m2\]. We can see that a uniform mesh is more efficient when low accuracy is desired while mesh adaptation shows advantages for high accuracy. This is consistent with our limited experience with adaptive moving mesh computation (also see [@HR11 Page 17]). The location of the break-even point depends on specific problems and specific mesh adaptation strategies. For the current situation, we have $(N, \|e_h\|_{L^2}) \approx (300, 3\times 10^{-4})$ for Hessian-based adaptation and $(5000, 10^{-4})$ for arclength-based adaptation. We now examine the effects of the parameter $\tau$ on the accuracy. Recall that $\tau$ is used in the moving mesh equation (\[mmpde-1\]) to adjust the response time of the mesh movement to the changes in the metric tensor. The smaller $\tau$ is, the faster the response is. On the other hand, for smaller $\tau$, the mesh equation (\[mmpde-2\]) becomes stiffer and harder to integrate. Fortunately, this only causes a slight increase in the cost when an implicit solver (Matlab solver [*ode15s*]{} in our computation) is used for the mesh equation. The convergence history is shown in Fig. \[fig:pme4-tauDelta\] for Hessian-based adaptive meshes for $\tau = 10^{-2}, 10^{-3}$, and $10^{-4}$. We can see that for both cases with $m=1$ and $m=2$, the convergence with $\tau = 10^{-2}$ and $10^{-3}$ slows down when the mesh is becoming finer whereas that with $\tau=10^{-4}$ stays second order at least for the same considered range of the number of mesh elements. This indicates that the mesh concentration needs to follow up the movement of the free boundary very closely or otherwise we may lose the accuracy improvements gained with mesh adaptation.
Next, we consider the effects of the physical parameter $m$. As seen in Section \[SEC:PME-theory\], the solution at the free boundary becomes steeper for larger $m$. It is not surprising that PME will also become more difficult to solve numerically. Indeed, as we can see in Fig. \[fig:pme4-m-vary:uniform\], the convergence rate for the uniform mesh decreases as $m$ increases. This is qualitatively consistent with the theoretical analysis for various finite element approximations for PME on quasi-uniform meshes which also shows a decrease in convergence order with $m$; cf. (\[rose-1\]), (\[nochetto-1\]), (\[ebmeyer-1\]), and (\[wei-1\]) and e.g., see [@Ebmeyer-1998; @Ebmeyer-2008; @Nochetto-1988; @Rose1983]. On the other hand, for Hessian-based adaptive meshes the convergence order is second for $m=1$, 2, and 3, although the error is larger for larger $m$; see Fig. \[fig:pme4-m-vary:adaptive\]. The final mesh and computed solution obtained for $m=3$ with the Hessian-based mesh adaptation are shown in Fig. \[fig:pme4-m3\].
It is worth pointing out that there are small oscillations around the free boundary in computed solutions; e.g., see Fig. \[fig:oscillations-cross-section\]. This is due to the nature of the problem where the solution is steep and has a corner shape near the free boundary and the loss of the maximum principle in the discretization. A standard finite element discretization like the one we used here typically leads to solutions with oscillations for this type of problems (also see Zhang and Wu [@Zhang-2009] for the case with the one-dimensional PME). The oscillations may be suppressed using, for instance, monotone schemes (e.g., see [@BaSo1991; @NH2015; @Oberman2006]) or structure-preserving schemes (e.g., see [@LSSV07; @LiHu2013; @Le05; @YuSh2008; @Zhang-2009; @ZZS2013]). These schemes and their combination with adaptive mesh movement for PME are worth future investigations.
[0.35]{}![Example \[exam4.1\] with $m=2$. The meshes (closer view near (-0.35, -0.35)) and computed solutions at $t = T$ obtained with uniform and arclength- and Hessian-based adaptive meshes ($N = 25600$).[]{data-label="fig:mesh-soln-compare"}](pme4_mesh_0.pdf "fig:")
[0.35]{}![Example \[exam4.1\] with $m=2$. The meshes (closer view near (-0.35, -0.35)) and computed solutions at $t = T$ obtained with uniform and arclength- and Hessian-based adaptive meshes ($N = 25600$).[]{data-label="fig:mesh-soln-compare"}](pme4_soln_0.pdf "fig:")
\
[0.35]{}![Example \[exam4.1\] with $m=2$. The meshes (closer view near (-0.35, -0.35)) and computed solutions at $t = T$ obtained with uniform and arclength- and Hessian-based adaptive meshes ($N = 25600$).[]{data-label="fig:mesh-soln-compare"}](pme4_mesh_1.pdf "fig:")
[0.35]{}![Example \[exam4.1\] with $m=2$. The meshes (closer view near (-0.35, -0.35)) and computed solutions at $t = T$ obtained with uniform and arclength- and Hessian-based adaptive meshes ($N = 25600$).[]{data-label="fig:mesh-soln-compare"}](pme4_soln_1.pdf "fig:")
\
[0.35]{}![Example \[exam4.1\] with $m=2$. The meshes (closer view near (-0.35, -0.35)) and computed solutions at $t = T$ obtained with uniform and arclength- and Hessian-based adaptive meshes ($N = 25600$).[]{data-label="fig:mesh-soln-compare"}](pme4_mesh_2.pdf "fig:")
[0.35]{}![Example \[exam4.1\] with $m=2$. The meshes (closer view near (-0.35, -0.35)) and computed solutions at $t = T$ obtained with uniform and arclength- and Hessian-based adaptive meshes ($N = 25600$).[]{data-label="fig:mesh-soln-compare"}](pme4_soln_2.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. Convergence history (in $L^2$ norm) for the three meshing strategies as $N$ (the number of the elements) increases.[]{data-label="fig:pme4-adaptivity"}](pme4-adaptivity-m1.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. Convergence history (in $L^2$ norm) for the three meshing strategies as $N$ (the number of the elements) increases.[]{data-label="fig:pme4-adaptivity"}](pme4-adaptivity-m2.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. Convergence history (in $L^1$ norm) for the three meshing strategies as $N$ increases.[]{data-label="fig:pme4-adaptivity-L1"}](L1pme4-adaptivity-m1.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. Convergence history (in $L^1$ norm) for the three meshing strategies as $N$ increases.[]{data-label="fig:pme4-adaptivity-L1"}](L1pme4-adaptivity-m2.pdf "fig:")
![Example \[exam4.1\] with $m=2$. The $L^2$ norm of the error is plotted against the CPU time in seconds for the computation corresponding to Fig. \[subfig:pme4-adaptivity-m2\].[]{data-label="fig:pme4-cpu-m2"}](CPU-PME4-m2.pdf)
[0.35]{} ![Example \[exam4.1\]. Convergence history for different values of $\tau$.[]{data-label="fig:pme4-tauDelta"}](pme4-num-tauDelta-m1.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. Convergence history for different values of $\tau$.[]{data-label="fig:pme4-tauDelta"}](pme4-num-tauDelta-m2.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. Convergence history for different values of $m$.[]{data-label="fig:pme4-m-vary"}](pme4-num-m-uniform.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. Convergence history for different values of $m$.[]{data-label="fig:pme4-m-vary"}](pme4-num-m.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. The final mesh (close view near (-0.35, -0.35)) and computed solution for $m = 3 $ with the Hessian-based mesh adaptation ($N = 25600$).[]{data-label="fig:pme4-m3"}](pme4_mesh_m4.pdf "fig:")
[0.35]{} ![Example \[exam4.1\]. The final mesh (close view near (-0.35, -0.35)) and computed solution for $m = 3 $ with the Hessian-based mesh adaptation ($N = 25600$).[]{data-label="fig:pme4-m3"}](pme4_soln_m4.pdf "fig:")
[0.25]{}![Example \[exam4.1\] with $m=2$. The cross section at $y = 0$ of the computed solutions at $t=T$ obtained with a uniform mesh and arclength- and Hessian-based adaptive meshes ($N = 102,400$).[]{data-label="fig:oscillations-cross-section"}](PME4-oscillation-m_2-uniform-N_102400.pdf "fig:")
[0.25]{}![Example \[exam4.1\] with $m=2$. The cross section at $y = 0$ of the computed solutions at $t=T$ obtained with a uniform mesh and arclength- and Hessian-based adaptive meshes ($N = 102,400$).[]{data-label="fig:oscillations-cross-section"}](PME4-oscillation-m_2-arclength-N_102400.pdf "fig:")
[0.25]{}![Example \[exam4.1\] with $m=2$. The cross section at $y = 0$ of the computed solutions at $t=T$ obtained with a uniform mesh and arclength- and Hessian-based adaptive meshes ($N = 102,400$).[]{data-label="fig:oscillations-cross-section"}](PME4-oscillation-m_2-hessian-N_102400.pdf "fig:")
\[exam4.2\]
We now consider examples with complex solution support. The first example models the movement and interaction of two columns of a substance, which start out with the same height. It has $$m = 5, \quad \Omega = (-5.5,5.5) \times (-5.5,5.5),$$ $$u_0(x,y) =
\begin{cases}
1 , & \quad \text{for} \quad (x,y) \in (0.5,3) \times (0.5,3) \\
1 , & \quad \text{for} \quad (x,y) \in (-3,-0.5) \times (-3,-0.5) \\
0 , & \quad \text{otherwise} .
\end{cases}
\label{two-box-1}$$ A typical adaptive mesh and the corresponding computed solution obtained with the Hessian-based mesh adaptation are shown in Fig. \[fig:Two-Box-I-soln-mesh\]. It can be seen that as time evolves, the support of the solution expands from the two boxes, and then merges into one big region. The mesh adaptation strategy works nicely for the current example, with the mesh points moving to concentrate around the free boundary. Particularly, the mesh stays concentrated and nonsingular even during the merging process of the two separated support regions. Moreover, the numerical results show that the support of the solution becomes smoother as time evolves, consistent with the theoretical prediction (e.g., see [@Shmarev2005]).
\[exam4.3\]
The next example is similar to the previous one except that the initial solution has different heights in the two boxes, $$u_0(x,y) =
\begin{cases}
1 , & \quad \text{for} \quad (x,y) \in (0.5,3) \times (0.5,3) \\
1.5 , & \quad \text{for} \quad (x,y) \in (-3,-0.5) \times (-3,-0.5) \\
0 , & \quad \text{otherwise} .
\end{cases}
\label{two-box-2}$$ A typical adaptive mesh and the corresponding solution are shown in Fig. \[fig:Two-Box-II-soln-mesh\]. Once again, the mesh is concentrated correctly around the free boundaries as they evolve with time. Moreover, the region with larger initial solution values expands faster than the region with smaller values. Overall, the support of the solution for this example expands faster than that of the previous example. At $t=50$, the two boxes have already merged into a single region of calabash shape.
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-mesh-twoBox1_tf-0.pdf "fig:")
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-soln-twoBox1_tf-0.pdf "fig:")
\
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-mesh-twoBox1_tf-51.pdf "fig:")
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-soln-twoBox1_tf-51.pdf "fig:")
\
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-mesh-twoBox1_tf-10001.pdf "fig:")
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-soln-twoBox1_tf-10001.pdf "fig:")
\
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-mesh-twoBox1_tf-50000.pdf "fig:")
[0.35]{}![Example \[exam4.2\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-I-soln-mesh"}](pme4-soln-twoBox1_tf-50000.pdf "fig:")
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-mesh-twoBox2_tf-0.pdf "fig:")
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-soln-twoBox2_tf-0.pdf "fig:")
\
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-mesh-twoBox2_tf-50.pdf "fig:")
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-soln-twoBox2_tf-50.pdf "fig:")
\
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-mesh-twoBox2_tf-10000.pdf "fig:")
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-soln-twoBox2_tf-10000.pdf "fig:")
\
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-mesh-twoBox2_tf-50000.pdf "fig:")
[0.35]{}![Example \[exam4.3\]. An adaptive mesh and the corresponding computed solution at various time instants ($N = 14400$).[]{data-label="fig:Two-Box-II-soln-mesh"}](pme4-soln-twoBox2_tf-50000.pdf "fig:")
\[exam4.4\] From Section \[SEC:PME-theory\] we recall that PME exhibits the waiting-time phenomenon for a certain type of initial solutions. To see this, we consider $$m = 8, \quad \Omega = (-\pi, \pi) \times (-\pi, \pi) ,$$ $$u_0(x,y) =
\begin{cases}
\cos(\sqrt{x^2 + y^2}) , & \quad \text{for} \quad \sqrt{x^2 + y^2} \leq \frac{\pi}{2} \\
0 , & \quad \text{otherwise} .
\end{cases}
\label{waiting-time-1}$$ We have $$\grad \cos^m(\sqrt{x^2+y^2}) = -\frac{m\, \cos^{m-1}(\sqrt{x^2+y^2})\, \sin(\sqrt{x^2+y^2})}{\sqrt{x^2+y^2}}
\begin{bmatrix} x \\ y \end{bmatrix} ,$$ which diminishes at $\sqrt{x^2 + y^2} = \frac{\pi}{2}$. From Darcy’s law (\[Darcy-law\]), we do not anticipate that the free boundary moves initially.
In Figs. \[fig:waiting-2D-cross-section\] and Fig. \[fig:waiting-2D\] we show the cross section at $y = 0$ of a computed solution and the solution itself. The results show that the free boundary of the solution does not move until around $t = 10$. Before this time, the solution is steepening. Interestingly, the steepening does not occur on the whole initial support. Instead, it first occurs on a smaller region inside the support and then this region is expanding until it fills the whole initial support. After that, the free boundary waits until it becomes sufficiently steep and then moves.
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-6.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-11.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-26.pdf "fig:")
\
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-51.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-111.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-501.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1001.pdf "fig:")
\
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1101.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1201.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1301.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1401.pdf "fig:")
\
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1501.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1601.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1701.pdf "fig:")
[0.25]{}![Example \[exam4.4\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D-cross-section"}](pme4-waiting_tf-1801.pdf "fig:")
\
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-mesh_tf-10.pdf "fig:")
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-soln_tf-10.pdf "fig:")
\
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-mesh_tf-50.pdf "fig:")
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-soln_tf-50.pdf "fig:")
\
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-mesh_tf-500.pdf "fig:")
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-soln_tf-500.pdf "fig:")
\
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-mesh_tf-1801.pdf "fig:")
[0.35]{}![Example \[exam4.4\]. A computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-2D"}](pme4-waiting-soln_tf-1801.pdf "fig:")
Numerical experiment for PME with variable exponents and absorption {#SEC:PME-numerics-2}
===================================================================
To demonstrate the robustness of the moving mesh finite element method described in Section \[SEC:mmfem\], we consider its application to PME with absorption and/or variable exponents, $$u_t = \grad \cdot (|u|^{\gamma} \grad u) - \lambda u^{\sigma}, \quad \Omega \times (t_0, T]
\label{PME-2}$$ subject to a homogeneous Dirichlet boundary condition and an initial condition. Here, $\gamma = \gamma ({\boldsymbol{x}}, t)$ and $\sigma = \sigma ({\boldsymbol{x}}, t)$ are nonnegative bounded functions and $\lambda$ is a constant. PME in the form of (\[PME-2\]) arises in continuum mechanics to model the motion of a barotropic gas through a porous medium, where the pressure is considered to depend on the density and temperature [@Antontsev2005]. Like the standard PME, (\[PME-2\]) with constant exponents (i.e., PME with absorption) has been studied extensively; e.g., see [@Knerr1979; @Shmarev2005]. However, there are very few theoretical results for the case with variable exponents [@Antontsev2005; @Lian2008]. For example, there is no theoretical result on the movement of the free boundary (cf. (\[Darcy-law\])) although the solution to (\[PME-2\]) is known to have the property of finite speed of propagation. Neither is there much numerical work on this situation; see [@Duque2013; @Duque2014; @Duque2015].
\[exam5.1\] We first consider an example with an absorption term, $$\begin{aligned}
& \lambda = 1, \quad \gamma = 2,\quad \sigma = 0.1,
\quad \Omega = (-1.5 \pi, 1.5 \pi) \times (-1.5 \pi, 1.5 \pi) , \\
&
u_0 =
\begin{cases}
| \sin(\sqrt{x^2 + y^2}) | , & \quad \text{for} \quad \sqrt{x^2 + y^2} \in (\frac{\pi}{6}, \pi) \\
0.5 , & \quad \text{for} \quad \sqrt{x^2 + y^2} \in [0, \frac{\pi}{6}) \\
0 , & \quad \text{otherwise} .
\end{cases}\end{aligned}$$ This example is the two-dimensional generalization of a one-dimensional example in [@Zhang-2009] that shows a splitting phenomenon in the middle after a finite time.
An adaptive mesh and the corresponding computed solution are shown in Fig. \[fig:Splitting-soln-mesh\]. We can see that as time evolves, the solution is becoming lower and the support is expanding on the outer boundary. Meanwhile, the solution is being “punched through” with a hole at the middle of the support. This is an additional feature with the absorption term.
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-mesh_tn-0.pdf "fig:")
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-soln_tf-0.pdf "fig:")
\
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-mesh_tn-40.pdf "fig:")
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-soln_tf-40.pdf "fig:")
\
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-mesh_tn-64.pdf "fig:")
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-soln_tf-64.pdf "fig:")
\
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-mesh_tn-80.pdf "fig:")
[0.35]{}![Example \[exam5.1\]. An adaptive mesh and the corresponding solution at various time instants ($N = 40000$).[]{data-label="fig:Splitting-soln-mesh"}](pme4-splitting-soln_tf-80.pdf "fig:")
\[exam5.2\] For this example, $$\begin{aligned}
& \lambda = 0,\quad \gamma = \left (\frac{x}{2}\right )^2 + \left (\frac{y}{2}\right )^2 + 1.1,
\quad \Omega = (-2,2)\times (-2,2),
\\
& u_0 = \begin{cases} - \sin(2 \pi \sqrt{x^2 + y^2}),& \text{for} \quad 0.5 < \sqrt{x^2 + y^2} < 1 \\
0,& \text{otherwise}.
\end{cases}\end{aligned}$$ This example has been studied in [@Duque2013; @Duque2015]. The support of the solution has a hole in the middle which disappears in a finite time. We take $t \in [0, 0.2]$ in the computation. An adaptive mesh and the corresponding numerical solution is shown in Fig. \[fig:VarExp1-example5.2\]. The result appears to have better resolution than that in [@Duque2013] where a uniform mesh has been used. Moreover, our method works just fine through the closing of the inside hole (cf. Fig. \[fig:VarExp1-example5.2\]) whereas the method in [@Duque2015] which explicitly traces the free boundary encounters the mesh singularity problem near the time when the hole is closing.
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-mesh-varExp-1_tf-0.pdf "fig:")
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-varExp-1_tf-0.pdf "fig:")
\
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-mesh-varExp-1_tf-2.pdf "fig:")
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-varExp-1_tf-2.pdf "fig:")
\
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-mesh-varExp-1_tf-15.pdf "fig:")
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-varExp-1_tf-15.pdf "fig:")
\
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-mesh-varExp-1_tf-70.pdf "fig:")
[0.35]{}![Example \[exam5.2\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp1-example5.2"}](pme4-varExp-1_tf-70.pdf "fig:")
\[exam5.3\] For this example, $$\begin{aligned}
& \lambda = 0,\quad \gamma = 2-x-y,\quad \Omega = (-1.5,1.5)\times (-1.5,1.5),
\\
& u_0 = \begin{cases} 5 (0.25-x^2-y^2), & \text{for} \quad \sqrt{x^2 + y^2} < 0.5 \\
0,& \text{otherwise}.
\end{cases}\end{aligned}$$ This example has been studied in [@Duque2015]. We take $t \in [0, 0.05]$. The free boundary of the solution does not move until $t \approx 0.02$. A moving mesh and the corresponding computed solution are shown in Fig. \[fig:VarExp2-example5.3\]. We can see that the variation of the exponent causes the free boundary to expand anisotropically and the solution to have different steepness along the free boundary. Moreover, a closer examination of the results confirms the waiting time phenomenon, where the interface in the region $\{(x,y): x+y \leq 0 \}$ does not move until a finite time has elapsed. Fig. \[fig:waiting-varExp2-2D-cross-section\] show the cross sections of the numerical solutions with the plane $y = x$ at various instants of time. In the figure, the red dashed line refers to the position of the initial interface, where the waiting time phenomenon subsequently occurs.
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-mesh-varExp-2_tf-0.pdf "fig:")
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-varExp-2_tf-0.pdf "fig:")
\
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-mesh-varExp-2_tf-6.pdf "fig:")
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-varExp-2_tf-6.pdf "fig:")
\
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-mesh-varExp-2_tf-28.pdf "fig:")
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-varExp-2_tf-28.pdf "fig:")
\
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-mesh-varExp-2_tf-90.pdf "fig:")
[0.35]{}![Example \[exam5.3\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp2-example5.3"}](pme4-varExp-2_tf-90.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-0.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-1.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-2.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-3.pdf "fig:")
\
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-4.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-5.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-6.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-7.pdf "fig:")
\
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-15.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-20.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-30.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-40.pdf "fig:")
\
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-50.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-60.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-70.pdf "fig:")
[0.25]{}![Example \[exam5.3\]. The cross section at $y = 0$ of a computed solution is shown at various time instants ($N = 40000$).[]{data-label="fig:waiting-varExp2-2D-cross-section"}](pme4-varExp2-waiting_tf-95.pdf "fig:")
\
\[exam5.4\] The last example, taken from [@Duque2014], has time dependent exponents, i.e., $$\begin{aligned}
& \lambda = 1,\quad \gamma = \frac{x^2+y^2}{t^2+1},\quad \sigma = x^2+y^2 + 1 + e^{-t},
\quad \Omega = (-1.5,1.5)\times (-1.5,1.5),
\\
& u_0 = \begin{cases} \cos(2 \pi (x^2+y^2)), & \text{for} \quad \sqrt{x^2 + y^2} < 0.5 \\
0,& \text{otherwise}.
\end{cases}\end{aligned}$$ We take $t \in [0, 0.1]$. The numerical results are shown in Fig. \[fig:VarExp3-example5.4-mesh\]. They are comparable with those in [@Duque2014].
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-mesh-varExp-3_tf-0.pdf "fig:")
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-varExp-3_tf-0.pdf "fig:")
\
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-mesh-varExp-3_tf-3.pdf "fig:")
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-varExp-3_tf-3.pdf "fig:")
\
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-mesh-varExp-3_tf-6.pdf "fig:")
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-varExp-3_tf-6.pdf "fig:")
\
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-mesh-varExp-3_tf-50.pdf "fig:")
[0.35]{}![Example \[exam5.4\]. An adaptive mesh and the corresponding solution at various time instants ($N = 25600$).[]{data-label="fig:VarExp3-example5.4-mesh"}](pme4-varExp-3_tf-50.pdf "fig:")
Conclusions and further remarks {#SEC:conclusion}
===============================
In the previous sections we have studied an adaptive moving mesh finite element method for the numerical solution of PME. The method is based on the MMPDE moving mesh strategy and its new implementation and uses a linear finite element method and the fifth-order Radau IIA scheme for the spatial and temporal discretization. Numerical results show that the method is able to produce correct mesh concentration around the free boundary and deal with problems having complex solution support. Three types of mesh have been considered, uniform and acrlength- and Hessian-based adaptive meshes. The method shows a first-order convergence behavior as the mesh is refined for uniform and arclength-based adaptive meshes and improves to a second-order convergence when Hessian-based adaptive meshes are used. This indicates that mesh concentration around the free boundary is important to the accuracy of the method. Moreover, the prompt response of the mesh movement to the changes in the solution is also crucial, requiring that a small value of the parameter $\tau$ in mesh movement (cf. (\[mmpde-1\])) be used especially for the computation with fine meshes. We have also studied the application of the method to PME with variable exponents and absorption for which there are very few theoretical results available. Numerical results demonstrate that the method is robust and able to deal with PDEs having more complicated structures.
It should be pointed out that there are small oscillations around the free boundary in computed solutions; see the discussion in Sect. \[SEC:PME-numerics\]. How to suppress these oscillations using a monotone or structure-preserving scheme (e.g., see [@BaSo1991; @LSSV07; @LiHu2013; @NH2015; @Oberman2006; @Le05; @YuSh2008; @Zhang-2009; @ZZS2013]) and to combine them with adaptive mesh movement for PME are worth future investigations.
[**Acknowledgment.**]{} Support from US Army Research Office under grant W911-NF-1510377 is gratefully acknowledged. The authors would also like to thank the anonymous referees for their valuable comments in improving the quality of the paper.
[10]{}
S. N. Antontsev and S. I. Shmarev. A model porous medium equation with variable exponent of nonlinearity: existence, uniqueness and localization properties of solutions. , 60(3):515–545, 2005.
D. G. Aronson. Regularity properties of flows through porous media. , 17:461–467, 1969.
M. J. Baines. . Oxford University Press, Oxford, 1994.
M. J. Baines, M. E. Hubbard, and P. K. Jimack. A moving mesh finite element algorithm for fluid flow problems with moving boundaries. , 47:1077–1083, 2005. 8th ICFD Conference on Numerical Methods for Fluid Dynamics. Part 2.
M. J. Baines, M. E. Hubbard, and P. K. Jimack. A moving mesh finite element algorithm for the adaptive solution of time-dependent partial differential equations with moving boundaries. , 54:450–469, 2005.
M. J. Baines, M. E. Hubbard, and P. K. Jimack. Velocity-based moving mesh methods for nonlinear partial differential equations. , 10:509–576, 2011.
M. J. Baines, M. E. Hubbard, P. K. Jimack, and A. C. Jones. Scale-invariant moving finite elements for nonlinear partial differential equations in two dimensions. , 56:230–252, 2006.
G. Barles and P. E. Souganidis. Convergence of approximation schemes for fully nonlinear second order equations. , 4:271–283, 1991.
W. Bo and M. Shashkov. R-adaptive reconnection-based arbitrary Lagrangian Eulerian method-R-ReALE. , 48:125–167, 2015.
C. Budd, G. Collins, W. Huang, and R. D. Russell. Self-similar numerical solutions of the porous medium equation using moving mesh methods. , 357:1047–1078, 1999.
C. J. Budd, W. Huang, and R. D. Russell. Adaptivity with moving grids. , 18:111–241, 2009.
C. J. Budd and J. F. Williams. Parabolic Monge-Ampére methods for blow-up problems in several spatial dimensions. , 39:5425–5444, 2006.
L. A. Caffarelli and A. Friedman. Regularity of the free boundary of a gas flow in an [$n$]{}-dimensional porous medium. , 29:361–391, 1980.
W. Cao, W. Huang, and R. D. Russell. A moving mesh method based on the geometric conservation law. , 24:118–142, 2002.
P. Daskalopoulos and R. Hamilton. Regularity of the free boundary for the porous medium equation. , 11:899–965, 1998.
J. C. M. Duque, R. M. P. Almeida, and S. N. Antontsev. Convergence of the finite element method for the porous media equation with variable exponent. , 51(6):3483–3504, 2013.
J. C. M. Duque, R. M. P. Almeida, and S. N. Antontsev. Numerical study of the porous medium equation with absorption, variable exponents of nonlinearity and free boundary. , 235:137–147, 2014.
J. C. M. Duque, R. M. P. Almeida, and S. N. Antontsev. Application of the moving mesh method to the porous medium equation with variable exponent. , 118:177–185, 2015.
C. Ebmeyer. Error estimates for a class of degenerate parabolic equations. , 35:1095–1112, 1998.
C. Ebmeyer and W. B. Liu. Finite element approximation of the fast diffusion and the porous medium equations. , 46:2393–2410, 2008.
E. Emmrich and D. [Š]{}i[š]{}ka. Full discretization of the porous medium/fast diffusion equation based on its very weak formulation. , 10:1055–1080, 2012.
S. Gonz[á]{}lez-Pinto, J. I. Montijano, and S. P[é]{}rez-Rodr[í]{}guez. Two-step error estimators for implicit [R]{}unge-[K]{}utta methods applied to stiff systems. , 30(1):1–18, 2004.
E. Hairer, S. P. N[ø]{}rsett, and G. Wanner. Springer-Verlag, Belin, second edition, 1993.
E. Hairer and G. Wanner. , volume 14 of [*Springer Series in Computational Mathematics*]{}. Springer-Verlag, Berlin, second edition, 1996. Stiff and differential-algebraic problems.
C. W. Hirt, A. A. Amsden, and J. L. Cook. An arbitrary Lagrangian-Eulerian computing method for all flow speeds. , 14:227–253, 1974.
W. Huang. Variational mesh adaptation: isotropy and equidistribution. , 174:903–924, 2001.
W. Huang. Mathematical principles of anisotropic mesh adaptation. , 1:276–310, 2006.
W. Huang and L. Kamenski. A geometric discretization and a simple implementation for variational mesh generation and adaptation. , 301:322–337, 2015. (arXiv:1410.7872).
W. Huang and L. Kamenski. On the mesh nonsingularity of the moving mesh [PDE]{} method. submitted, 2015. (arXiv:1512.04971).
W. Huang, Y. Ren, and R. D. Russell. Moving mesh partial differential equations ([MMPDEs]{}) based upon the equidistribution principle. , 31:709–730, 1994.
W. Huang and R. D. Russell. . Springer, New York, 2011. Applied Mathematical Sciences Series, Vol. 174.
W. Huang and W. Sun. Variational mesh adaptation [II]{}: error estimates and monitor functions. , 184:619–648, 2003.
P. K. Jimack and A. J. Wathen. Temporal derivatives in the finite-element method on continuously deforming grids. , 28:990–1003, 1991.
A. S. Kala[š]{}nikov. Formation of singularities in solutions of the equation of nonstationary filtration. , 7:440–444, 1967.
L. Kamenski. . PhD thesis, TU Darmstadt, 2009.
L. Kamenski and W. Huang. How a nonconvergent recovered [H]{}essian works in mesh adaptation. , 52:1692–1708, 2014. (arXiv:1211.2877).
B. F. Knerr. The behavior of the support of solutions of the equation of nonlinear heat conduction with absorption in one dimension. , 249:409–424, 1979.
R. Li, T. Tang, and P. W. Zhang. Moving mesh methods in multiple dimensions based on harmonic maps. , 170:562–588, 2001.
S. Lian, W. Gao, C. Cao, and H. Yuan. Study of the solutions to a model porous medium equation with variable exponent of nonlinearity. , 342:27–38, 2008.
G. J. Liao and D. Anderson. A new approach to grid generation. , 44:285–298, 1992.
K. Lipnikov, M. Shashkov, D. Svyatskiy, and Y. Vassilevski. Monotone finite volume schemes for diffusion equations on unstructured triangular and shape-regular polygonal meshes. , 227:492–512, 2007.
X. Li and W. Huang. Maximum principle for the finite element solution of time dependent anisotropic diffusion problems. , 29:1963–1985, 2013.
K. Miller and R. Miller. Moving finite elements I. , 18:1019–1032, 1981.
C. Ngo and W. Huang. Monotone finite difference schemes for anisotropic diffusion problems via nonnegative directional splittings. , 19:473–495, 2016.
R. H. Nochetto and C. Verdi. Approximation of degenerate parabolic problems using numerical integration. , 25:784–814, 1988.
A. M. Oberman. Convergent difference schemes for degenerate elliptic and parabolic equations: [H]{}amilton-[J]{}acobi equations and free boundary problems. , 44:879–895 (electronic), 2006.
O. A. Ole[ĭ]{}nik, A. S. Kala[š]{}inkov, and Y. [Č]{}[ž]{}ou. The [C]{}auchy problem and boundary problems for equations of the type of non-stationary filtration. , 22:667–704, 1958.
C. Le Potier. A finite volume method for the approximation of highly anisotropic diffusion operators on unstructured meshes. , IV:401-412, 2005.
M. E. Rose. Numerical methods for flows through porous media. [I]{}. , 40:435–467, 1983.
M. E. Rose. Numerical methods for flows through porous media. [I]{}. , 40:435–467, 1983.
J. Rulla and N. J. Walkington. Optimal rates of convergence for degenerate parabolic problems in two dimensions. , 33:56–67, 1996.
S. Shmarev. Interfaces in solutions of diffusion-absorption equations in arbitrary space dimension. In [*Trends in partial differential equations of mathematical physics*]{}, volume 61 of [*Progr. Nonlinear Differential Equations Appl.*]{}, pages 257–273. Birkhäuser, Basel, 2005.
S. I. Shmarev. Interfaces in multidimensional diffusion equations with absorption terms. , 53:791–828, 2003.
T. Tang. Moving mesh methods for computational fluid dynamics flow and transport. In [*Recent Advances in Adaptive Computation (Hangzhou, 2004)*]{}, volume 383 of [*AMS Contemporary Mathematics*]{}, pages 141–173. Amer. Math. Soc., Providence, RI, 2005.
J. L. V[á]{}zquez. . Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, Oxford, 2007. Mathematical theory.
D. Wei and L. Lefton. A priori [$L^\rho$]{} error estimates for [G]{}alerkin approximations to porous medium and fast diffusion equations. , 68:971–989, 1999.
G. Yuan and Z. Sheng. Monotone finite volume schemes for diffusion equations on polygonal meshes. , 227:6288–6312, 2008.
Q. Zhang and Z.-L. Wu. Numerical simulation for porous medium equation by local discontinuous [G]{}alerkin finite element method. , 38:127–148, 2009.
Y. Zhang, X. Zhang, and C.-W Shu. Maximum-principle-satisfying second order discontinuous Galerkin schemes for convection-diffusion equations on triangular meshes. , 234:295–316, 2013.
[^1]: Department of Mathematics, University of Kansas, Lawrence, KS 66045, U.S.A. ([](mailto:[email protected])).
[^2]: Department of Mathematics, University of Kansas, Lawrence, KS 66045, U.S.A. ([](mailto:[email protected])).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- Dominik Gerstung
- Norbert Kaiser
- Wolfram Weise
title: 'Hyperon-nucleon three-body forces and strangeness in neutron stars'
---
Introduction {#sec:intro}
============
The existence of heavy neutron stars with masses around 2$M_\odot$ [@Demorest2010; @Antoniadis2013; @Fonseca2016; @Cromartie2019] sets strong constraints on the equation of state (EoS) of dense baryonic matter. This EoS must be sufficiently stiff, i.e. the pressure $P(\cal{E})$ at energy densities $\cal{E}\sim$ 1 GeV/fm$^3$ must be large enough to support such massive compact objects against gravitational collapse. The detection of gravitational wave signals from two merging neutron stars [@Abbott2017] adds further important information on the EoS, by providing limits for the tidal deformability and for neutron star radii [@Most2018; @De2018].
The composition and properties of strongly interacting matter at high baryon densities is a topic of continuing interest. Various options are under discussion. A time-honored description of matter in the core of neutron stars uses hadronic degrees of freedom (baryons and mesons) with strong many-body correlations [@APR1998]. A modern version that arrives at a similar EoS, consistent with observations, is based on a chiral nucleon-meson field theory combined with functional renormalization group methods [@DW2015; @DW2017]. In this latter approach neutron star matter can be represented as a relativistic Fermi liquid in the sense of Landau theory [@FW2019]. The density-dependent leading Landau parameters show the characteristic behaviour of a strongly correlated fermionic many-body system, albeit less extreme in comparison with another well-known Fermi system at low temperature, namely liquid $^3$He. Alternative descriptions of neutron star matter involve a (possibly smooth) transition from hadronic matter to some form of quark matter [@Baym2018; @Baym2019; @McLerran2019; @FFM2019].
The required stiffness of the EoS implies strong restrictions on the appearance of hyperons in neutron star matter and for the underlying hyperon-nuclear interactions. A naive introduction of $\Lambda$ hyperons, with only two-body $\Lambda N$ interactions, would appear to be energetically favorable by replacing neutrons at baryon densities around 2-3 $\rho_0$ (in terms of the equilibrium density of normal nuclear matter, $\rho_0 = 0.16$ fm$^{-3}$). However, then the EoS of neutron star matter would become far too soft to satisfy the two-solar-mass constraint [@Djapo2010; @Hell2014; @Lonardoni2015]. A phenomenological analysis [@BS2015], taking into account constraints from both hypernuclear physics and neutron star observations, points out indeed that the interaction between $\Lambda$ hyperons and dense matter has to become repulsive at densities below three times $\rho_0$.
The aim of the present work is to investigate whether and to what extent a microscopic description of hyperon-nucleon two- and three-body forces is capable of providing the necessary repulsion in dense baryonic matter. The key quantities to be calculated are the hyperon ($\Lambda$ and $\Sigma$) single-particle potentials $U_{\Lambda,\Sigma}(\rho)$ in nuclear and neutron matter. The density dependence of the resulting chemical potential $\mu_\Lambda(\rho)= M_\Lambda +U_{\Lambda}(\rho)$ of a $\Lambda$ in neutron matter indicates whether it is favourable to replace neutrons by $\Lambda$ hyperons in the core of neutron stars.
Our starting point is the hyperon-nucleon interaction derived from SU(3) chiral effective field theory (ChEFT) at next-to-leading order (NLO) [@Haidenbauer2013]. This interaction includes exchanges of one and two pseudoscalar octet mesons and four-baryon contact terms with SU(3)-symmetric low-energy constants fitted to the (admittedly scarce) hyperon-nucleon scattering data. Three-body forces enter first at next-to-next-to-leading order (NNLO) in this scheme [@Petschauer2016]. Estimating the strength of contact terms in these forces through the contributions of explicit baryon decuplet resonances in intermediate states, these three-body interactions are promoted from NNLO to NLO [@Petschauer2017] due to the small decuplet-octet baryon mass-splitting. We shall adopt here the same strategy. Previous calculations of hyperon-nuclear potentials using Brueckner theory with these interactions as input have been reported in refs.[@Kohno2018; @Petschauer2016b; @Haidenbauer2017]. The present study builds on these results, with extensions in several directions. Special attention will be paid to the importance of the $\Lambda N\leftrightarrow\Sigma N$ transition potential, a necessary ingredient of any such calculation. In particular, the $\Lambda NN\leftrightarrow\Sigma NN$ coupled channels in the three-body sector, translated into density-dependent effective two-body potentials as in ref.[@Petschauer2017], will be treated explicitly when solving coupled-channel Bethe-Goldstone equations.
Basic baryonic interactions {#sec:int}
===========================
Hyperon-nucleon two-body interactions
-------------------------------------
For the description of the hyperon-nucleon two-body interaction, SU(3) chiral effective field theory is used up to NLO with the Weinberg power counting applied to the potential, as reported in detail in ref.[@Haidenbauer2013]. The leading order potential involves the exchange of a single pseudoscalar octet meson $(\pi, K, \eta)$ and non-derivative four-baryon contact terms. At NLO two-meson exchange diagrams at one-loop level arise together with additional contact terms with explicit momentum dependence of order $p^2$ (see fig.\[fig:pwrcounting\]). The set of SU(3)-symmetric contact terms represents unresolved short-distance dynamics. The corresponding low-energy constants (LECs) are fitted to low-energy hyperon-nucleon scattering data and the hypertriton ${}^3_\Lambda H$ binding energy (and partly to elastic $NN$-scattering phase shifts). SU(3) symmetry breaking is none- theless incorporated through the physical masses of the baryons and of the exchanged pseudoscalar mesons. Further details on input parameters can be found in ref.[@Haidenbauer2013].
![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBcont "fig:"){width="\feynwidth"} ![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBtree "fig:"){width="\feynwidth"}\
![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBcontNLO "fig:"){width="\feynwidth"} ![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBpb "fig:"){width="\feynwidth"}![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBcb "fig:"){width="\feynwidth"}![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBtriR "fig:"){width="\feynwidth"}![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBtriL "fig:"){width="\feynwidth"}![Leading and next-to-leading order diagrams representing the baryon-baryon interaction potential. Solid and dashed lines denote octet baryons ($N,\Lambda,\Sigma$) and mesons ($\pi,K,\eta$), respectively.[]{data-label="fig:pwrcounting"}](FBBfoot "fig:"){width="\feynwidth"}
Solving the coupled-channel Lippmann-Schwinger equations using these chiral potentials involves a regulator function of the form $\exp[-(p'^4+p^4)/\lambda^4]$ depending on the momenta, $p$ and $p'$, of the in- and outgoing baryons. The cutoff $\lambda$ is chosen in a reasonable range 500 ... 650 MeV.
A characteristic feature of the resulting $\Lambda N$ S-wave phase shifts that distingushes the chiral EFT-based potential from all earlier phenomenological interactions, such as Jülich’04 [@HM2005] and Nijmegen NSC97f [@Rijken1999], is the following. As there is no (leading-order) one-pion exchange $\Lambda N$ interaction term, the driving intermediate range attraction in this channel comes from the iterative process $\Lambda N \rightarrow \Sigma N \rightarrow \Lambda N$, involving the exchange of two pions. The $\Lambda\Sigma\pi$ coupling as given by SU(3) symmetry implies strong attraction through this second-order process that is partly balanced by the strong short-distance repulsion. The net attraction is seen in the low-energy $^3S_1$ $\Lambda N$ phase shift. The $^1S_0$ phase shift also shows this attractive behavior at low energies but changes sign around a lab momentum of 600 MeV and becomes repulsive. This momentum-dependent repulsion is generated by the short-distance terms at NLO, as can be clearly seen by selectively turning off the $\Lambda\Sigma\pi$ coupling [@Haidenbauer2017]. The repulsive effects in the $\Lambda N$ interaction at high momenta have their direct impact on the $\Lambda$ single-particle potentials in nuclear and neutron matter which change from attraction to repulsion at densities around or below $2\,\rho_0$, a feature not matched by any of the phenomenological potentials. However, as we shall see, such two-body repulsion is still not strong enough to prevent $\Lambda$ hyperons from occuring at baryon densities encountered in neutron stars and thereby softening the EoS to an unwanted degree.
Three-body interactions of hyperons with nucleons
-------------------------------------------------
Three-nucleon forces are an important element in the quantitative understanding of nuclear few- and many-body systems. Likewise, the $\Lambda NN$ three-body interactions is expected to play a significant role in hypernuclear systems. Our interest in the present work is directed, in particular, to the behaviour of $\Lambda NN$ and $\Sigma NN$ three-body forces (3BF) in dense baryonic matter, where their influence is expected to grow continuously with increasing density.
In the systematic expansion of SU(3) chiral EFT, baryonic three-body forces appear at NNLO. Their detailed derivation is given in ref.[@Petschauer2016] on which we shall build. The leading 3BF diagrams are shown in fig.\[fig:3BF\]. They fall into three classes: six-baryon contact terms, one-meson exchange pieces, and two-meson exchange pieces. The corresponding potentials are denoted by $V^{(0)}, V^{(1)}$ and $V^{(2)}$, respectively. The contact potential $V^{(0)}$ involves all possible combinations of baryon spin operators $\vec \sigma_1, \vec \sigma_2, \vec \sigma_3$. The one-meson exchange 3BF potential has the following generic form: $$V^{(1)}_i = \frac{1}{2f^2} \frac{\vec\sigma_1\cdot\vec q}{\vec q^{\,2}+m_i^2} (
A \vec\sigma_3+ B \, \mathrm i \vec\sigma_2\times\vec\sigma_3)\cdot\vec q
\,,$$ where the inverse squared pseudoscalar meson decay constant $f^{-2} 4.8\,$fm$^2$ determines the interaction strength. The index $i\in \{\pi^0,\pi^+,\pi^-,K^+,K^-,K^0,\bar K^0,\eta\} $ refers to the exchanged meson with mass $m_i$, carrying the momentum transfer $\vec q$. The two parameters $A$ and $B$ are combinations of low-energy constants.
![ Leading three-baryon interactions: contact term, one-meson exchange and two-meson exchange. \[fig:3BF\] ](FBBBcont "fig:")![ Leading three-baryon interactions: contact term, one-meson exchange and two-meson exchange. \[fig:3BF\] ](FBBB1ME "fig:")![ Leading three-baryon interactions: contact term, one-meson exchange and two-meson exchange. \[fig:3BF\] ](FBBB2ME2 "fig:")
On the other hand, the two-meson exchange potential (depicted in fig.\[fig:3BF\]) has following generic form: $$\begin{aligned}
V^{(2)}_{ij} ={} & -\frac{1}{4f^4} \frac{\vec\sigma_1\cdot\vec q\ ~~\vec\sigma_3\cdot\vec q'}{(\vec q^{\,2}+m_i^2)(\vec q'^{\,2}+m_j^2)} \\
&\qquad\times\Big(A' + B'\vec q\cdot\vec q' +D'\, \mathrm i\vec\sigma_2\cdot(\vec q\times\vec q')\Big) \,. {\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ The parameters $A',B'$ and $D'$ are combinations of low-energy constants determined in the meson-baryon subsector. The transferred momenta $\vec q$ and $\vec q'$ are carried by the mesons with masses $m_i$ and $m_j$, respectively. The complete three-body potential is constructed by summing the contributions from all distinguishable diagrams with all possible exchanged pseudoscalar mesons, including baryon exchange operations where needed.
As they stand, the 3BF terms at NNLO introduce a prohibitively large number of low-energy constants in the contact potentials $V^{(0)}$. The existing data-base is too limited and does not permit a meaningfully constrained determination of these parameters. However, their number can be reduced substantially by employing an approximate scheme referred to as decuplet saturation: contact vertices are resolved by propagating explicit decuplet baryons in intermediate states. Furthermore, it has the welcome feature of promoting three-baryon forces from NNLO to NLO within the chiral hierarchy. In essence, the NNLO diagrams in fig.\[fig:3BF\] are replaced by the NLO diagrams of fig.\[fig:3BFdec\]. In our subsequent application, two of the in- or outgoing baryons are nucleons (proton, neutron), while the third one is a strangeness $S=-1$ hyperon ($\Lambda$ or $\Sigma$) and only the (long-range) pion-exchange is considered. The decuplet intermediate states are then either $\Sigma^*(1385)$ or $\Delta(1232)$.
![Three-baryon forces with explicit decuplet baryons in the intermediate states (represented by double lines) [@Petschauer2017].[]{data-label="fig:3BFdec"}](FBBBdeccont "fig:")![Three-baryon forces with explicit decuplet baryons in the intermediate states (represented by double lines) [@Petschauer2017].[]{data-label="fig:3BFdec"}](FBBBdec1MEu "fig:")![Three-baryon forces with explicit decuplet baryons in the intermediate states (represented by double lines) [@Petschauer2017].[]{data-label="fig:3BFdec"}](FBBBdec1MEd "fig:")![Three-baryon forces with explicit decuplet baryons in the intermediate states (represented by double lines) [@Petschauer2017].[]{data-label="fig:3BFdec"}](FBBBdec2ME "fig:")
In what follows we use the detailed formalism developed in refs.[@Petschauer2016; @Petschauer2017]. Within the decuplet dominance approximation there are only three remaining constants to be determined: the coupling strength $C$ of the transition vertex between an octet baryon and a decuplet baryon with a pseudoscalar meson absorbed or emitted; and two coupling constants, $H_1$ and $H_2$, for the four-point vertices connecting three octet baryon lines and one decuplet baryon [@Petschauer2016; @Petschauer2017]. The coupling $C$ is uniquely determined by the $\pi N\rightarrow\Delta$ transition vertex which is in turn constrained by the decay width $\Gamma(\Delta\rightarrow \pi N)\simeq 115$ MeV. The (large-$N_c$) value $C = 3g_A/4 \simeq 0.95$ with the nucleon axial vector coupling constant, $g_A = 1.26$, is well compatible with this constraint. The remaining two constants, $H_{1,2}$ of dimension $(length)^2$, are still free to choose. We use hypernuclear phenomenology in order to restrict their possible values.
Next, the three-body interactions in the $\Lambda NN\!\rightarrow\!\Lambda NN$, $\Lambda NN\leftrightarrow\Sigma NN$, and $\Sigma NN \rightarrow\Sigma NN$ coupled channels are translated into effective density-dependent hyperon-nucleon potentials, which additionally enter the coupled Bethe-Goldstone equations. This is done as in refs.[@Petschauer2016; @Petschauer2017; @Haidenbauer2017; @Holt2010] by integrating one of the two nucleons over the filled Fermi sea, as illustrated schematically in fig.\[fig:3BFeff\]:
![Density-dependent in-medium potentials generated from 3-body forces. Upper left: contact term; upper right: one-meson exchange term; lower part: two-meson exchange term. Loops with arrows indicate integration over the filled nucleon Fermi sea. []{data-label="fig:3BFeff"}](FBBeff)
$$\label{eq:red}
V_{12}^{\text{eff}}(\rho_p,\rho_n) = \sum_{N=p,n}\int\limits_{|\vec k|\leq k_F^{(N)}}\frac{\mathrm d^3k}{(2\pi)^3}\,{\mathrm tr}_3V_{123} \,,$$
where ${\mathrm tr}_3$ denotes the spin-trace over the third particle and the summation over nucleons (protons and neutrons) in the Fermi sea is performed. The resulting effective two-body potentials depend explicitly on the proton and neutron densities, $$\rho_p = {(k_F^{(p)})^3\over 3\pi^2}~,~~~\rho_n = {(k_F^{(n)})^3\over 3\pi^2}~.$$ As a first example, we present expressions for the $\Lambda N\to \Lambda N$ effective two-body potential in (asymmetric) nuclear matter with density $\rho = \rho_p + \rho_n$. Only the $\Lambda n $ effective potential in the nuclear medium needs to be given since the $\Lambda p$ potential follows by simply interchanging $\rho_n\leftrightarrow\rho_p$. The three parts have the following structure [@Petschauer2017]:\
Contact term: $$V^{(0)}_{\text{eff}}(\Lambda n) = {(H_1+3H_2)^2\over 18 \Delta}\,(\rho_n+2\rho_p)~,$$ with the decuplet-octet baryon mass difference $\Delta$, for which we take an average value of $\Delta = 270$ MeV.\
One-pion exchange part: $$V^{(1)}_{\text{eff}}(\Lambda n) =
{g_A C\over 9f^2 \Delta}(H_1+3H_2)\left(\rho_n+2\rho_p - {m_\pi^2\over 2\pi^2}\Gamma_0\right),$$ where $\Gamma_0(p;k_F^{(N)})$ is a function of the $\Lambda N$ center-of-mass momentum $p$, and the proton or neutron Fermi momentum $ k_F^{(N)}$, given explicitly in ref.[@Petschauer2017].\
Two-pion exchange part: $$V^{(2)}_{\text{eff}}(\Lambda n) = {g_A^2 C^2\over 6f^4 \Delta}\left[\rho_n+2\rho_p+ {\cal F}(p,q;k_F^{(n)},k_F^{(p)})\right]\,,$$ where the function ${\cal F}$ depends additionally on the momentum transfer $q$ in the reduced two-body system. It is a lengthy expression involving the spin-orbit operator $\mathrm i \vec \sigma_2\cdot(\vec q\times \vec p\,)$, that is explicitly given in eq.(46) of ref.[@Petschauer2017].
Next, we present some selected expressions for the effective $\Lambda N \rightarrow \Sigma N$ interaction. As representative examples, consider the direct terms, shown in fig.\[fig:3BFeff\], of the $\Lambda n\rightarrow \Sigma^- p$ effective transition potential. Exchange terms are also included in the actual calculations but are not displayed here for simplicity. In the case of symmetric nuclear matter, the corresponding lengthy formulas can be found in subsection III.D of ref.[@Petschauer2017].\
Contact term: $$\begin{aligned}
&& V^{(0)}_\text{{eff}}(\Lambda n\!\rightarrow\!\Sigma^- p) = {1\over 6\sqrt{6}\Delta}\bigg\{(\rho_p\!+\!\rho_n) \big[4H_2^2\!-\!(H_1\!+\!H_2)^2\big]\nonumber\\ &&+ \vec\sigma_1\!\cdot\!\vec\sigma_2\bigg[\bigg(H_1^2\!+\!H_2^2\!+\!{10\over 3}H_1 H_2\bigg)\rho_n+ (H_1\!+\!3H_2)^2{ \rho_p\over 3}\bigg]\bigg\}\,. \nonumber\\ &&\end{aligned}$$ One-pion exchange part: $$\begin{aligned}
&&V^{(1)}_\text{{eff}}(\Lambda n\!\rightarrow\!\Sigma^- p)={\sqrt{6}g_A C \over 27f^2\Delta}\, {\vec\sigma_1\cdot\vec q\,\vec\sigma_2\cdot\vec q\over m_\pi^2+\vec q^2}\nonumber\\ && \qquad\qquad \times \Big[(2H_1-3H_2)\rho_n-(2H_1+9H_2)\rho_p\Big]\,.\end{aligned}$$ Two-pion exchange part: $$\begin{aligned}
V^{(2)}_\text{{eff}}(\Lambda n\!\rightarrow\!\Sigma^- p)=-{16g_A D C^2\over 9\sqrt{6}f^4\Delta}{\vec\sigma_1\!\cdot\!\vec q\,\vec\sigma_2\!\cdot\!\vec q\over( m_\pi^2+\vec q^2)^2}(\rho_p+\rho_n) \vec q^2 , \nonumber\\ &&\end{aligned}$$ where the SU(3) axial vector coupling constant $D$ stems from the $\Lambda\Sigma\pi$ vertex. For the $\Lambda p\rightarrow \Sigma^+ n$ channel, corresponding expressions hold with $\rho_n\leftrightarrow\rho_p$ interchanged.
Finally, the ladder summations in the coupled hyperon-nucleon channels require also as input the $\Sigma N\leftrightarrow\Sigma N$ effective two-body potentials in different charge combinations. As representative examples we list here expressions for the in-medium $\Sigma^-n\rightarrow\Sigma^- n$ potential:\
Contact term: $$\begin{aligned}
&&V^{(0)}_\text{{eff}}(\Sigma^- n) =
{1\over 6\Delta}\,\Big\{\big[(H_1+H_2)^2\rho_p + 4H_1^2\rho_n\big]
\nonumber\\ && \qquad \qquad\qquad\qquad -{1\over 3}\,\vec\sigma_1\cdot\vec\sigma_2\,(H_1+H_2)^2\,\rho_p\Big\}\, .\end{aligned}$$ One-pion exchange part: $$\begin{aligned}
&&V^{(1)}_\text{{eff}}(\Sigma^- n)=
-{4g_A C\over 9f^2\Delta}\,{\vec\sigma_1\!\cdot\!\vec q\,\vec\sigma_2\!\cdot\!\vec q\over m_\pi^2+\vec q^2}\Big[H_1\rho_n \!+\!(H_1\!+\!H_2)\rho_p\Big]\,,\nonumber\\ &&\end{aligned}$$ Two-pion exchange part: $$V^{(2)}_\text{{eff}}(\Sigma^- n)=-{8g_A F C^2\over 9f^4\Delta}\,{\vec\sigma_1\cdot\vec q\,\vec\sigma_2\cdot\vec q\over( m_\pi^2+\vec q^2)^2}\,(\rho_n+\rho_p)\,\vec q^2\,,$$ where now the SU(3) axial vector coupling constant $F$ enters. The $\Sigma^+ p\rightarrow\Sigma^+ p$ effective potential is obtained by interchanging $\rho_n\leftrightarrow\rho_p$. The $\Sigma^0 n\rightarrow\Sigma^0 n$ and $\Sigma^0 p\rightarrow\Sigma^0 p$ effective potentials have similar expressions, except that for their two-pion exchange parts the direct term as shown in fig.\[fig:3BFeff\] vanishes and only an exchange diagram contributes in this sector.
The examples just discussed demonstrate the following. In the diagonal $\Lambda N\rightarrow\Lambda N$ potential the two low-energy constants, $H_1$ and $H_2$, those associated with the three-body contact term using decuplet dominance, appear in the combination $H_1 + 3H_2$. This leaves only one parameter in this channel, as it has been the case in previous calculations of $\Lambda$ single-particle potentials in nuclear and neutron matter performed in refs.[@Kohno2018; @Haidenbauer2017]. In contrast, the explicit inclusion of $\Lambda NN\leftrightarrow\Sigma NN$ coupled channels introduces $H_1$ and $H_2$ in various different combinations so that these parameters have to be constrained independently.
Brueckner-Hartree-Fock approach to hyperon single-particle potentials
=====================================================================
The two- and three-body potentials described in the previous section are used[^1] as input for computing a hyperon-nucleon $G$-matrix within Brueckner theory at first order in the hole-line expansion (the Brueckner-Hartree-Fock approximation [@Day1967]). We focus on the single-particle potentials, i.e. the self-energies of hyperons in nuclear matter. The relevant formalism is briefly summarized below. For more details we refer to refs.[@Schulze1998; @Rijken1999; @Vidana2000; @Kohno2000].
The Brueckner reaction matrix or G-matrix is determined by solving the coupled-channel Bethe-Goldstone equation (in symbolic form) $$G(\omega) = V + V \frac Q{e(\omega)+\mathrm i\epsilon} G(\omega)\,.$$ The energy denominator $e(\omega)$ depends on the starting energy $\omega$. The Pauli blocking operator $Q$ excludes particles in intermediate states from scattering into the filled Fermi sea. The potential $V$ is a matrix (labeled by the outgoing and ingoing two-baryon channels) including two-body and three-body contibutions, with the latter given in their density-dependent effective two-body form.
After angle-averaging (see appendix A for details) the Bethe-Goldstone equation decomposes into partial waves with total angular momentum $J$: $$\begin{aligned}
\label{eq:BGE}
&&G^{J}_{\alpha\beta}(p',p; P,\omega) = V^{J}_{\alpha\beta}(p',p) +\\
&&\sum_{\nu}\int_0^\infty\frac{\mathrm d k\, k^2}{(2\pi)^3} \, V^{J}_{\alpha\nu}(p',k)
\frac{\bar Q_{\nu}(P,k)}{\bar e_{\nu}(P,k;\omega)+\mathrm i \epsilon}G^{J}_{\nu\beta}(k,p;P,\omega) \,.\nonumber \end{aligned}$$ Here, $P$ is the total momentum of the two baryons. The indices $\alpha, \beta, \nu$ represent complete sets of channels, including partial wave quantum numbers and pairs of interacting baryons, $(B_1,B_2)$. In our case of interest, $B_1$ in the initial state is a $\Lambda$ hyperon, while $B_2$ is a nucleon within its Fermi sea. The channel coupling turns $B_1$ into a $\Sigma$ which subsequently interacts with the nuclear medium and turns back into a $\Lambda$ in the final state.
The $G$-matrix elements of eq.(\[eq:BGE\]) are calculated at the on-shell starting energy $$\begin{aligned}
\omega &=& E_{B_1}(p_1) + E_{B_2}(p_2)~,\nonumber\\
\mathrm{with}~~~~~~E_{B_i}(p) &=& M_i + {p^2\over 2M_i} + {\operatorname{Re}}U_{B_i}(p)~,\end{aligned}$$ where $M_i$ is the mass of baryon $B_i$. Hence, the energy denominator $e(\omega)$ (see appendix A) requires the single-particle potentials for both hyperons and nucleons as an input.
In Brueckner-Hartree-Fock approximation the single-particle potential for a baryon $B_1\in\{\Lambda,\Sigma^+,\Sigma^0,\Sigma^-,p,n\}$ interacting with the Fermi sea of nucleons $B_2\in\{p,n\}$ is then given by: $$\begin{aligned}
&U_{B_1}(p_1) =
\sum_{\alpha}\left[1+\delta_{B_1B_2}(-1)^{L+S}\right]\frac{(1+M_2/M_1)^3}{2}~\times\nonumber \\
& \sum_{J}(2J+1)
\int_{k_\mathrm{min}}^{k_\mathrm{max}}\! \frac{\mathrm dk\, k^2}{(2\pi)^3}\,W(p_1,k) \, G_{\alpha\alpha}^{J}(k,k;\bar P,\omega) \,.
{\addtocounter{equation}{1}\tag{\theequation}}\label{eq:U}\end{aligned}$$ The weight function $W(p_1,k)$ resulting from the angular averaging procedure is specified in appendix A, together with the integration boundaries $k_\mathrm{min}$ and $k_\mathrm{max}$. Evidently, the calculation of the single-particle potentials depends on the single-particle potentials themselves as they appear in the energy denominator. Therefore eqs. and must be solved self-consistently by iteration until convergence is reached. For the nucleon sector itself the $NN$ input potential is taken from ref.[@Entem2003], employing SU(2) chiral perturbation theory at fourth order (N3LO). Chiral three-nucleon forces at order N2LO are also included.
The computations are carried out using the so-called continuous choice, where single-particle potentials enter the energy denominator of intermediate states for all momenta, below and above the Fermi sea). From previous work [@Schulze1998] it is known that the continuous choice for intermediate states is preferable over the simpler gap choice (with potentials set to zero for momenta above the Fermi sea), because it allows for a reliable determination of the single-particle potentials including their imaginary parts. Some further technical details are given in Appendix B.
Results {#sec:res}
=======
Single-particle potentials of a $\Lambda$ hyperon in nuclear and neutron matter {#subsec:spp}
-------------------------------------------------------------------------------
The next step is now to calculate the single-particle potential, $U_\Lambda(p_1=0,\rho)$, for a $\Lambda$ at rest in symmetric nuclear matter as a function of baryon density $\rho=\rho_p+\rho_n$. Using as input the NLO hyperon-nucleon potentials together with the three-body interactions as specified in Section \[sec:int\], it is still necessary to constrain the two free parameters of the three-body contact terms. We recall that the calculation explicitly includes the $\Lambda NN \leftrightarrow\Sigma NN$ coupled channels, so the respective constants $H_1$ and $H_2$ enter independently.
A key constraint comes from hypernuclear phenomenology. Bulk properties and the shell structure of $\Lambda$ hypernuclei are well described by a $\Lambda$-nuclear single-particle potential parametrized in Wood-Saxon form, with an accurately determined potential depth at nuclear central densities [@Gal2016]: $$U_\Lambda(\rho \simeq \rho_0) = - 30~ \mathrm{MeV}~,
\label{eq:USNM}$$ corresponding to about half of the attractive strength of the nucleon single-particle potential at nuclear matter saturation density, $\rho_0 = 0.16$ fm$^{-3}$. This constraint will be satisfied by any of the calculations to be described below.
Fig.\[fig:ULambda1\] shows instructive examples of the calculated\
$U_\Lambda(p_1=0;\rho)$ in nuclear matter at low densities. Notably, the result using only the two-body chiral NLO potential would actually produce too much attraction in the $\Lambda$ potential, causing an overbinding of hypernuclei. This calls for repulsive effects beyond two-body interactions, with three-body forces being the natural extension. The three-body effects are moderate (a few MeV) at densities $\rho\simeq\rho_0$, but their relative importance grows continuously as the density increases. At $\rho \simeq 2\,\rho_0$ the repulsive $\Lambda NN$ 3BF is essential in order to turn the sign of $U_\Lambda$ from attractive to repulsive. The result in fig.\[fig:ULambda1\] is produced using a combination of three-body constants, $(H_1,H_2)$, that evidently satifies the constraint eq.(\[eq:USNM\]). However, this choice of $H_1$ and $H_2$ is not uniquely determined. It turns out that any pair $(H_1,H_2)$ that lies on either of the two solid curves in fig.\[fig:H1H2\] is compatible with the condition, $U_\Lambda(\rho=\rho_0) = - 30$MeV. On these two constrained curves one can identify points $(H_1,H_2)$ for which the $\Lambda$ single-particle potential at a higher density, $U_\Lambda(\rho=3\rho_0)$, becomes maximally repulsive (while avoiding instabilities of the numerical procedure): $(H_1,H_2) = (-2.2,0)$ and $(-2.47,1.2)$ in units of the squared inverse pion decay constant, $f^{-2} =4.8\,$fm$^2$. These are the parameters of choice for extrapolations to higher densities, as we are seeking mechanisms and conditions for avoiding a softening of the equation-of-state that would be in conflict with neutron star observations.
The computations of $U_\Lambda$ based on self-consistent solutions of the Bethe-Goldstone equation (\[eq:BGE\]) together with eq.(\[eq:U\]) involve a momentum cutoff not only in the input hyperon-nucleon potentials but also in the momentum-dependent single-particle potentials $U(k)$ that appear in energy denominators and are integrated over intermediate momenta $k$. We apply a cutoff by a factor $\exp[-(k/\lambda_{spp})^6]$ and choose $\lambda_{spp}$ between 600 and 650 MeV, close to the cutoff in the NLO hyperon-nucleon potential, while not altering the low-energy behavior. The ranges of these cutoffs, together with the two possible choices for the pair of constants ($H_1,H_2$), are reflected in the uncertainty estimates shown in the figures.
The computations of $U_\Lambda$ for both symmetric nuclear matter and pure neutron matter can be carried out safely (i.e. with well converging and numerically stable results) up to densities $\rho\simeq \rho_c=3.5\,\rho_0$. At even higher densities the occurence of numerical instabilities becomes prohibitive. One might argue that, in any case, performing calculations at $\rho\gtrsim 2\rho_0$ is not legitimate given the framework of chiral effective field theory with its limited range of applicability. On the other hand, the non-perturbative ladder summations of the Brueckner calculation lead beyond chiral perturbation theory, and the input interactions in the hyperon-nucleon sector are tested for total momenta $P$ up to and even above 800 MeV, whereas the neutron Fermi momentum even at central neutron star densities around $5\,\rho_0$ does not exceed 600 MeV, the typical cutoff scale in the calculations. We can therefore assume that at least qualitative extrapolations towards such high densities can be performed.
For extrapolations to densities beyond the technically (numerically) accessible $\rho_c \simeq 3.5\,\rho_0$, we argue as follows. Around $\rho_c$, one finds clear evidence in both symmetric nuclear matter and neutron matter that the density dependence of $U_\Lambda$ has a prominent quadratic behaviour, $U_\Lambda(\rho) \propto \rho^2$. This is expected as the three-body terms begin to dominate at high densities and grow as $\rho^2$. The two-body terms at high density increase linearly with $\rho$, and this expected subleading behaviour is also observed in the numerical results already at densities beyond $2\,\rho_0$. Therefore, assuming continuity an extrapolation of $U_\Lambda$ to high denstites is supposed to be well justified using the following ansatz for $\rho > \rho_0$: $$U_\Lambda(\rho) = u_0 + u_1\left({\rho\over\rho_0}-1\right) + u_2\left({\rho\over\rho_0}-1\right)^2\,.
\label{eq:extrap}$$ The parameters $u_1$ and $u_2$ are determined by fits to the calculated single-particle potentials at densities below $\rho_c$. With $u_0$ fixed at $\rho=\rho_0$, it turns out that the high-density behaviour is indeed governed by the quadratic term, with $u_2 \gg u_1$.
Results for the high-density extrapolation of $U_\Lambda$ in symmetric nuclear matter and neutron matter are displayed in fig.\[fig:ulambda\]. With two-body $YN$ interactions only, the linearly increasing $U_\Lambda$ at densities $\rho\gtrsim 2\rho_0$ is reminiscent of a Hartree potential. When increasingly repulsive three-body interactions are added, the leading $\rho^2$ dependence of the 3BF takes over at high densities. Note that in neutron matter, the strong repulsion in $U_\Lambda$ exceeds 100 MeV already at $\rho \simeq 3.5 \,\rho_0$ and reaches $U_\Lambda > 300$ MeV at $\rho\simeq 5\,\rho_0$, the density range characteristic of the inner core of neutron stars.
Hyperon and neutron chemical potentials in neutron stars
--------------------------------------------------------
If the chemical potentials of the $\Lambda$ hyperon and neutron fulfill the condition $\mu_\Lambda = \mu_n$ at some density in neutron star matter, it becomes energetically favourable to replace neutrons by $\Lambda$ hyperons via weak interactions. For a baryonic species $j$, the chemical potential $\mu_j$ is determined by the derivative of the energy density ${\cal E}$ with respect to the partial density $\rho_j$ of that species: $$\mu_j = {\partial{\cal E}\over\partial\rho_j}~.$$ Given a realistic neutron star equation-of-state, $P({\cal E})$, the Gibbs-Duhem relation for multicomponet sytems $$P({\cal E}) + {\cal E} = \rho_n\,\mu_n + \rho_p\,\mu_p + \rho_\Lambda\,\mu_\Lambda\,+ \dots~~.$$ determines the chemical potentials. In practice we use the microscopic EoS computed from a chiral nucleon-meson field theory in combination with functional renormalization group methods [@DW2017; @FW2019]. This EoS is consistent with all important nuclear physics constraints and neutron star observations. For simplicity we neglect the small (few percent) proton fraction from beta equilibrium which is part of the neutron star EoS, but its effect is marginal in the present context.
At the possible onset of $\Lambda$ hyperons, their kinetic energy vanishes as there is no $\Lambda$ Fermi sea to start with. Therefore the minimal $\Lambda$ chemical potential at that point is simply given by $$\mu_\Lambda(\rho) = M_\Lambda + U_\Lambda(\rho)~,
\label{eq:mulambda}$$
where $M_\Lambda$ is the $\Lambda$ hyperon mass and $\rho\simeq\rho_n$ is the neutron density. In the following the input potential $U_\Lambda(\rho)$ is the one presented in fig.\[fig:ulambda\] for neutron matter.
The comparison of $\mu_\Lambda$ and $\mu_n$ is shown in fig.\[fig:chempot\]. The uncertainty band of the neutron chemical potential is related primarily to the range of possible values of the nuclear symmetry energy, $E_\text{{sym}} = (32\pm 3)$ MeV. We note that this uncertainty band also includes $\mu_n$ as given in Ref.[@APR1998] for their maximally repulsive interaction (AV18+$\delta v$ + UIX\*).
Fig.\[fig:chempot\] points out that the combined repulsion from two- and three-body hyperon-nuclear interactions is indeed potentially strong enough to avoid the appearance of $\Lambda$ hyperons in neutron stars: one finds $\mu_\Lambda > \mu_n$ throughout the neutron star density range, with a set of three-body parameters that is consistent with basic constraints from hypernuclei. However, two-body $\Lambda N$ interactions alone would not be sufficient to suppress the occurrence of $\Lambda$ hyperons. Their onset would appear at densities around $3\,\rho_0$. The resulting softening of the EoS would be prohibitive.
The Brueckner calculations underlying eq.(\[eq:mulambda\]) are based on the $YN$ potential taken from ref.[@Haidenbauer2013] (referred to as NLO13) which features a strong $\Lambda N\leftrightarrow\Sigma N$ coupling. This channel coupling is a major source of the $\Lambda N$ attraction at intermediate distances. Its partial suppression by the Pauli principle acting on intermediate nucleon states in nuclear or neutron matter is responsible for the turnover from attraction to repulsion in the $\Lambda$-nuclear two-body force at a density well below $3\,\rho_0$ in nuclear matter. From fig.\[fig:ULambda1\] we also recall that the $\Lambda N$ attraction provided by NLO13 is slightly stronger than what is needed for the empirical hypernuclear shell-model potential. This sets the frame for constraining unknown pieces of the repulsive $YNN$ three-body force which in turn governs the behaviour of $\mu_\Lambda$ at high density. One might then ask whether alternative options exist which could imply different constraints and extrapolations for the 3BF sector. Such an alternative $YN$ potential (NLO19) has recently been studied [@Haidenbauer2019]. This potential is equivalent to NLO13 with respect to $\Lambda N$ and $\Sigma N$ scattering data, but it produces a stronger attractive $\Lambda$-nuclear potential than NLO13 at $\rho\simeq\rho_0$. This difference needs to be balanced by a more strongly repulsive three-body $\Lambda NN$ interaction. We have thus performed Brueckner calculations (for continuous choice) using NLO19, together with a 3BF readjusted to meet the hypernuclear $U_\Lambda(\rho=\rho_0) = -30$ MeV constraint. The effect at high density is to increase the $\Lambda$ chemical potential further in comparison with the NLO13 result plotted in fig.\[fig:chempot\], so as to widen the gap between $\mu_\Lambda$ and $\mu_n$ even more.
A further comment concerns $\Sigma$ hyperons in neutron star matter. At high density it is in principle possible to consider the condition $$\mu_{\Sigma^-} = \mu_n+\mu_e = 2\mu_n -\mu_p$$ for the occurrence of $\Sigma^-$ replacing an electron together with a neutron. However, this turns out not to be an option because the $\Sigma^-$ potential in neutron matter is strongly repulsive [@Petschauer2016b], so that $\mu_{\Sigma^-}$ remains separated by at least 50 MeV from $\mu_n + \mu_e$ at all relevant densities.
Summary and outlook {#sec:sum}
===================
Investigations of the properties of hyperons in nuclear and neutron matter have been extended to high baryon densities with the quest of exploring the possible occurrence of hyperons in neutron stars. Previous calculations based entirely on two-body hyperon-nucleon interactions suggested an onset for hyperons already at relatively low densities (between two and three times $\rho_0$). As a consequence the resulting equation-of-state is significantly softened and cannot support two-solar-mass neutron stars. The present work addresses the question whether hyperon-nucleon three-body forces (3BF) can provide the necessary repulsion to stop $\Lambda$ hyperons from replacing neutrons, maintaining the required stiffness of the neutron star EoS.
Calculations have been carried out within the self-consistent Brueckner-Hartree-Fock (BHF) approach using the continuous choice for intermediate spectra. The employed microscopic potentials for the hyperon-nucleon interaction are constructed from SU(3) chiral effective field theory at next-to-leading order. In addition, the pertinent N3LO nucleon-nucleon interaction is taken from chiral EFT.
Our main focus lies on the role of three-body forces, their reduction to density-dependent effective two-body interactions and the explicit treatment of $\Lambda NN\leftrightarrow\Sigma NN$ coupled channels in the BHF matrix equations. Unknown 3BF parameters have been constrained by reproducing the depth, $U_\Lambda(\rho=\rho_0) \simeq -30$ MeV, of the $\Lambda$ single-particle potential in nuclear matter as deduced from the phenomenology of hypernuclei.
The detailed discussion of $\Lambda$ and neutron chemical potentials in neutron matter then indicates that the $\Lambda NN$ three-body forces can indeed potentially develop sufficient repulsion to avoid the occurrence of $\Lambda$ hyperons in neutron stars, without the need to introduce strong extra assumptions in performing the extrapolation to high densities. A comparison using different versions of the chiral EFT-based $YN$ interactions (NLO13 versus NLO19) underlines this expectation further. While we regard these findings as a promising step towards a possible clarification of the so-called hyperon puzzle in neutron stars, we can nevertheless not yet claim that this problem has been “solved”. Detailed properties of the hyperon-nuclear three-body interactions must still be better constrained by data, e.g. in hypernuclear few-body systems. However, one can begin to consider the steadily increasing number of observational facts about neutron stars as part of a developing empirical data base that sets restrictive conditions for mechanisms involving strangeness in dense baryonic matter.\
\
\
\[app:angularaverage\]
Here, we present details of the angular averaging procedure that preceeds the derivation of eq.(\[eq:BGE\]). Total and relative momenta for two baryons $B_1$ and $B_2$ are introduced by $$\vec P = \vec p_1 + \vec p_2\,,\quad
\vec k = \frac{\xi_{12}\vec p_1-\vec p_2}{1+\xi_{12}}\,,\quad
\xi_{12} = \frac{M_{2}}{M_{1}}\,.$$
A standard approximation is applied to replace $Q/e$ by the ratio of angle-averaged quantities, $\bar Q/\bar e$. The Pauli operator, averaged over the angle $\theta$ between $\vec P$ and $\vec k$, is given by $$\begin{aligned}
\bar Q_{\nu}(P,k) & =
\frac12\int_{-1}^1\!\mathrm d\cos\theta\
\Theta\big(|\vec p_1| - k_F^{(1)}\big)\Theta\big(|\vec p_2| - k_F^{(2)}\big) \\
& = [0|\frac{[-1|z_1|1]+[-1|z_2|1]}2|1] \,,
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ with $$\begin{aligned}
z_1 & = \frac{1+\xi_{12}}{2kP}\bigg\{\left(\frac1{1+\xi_{12}}P\right)^2+k^2-(k_F^{(1)})^2\bigg\}\,, \\
z_2 & = \frac{1+1/\xi_{12}}{2kP} \bigg\{\left(\frac{\xi_{12}}{1+\xi_{12}}P\right)^2 + k^2-(k_F^{(2)})^2\bigg\} \,.
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ and the notation $[a|b|c]\equiv\operatorname{max}(a,\operatorname{min}(b,c))$ introduced in ref.[@Schulze1998]. The angle-averaged energy denominator is given by $$\begin{aligned}
\bar e_{\nu}(P,k;\omega) = \omega & - \frac{P^2}{2M_\nu}-\frac{k^2}{2\mu_\nu} -M_\nu \\
& - {\operatorname{Re}}U_{B_1}(\bar p_1)- {\operatorname{Re}}U_{B_2}(\bar p_2) \,,
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ with $M_{\nu}=M_1+M_2$ and $\mu_{\nu}=M_1 M_2/(M_1+M_2)$. The angle-average is done for the arguments of the single particle potentials $U_{B_i}$ of the intermediate baryons: $$\begin{aligned}
\bar p_1 &= \left(\tfrac1{(1+\xi_{12})^2} P^2 + k^2 + 2\tfrac1{1+\xi_{12}} P k\, \overline{\cos\theta}\right)^{1/2} \,, \\
\bar p_2 &= \left(\tfrac{\xi^2_{12}}{(1+\xi_{12})^2} P^2 + k^2 - 2\tfrac{\xi_{12}}{1+\xi_{12}} P k\, \overline{\cos\theta}\right)^{1/2} \,,
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ with $$\begin{aligned}
\overline{\cos\theta} & = \frac{ \int_{-1}^1\!\mathrm d\cos\theta\, \cos\theta\, Q(\vec P,\vec k) }{ \int_{-1}^1\!\mathrm d\cos\theta\, Q(\vec P,\vec k) } \\
& = \frac12\big([-1|z_2|1]-[-1|z_1|1]\big) \,.
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ where $Q(\vec P,\vec k)$ is the exact Pauli blocking operator.
It is common practice to introduce a further simplification, replacing the squared momenta $P^2=P^2(\vec p_1, \vec k)$ and $p^2_2=p^2_2(\vec p_1, \vec k)$ entering the Bethe-Goldstone equation by their angle averages: $$\begin{aligned}
\bar{P}^2(p_1,k) & = \frac{\int_{|\vec p_2|\leq k_F^{(2)}}\!\mathrm d\cos\vartheta\, P^2(p_1,k,\cos\vartheta)}{\int_{|\vec p_2|\leq k_F^{(2)}}\!\mathrm d\cos\vartheta} \\
&=
(1+\xi_{12})^2\left[p_1^2+k^2-p_1k(1+[-1|x_0|1])\right] \,, \\
\bar p_2^2(p_1,k)
&=
\frac{\xi_{12}}{1+\xi_{12}}\bar P^2(p_1,k) + (1+\xi_{12})k^2 -\xi_{12}\,p_1^2 \,,
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ where $\vartheta$ is the angle between $\vec p_1$ and $\vec k$, and one finds $$x_0 = \frac{\xi_{12}^2p_1^2+(1+\xi_{12})^2k^2-(k_F^{(2)})^2}{2\xi_{12}(1+\xi_{12})p_1k} \,.
{\addtocounter{equation}{1}\tag{\theequation}}$$ Note that baryon $B_2$ in the initial state is within its Fermi sea.
The weight function $W(p,k)$ that appears in the calculation of the single-particle potential $U_B(p)$ via eq.(\[eq:U\]) is given by $$\begin{aligned}
W(p,k) &= \frac1{4\pi}\int_{|\vec p_2|\leq k_F^{(2)}}\!\mathrm d\Omega_k = \frac12(1-[-1|x_0|1]) \,.
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$ The integration boundaries, $k_\mathrm{min}$ and $k_\mathrm{max}$, of the relative momentum are determined by the condition $W(k_1,k) = 0$, which leads to $$\begin{aligned}
k_\mathrm{min} &= \operatorname{max}\left( 0, \frac{-k_F^{(2)}+\xi_{12} p}{1+\xi_{12}} \right) \,,\
k_\mathrm{max} = \frac{k_F^{(2)}+\xi_{12} p}{1+\xi_{12}} \,.
{\addtocounter{equation}{1}\tag{\theequation}}\end{aligned}$$
\
Eqs. (\[eq:BGE\],\[eq:U\]) are solved numerically by alternately iterating both equations until the potentials $U_B$ converge. For certain kinematical conditions, e.g. involving large momenta in the initial state, the energy denominator in eq.(\[eq:BGE\]) vanishes, giving rise to a pole in the Bethe-Gold- stone equation. To make the integral numerically manageable, the principal value prescription is utilized: $$-\!\!\!\!\!\!\int_0^\infty \!dk \frac{N(k)}{D(k)} =
\int_0^\infty \!dk \left(\frac{N(k)}{D(k)}-\frac{N(k_0)}{D^\prime(k_0)} \frac{2k_0}{k^2-k_0^2}\right)\,,$$ for a simple pole at $k_0$, i.e. $D(k_0)=0$ and $D'(k_0)\neq0$. This prescription eventually meets its limits at densities beyond $3.5\,\rho_0$, once second-order poles begin to show up.
For densities $\rho > 2\rho_0$ in pure neutron matter, slow convergence of $U_B$ occurs because of increasingly strong input potentials. Convergence is improved by averaging $U_B(k)$ over subsequent iterations. High-density calculations occasionally require intervention by estimating an appropriate starting point for iterations from lower-density results. However, numerical stability is generally not guaranteed any more for $\rho > \rho_c \simeq 3.5\,\rho_0$. For densities $\rho > \rho_c$ we use the power series extrapolation in eq.(\[eq:extrap\]), fitted to numerically stable results at $\rho \lesssim \rho_c$ as described in the text.
As discussed in ref. [@Kohno2018], the cutoff $\lambda$ in the chiral input potential causes numerical oscillations of $U_B(k)$ for momenta $k > \lambda$. These oscillations tend to slow down the convergence of $U_B$ and are of no physical relevance. They are suppressed by the additional cutoff factor $\exp[-(k/\lambda_{spp})^6]$ attached to the single-particle potentials as described in Section \[subsec:spp\].\
[99]{}
P. Demorest, T. Pennucci, S. Ransom, M. Roberts, and J. Hessels, Nature [**467**]{}, 1081 (2010).
J. Antoniadis [*et al.*]{}, Science [**340**]{}, 6131 (2013).
E. Fonseca [*et al.*]{}, Astroph. J. [**832**]{}, 167 (2016).
H.T. Cromartie [*et al.*]{}, Nat. Astron. [**4**]{}, 72 (2019).
B.P. Abbott [*et al.*]{} (LIGO & Virgo Collaborations), Phys. Rev. Lett. [**119**]{}, 161101 (2017).
E.R. Most, L.R. Weih, L. Rezzolla, and J. Schaffner-Bielich, Phys. Rev. Lett. [**120**]{}, 261103 (2018).
S. De, D. Finstad, J.M. Lattimer, D.A. Brown, E. Berger, and C.M. Biwer, Phys. Rev Lett. [**121**]{}, 091102 (2018).
A. Akmal, V.R. Pandharipande, and D.G. Ravenhall, Phys. Rev. [**C58**]{}, 1804 (1998).
M. Drews and W. Weise, Phys. Rev. [**C91**]{}, 035802 (2015).
M. Drews, and W. Weise, Prog. Part. Nucl. Phys. [**93**]{}, 69 (2017).
B. Friman and W. Weise, Phys. Rev. [**C 100**]{}, 065807 (2019).
G. Baym, T. Hatsuda, T. Kojo, P.D. Powell, Y. Song, and T. Takatsuka, Rept. Prog. Phys. [**81**]{}, 056902 (2018).
G. Baym, S. Furusawa, T. Hatsuda, T. Kojo, and H. Tagashi, Astrophys. J. [**885**]{}, 42 (2019).
L. McLerran and S. Reddy, Phys. Rev. Lett. [**122**]{}, 122701 (2019).
Y. Fujimoto, K. Fukushima, and K. Murase, arXiv:1903.03400 \[nucl-th\].
H. Djapo, B.J. Schaefer, and J. Wambach, Phys. Rev. [**C 81**]{}, 035803 (2010).
T. Hell and W. Weise, Phys. Rev. [**C 90**]{}, 045801 (2014).
D. Lonardoni, A. Lovato, S. Gandolfi, and F. Pederiva, Phys. Rev. Lett. [**114**]{}, 092301 (2015).
P.F. Bedaque and A.W. Steiner, Phys. Rev. [**C 92**]{}, 025803 (2015).
J. Haidenbauer, S. Petschauer, N. Kaiser, U.-G. Mei[ß]{}ner, A. Nogga, and W. Weise, Nucl. Phys. [**A 915**]{}, 24 (2013).
S. Petschauer, N. Kaiser, J. Haidenbauer, U.-G. Mei[ß]{}ner, and W. Weise, Phys. Rev. [**C 93**]{}, 014001 (2016).
S. Petschauer, J. Haidenbauer, N. Kaiser, U.-G. Mei[ß]{}ner, and W. Weise, Nucl. Phys. [**A 957**]{}, 347 (2017).
M. Kohno, Phys. Rev. [**C 97**]{}, 035206 (2018).
S. Petschauer, J. Haidenbauer, N. Kaiser, U.-G. Mei[ß]{}ner, and W. Weise, Eur. Phys. J. [**A 52**]{}, 15 (2016).
J. Haidenbauer, U.-G. Mei[ß]{}ner, N. Kaiser, and W. Weise, Eur. Phys. J. [**A 53**]{}, 121 (2017).
J. Haidenbauer and U.-G. Mei[ß]{}ner, Phys. Rev. [**C 72**]{}, 044005 (2005).
T.A. Rijken, V.G.J. Stoks, and Y. Yamamoto, Phys. Rev. [**C 59**]{}, 21 (1999).
J.W. Holt, N. Kaiser, and W. Weise, Phys. Rev. [**C 81**]{}, 024002 (2010).
B. Day, Rev. Mod. Phys. [**39**]{}, 719 (1967).
H.-J. Schulze, M. Baldo, U. Lombardo, J. Cugnon, and A. Lejeune, Phys. Rev [**C 57**]{}, 704 (1998).
T.A. Rijken, V.G.J. Stoks, and Y. Yamamoto, Phys. Rev [**C 59**]{}, 21 (1999).
I. Vidana, A. Polls, A. Ramos, M. Horth-Jensen, and V.G.J. Stoks, Phys. Rev [**C 61**]{}, 025802 (2000).
M. Kohno, Y. Fujiwara, T. Fujita, C. Nakamoto, and Y. Suzuki, Nucl. Phys. [**185**]{}, 229 (2000).
D.R. Entem and R. Machleidt, Phys. Rev. [**C 68**]{}, 041001 (2001).
A. Gal, E.V. Hungerford, and D.J. Millener, Rev. Mod. Phys. [**88**]{}, 035004 (2016).
J. Haidenbauer, U.-G. Mei[ß]{}ner, and A. Nogga, arXiv:1906.11681 \[nucl-th\].
[^1]: We remind that due to direct and exchange contributions to the density-dependent effective $YN$ potentials, these have to be weighted with a statistical factor $1/2$ when added to the two-body potentials [@Kohno2018].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The paper is devoted to the analysis of Fermi acceleration of protons in GRBs and its neutrino signature. We have compared the consequences of Bohm scaling and those of a Kolmogorov scaling, the latter being more reliable. The predictions about the energy limitation of UHE-protons by the various losses and the neutrino emissions turn out to be very sensitive to these scalings. We consider Kolmogorov scaling as the most realistic and predict a reasonable pp-neutrino emission around 100 GeV at the end of the radiative stage of the fireball expansion, for a large number of GRBs pending on their baryonic load. A second p$\gamma$-neutrino emission is expected with the acceleration of protons in the radiation free stage, but with a synchrotron loss limitation immediately followed by a severe expansion loss limitation. According to the Kolmogorov scaling, the protons could not reach the UHE-range. Anyway the large possibility of a two component neutrino emission would be an interesting clue of cosmic ray physics.'
author:
- |
**by D. Gialis$^{1}$ and G. Pelletier$^{1,2}$\
[1 Laboratoire d’Astrophysique de Grenoble]{}\
[2 Institut Universitaire de France]{}**
bibliography:
- '/gagax1/ur1/dgialis/THESE/BIBFILES/biblio.bib'
title: 'Cosmic Rays and Neutrinos from GRBs: Predictions versus Acceleration Modeling'
---
Keywords: GRBs, Cosmic Rays, Neutrinos, Fermi processes
Introduction
============
Gamma Ray Bursts (hereafter GRBs) are unique high energy phenomena in astrophysics because of their possibility to manifest all the interesting “astroparticles” processes, such as generation of high energy gamma rays, ultra high energy cosmic rays, high energy neutrinos and gravitational waves (see @Dermer01 and @Meszaros02). The “fireball” model [@ReesMeszaros92] has been successful in explaining the afterglow stage, and its more elaborated form with the addition of internal shocks (@ReesMeszaros94, @PacXu94) during the “free” expansion has been successful in explaining the light curve and gamma spectra (@Fishman95, @Beloborodov00). Because of the highly relativistic dynamics necessarly involved in the GRB phenomenon, the generation of very high energy particles is expected through strong shocks and strong magnetic perturbations. Indeed GRB population could be a main source of UHE-cosmic rays which could be generated by the external shocks [@Vietri95] or by the internal shocks [@Waxman95], and a flux of neutrinos produced by the collisions of the UHE-cosmic rays with the GRB gamma photon is reasonably expected [@WaxBac99].
In this paper, our intend is to look at the sensitivity of the predictions of cosmic ray and neutrino generation to the description of the Fermi acceleration process together with opacity effects for protons. We have analysed the consequences of two assumptions, one is the so called “Bohm scaling” assumption, the other is what we called “Kolmogorov scaling”. The former consists in the statement that the Fermi acceleration time is proportional to the Larmor time of the accelerated particle, with a constant proportionality factor larger than unity (often chosen between 1 and 10 in the litterature). The latter consists in taking into account that the ratio between these two characteristic times depends on the rigidity of the particle through a law governed by the turbulence spectrum, as is confirmed by numerical works [@Casse01] , where no Bohm scaling has been found. In astrophysical media such as the solar wind, the interstellar medium, the turbulence spectrum is likely consistent with the Kolmogorov law.
The Bohm scaling is very convenient to make a first investigation of the high energy physics performances of the objects, and our prejudices are often grounded on this. However, in this paper, we will show that this scaling leads to unrealistic results. The analysis based on the Kolmogorov scaling is more reliable, and it turns out that it leads to very different conclusions, as we will present with details in the paper. The analysis is developed both in the radiation free stage and in the radiative stage. We especially emphasise the regime when the fireball is opaque to pp-collisions during the beginning of the internal shocks stage because of its baryonic load, which should occur for a large fraction of GRBs population. We make unusual predictions about the performance of cosmic ray generation, and about neutrino emission that could have two components, namely a non-thermal pp-neutrino emission and a p$\gamma$-neutrino emission.
The paper is organised as follows. In section 2, we briefly present the dynamical description we need for the estimation of the opacity effects, not only for photons but also for protons. In section 3, we analyse the consequence of the fast acceleration regime governed by the Bohm scaling. In section 4, we present the properties of the more progressive acceleration process governed by the Kolmogorov scaling and predict a reasonable non-thermal pp-neutrino emission with its spectrum. We investigate, more briefly in section 5, the radiation free stage, where UHE-cosmic rays are expected to be accelerated and p$\gamma$-neutrinos generated. We end the paper with a discussion that summaries our conclusions about cosmic ray and neutrino generations and the sensitivity of the predictions to the acceleration model.
Preliminary considerations
==========================
Dynamics of the fireball
------------------------
In this subsection, we summarize all the results we need for this paper that describe the expansion of the fireball [@Meszaros93]. The wind flow is considered to be a set of discrete shells which are successively emitted with an energy $E_{s}=E/N_{s}$, where $N_{s}$ is the total number of shells. The duration, $t_{w}$, of this wind flow provides with an interval of shell number, namely, $1 \leq N_{s} \leq
c\, t_{w}/r_{0}$ where $r_{0}$ is the size of the central object.\
At the very beginning of the expansion of a shell, the pressure is supposed to be dominated by the radiative pressure. The temperature, $T$, of the plasma which is mainly composed by electron-positron pairs is equal to the photon temperature. Considering a shell is initially spherical with a radius, $r_{0}$, and with an energy emitted in $\gamma$-rays equal to $E_{s}$, we have $$\begin{aligned}
\frac{E_{s}}{(4/3)\pi\, r_{0}^{3}}=a\, T^{4}\, ,\end{aligned}$$ where $a$ is the Stefan constant. Thus, the plasma temperature is $$\begin{aligned}
T=\left( \frac{3\,E_{s}}{4\pi\, a\, r_{0}^{3}} \right)^{1/4}\,.\end{aligned}$$
For an energy, $E$, in the neighborhood of $10^{51}$ ergs, a number of shells, $N_{s}$, of 20 and a radius, $r_{0}$, of $10^{7}$ cm, the temperature is below 10 MeV.\
When a shell starts, the energy $E_{s}$ is very upper to the baryon mass energy. We can define the ratio, $\eta$, between these two energies which is $$\begin{aligned}
\eta = \frac{E_{s}}{(M_{b}/N_{s})\, c^{2}} = \frac{E}{M_{b}\, c^{2}}
\gg 1\end{aligned}$$ where $M_{b}$ is the total baryonic mass ejected.\
In the observer frame, the shell thickness, $\Delta r$, is supposed to remain constant and equal to $r_{0}$ until the broadening radius $r_{b}$ (@Goodman86, @Meszaros93). Beyond this radius, defining the Lorentz factor, $\Gamma$, of the baryonic matter, the thickness becomes $$\begin{aligned}
\Delta r \simeq (r/c)\, \Delta v \simeq r/2\Gamma^{2}
\label{shell}\end{aligned}$$
In the same frame, we also define a radius, $r_{s}$, where the kinetic energy of baryonic matter reaches its saturation value. At this moment, the Lorentz factor $\Gamma=\Gamma_{max}$ is close to $\eta$.\
In the co-moving frame, according to the Lorentz transformation, the shell thickness is given by $$\begin{aligned}
\Delta R = \Gamma\, \Delta r\, .\end{aligned}$$
If we consider an adiabatic expansion of the shell, conservation of entropy in the co-moving frame for a radius lower to $r_{s}$ is given by $$\begin{aligned}
r^{2} \, \Gamma \, r_{0} \, T^{3} = constant \, .
\label{ent}\end{aligned}$$ At the same time, conservation of energy is such that $$\begin{aligned}
r^{2} \, \Gamma^{2} \, r_{0} \, T^{4} = constant\, .
\label{enr}\end{aligned}$$ From (\[ent\]) and (\[enr\]), we deduce two laws of evolution for $r\leq r_{s}$ which are $\Gamma(r) \propto r$ and $T(r)
\propto r^{-1}$. Then, the saturation radius $r_{s}$ can be defined by $$\begin{aligned}
r_{s}=\eta \, r_{0}\, .\end{aligned}$$ For $\eta$ of the order of 300, we obtain $r_{s}\simeq 3\times
10^{9}$ cm.\
Beyond the radius $r_{s}$, the equation (\[ent\]) is always valid but the Lorentz factor of the shell remains constant. In this case, the temperature is such that $T(r) \propto r^{-2/3}$. Considering $\Delta r \geq r_{0}$, the radius $r_{b}$, according to (\[shell\]), satisfies $r_{b}\geq \Gamma^{2}\, r_{0}$. At last, $\Gamma$ is supposed to reach its saturation value, $\eta$, around $r_{s}$ well before $r_{b}$. In fact, $\eta$ is the average value of the saturation bulk Lorentz factor, and we have to bear in mind that deviations from this average is expected to generate internal shocks. The radius $r_{b}$ is so defined by $$\begin{aligned}
r_{b}\simeq \eta^{2}\, r_{0}\, .
\label{rb}\end{aligned}$$ For $\eta$ of the order of 300, we obtain $r_{b}\simeq 9\times
10^{11}$ cm.\
To conclude, all the different parameters of a shell we need, can be summarized by the next following expressions :\
In the observer frame, the thickness of a shell is $$\begin{aligned}
\Delta r \simeq \left\{ \begin{array}{ccl}
& r_{0} & \hspace{0.8cm} {\rm for} \hspace{0.3cm} r\leq r_{b} \\
& r/\Gamma_{max}^{2} \simeq r/\eta^{2}& \hspace{0.8cm} {\rm for}
\hspace{0.3cm} r\geq r_{b}
\end{array} \right.\, .\end{aligned}$$ The Lorentz factor of a shell is such that $$\begin{aligned}
\Gamma(r) \simeq \left\{ \begin{array}{ccl}
& r/r_{0} & \hspace{0.8cm} {\rm for} \hspace{0.3cm} r\leq r_{s} \\
& \Gamma_{max}\simeq \eta & \hspace{0.8cm} {\rm for} \hspace{0.3cm}
r\geq r_{s}
\end{array} \right.\, .\\\end{aligned}$$
In the co-moving frame, the thickness of a shell becomes equal to $$\begin{aligned}
\Delta R \simeq \left\{ \begin{array}{ccl}
& \Gamma\, r_{0} \simeq r& \hspace{0.8cm} {\rm for}
\hspace{0.3cm}
r\leq r_{s} \\
& \Gamma_{max}\, r_{0} \simeq \eta \, r_{0}& \hspace{0.8cm}
{\rm for} \hspace{0.3cm} r_{s}\leq r\leq r_{b} \\
& r/\Gamma_{max} \simeq r/\eta& \hspace{0.8cm} {\rm for}
\hspace{0.3cm}
r\geq r_{b}
\end{array} \right.\, .
\label{DELTAR}\end{aligned}$$ The temperature of the pair electron-positron plasma is such that $$\begin{aligned}
T(r)\simeq \left\{ \begin{array}{ccl}
& 10 \times \left( \frac{T(r_{0})}{10\, MeV} \right) \left(
\frac{r_{0}}{r} \right) \, \, MeV & \hspace{0.8cm} {\rm for}
\hspace{0.3cm} r\leq r_{s} \\
& 10 \times \left( \frac{T(r_{0})}{10\, MeV} \right) \left(
\frac{r_{0}}{r_{s}} \right) \left( \frac{r_{s}}{r} \right)^{2/3}\, \,
MeV
& \hspace{0.8cm} {\rm for} \hspace{0.3cm} r\geq r_{s}
\end{array} \right.\, ,\end{aligned}$$ and the co-volume of a shell, $V_{c}(r)$, is $$\begin{aligned}
V_{c}(r) \simeq \left\{ \begin{array}{ccl}
& \Omega \, r^{3} & \hspace{0.8cm} {\rm for} \hspace{0.3cm}
r\leq
r_{s} \\
& \Omega\, \eta \, r_{0}\, r^{2}& \hspace{0.8cm} {\rm for}
\hspace{0.3cm} r_{s}\leq r\leq r_{b} \\
& (\Omega/\eta)\, r^{3} & \hspace{0.8cm} {\rm for} \hspace{0.3cm}
r\geq r_{b}
\end{array} \right.,
\label{COVOL}\end{aligned}$$ where $\Omega$ ($\simeq 4\pi/500$) is the opening angle of the emission.\
The collision date (if any…), $t_{c}$, between two shells of Lorentz factor $\gamma_{1}$ and $\gamma_{2}$, which started at two different times separated by $\Delta t_{0}$ is such that $t_{c} \simeq
2\frac{\gamma_{1}^{2}\gamma_{2}^{2}}{\vert
\gamma_{1}^{2}-\gamma_{2}^{2}\vert}
\Delta t_{0}$, which leads to collisions at various distances $r_{c}$ such that $r_{c}/r_{b} \sim \Delta t_{0}/t_{0}$ (see @Daigne98). For a long GRB having a duration $t_{w} \sim 10$ s, the maximum collision radius reaches the deceleration radius $r_{d}$. The light curve duration could be due to either the duration of the wind $t_{w}$ or to the time spread of the shell at the most remote collision, namely, $\Delta
t_{d} = t_{d}/\eta^{2}$. It turns out that this latter time is comparable to the duration of the wind of a long GRB, namely, $1-10$ s. The observed millisecond variations should come from internal collisions located at a few $r_{b}$ where $\Delta r_{b} \sim
r_{0}$.
Anyway after a few $r_{b}$, all the shells mix up and form a single jet. That is the reason why volume and thickness of a shell can really be defined only before the broadening radius. After this radius, in the equations (\[DELTAR\]) and (\[COVOL\]), they are undervalued.
Conversion of the GRB energy into cosmic rays {#subsec.}
---------------------------------------------
During the evolution, the energy of the fireball is shared between several forms : the thermal energy $E_{th}$, the magnetic energy $E_{m}$, the bulk kinetic energy, the cosmic ray energy $E_{\star}$ and the energy in the form of hydromagnetic perturbations $E_{m}^{\star}$. The energy share $E_{m}^{\star}$ is the reservoir for particle acceleration.\
Until the saturation radius $r_{s}$, the ratio of the pair thermal energy over the fireball energy, $(a\,T^4\,
V_{c})/E_{s}$, starts with a value close to unity and then decays as $1/r$. In this stage, the proton population is a tiny contribution both in number of particles and in energy. Assuming that the protons dominate the baryonic load, we get the total number of protons $$N_{p} \simeq \frac{M_{b}}{m_{p}} \simeq 0.67 \times
10^{51}
\left(\frac{\eta}{300}\right)^{-1}\left(\frac{E}{10^{51}
erg}\right)\,.
\label{eq:NP}$$ Thus, the density of protons by shell, $n_{p}(r)$, in the co-moving frame and before the broadening radius, is such that $$\begin{aligned}
n_{p}(r) \geq 10^{-4}\left(\frac{t_{0}}{1\,ms}\right)
\left(\frac{t_{w}}{10\,s}\right)^{-1}\frac{N_{p}}{V_{c}(r)}\,,\end{aligned}$$ and the maximum value of $n_{p}(r)$ is equal to $N_{p}/V_{c}(r)$.\
For $t_{0}=r_{0}/c=1$ ms, we can write $$\begin{aligned}
n_{p}(r) \geq \left\{ \begin{array}{ccl}
& 2.7\times 10^{27} \left(
\frac{t_{w}}{10\,s}\right)^{-1}\left(\frac{\eta}{300}\right)^{-1}
\left(\frac{\Omega/4\pi}{2\times
10^{-3}}\right)^{-1}\left(\frac{E}{10^{51}
erg}\right) \left(\frac{r}{r_{0}}\right)^{-3} cm^{-3} \\
& 1.0\times 10^{20} \left( \frac{t_{w}}{10\,s}\right)^{-1}
\left(\frac{\eta}{300}\right)^{-2} \left(\frac{\Omega/4\pi}{2\times
10^{-3}}\right)^{-1}\left(\frac{E}{10^{51}
erg}\right) \left(\frac{r}{r_{s}}\right)^{-2} cm^{-3}
\end{array} \right.\end{aligned}$$ respectively for $r\leq r_{s}$ and $r_{s}\leq r\leq r_{b}$.\
Considering $\eta=300$, even if the GRB has a long duration about 10 s, the density of protons, at $r=r_{s}$, is larger than $10^{20}$ cm$^{-3}$. We note that, up to the saturation radius, the density of a shell is comparable to the density of a solid.\
The corresponding ratio of the thermal energy of protons, $3
T/(2\eta\,
m_{p}c^{2})$, starts with a very low value ($\sim 10^{-4}$) and decays like $1/r$ as well. The cosmic ray component is supposed to develope out of the proton thermal component and would grow up to several percents of $E$ to account for a significant contribution of the UHE cosmic ray generation in the Universe. Assume that a fraction $\xi_{\star}$ of the proton population is injected into the cosmic ray component (i.e. $N_{\star} = \xi_{\star}N_{p}$) and that they reach a mean energy $\bar \epsilon(r) = \bar \gamma(r)\, m_{p}\,c^{2}$ in the co-moving frame. Then, the contribution to the energy of the fireball, in the observer frame, is given by $$E_{\star}(r) = \frac{\xi_{\star} \bar
\gamma(r)\,\Gamma(r)}{\eta}\, E
\label{ESTAR}$$ When the radius $r$ reaches the saturation radius $r_{s}$, the energy of the cosmic ray component is equal to $\xi_{\star} \bar
\gamma(r)\,E/(1+4\xi_{\star}\bar \gamma(r)/3)$. For a cosmic ray spectrum in $\epsilon^{-2}$, $\bar \gamma = \log
(\gamma_{max})$, which clearly shows that the goal of converting about $10$ percents of the fireball energy into cosmic rays energy is achieved when a sizeable fraction of the protons are injected in the cosmic ray population.
Regarding the magnetic energy of the fireball, two points of view can be considered; either the magnetic field behaves like in a jet and the poloidal component decreases like $1/r^{2}$ thus its pressure decreases like the relativistic thermal pressure in $1/r^{4}$; whereas the toroidal component decreases more slowly; or the magnetic field of the shell disconnects from the central source; which is very likely. In this latter case, the magnetic energy of the shell is conserved as long as the dissipation is negligible (e.g. $B \propto
V_{c}^{-1/2}$) and decays when the Fermi processes and/or reconnections become efficient in accelerating particles. Initially, the intensity of magnetic field is usually considered as of the order of the equipartition value, which means that the magnetic energy is a sizeable fraction of the fireball energy ($E_{m}(r_{0}) = \xi_{m} E$).
The importance of pp-collisions during the primeval stage
---------------------------------------------------------
We begin with defining the radius $r_{\star}$ where a shell becomes optically thin with respect to Compton scattering. It can easily be checked that a typical shell width $\Delta R$ becomes smaller than the flow transverse radius after a short while, when $r> \eta
\sqrt{\pi/4\Omega}\,r_{0}$ which is comparable to $r_{s}$. It will turn out that the photosphere is located at a much larger distance for large enough $\eta$ (see (\[RST\])) and therefore the opacity of a shell is determined by its width. The Compton opacity is $\tau_{\star} = \sigma_{T}\, n_{e}\,\Delta
R$ where $\sigma_{T}$ is the Thompson cross section. We assume $n_{e}\simeq n_{p} \simeq n_{b}$ and thus the co-moving baryon density is related to the wind mass flux $\dot M_{w}$ by $$n_{b} = \frac{\dot M_{w}}{\Gamma\, \Omega\, r^{2}\, m_{p}\,c}
\label{eq:NB}$$ Moreover for $r>r_{s}$, the GRB energy is converted into relativistic kinetic energy such that the kinetic luminosity (kinetic energy flux) of the wind $L_{w} = E/t_{w} = \eta \,\dot M_{w}\,c^{2}$. For $r>r_{s}$, $\Gamma \simeq \eta$ and the Compton opacity is thus $$\tau_{\star} = \frac{\sigma_{T}\,L_{w}}{\Omega
\,m_{p}\,c^{3}\,\eta^{2}}\,\frac{\Delta R}{r^{2}} \, .
\label{eq:TAUS}$$ We can define a critical value for $\eta$ such that the photospheric radius is located at $r_{b}$, where shock acceleration starts. This critical value $\eta_{\star}$ is given by $$\eta_{\star} \equiv \left(\frac{\sigma_{T}L_{w}}{\Omega
m_{p}c^{3}r_{0}}\right)^{1/5} \simeq 570 \times
\left(\frac{E}{10^{51}erg}\right)^{1/5}\left(\frac{t_{w}}{1s}\right)^{
-1/5}\left(\frac{\Omega/4\pi}{2\times 10^{-3}}\right)^{-1/5}
\label{eq:ETAS}$$ The photospheric radius is such that $$\begin{aligned}
r_{\star} = \left\{ \begin{array}{ccl}
r_{b} \,\left(\eta_{\star}/\eta\right)^{5} & \hspace{0.5cm} {\rm
for} \hspace{0.3cm} \eta \leq \eta_{\star} \\
r_{b} \,\left(\eta_{\star}/\eta\right)^{5/2} & \hspace{0.5cm} {\rm
for}
\hspace{0.3cm} \eta \geq \eta_{\star}
\end{array} \right.\, .
\label{RST}\end{aligned}$$ These simple formulae illustrate the requirement of a large value of the asymptotic bulk Lorentz factor $\eta$ in order to observe an optically thin X-ray spectrum. For $\eta$ larger than the critical value, the internal shocks generate only a non-thermal spectrum just after $r_{b}$, whereas, for $\eta < \eta_{\star}$, the internal shocks start accelerating particles in an optically thick plasma. Assuming that the GRB energy varies within two orders of magnitude, the wind time by two orders of magnitude and the solid angle by one order of magnitude, the possible values of $\eta_{\star}$ extend within an interval of one order of magnitude. However it is worth mentioning that this usual approach of the opacity issue is an underestimate of the opacity effects, because the coalescence of shells and ultimately their dissolution into a single jet increase the opacity.\
The possibility for the internal shock scenario to partially operate in an optically thick regime is important when one consider pp-collisions. Indeed, any relativistic protons can collide with other protons (or neutrons) producing pions and thus neutrinos. The cross section, $\sigma_{pp}$, for this production is constant and equal to $2.7\times 10^{-26}$ cm$^{2}$ when the kinetic energy of protons is larger than 1 GeV. The sheets become thin to pp-collisions always before the photosphere since $\tau_{pp} =
n_{p}\,\sigma_{pp}\,\Delta R = 1$ at $r_{pp}$ such that $r_{pp}
=r_{\star}\sqrt{\sigma_{pp}/\sigma_{T}}<r_{\star}$ for $r_{pp}<r_{b}$ and $r_{pp}
=r_{\star}{\sigma_{pp}/\sigma_{T}}<r_{\star}$ for $r_{pp}>r_{b}$. In this latter case, $r_{pp} \simeq 0.04\,r_{\star}$; which occurs when $\eta \leq \eta_{\star}/2$. In this paper, we intend to emphasize the importance of this energy limitation to proton acceleration and estimate the resulting neutrino emission.
Consequences of a fast acceleration regime
==========================================
In this section, we analyse the consequences of the usual assumption of a Bohm scaling for the Fermi process for both the energy distribution cut off and the depletion of the energy reservoir for acceleration.
Bohm scaling and energy losses
------------------------------
We first consider a Fermi acceleration process that is assumed to follow the so called Bohm’s scaling namely characterized by its time scale proportional to the Larmor time, i.e. : $$t_{acc} = \kappa_{0}\, t_{L} = \kappa_{0}\, \frac{10^{-4}}{Z}
\left(\frac{\epsilon}{1\,GeV}\right)\left(\frac{B}{1\,G}\right)^{-1}
\, s \ ,
\label{eq:TACC}$$ where we take $\kappa_{0} =10$, as often assumed in astroparticle physics.\
In the early phase, the cut off of the proton energy distribution can be caused by either the synchrotron loss or the pp-collisions; which is obtained by equating the acceleration time and the loss time. The synchrotron time $t_{syn}$ is given by $$t_{syn}^{-1} =
\frac{4Z^{4}\sigma_{T}\,\gamma}{3m\,c}\,\left(
\frac{m_{e}}{m} \right)^{2}\,W_{m} ,
\label{VM}$$ where $W_{m}$ is the density of magnetic energy.\
This expression leads to the cut off energy : $$\epsilon_{syn} \simeq 2.4\times 10^{11}
\frac{1}{\sqrt{\kappa_{0}}}
\left(\frac{B}{1\,G}\right)^{-1/2} \,GeV \ .
\label{VM}$$
We consider now the limitation caused by pp-collisions. The cut off energy $\epsilon_{pp}$ is such that $(n_{p}\,\sigma_{pp}\,c)
\,\kappa_{0}\,t_{L} =1$, and thus $$\epsilon_{pp} \simeq
\frac{10^4}{\kappa_{0}}
\left(\frac{B}{1\,G}\right)
\left(\frac{\eta^{6}}{\eta_{\star}^{5}}\right)
\left(\frac{r_{0}}{c}\right)
\, GeV \ .
\label{VM}$$ At the broadening radius ($r_{b}$), where the magnetic field could be as high as $10^{7}$ G, these two limitations are comparable : $$\epsilon_{syn} \simeq 2.4\times 10^{7}
\left(\frac{\kappa_{0}}{10}\right)^{-1/2}
\left(\frac{B}{10^{7}\,G}\right)^{-1/2}\, GeV \ ,
\label{VM}$$ and $$\epsilon_{pp} \simeq 3\times 10^{9}
\left(\frac{\kappa_{0}}{10}\right)^{-1}
\left(\frac{B}{10^{7}\,G}\right)
\left(\frac{\eta}{300}\right)
\left(\frac{\eta}{\eta_{*}}\right)^{5}\, GeV
\label{VM}$$
During the expansion the ratio between these two limitations evolves such that $$\frac{\epsilon_{pp}}{\epsilon_{syn}} \simeq 1.2\times 10^{3}\,
\left(\frac{\kappa_{0}}{10}\right)^{-1/2}
\left(\frac{\eta}{300}\right)
\left(\frac{\eta}{\eta_{*}}\right)^{5}
\left(\frac{B}{10^{7}G}\right)^{3/2}
\left(\frac{r}{r_{b}}\right)^{2} \;
\label{eq:*}$$ which suggests that for any decrease of the magnetic field in $r^{-\alpha}$ with $\alpha$ between 1 and 2, the most severe limitation would be due to synchrotron losses. However we will show in the next subsection that such efficient Fermi acceleration would dissipate the magnetic energy very rapidly.
The depletion of the energy reservoir for acceleration
------------------------------------------------------
Such an efficient acceleration produces a strong depletion of the energy reservoir for particle acceleration. The previous results would make sense only if the depletion of the energy reservoir for particle acceleration is slow. This is exactly the purpose of this subsection.\
Let $E_{m}^{\star}$ be the energy in the hydromagnetic perturbations (magnetic and kinetic energy) involved in the acceleration of particles. We assume that this is a sizeable fraction $\xi_{m}^{\star}$ of the total magnetic energy and that even the large scale magnetic field in the shell is mostly disconnected from the central source after a short while and thus is rather tangle. Therefore we assume that the total magnetic energy dissipates at the same rate than $E_{m}^{\star}$. A minimum dissipation is assumed by considering its energy loss by particle acceleration only. According to (\[eq:TACC\]), the acceleration power is $Q^{+} =
\frac{N_{\star}}{\kappa_{0}t'_{L0}}m_{p}c^{2}$, where $t'_{L0}$ is the Larmor time for a proton of 1 GeV in the co-moving frame. The minimum depletion is then governed by $\dot E_{m}^{\star} = -Q^{+}$, which reads : $$\dot E_{m} = - \frac{E}{t'_{\star}}\, \frac{\bar
B}{B_{\star}} \, ,
\label{eq:DEPL}$$ where we define a characteristic time $t'_{\star}$ at a given radius $\tilde r$ where the magnetic field has an intensity $B_{\star}$, such that $$t'_{\star} \equiv (\eta\kappa_{0}\,
\xi_{m}^{\star}/\xi_{\star})\, t'_{L0}(B_{\star}) \ .
\label{eq:TDEP}$$ Clearly, this time measures the rate of depletion at the considered radius $\tilde r$ and must be compared to the co-moving dynamical time $t'(\tilde r) = \tilde r/\Gamma c$. Let us make an estimate at $r_{b}$, the late stage actually; for $B = 10^{7}$ G and $\epsilon =1$ GeV, $t'_{\star} \simeq 3\times 10^{-8}$ s, whereas $t'(r_{b}) \simeq
10^{-2}$ s. This perturbation burning out within a few nanoseconds is not realistic at all. Before this stage, the depletion time is much shorter than the fireball dynamical time by a factor proportional to $(B\,r)^{-1}$ that goes at best like $r^{1/2}$ for $B
\propto r^{-3/2}$ corresponding to magnetic energy conservation and the perturbations are burnt out almost immediately, their energy being radiated by photons and neutrinos. The observation of these perturbations in the light curve of gamma emissions suggests a much slower depletion; which we will analyse in the next section.
The interest of a progressive acceleration regime
=================================================
The Bohm scaling is often used in astroparticle physics, which sometimes can provide some rough estimate of the high energy cut off. But the true scaling [@Casse01] depends on the turbulence spectrum. We will assume that the perturbations are distributed according to the Kolmogorov law. This will significantly change the performance of the Fermi process and the conclusions about cosmic ray generation.
Kolmogorov scaling and energy losses
------------------------------------
In the co-moving frame, the acceleration time depends on the speed $\beta_{\star}\,c$ of magnetic perturbations that scatter particles on both sides of the shocks. Considering the same average velocity, we have the relation $$\begin{aligned}
t_{acc}=\frac{t_{s}}{\beta_{\star}^{2}}\, ,\end{aligned}$$ where the scattering time, $t_{s}$, can be expressed like $$\begin{aligned}
t_{s}=(\eta_{b}\,\rho^{\beta-1}\,\omega_{L})^{-1}\end{aligned}$$ with $\eta_{b}=\frac{<\delta B^{2}>}{\bar{B}^{2}}$ and $\rho=\frac{r_{L}}{l_{c}}\leq 1$ where $\l_{c}$ is the mean coherence length of the magnetic perturbations whose spectrum is supposed to be a power law of index $\beta$. According to the Kolmogorov theory, one can take $\beta=5/3$ and, with $\kappa_{0}=1/(\beta_{\star}^{2}\,\eta_{b})$, we get $$\begin{aligned}
\kappa=\kappa_{0}\,\rho^{-2/3} \ .\end{aligned}$$
Considering $l_{c}\lesssim \Delta R$, $$\begin{aligned}
\rho=\frac{r_{L}}{\Delta R}\simeq\frac{\epsilon}{Z\,e\,B\,\Delta R\,c}\end{aligned}$$ and we have $$\begin{aligned}
\kappa\simeq 2.2\times 10^{-3}\,
\left(\frac{\kappa_{0}}{10}\right)\,
\left(\frac{\epsilon}{1\,GeV}\right)^{-2/3}
\left(\frac{B}{1\,G}\right)^{2/3}
\left(\frac{\Delta R}{1\,cm}\right)^{2/3}.\end{aligned}$$ The synchrotron limitation becomes $$\begin{aligned}
\epsilon_{syn}\simeq 1.2\times 10^{19}\,
\left(\frac{\kappa_{0}}{10}\right)^{-3/4}\,
\left(\frac{B}{1\,G}\right)^{-5/4}
\left(\frac{\Delta R}{1\,cm}\right)^{-1/2}\,GeV.
\label{esyn}\end{aligned}$$ For $r=r_{b}$ ($\simeq 10^{12}$ cm), $B=10^{7}$ G and $\Omega=4\pi/500$, $\epsilon_{syn}\simeq 3.8\times 10^{5}$ GeV.\
As for the pp-collision limitation, $$\begin{aligned}
\epsilon_{pp}\simeq
\left(\frac{\kappa_{0}}{10}\right)^{-3}
\left(\frac{\eta}{300}\right)^{3}
\left(\frac{\eta}{\eta_{\star}}\right)^{15}\left(\frac{B}{10^{7}\,G}\right)
\left(\frac{\Delta R}{10^{12}\,cm}\right)^{-2}
\left(\frac{r}{r_{b}}\right)^{6}\,GeV.\end{aligned}$$ For $r\geq r_{b}$, $\Delta R = r/\eta$, defining $r_{b\star} \equiv
\eta_{\star}^{2}r_{0}$, we obtain $$\epsilon_{pp}\simeq 2.0\times
10^{5}\left(\frac{\kappa_{0}}{10}\right)^{-3}
\left(\frac{B}{10^{7}\,G}\right)
\left(\frac{\eta_{\star}}{570}\right)
\left(\frac{\eta}{\eta_{\star}}\right)^{8}
\left(\frac{r}{r_{b\star}}\right)^{4} \, GeV.
\label{eq:EPPB}$$ This energy increases up to a maximum value reached at $r_{pp}$ (when it’s larger than $r_{b}$), namely, $$\epsilon_{pp}\simeq 3.5 \times
10^{2}\left(\frac{\kappa_{0}}{10}\right)^{-3}
\left(\frac{B}{10^{7}\,G}\right)
\left(\frac{\eta_{\star}}{570}\right)
\left(\frac{\eta_{\star}}{\eta}\right)^{4}
\, GeV.
\label{eq:EPPR}$$ As long as $r<r_{pp}$, the energy limitation is clearly due to pp-collisions rather than synchrotron losses. These results show, first, that it is not possible to accelerate protons beyond 100 GeV before the fireball becomes thin for protons at $r_{pp}$, for $\eta> \eta_{\star}$; second, that, for $\eta
<\eta_{\star}/2$, the proton energy could increase above 100 GeV. For instance, if $\eta = \eta_{\star}/3$, $r_{\star} \simeq
240\, r_{b}$, $r_{pp} \simeq 10\, r_{b}$, $B$ weakens by a factor $(r_{pp}/r_{b})^{-\alpha} \simeq 3.2 \times 10^{-2}$ with $\alpha = 3/2$ and thus compensates the increase due to $\eta$. Therefore the energy remains of order 100 GeV. We have to examine how this regime changes the rate of perturbation energy depletion.\
In the opposite situation where the protons are not accelerated beyond GeV in the opaque stage, for $\eta>\eta_{\star}$ and/or $B<10^{5}$ G at $r_{b}$, nucleosynthesis is possible as shown by Lemoine [@Lemoine02].
The depletion of the energy reservoir for acceleration
------------------------------------------------------
Taking into account the variation of $\kappa$ in the acceleration power, we define a new characteristic time $t'_{\star}$ which is quite different from the last one : $$t'_{\star} \equiv (\eta\kappa_{0}<\rho^{2/3}>^{-1}\,
\xi_{m}^{\star}/\xi_{\star})\, t'_{L0}(B_{\star}) \ .
\label{eq:TDEP}$$ For a power law distribution, not harder than $\epsilon^{-2}$, $<\rho^{2/3}> \sim \rho_{0}^{2/3}$, $\rho_{0}$ being the rigidity for $1$ GeV. For bell type distribution of standard deviation $\bar
\epsilon$, $<\rho^{2/3}> \sim <\rho(\bar \epsilon)^{2/3}>$. Compared to the Bohm regime, the Kolmogorov regime increases this time by a factor $\rho_{0}^{-2/3}$. At $r_{b}$ we get a more extended time of the order of $0.1$ s, which is much more realistic than in the Bohm scaling case. If $r_{pp}$ is significantly larger, the depletion behaviour is changed such that $t'_{\star}/t'$ decreases instead of increasing like in Bohm regime. Indeed in Kolmogorov regime, this ratio scales like $B^{-1/3}r^{-5/9}$, which decreases when $B$ decreases less slowly than $r^{-5/3}$.
For a more detailed estimate, we integrated the differential equation (\[eq:DEPL\]) which, after some algebra, leads to the following result : $$\frac{E_{m}}{E_{m}(r_{b})} =
\left[1-\frac{E}{E_{m}(r_{b})}
\frac{t'(r_{b})}{5t'_{\star}(r_{b})}
\left(1-\left(\frac{r}{r_{b}}\right)^{-5/6}\right)
\right]^{6} \ .
\label{DEM}$$ In fact, the perturbations are burnt out in a fraction $\delta t'$ of $t'_{\star}(r_{b})$, which can be obtained directly from the differential equation (\[eq:DEPL\]) : $$\delta t' = t'_{\star}(r_{b})\frac{E_{m}(r_{b})}{E} \ ,
\label{DET}$$ which confirms the previous statement of a reasonable depletion time.
The pp-neutrino emission {#subsec.NU}
------------------------
In this subsection, we give an evaluation of the number of emitted pp-neutrinos and we calculate their energy spectrum.\
The pp-collisions produce neutrinos after some reactions which are $$\begin{aligned}
p + p \longrightarrow & D + \pi^{+} \\
& p + p + a (\pi^{+} + \pi^{-}) + b \pi^{0} \\
& p + n + \pi^{+} + a (\pi^{+} + \pi^{-}) + b \pi^{0} \\
& 2n + 2\pi^{+} + a (\pi^{+} + \pi^{-}) + b \pi^{0}\end{aligned}$$ The production of mesons $\pi^{-}$ and $\pi^{+}$ gives neutrinos et $\mu$-mesons through the decay reactions $$\begin{aligned}
\pi^{-} \longrightarrow \mu^{-} + \overline\nu_{\mu} \nonumber\\
\pi^{+} \longrightarrow \mu^{+} + \nu_{\mu}
\label{Pi}\end{aligned}$$ and other neutrinos are produced after $$\begin{aligned}
\mu^{-} \longrightarrow e^{-} +\overline\nu_{e} + \nu_{\mu}
\nonumber\\
\mu^{+} \longrightarrow e^{+} + \nu_{e} + \overline\nu_{\mu}\,\,.
\label{Mu}\end{aligned}$$ Let’s write energy of different particles $$\begin{aligned}
m_{\pi^{\pm}} & \simeq & 140 \, MeV/c^{2} \\
m_{\mu} & \simeq & 105 \, MeV/c^{2} \,\, .\\
m_{e^{\pm}} & \simeq & 0.5 \, MeV/c^{2}\end{aligned}$$ Thus, neutrinos coming from (\[Pi\]) have a minimum energy equal to $$\begin{aligned}
\epsilon_{\nu_{\mu}} \simeq \frac{m_{\pi^{\pm}}^{2}-m_{\mu}^{2}}{2\,m_{\pi^{\pm}}}
\simeq 30\, MeV\end{aligned}$$ and decay of $\mu$-mesons (\[Mu\]) gives neutrinos having a minimum energy which varies from 25 to 50 MeV.\
Let $f_{\nu}(\epsilon_{\nu})$ be the energy distribution of neutrinos, normalised such that $\int
f_{\nu}(\epsilon_{\nu})\,d\epsilon_{\nu}=n_{\nu}$, the number density of neutrinos; and $f_{\star}(\gamma)$ the Lorentz factor distribution of relativistic protons, normalised such that $\int f_{\star}(\gamma)\,d\gamma=n_{\star}$, the number density of relativistic protons. The kinetic equation for neutrinos can be written in a simplified way as follows : $$\frac{\partial}{\partial t}f_{\nu}+c\,\vec n.\nabla f_{\nu}=
\xi_{\nu}\,\nu_{pp}\int f_{\star}(\gamma)\,
\delta(\epsilon_{\nu}-\gamma_{c}\epsilon_{0})\,d\gamma \ ,
\label{eq:EENU}$$ where $\epsilon_{0}$ is the average energy of the neutrinos generated by the pp-collision in the center of mass frame (above estimated), $\xi_{\nu}$ the average number of produced neutrinos at each collision and $\gamma_{c}$ the Lorentz factor of the collision frame. The energy of the emitted neutrinos remain close to the minimum values because the pions do not take all the energy available above the threshold, but just a little excess above it, the remaining energy being kept by the proton or the neutron. Therefore the number $\xi_{\nu}$ is just a few. The Lorentz factor $\gamma_{c} \sim \sqrt{\gamma}$, where $\gamma$ is the Lorentz factor of the relativistic proton. So we can easily deduce that, for $f_{\star} \propto \gamma^{-s}$, the energy flux of neutrinos is in $\epsilon_{\nu}^{-2s+2}$, between $\epsilon_{0}$ and $\epsilon_{0}\sqrt{\epsilon_{pp}/m_{p}c^{2}}$, in the co-moving frame. For $s=2$, a neutrino spectrum in $\epsilon_{\nu}^{-2}$ is expected between 5 GeV and 150 GeV, typically, for the observer frame seing the GRB flow coming front.
Assuming an isotropic neutrino emission in the co-moving frame, the power emitted is $$\dot E'_{\nu} = \Omega \,\xi_{\nu}\, \epsilon_{0} \int
\nu_{pp}\,n_{\star}\, r^{2}\,d(\Delta R) \ ,
\label{eq:ENU0}$$ where the integral over the width extends down to the lower bound at $r_{pp}$, corresponding to $\tau_{pp}=1$. Thus $E'_{\nu}= t'_{w}\,
\Omega\, \xi_{\nu}\, \epsilon_{0}\, c\,n_{\star}\, r_{pp}^{2}$. The quantity $t'_{w}\,\Omega\, c\,n_{\star}\,r_{pp}^{2}$ is the number of relativistic protons $\xi_{\star}N_{p}$ which have flown through the “proto”-sphere during the GRB event. Finally, we get a simple formula giving the amount of neutrino energy in the co-moving frame : $$E'_{\nu} =
\xi_{\star}\,\xi_{\nu}\,\frac{\epsilon_{0}}{m_{p}c^{2}}\,
\frac{E}{\eta} \ .
\label{eq:ENUP}$$ For an observer seing the GRB shells coming front, the neutrino energy emission is multiplied by the bulk Lorentz factor, thus $E_{\nu} = \xi_{\star}\xi_{\nu}(\epsilon_{0}/m_{p}c^{2})E$ which reasonably leads to $E_{\nu} \sim 10^{-3}-10^{-2}E$. The number of emitted neutrinos by a GRB is simply $\xi_{\nu}N_{\star}$. A neutrino telescope of collecting surface $A$ can receive $N_{\nu}A/4\pi D^{2}$ from a GRB exploding at a distance $D$. For example, a GRB occuring at 100 Mpc can provide with $10^{5}$ neutrinos crossing a $km^{2}$-detector. However, the number of events recorded by the detector is obtained by multiplying this number by the detection probability which is quite low.\
In the same process, $\pi^{0}$-decay occurs and generates gamma photons. An excess of these photons could be seen in the gamma spectrum; they superimpose a spectrum in $\epsilon_{\gamma}^{-2s+2}$ on the synchrotron spectrum due to the electrons. Because we focused on proton acceleration and neutrino emission, we disregarded the interesting issue of the neutron component which can decouple under some conditions [@Derishev99].
The stage of UHE Cosmic Rays acceleration
=========================================
Because of their magnetic field intensity and the size of their wind, the GRBs are considered as possible accelerators of UHE-cosmic rays (@Vietri95, @Waxman95) : for $r>r_{b}$, particles could reach the maximal energy beyond which they are no longer confined, $\epsilon_{max}=Z\,e\,B\,r$, namely $$\begin{aligned}
\epsilon_{max}=7.7\times 10^{21}\left(\frac{B}{10^{7}\,G}\right)\left(\frac{r}{10^{12}\,cm}\right)\,eV,\end{aligned}$$ for a proton ($Z=1$).
Proton energy in the transparent stage
--------------------------------------
We examine the loss limitation suffered by the protons after crossing the “proto”-sphere. After the equation (\[esyn\]), we have already given the value of the maximum energy due to the synchrotron loss at $r_{b}$; thus, it scales like : $$\epsilon_{syn} \simeq 3.8\times 10^{5}
\left(\frac{B(r_{b})}{10^{7}G}\right)^{-5/4}
\left(\frac{r}{r_{b}}\right)^{(5\alpha -2)/4}\, GeV \ .
\label{eq:ESYNP}$$ It will turn out that this value is above the threshold of the p$\gamma$-process, but is far below the range expected for getting UHE-Cosmic Rays. For $\alpha = 3/2$, this range would be reached at a few 100 $r_{b}$.\
The expansion loss can be significant. Indeed, the cut off energy is obtained by setting $t_{acc} = t'=t/\Gamma$ in the co-moving frame, and the limitation, $\epsilon_{exp}$, is such that $$\epsilon_{exp} \simeq 10^{5}\left(\frac{\kappa_{0}}{10}\right)^{-3}\left(\frac{\eta}{300}\right)
\left(\frac{B(r_{b})}{10^{7}G}\right)
\left(\frac{r}{r_{b}}\right)^{1-\alpha}\, GeV \ .
\label{eq:EEXP}$$ Whereas the synchrotron loss diminishes with distance, the expansion limitation becomes more and more severe (see figure \[fig\]). Moreover, it can easily be checked that the escape time due to transverse diffusion is always much longer than the expansion time.\
The p$\gamma$-neutrino emission
-------------------------------
A signature of VHE-protons acceleration is expected with the neutrino production resulting from the photo-production of pions, which is efficient to produce neutrinos through the $\Delta-resonance$ (the so-called GZK effect) : $$\begin{aligned}
& & \stackrel{(2/3)}{\rightarrow} p + \pi^0
\rightarrow p
+
\gamma + \gamma \nonumber \\
p + \gamma & \rightarrow \Delta^+ & \label{GZK} \\
& & \stackrel{(1/3)}{\rightarrow} n + \pi^+
\rightarrow
...
\rightarrow p + e^+ + e^- + \nu_e + \bar \nu_e +
\nu_{\mu} +
\bar \nu_{\mu} \nonumber\end{aligned}$$ The threshold of the process is such that $\epsilon_{p}\,\epsilon_{\gamma}\geq \frac{1}{2}m_{n}m_{\pi}
-\frac{1}{4}m_{\pi}^{2}$, which leads to a threshold for the proton Lorentz factor $\gamma_{th} \sim 10^{4}$, the energy of the photons being in the keV range in the co-moving frame. As previously seen, $\gamma_{max}$ is at least ten times larger than $\gamma_{th}$. The GRBs are likely opaque to this process until the shells reach the radius $r_{p\gamma}$ such that $$r_{p\gamma} = \left(\frac{\sigma_{p\gamma}E_{\gamma}}{\Omega\,\bar
\epsilon_{\gamma}}\right)^{1/2} \ ,
\label{eq:RPG}$$ where $\bar \epsilon_{\gamma}$ is the average energy of the target photons in the observer frame (typically 1 MeV); which puts this p$\gamma$-“proto”-sphere at a rather large radius of $1.5 \times
10^{15}$ cm for $E_{\gamma}=0.1\, E$.\
In the collision frame, the energy of a pion, generated sufficiently above the threshold, is $\epsilon'_{\pi} \simeq \epsilon'_{p}
\simeq \epsilon'_{\gamma} \simeq \gamma \epsilon_{\gamma}(1-\cos
\theta)$, where $\gamma$ is the Lorentz factor of the proton. The emitted neutrinos have an energy which is a fraction $\alpha_{0}$ ($\simeq 5\%$) of the pion energy. Similarly to the statistical treatment of the pp-collisions, we write a simplified (with delta approximation instead of a function smoothed by angle averaging) kinetic equation for the isotropic distribution of the neutrinos generated by p$\gamma$-collisions in the co-moving frame, which is : $$\frac{\partial}{\partial t}f_{\nu}+c\,\vec n.\nabla f_{\nu}=
\xi_{\nu}\, c\,\sigma_{p\gamma}\int_{\gamma>\gamma_{th}}
f_{\star}(\gamma)\, d\gamma \int
f_{\gamma}(\epsilon_{\gamma}) \,\delta(\epsilon_{\nu}-
\alpha_{0}\gamma^{2} \epsilon_{\gamma})\, d\epsilon_{\gamma}\ .
\label{KENUG}$$ Integrating both sides of the equation over the co-volume and the proper time, we easily get the number of emitted neutrinos : $$N_{\nu} =\xi_{\nu}N_{\star}^{>}=
\xi_{\nu}N_{\star}\left(\frac{1}{\gamma_{th}}-
\frac{1}{\gamma_{max}}\right) \ ,
\label{eq:NUG}$$ where $N_{\star}^{>}$ is the number of cosmic rays above the p$\gamma$-threshold energy, the last result being obtained for a $\gamma^{-2}$ proton distribution. The integration, like in the case of pp-collisions, starts at the p$\gamma$-“proto”-sphere. The total energy radiated by neutrinos in the co-moving flow is also derived easily from the kinetic equation : $$E'_{\nu} \sim
\alpha_{0}\,\xi_{\nu}N_{*}\log\left(\frac{\gamma_{max}}{\gamma_{th}}\right)\,\bar \epsilon'_{\gamma} \ .
\label{eq:EPNU}$$ The logarithm factor is obtained for a $\gamma^{-2}$ distribution of protons and can easily be modified. This result can also be rewritten in the observer frame as follows : $$E_{\nu} \sim \alpha_{0} \,\gamma_{th}\,\bar \epsilon_{\gamma}
\log\left(\frac{\gamma_{max}}{\gamma_{th}}\right) N_{\nu} \ .
\label{eq:ENU}$$ The energy radiated in the form of p$\gamma$-neutrinos is $E_{\nu} \sim 10^{-7}E$. Numerical computation of such spectra has been done (see @Mucke98). The kinetic equation also provides with the neutrino spectrum. Assuming $f_{\star}$ is a power law distribution and that the energy distribution of target photons is $\epsilon_{\gamma}\, f_{\gamma}
\propto \epsilon_{\gamma}^{-\alpha}$, we get a power law energy spectrum for the neutrinos, namely $$\epsilon_{\nu}^{2}\,\frac{dN_{\nu}(>\epsilon_{\nu})}{d\epsilon_{\nu}} \propto
\epsilon_{\nu}^{-\frac{s-1}{2}} \ ,$$ in an energy range depending on the energy range of the protons since $\epsilon_{\nu} \simeq \alpha_{0}\, \gamma^{2}
\epsilon_{\gamma}$. These spectra are given in the co-moving frame. For the observer, they are Doppler beamed with the bulk Lorentz factor of the relativistic wind. However, the number of events is so low that it is still “virtual” to talk about a spectrum...
Discussion
==========
In order to account for the non-thermal and highly variable gamma emission of GRBs, the fireball model and the internal shock model have been designed with baryonic load parameter $\eta$ that has been supposed large enough to get a relativistic wind achieving a large Lorentz factor $\Gamma \sim \eta$. However, it turns out that opacity effects could easily be significant at the beginning of the emission. We stress that point through a discussion involving a critical value $\eta_{\star}$ of the baryon load parameter. Indeed one emphasis of this paper is to analyse the opacity of the GRB to relativistic protons with respect to pion production by pp-collisions. Actually, the opacity condition relative to pp-collision is not far from the Compton opacity condition. Therefore it is reasonable to think that a significant fraction of the GRBs experience proton Fermi acceleration with efficient pp-collisions revealed by neutrino radiation.
During a stage of pp-opacity, we have shown that Fermi acceleration leads to conclusions about the pp-process that are very sensitive to the choice of the efficiency law of the acceleration. We have shown that the Bohm scaling assumption leads to an efficient proton acceleration that would be limited by synchroton loss, whereas the correct law governed by the turbulence spectrum (Kolmogorov law was used) leads to very different estimates and the pp-collisions process turns out to be the main limitation of the proton energy. Moreover, the excessive efficiency of the Bohm scaling would make the acceleration to deplete its energy reservoir in a time too short to maintain the gamma emission. The most important surprise raised in estimating the energy limitation in the radiation free stage. Indeed, the expansion limitation turns out to be drastic with the Kolmogorov law and maintains the proton energy below a few $10^{5}$ GeV.
The analysis of the paper indicates that a double neutrino emission can be expected with many GRBs, namely a stage of pp-neutrino emission followed by a stage of p$\gamma$-neutrinos. The number of emitted pp-neutrinos gives the amount of relativistic protons; and the number of p$\gamma$-neutrinos gives the number of protons above the threshold, which is about $10^{-4}$ less. We proposed an analytical shape of the neutrinos spectra, as well as the photon spectrum generated by the $\pi^{0}$-decay.\
The magnetic field intensity is an important parameter that controls the proton acceleration and the synchrotron losses. The neutrino emissions significatively depend on its value at the crucial distance $r_{b}$. We took a high but still reasonable value of $10^{7}G$ at this distance; if we take less, the synchrotron limitation is less important, but the expansion limitation, which controls the highest energy of the protons in the GRB, becomes more severe, and also the pp-neutrino emission becomes less energetic. If we unduly take more, synchrotron losses dominate over pp-collisions and the acceleration is more efficient against the expansion losses, however not sufficient to get UHE-cosmic rays…This paper does not exclude the possibility of UHE-cosmic ray generation in GRBs. It simply states that its achievement with Bohm scaling is not reliable and leads to observational inconsistencies and that its achievement with Kolmogorov scaling is impossible…We think that there is another possibility (@Pelletier99, @PelletierK00) that deserves a detailed investigation that we will present in a forthcoming paper.\
Acknowledgement : The authors are grateful to Frédéric Daigne, Gilles Henri, Martin Lemoine and Eli Waxman for fruitful discussions.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We discuss the problem of deciding when a metrisable topological group $G$ has a canonically defined local Lipschitz geometry. This naturally leads to the concept of [*minimal*]{} metrics on $G$, that we characterise intrinsically in terms of a linear growth condition on powers of group elements.
Combining this with work on the large scale geometry of topological groups, we also identify the class of metrisable groups admitting a canonical global Lipschitz geometry.
In turn, minimal metrics connect with Hilbert’s fifth problem for completely metrisable groups and we show, assuming that the set of squares is sufficiently rich, that every element of some identity neighbourhood belongs to a $1$-parameter subgroup.
address: |
Department of Mathematics, Statistics, and Computer Science (M/C 249)\
University of Illinois at Chicago\
851 S. Morgan St.\
Chicago, IL 60607-7045\
USA
author:
- Christian Rosendal
title: Lipschitz structure and minimal metrics on topological groups
---
[^1]
The present note deals with the problem of deciding which metrisable topological groups have a well-defined local geometry intrinsic to the topological group structure. To make this problem more precise, let us recall that a [*metrisable topological group*]{} is a topological group $G$ whose topology may be induced by some metric, which then is said to be [*compatible*]{} with the topology on $G$. Thus, the metric itself is not part of the given data. These groups where characterised in fundamental papers by G. Birkhoff [@birkhoff] and S. Kakutani [@kakutani], namely, a Hausdorff topological group $G$ is metrisable if and only if it is first countable. Moreover, such a group necessarily admits a compatible [*left-invariant*]{} metric $d$, i.e., so that $d(hg,hf)=d(g,f)$ for all $g,f,h\in G$.
An easy calculation shows that, if $d$ and $\partial$ are compatible left-invariant metrics on a topological group $G$, then the identity map ${\rm id}\colon (G,\partial){\rightarrow}(G,d)$ is always uniformly continuous and hence, by symmetry, a uniform homeomorphism. This is of course also a reflection of the fact that both $d$ and $\partial$ will metrise the left-uniform structure on $G$. However, unless further assumptions are added, there is in general no control on the modulus of uniform continuity of the mapping. The problem is thus to decide which, if any, of the compatible left-invariant metrics on $G$ determine a canonical local geometric structure on the group. At least up to local bi-Lipschitz equivalence, this is solved if $G$ admits a [*minimal*]{} metric in the following sense.
\[defi min\] A metric $d$ on a topological group $G$ is said to be [*minimal*]{} if it compatible, left-invariant and, for every other compatible left-invariant metric $\partial$ on $G$, the map $${\rm id}\colon (G,\partial){\rightarrow}(G,d)$$ is Lipschitz in a neighbourhood of the identity, i.e., if there is an identity neighbourhood $U$ and a constant $K$ so that $$d(g,f)\leqslant K\cdot \partial(g,f)$$ for all $g,f\in U$.
Let us first observe that, if $U$ and $K$ are as above, then ${\rm id}\colon (G,\partial){\rightarrow}(G,d)$ is locally $K$-Lipschitz. For given $h\in G$ and $v,w\in U$, note that $$d(hv,hw)=d(v,w)\leqslant K\cdot \partial(v,w)=\partial(hv,hw),$$ so the identity map is $K$-Lipschitz on the neighbourhood $hU$ of $h$. It follows immediately that any two minimal metrics on $G$ are locally bi-Lipschitz and thus identify a canonical local geometric or, more specifically, Lipschitz structure on $G$.
The concept of an intrisic Lipschitz structure on a topological object is of course common to other areas. For example, a well-known result due to D. Sullivan [@sullivan] states that, except for $n=4$, any topological $n$-manifold $M$ admits a Lipschitz structure, that is, an atlas $\{\phi_i\colon U_i\to {\mathbb R}^n\}$ whose transition maps are locally Lipschitz. Moreover, any two such Lipschitz structures are related by a locally bi-Lipschitz homeomorphism of $M$. The local Lipschitz structure identified by a minimal metric is even more rigid, since any two minimal metrics are locally bi-Lipschitz by the identity map.
We remark that, unless we accept to force the metric $d$ to be bounded, the local minimality of Definition \[defi min\] really describes the strongest notion of minimality possible. Indeed, if $d$ is unbounded, then $\sqrt d$ is a compatible left-invariant metric, while ${\rm id}\colon (G,\sqrt d){\rightarrow}(G,d)$ is not Lipschitz for large distances.
Note also that, at least for short distances, there is no maximal metric unless $G$ is discrete. That is, if $G$ is non-discrete and $d$ is any compatible left-invariant metric, then the mapping ${\rm id}\colon (G,d){\rightarrow}(G,\sqrt d)$ will not be Lipschitz for short distances.
Another helpful observation is that, if $d$ is minimal and $U\ni1 $ is a fixed $d$-bounded open neighbourhood, then, for every compatible left-invariant $\partial$, the map $${\rm id}\colon (U,\partial){\rightarrow}(U,d)$$ is Lipschitz, i.e., the localising set $U$ is independent of $\partial$. It follows immediately that any two bounded minimal metrics are bi-Lipschitz equivalent. Nevertheless, in many cases, a better global Lipschitz structure can be identified that captures the large scale geometry of the group. We return to this in Theorem \[min-max\].
One way to think about minimal metrics is via the growth or rather decay of the balls $ B_d(\alpha)=\{g\in G{ \; \big| \;}d(g,1)\leqslant \alpha\}$ as $\alpha{\rightarrow}0$. Namely, minimality of $d$ simply expresses that, if $\partial$ is another compatible left-invariant metric on $G$, then, for some $K=K(\partial)\geqslant 1$, we have $$B_\partial(\alpha)\subseteq B_d({K\alpha})$$ whenever $\alpha\leqslant 1$. So, up to the rescaling by $K$, the $d$-balls $B_d(\beta)$ are as large as possible as $\beta$ decreases to $0$.
Whereas minimality of a metric is a relative notion, i.e., defined in terms of comparisons with other compatible left-invariant metrics on the group, the main result of our note furnishes an internal characterisation of minimality without reference to other metrics. Namely, we characterise the minimal metrics as those satisfying a certain linear growth condition on powers in a neighbourhood of $1$. This condition in turn has already been studied in the literature in the context of locally compact groups, where it turned out to be central to the solution to Hilbert’s fifth problem. We shall discuss this connection after our result.
\[main\] The following conditions are equivalent for a compatible left-invariant metric $d$ on a topological group $G$.
1. $d$ is minimal,
2. \[strongest cond\] there is an open set $U\ni1 $ so that $$g, g^2,g^3, \ldots, g^n\in U\;\;{\Rightarrow}\;\; d(g,1)\leqslant \frac 1{n},$$
3. \[gleason cond\] there are constants ${\epsilon}>0$ and $K\geqslant 1$ so that $$d(g,1)\leqslant \frac{\epsilon}n\;\;{\Rightarrow}\;\; n\cdot d(g,1) \leqslant K\cdot d(g^n,1),$$
4. there are an open set $U\ni 1$ and a constant $K\geqslant 1$ so that $$g,g^2,g^4, g^8,\ldots, g^{2^n}\in U\;\;{\Rightarrow}\;\; 2^n\cdot d(g,1) \leqslant K\cdot d(g^{2^n},1).$$
The above result may be said to provide a satisfying description of minimal metrics on the group, indeed, the criterion only involves computations with powers of single elements. On the other hand, we have no informative reformulation of which metrisable groups admit minimal metrics. One would like to know if there is such a description that does not directly involve asking for an object as complicated as a minimal metric itself. More precisely, the following problem remains open.
Let $G$ be a universal Polish group, e.g., $G={\rm Homeo}([0,1]^{\mathbb N})$. Is the collection $$\{H\leqslant G{ \; \big| \;}H \text{ is a closed subgroup admitting a minimal metric}\}$$ Borel in the standard Borel space of closed subgroups of $G$?
Before commencing the proof of Theorem \[main\], we recall some procedures for constructing compatible left-invariant metrics on a topological group. The main result in this area is the above mentioned theorem independently due to Birkhoff and Kakutani. Of the two proofs, Birkhoff’s is the simplest and relies on a memorable little trick.
\[birkhoff\] Let $G$ be a topological group and $\{V_{3^n}\}_{n\in{\mathbb Z}}$ a neighbourhood basis at the identity consisting of symmetric open sets so that $G=\bigcup_{n\in {\mathbb Z}}V_{3^n}$ and $\big(V_{3^n}\big)^3\subseteq V_{3^{n+1}}$. Define $\delta(g,f)=\inf\big(2^n{ \; \big| \;}f{^{-1}}g\in V_{3^n}\big)$ and put $$d(g,f)=\inf \Big(\sum_{i=0}^{k-1}\delta(h_i,h_{i+1}){ \; \big| \;}h_0=g,
h_k=f\Big).$$ Then $\delta(g,f)\leqslant 2\cdot d(g,f)\leqslant 2\cdot\delta(g,f)$ and $d$ is a compatible left-invariant metric on $G$.
However, the metric $d$ produced by Birkhoff’s construction decreases exponentially faster than needed for our purposes due to a factor $\big(\frac 32)^n$. For this, we shall instead rely on the construction of Kakutani from which a better estimate can be extracted (see [@diestel] for a proof of the exact statement of Lemma \[kakutani\] below).
\[kakutani\] Let $G$ be a topological group and $\big\{V_{2^{-n}}\big\}_{n\in {\mathbb N}}$ a neighbourhood basis at the identity consisting of symmetric open sets satisfying $\big(V_{2^{-n}}\big)^2\subseteq V_{2^{-n+1}}$. Then there is a compatible left-invariant metric $d$ on $G$ so that $$B_d({2^{-n}})\subseteq V_{2^{-n}}\subseteq B_d({4\cdot 2^{-n}})$$ for all $n\in {\mathbb N}$.
We now turn to the proof of Theorem \[main\].
(1)${\Rightarrow}$(2): We claim first that there is an open neighbourhood $U\ni 1$ so that, for all $k\geqslant 1$ and $g\in G$, $$g, g^2, g^3, \ldots, g^{2^k}\in U\;\;{\Rightarrow}\;\; g\in B_d(2^{-k}).$$
In order to see this, we assume the contrary. Let $V_{2^{-0}}=G$ and, for $m\geqslant 1$, inductively define symmetric open sets $V_{2^{-m}}\ni 1$ as follows.
Assume that $V_{2^{-m}}$ is the last term that has been defined thus far and let $n\geqslant m$ be large enough so that $B_d(n{^{-1}})\subseteq V_{2^{-m}}$. Since the claim fails for $U=B_d(n{^{-1}})$, there are $k\geqslant 1$ and $g\notin B_d(2^{-k})$ so that $$g, g^2, g^3, \ldots, g^{2^k}\in B_d(n{^{-1}}).$$ Let $F=\{1,g, g{^{-1}}\}$ and note that, since $B_d(n{^{-1}})$ is symmetric, $F^{2^k}\subseteq B_d(n{^{-1}})$. Therefore, as $F$ is finite, we can pick a sufficiently small symmetric open $W\ni1$ so that also $(WFW)^{2^k}\subseteq B_d(n{^{-1}})\subseteq V_{2^{-m}}$. We then set $$V_{2^{-m-k}}=WFW,\quad V_{2^{-m-k+1}}=(WFW)^2,\quad \ldots\quad , V_{2^{-m-1}}=(WFW)^{2^{k-1}}.$$ At the next stage, we begin with the term $V_{2^{-m-k}}$ and proceed as above.
Therefore, at the end of the construction, we have a sequence $G=V_{2^{-0}}\supseteq V_{2^{-1}}\supseteq \ldots$ of symmetric open sets forming a neighbourhood basis at $1$ so that $\big(V_{2^{-m}}\big)^2\subseteq V_{2^{-m+1}}$ for all $m\geqslant 0$. We now apply Lemma \[kakutani\] to the sequence $(V_{2^{-m}})_{m\geqslant 0}$ to obtain a compatible left-invariant metric $\partial$ satisfying $$B_\partial({2^{-m}})\subseteq V_{2^{-m}}\subseteq B_\partial(4\cdot2^{-m}).$$
Note now that there are infinitely many $m$ so that some stage in the construction began with the term $V_{2^{-m}}$. So fix such an $m$ and let $k$ and $g$ be as in the construction step. Then $g\in WFW= V_{2^{-m-k}}\subseteq B_\partial(4\cdot2^{-m-k})$ and $g\notin B_d(2^{-k})$, whence $$2^{m-2}\cdot \partial(g,1)\leqslant 2^{-k}\leqslant d(g,1).$$ Therefore, $\partial$ is a compatible left-invariant metric on $G$, but ${\rm id}\colon (G,\partial){\rightarrow}(G,d)$ is not Lipschitz for short distances, contradicting the minimality of $d$ and thus proving the claim.
So, using the claim, fix $U\ni1 $ open so that, for all $g\in G$ and $k\geqslant 1$, $$g, g^2, g^3, \ldots, g^{2^k}\in U\;\;{\Rightarrow}\;\; g\in B_d(2^{-k})$$ and pick some open $V\ni1$ so that $V^2\subseteq U$. Now suppose $g, g^2, g^3, \ldots, g^m\in V$ for some $m$ and let $k\geqslant 0$ be so that $2^k\leqslant m<2^{k+1}$. Then also $g^{2^k+n}=g^{2^k}g^n\in V^2\subseteq U$ for all $n\leqslant 2^k$ and so $g^i\in U$ for all $i\leqslant 2^{k+1}$. In particular, $g\in B_d(2^{-k-1})\subseteq B_d(m{^{-1}})$. In other words, $$g, g^2, g^3, \ldots, g^{m}\in V\;\;{\Rightarrow}\;\; d(g,1)\leqslant \frac 1m,$$ which proves (2).
(2)${\Rightarrow}$(3): Assume that $U\ni1 $ is an in (2). By shrinking $U$ if needed, we may assume that $U=B_d(p{^{-1}})$ for some integer $p\geqslant 1$. Now, choose a symmetric open $W\ni 1$ so that $W^{2p}\subseteq U$. We claim that $$g, g^2, g^3, \ldots, g^n\in W\;\;{\Rightarrow}\;\; n\cdot d(g^n,1)\leqslant 4p\cdot d(g^n,1),$$ which thus verifies (3).
Indeed, suppose $g, g^2, g^3, \ldots, g^n\in W$ and $g\neq 1$. Then also $g, g^2, g^3, \ldots, g^{2pn}\in W^{2p}\subseteq U$ and thus, if $m$ is minimal so that $g^{m+1}\notin U$, we have $m\geqslant 2pn$ and $d(g, 1)\leqslant m{^{-1}}$. Let now $k\geqslant 1$ be such that $kn\leqslant m<m+1\leqslant(k+1)n$. Then $$\begin{split}
d\big(g^{(k+1)n},1\big)
&\geqslant d\big(g^{m+1},1\big)-d\big(g^{(k+1)n}, g^{m+1}\big)\\
&= d\big(g^{m+1},1\big)-d\big(g^{(k+1)n-(m+1)},1\big)\\
&\geqslant p{^{-1}}-\big[(k+1)n-(m+1)\big]\cdot d(g,1)\\
&\geqslant p{^{-1}}-\frac nm,
\end{split}$$ whereby $$\begin{split}
d(g^n,1)
\geqslant \frac{d\big(g^{(k+1)n},1\big)}{k+1}
\geqslant \frac{\frac1p-\frac nm}{k+1}
\geqslant \frac{\frac1p-\frac n{2pn}}{2k}
= \frac1{4p}\cdot \frac 1k\geqslant \frac1{4p}\cdot \frac nm\geqslant n\frac1{4p}\cdot d(g,1)
\end{split}$$ as claimed.
(3)${\Rightarrow}$(4): Let ${\epsilon}>0$ and $K\geqslant 1$ be as in (3) and set $U=B_d(\frac {{\epsilon}}{2K})$. Now, suppose $g\neq 1$ and that $g,g^2,\ldots ,g^n\in U$ for some $n\geqslant 1$. Let $m\geqslant 1$ be so that $g\in B_d(\frac {\epsilon}m)\setminus B_d(\frac {\epsilon}{m+1})$. Then, by (3), we have $$K\cdot d(g^m,1)\geqslant m\cdot d(g,1)>{\epsilon}\frac m{m+1}\geqslant \frac {{\epsilon}}2,$$ i.e., $g^m\notin U$, whereby $m>n$. It thus follows that $d(g,1)\leqslant \frac {\epsilon}n$ and hence, by (3) again, that $n\cdot d(g,1)\leqslant K\cdot d(g^n,1)$. In other words, $$g,g^2,\ldots ,g^n\in U\;\;{\Rightarrow}\;\; n\cdot d(g,1)\leqslant K\cdot d(g^n,1).$$
Now, by shrinking $U$ and increasing $K$, we may suppose that $U=B_d(2^{-k}\big)$ and $K=2^k$ for some $k\geqslant 3$. Set also $V=B_d(2^{-2k})$, whereby $V^{2^k}\subseteq U$. We claim that, for all $m$ and $g$, $$\big(
{\forall}i\leqslant 2^m\colon g^i\in V
\big)
\;\;\&\; \;
\big(
{\forall}i\leqslant m+2k+1\colon g^{2^i}\in V
\big)
\;\;{\Rightarrow}\;\;
\big(
{\forall}i\leqslant 2^{m+1}\colon g^i\in V
\big).$$ Indeed, suppose that $\big({\forall}i\leqslant 2^m\colon g^i\in V\big)$ and $\big({\forall}i\leqslant m+2k+1\colon g^{2^i}\in V\big)$. Assume also that $2^m<i<2^{m+2k+1}$ is given. Write $i=2^{p_r}+2^{p_{r-1}}+\ldots+2^{p_1}+j$ for some $m\leqslant p_1<p_2<\ldots<p_r<m+2k+1$ and $j<2^m$ and note that, since $r\leqslant 2k+1\leqslant 2^k-1$, we have $$g^i=g^{2^{p_r}}\cdots g^{2^{p_1}}\cdot g^j\in V^rV\subseteq V^{2^k}\subseteq U.$$ Therefore, as also $g^{2^{m+2k+1}}\in V$, we see that $g^i\in U$ for all $i\leqslant 2^{{m+2k+1}}$. By the hypothesis on $U$, it follows that $$2^{{m+2k+1}}\cdot d(g,1)\leqslant 2^k\cdot d(g^{2^{m+2k+1}},1)\leqslant 2^k\cdot 2^{-k}=1,$$ i.e., $d(g,1)\leqslant 2^{-({m+2k+1})}$. So, using that $d(g^i,1)\leqslant i\cdot d(g,1)$, this shows that $g^i\in V$ for all $i\leqslant 2^{m+1}$, proving the claim.
Put now $W=B_d(2^{-4k})$ and assume that $n$ and $g$ are given so that $$g, g^2, g^4, g^8, \ldots, g^{2^n}\in W.$$ Then, since $W^{2^{2k}}\subseteq V$, we have that $g^{2^i}\in V$ for all $i\leqslant n+2k$. Using our claim to induct on $m=1,2, \ldots, n-1$, we see that $g^i\in V\subseteq U$ for all $i\leqslant 2^n$, and thus $2^n\cdot d(g,1)\leqslant 2^k\cdot d(g^{2^n},1)$, which proves (4).
(4)${\Rightarrow}$(1): Suppose that $\partial$ is another compatible left-invariant metric on $G$. Fix ${\epsilon},\eta>0$ and $K\geqslant 1$ so that $$g,g^2,g^4, g^8, \ldots, g^{2^n}\in B_d({\epsilon})\;\;{\Rightarrow}\;\; 2^n\cdot d(g,1) \leqslant K\cdot d(g^{2^n},1)$$ and $B_\partial(\eta)\subseteq B_d({\epsilon})$. Note then that, if $n\geqslant 0$ and $g\in B_\partial({\frac \eta {2^n}})$, we also have $g,g^2,g^4, g^8, \ldots, g^{2^n}\in B_\partial(\eta)\subseteq B_d({\epsilon})$, whence $$2^n\cdot d(g,1) \leqslant K\cdot d(g^{2^n},1)\leqslant K{\epsilon}$$ i.e., $g\in B_d({\frac{K{\epsilon}}{2^n}})$. In other words, $B_\partial({\frac \eta {2^n}})\subseteq B_d({\frac{K{\epsilon}}{2^n}})$ for all $n\geqslant 0$.
Now, if $g\in B_\partial(\eta)$ is any non-identity element, pick $n\geqslant 0$ so that $g\in B_\partial({\frac \eta {2^n}})\setminus B_\partial({\frac \eta {2^{n+1}}})\subseteq B_d({\frac{K{\epsilon}}{2^n}})$. Then, $$d(g,1)
\leqslant{\frac{K{\epsilon}}{2^n}}
=\frac{2K{\epsilon}}{\eta}\cdot \frac \eta {2^{n+1}}
\leqslant \frac{2K{\epsilon}}{\eta}\cdot \partial(g,1),$$ showing that the map ${\rm id}\colon (B_\partial(\eta),\partial){\rightarrow}(B_\partial(\eta),d)$ is $\frac{2K{\epsilon}}{\eta}$-Lipschitz.
Since a compatible left-invariant metric on a topological group $H$ need not extend to a compatible left-invariant metric on a supergroup $G$, it is far from clear from the definition of minimality that the restriction of a minimal metric on $G$ to a subgroup $H$ is also minimal on $H$. However, using instead the reformulations of Theorem \[main\], this becomes obvious, whence the following corollary.
The class of topological groups admitting minimal metrics is closed under passing to subgroups.
Using the spectral theorem, it is not hard to verify that the metric induced by the operator norm is minimal on the unitary group ${\mathcal}U({\mathcal}H)$ of separable infinite-dimensional Hilbert space. Indeed, set $$U=\{u\in {\mathcal}U({\mathcal}H){ \; \big| \;}{\lVertu-{{{\rm Id}}}\rVert}<1\}$$ and suppose $u, u^2, u^3, \ldots, u^n\in U$ for some fixed $u\in {\mathcal}U({\mathcal}H)$. Then, by the spectral theorem, there is a $\sigma$-finite measure space $(X,\mu)$ and a measurable function $\phi\colon X{\rightarrow}{\mathbb T}$ so that $u$ is unitarily equivalent to the multiplication operator $M_\phi$ on $L^2(X,\mu)$ defined by $$\big(M_\phi\xi\big)(x)=\phi(x)\xi(x).$$ It follows that, for $i=1,\ldots, n$, $${\rm ess}\sup\{|{\phi^i}(x)-1|{ \; \big| \;}x\in X\}= {\lVertM_{\phi^i-1}\rVert}_\infty={\lVertM_\phi^i-{{\rm Id}}\rVert}_\infty={\lVertu^i-{{\rm Id}}\rVert}<1$$ and hence that $|\phi(x)-1|<\frac 2n$ for almost all $x\in X$. In other words, ${\lVertu-{{\rm Id}}\rVert}<\frac 2n$. So $$u,u^2, \ldots, u^n\in U{\Rightarrow}{\lVertu-{{\rm Id}}\rVert}<\frac 2n,$$ showing that the metric is minimal.
Condition (\[gleason cond\]) of Theorem \[main\] has been studied earlier in the literature as part of the solution to Hilbert’s fifth problem due to A. Gleason, D. Montgomery, H. Yamabe and L. Zippin. Indeed, in the book [@tao] by T. Tao, metrics satisfying this condition are termed [*weak Gleason*]{} as they underlie A. Gleason’s results in [@gleason]. In particular, in [@tao] it is shown that a locally compact metrisable group is a Lie group if and only if it has a weak Gleason metric. Moreover, in the locally compact setting, every weak Gleason metric is actually [*Gleason*]{}, meaning that it satisfies a further estimate on commutators (cf. Theorem 1.5.5 [@tao]).
Let us also mention that, if $d$ is a minimal metric on $G$, then there is a constant $K$ and an open set $V\ni 1$ so that $$f,g,h\in V\;\;{\Rightarrow}\;\;d(fh,gh)\leqslant K\cdot d(f,g),$$ i.e., right-multiplication is $K$-Lipschitz in a neighbourhood of $1$. To see this, let $B_d({\epsilon})$ be so that $d(g,1)\leqslant 1/n$ whenever $g, g^2, \ldots, g^n\in B_d({\epsilon})$. Then, given distinct $g,f\in B_d(\frac{\epsilon}4)$, let $n\geqslant 1$ be so that $g{^{-1}}f\in B_d\big(\frac {\epsilon}{n2}\big)\setminus B_d\big(\frac {\epsilon}{(n+1)2}\big)$. So, if $h\in B_d(\frac{\epsilon}4)$, we have $(h{^{-1}}g{^{-1}}fh)^i=h{^{-1}}(g{^{-1}}f)^ih\in B_d({\epsilon})$ for $i=1,\ldots, n$, whereby $$d(fh,gh)=d(h{^{-1}}g{^{-1}}fh,1)\leqslant \frac 1n\leqslant \frac 2{n+1}\leqslant \frac 4{\epsilon}\cdot d(f,g).$$
By Condition (\[strongest cond\]) of Theorem \[main\], it is easy to see that, if $G$ is a group with a weak Gleason metric, then $G$ is [*NSS*]{}, i.e., has [*no small subgroups*]{}, which simply means that there is a neighbourhood $U\ni 1$ not containing any non-trivial subgroup. Moreover, in the locally compact metrisable case, being NSS is equivalent to being a Lie group and thus also to having a weak Gleason metric (see the exposition in [@montgomery] or [@tao]). By Theorem \[main\], weak Gleason and minimal metrics coincide, but we shall prefer the latter more descriptive terminology.
Whereas a minimal metric establishes a canonical local Lipschitz geometry on a topological group, we will now combine this with the analysis from [@coarse] of the corresponding problem at the large scale.
A compatible left-invariant metric $d$ on a topological group $G$ is said to be [*maximal*]{} if, for every other compatible left-invariant metric $\partial$ on $G$, the map $${\rm id}\colon (G, d)\to (G,\partial)$$ is Lipschitz for large distances, that is, $\partial \leqslant K\cdot d+ C$ for some constants $K,C$.
Clear, any two maximal metrics $d$ and $\partial$ on a topological group $G$ are [*quasi-isometric*]{}, that is $\frac 1K\cdot d-C\leqslant \partial \leqslant K\cdot d+ C$ for some constants $K,C$.
Suppose that $d$ and $\partial$ are both simultaneously minimal and maximal metrics on a topological group $G$. Then $d$ and $\partial$ are bi-Lipschitz equivalent, i.e., $$\frac 1L\cdot d\leqslant \partial \leqslant L\cdot d$$ for some constant $L$.
Since $d$ is minimal, there is an identity neighbourhood $V$ and a constant $K$ so that $${\rm id}\colon (V,\partial)\to (V,d)$$ is $K$-Lipschitz. On the other hand, as $\partial$ is maximal, there are constants $M,N$ so that $d\leqslant M\cdot \partial +N$. It thus follows that, for $x\in G$ and $v\in V$, $$d(xv,x)=d(v,1)\leqslant K\cdot \partial (v,1)=K\cdot \partial (v,1)=K\cdot \partial (xv,x),$$ while, for $a\notin V$, $$\begin{split}
d(xa,x)
&\leqslant M\cdot \partial(xa,x)+N\\
&\leqslant M\cdot \partial(xa,x)+\frac{N\cdot \partial(a,1)}{\inf(\partial(y,1)\mid y\notin V)}\\
&\leqslant \Big(M+\frac{N}{\inf(\partial(y,1)\mid y\notin V)}\Big)\cdot \partial(xa,x).
\end{split}$$ So $d\leqslant L\cdot \partial$, where $L=\max\Big\{K,M+\frac{N}{\inf(\partial(y,1)\mid y\notin V)}\Big\}$ and by symmetry we find that $d$ and $\partial$ are bi-Lipschitz equivalent.
Thus, if $G$ admits a metric that is simultaneously minimal and maximal, then this defines a canonical global Lipschitz geometric structure on $G$. To characterise this situation, we need a few new concepts from [@coarse].
A topological group is [*Baire*]{} if it satisfies the Baire category theorem, that is, if the intersection of countably many dense open sets is dense in $G$. Also, we say that $G$ is [*European*]{} if it is Baire and, for every identity neighbourhood $V$, there is is a countable set $D\subseteq G$ so that $G=\langle V\cup D\rangle$. Clearly every Polish group and every connected completely metrisable group, e.g., the additive group $(X,+)$ of a Banach space, is European. Also, a locally compact Hausdorff group is European if and only if it is $\sigma$-compact.
A subset $B$ of a topological group $G$ is [*coarsely bounded*]{} if it has finite diameter in every continuous left-invariant pseudometric on $G$. If $G$ is European, this is equivalent to asking that, for every identity neighbourhood $V$, there is a finite set $F\subseteq G$ and a $k$ so that $B\subseteq (FV)^k$.
Now, as opposed to minimal metrics, we do have a characterisation of the existence of maximal metrics. Namely, as shown in [@coarse], a metrisable European group $G$ admits a maximal metric $d$ if and only if it is algebraically generated by a coarsely bounded set, which furthermore may be taken to be an identity neighbourhood $V$. Moreover, in this case, the maximal metric $d$ will be quasi-isometric to the word metric $$\rho_V(g,f)=\min(k\mid {\exists}v_1, \ldots, v_k\in V^\pm\colon g=fv_1\cdots v_k)$$ associated with $V$.
\[min-max\] Let $G$ be a Polish or, more generally, a European topological group generated by a coarsely bounded set and assume that $G$ has a minimal metric. Then $G$ has a metric that is simultaneously maximal and minimal and hence $G$ has a well-defined Lipschitz geometric structure.
Fix a minimal metric $d$ on $G$ and a coarsely bounded identity neighborhood $V$ generating $G$. Now, as shown in [@coarse], the formula $$\partial(g,f)=\inf\big(\sum_{i=1}^nd(v_i, 1)\mid v_i\in V\;\&\; g=fv_1\cdots v_n\big)$$ defines a compatible left-invariant metric on $G$, which is quasi-isometric to the word metric $\rho_V$. It thus follows that $\partial$ is quasi-isometric to a maximal metric on $G$ and therefore maximal itself.
Observe now that, if $W$ is a symmetric identity neighbourhood so that $W^2\subseteq V$, then any two elements of $W$ differ on the right by an element of $V$ and so the two metric $d$ and $\partial$ agree on $W$. It therefore follows that $\partial$ is also minimal.
Outside of the class of locally compact groups, the problem of determining which groups admit a minimal metric is unsolved. However, as is evident from Condition (\[strongest cond\]) of Theorem \[main\], a group $G$ with a minimal metric must be [*uniformly NSS*]{} in the sense of the following definition.
A topological group $G$ is [*uniformly NSS*]{} group if there is an open set $U\ni 1$ so that, for every open $V\ni 1$, there is some $n$ for which $$g, g^2, \ldots, g^n\in U \;\;{\Rightarrow}\;\; g\in V.$$
As noted by P. Enflo [@enflo], a uniformly NSS group is metrisable. Indeed, let $U\ni 1$ be as in the definition of the uniform NSS property and pick open neighbourhoods $W_n\ni1 $ so that $(W_n)^n\subseteq U$. Now, suppose $V\ni 1$ is open and let $n$ be such that $g\in V$ whenever $g, g^2, \ldots, g^n\in U$. Then clearly $g\in W_n$ implies that $g\in V$, i.e., $W_n\subseteq V$. Thus, the sets $W_n$ form a countable neighbourhood basis at $1$ and $G$ is metrisable by the result of Birkhoff and Kakutani.
For other interesting facts about uniformly NSS groups, including that every Banach-Lie group is uniformly NSS, one may consult the paper [@pestov] by S. A. Morris and V. Pestov. In particular, the authors show that uniformly NSS groups are [*locally minimal*]{}, which we shall not define here. However, by essentially the same proof, we may prove the following.
Suppose $G\curvearrowright X$ is a continuous isometric action of a uniformly NSS topological group $G$, as witnessed by an identity neighbourhood $U$, on a metric space $(X,d)$. Assume also that, for some ${\epsilon}>0$ and $x\in X$, we have $$g\notin U{\Rightarrow}d(gx,x)>{\epsilon}.$$ Then the orbit map $$g\in G\mapsto gx\in X$$ is a uniform embedding of $G$ into $X$.
As the orbit map is easily uniformly continuous, to see that it is a uniform embedding of $G$ into $X$, it suffices to show that, for every identity neighbourhood $V$, there is $\eta>0$ so that $d(gx,x)\geqslant \eta$ whenever $g\notin V$. So let $V$ be given and pick $n$ so that $g\in V$ whenever $g, g^2, \ldots, g^n\in U$. Then, if $g\notin V$, there is $i\leqslant n$ so that $g^i\notin U$ and thus also $n\cdot d(gx,x)\geqslant d(g^ix,x)>{\epsilon}$. In other words, $d(g,1)\geqslant \frac {\epsilon}n$ for all $g\notin V$.
Recall that a topological group $G$ is said to be a [*SIN*]{} group (for [*small invariant neighbourhoods*]{}) if there a a neighbourhood basis at the identity consisting of conjugacy invariant sets. In the context of metrisable groups, these are, by a result of V. Klee [@klee], simply the groups admitting a compatible bi-invariant metric.
\[SIN+minimal\] Let $G$ be a SIN group with a minimal metric. Then $G$ admits a bi-invariant minimal metric.
Suppose $d$ is a left-invariant minimal metric on $G$ as witnessed by an open set $U\ni1 $ so that $$g,g^2,\ldots, g^n\in U\;{\Rightarrow}\; d(g,1)\leqslant \frac 1n.$$ Since $G$ is SIN, we may assume that $U$ is conjugacy invariant. Also, replacing $d$ with $\min\{d,1\}$, we can assume that $d\leqslant 1$. Define now a metric $\partial$ by $$\partial(g,h)=\sup_{f\in G}d(gf,hf),$$ and note that, as $G$ is SIN, $\partial$ is a compatible bi-invariant metric on $G$. We claim that $\partial$ is minimal. Indeed, supposing that $g,g^2, \ldots, g^n\in U$, then, for every $f\in G$, we have $f{^{-1}}gf, (f{^{-1}}gf)^2, \ldots, (f{^{-1}}gf)^n\in U$ and thus $d(gf,f)=d(f{^{-1}}gf, 1)\leqslant 1/n$, i.e., $\partial(g,1)\leqslant 1/n$.
$G$ is a [*locally SIN*]{} group if there is an identity neighbourhood ${\mathcal}O$ so that the sets $$V^{\mathcal}O=\{gfg{^{-1}}{ \; \big| \;}g\in {\mathcal}O\;\&\; f\in V\},$$ where $V$ varies over identity neighbourhoods, form a neighbourhood basis at the identity.
We claim that $G$ is locally SIN if and only if the inversion map $g\mapsto g{^{-1}}$ is left-uniformly continuous on an open symmetric set $W\ni1 $. Indeed, suppose first that inversion is left-uniformly continuous on $W$. This means that, for all open $V\ni 1$ there is an open $U\ni1 $ so that $$g,f\in W\;\;\&\;\; g{^{-1}}f\in U\;\;{\Rightarrow}\;\; gf{^{-1}}\in V.$$ We let ${\mathcal}O\ni1 $ be symmetric open so that ${\mathcal}O^2\subseteq W$. Then, for every open $V\ni 1$, pick $U\subseteq {\mathcal}O$ as above. Then, if $g\in {\mathcal}O\subseteq W$ and $h\in U$, note that also $f=gh\in {\mathcal}OU\subseteq W$ and $g{^{-1}}f=h\in U$, whereby $ghg{^{-1}}=gf{^{-1}}\in V$, showing that $U^{\mathcal}O\subseteq V$, whence $G$ is locally SIN.
Conversely, suppose that $G$ is locally SIN as witnessed by some symmetric open ${\mathcal}O\ni 1$. Then, if $V\ni1$ is symmetric open, find some open $U$ with $1\in U\subseteq {\mathcal}O$ and $U^{\mathcal}O\subseteq V$ and note that $$g,f\in {\mathcal}O\;\;\&\;\; g{^{-1}}f\in U\;{\Rightarrow}\; fg{^{-1}}=g\cdot g{^{-1}}f\cdot g{^{-1}}\in V\;{\Rightarrow}\; gf{^{-1}}=(fg{^{-1}}){^{-1}}\in V.$$ So inversion is left-uniformly continuous on ${\mathcal}O$.
Similarly, one may show that $G$ is locally SIN if and only if there is an open set $W\ni 1$ so that the map $(g,f)\in W\times W\mapsto gf\in W^2$ is left-uniformly continuous.
For the next Proposition, we recall that a topological group $G$ has [*property (OB)*]{} if, whenever $G\curvearrowright (X,d)$ is a continuous isometric action on a metric space $(X,d)$, every orbit is bounded. If $G$ is separable metrisable, property (OB) is equivalent to the following property (see [@OB]), which we may term the [*strong property (OB)*]{}: For every open ${\mathcal}O\ni1$ there are a finite set $F\subseteq G$ and a $k$ so that $G=(F{\mathcal}O)^k$.
Apart from compact groups, a surprisingly large number of topological groups have property (OB). Of particular interest to us is the unitary group $U({\mathcal}H)$ of separable infinite-dimensional Hilbert space, which has the strong property (OB) when equipped with the operator norm topology. This follows from the spectral theorem. More examples can be found, e.g., in [@OB].
\[SIN+OB\] Let $G$ be a metrisable locally SIN group with the strong property (OB). Then $G$ is SIN.
Let ${\mathcal}O$ be an open neighbourhood of $1$ witnessing that $G$ is locally SIN and pick $k$ and a finite subset $F\subseteq G$ so that $G=(F{\mathcal}O)^k$. Suppose now that $U\ni1 $ is open. Set $V=\bigcap_{f\in F}f{^{-1}}Uf$, which is an open neighbourhood of $1$, and let $W\ni 1$ be an open set so that $W^{\mathcal}O\subseteq V$, whence $W^{F{\mathcal}O}=(W^{\mathcal}O)^F\subseteq V^F\subseteq U$. Thus, by induction, we can choose open $U=W_0\supseteq W_1\supseteq W_2\supseteq \ldots\supseteq W_k\ni1 $ so that $W_{i}^{F{\mathcal}O}\subseteq W_{i-1}$, whereby $W_k^G=W_k^{(F{\mathcal}O)^k}\subseteq W_0=U$. In particular, $W_k^G$ is a conjugacy invariant neighbourhood of $1$ contained in $U$, which shows that $G$ admits a neighbourhood basis at $1$ consisting of conjugacy invariant sets.
Enflo [@enflo] showed that uniformly NSS groups are locally SIN (though he used the terminology [*locally uniform*]{} in place of [*locally SIN*]{}). To see this, let $U\ni 1$ be the open set given by the uniform NSS property and pick a symmetric open ${\mathcal}O\ni 1$ so that ${\mathcal}O^3\subseteq U$. Suppose now $W$ is an arbitrary neighbourhood of $1$ and find $n$ so that $$g, g^2, \ldots, g^n\in U\;\; {\Rightarrow}\;\;g\in W.$$ We now choose some open $V\ni 1$ so that $V^n\subseteq {\mathcal}O$, whence also ${\mathcal}OV^n{\mathcal}O{^{-1}}\subseteq U$. In particular, if $v\in V$ and $g\in {\mathcal}O$, then $(gvg{^{-1}})^m=gv^mg{^{-1}}\in U$ for $m=1,\ldots, n$, whereby $gvg{^{-1}}\in W$. In other words, $V^{\mathcal}O\subseteq W$, verifying that $G$ is locally SIN.
So combining this result of Enflo with Corollary \[SIN+minimal\] and Proposition \[SIN+OB\], we obtain the following.
Every group with a minimal metric and the strong property (OB) also has a bi-invariant minimal metric.
Recall that a sequence $(f_i)$ in a metrisable group $G$ is [*left-Cauchy*]{} if $f_i{^{-1}}f_j{\mathop{\longrightarrow}\limits_{i,j{\rightarrow}\infty}}1$ and [*right-Cauchy*]{} if $f_i f_j{^{-1}}{\mathop{\longrightarrow}\limits_{i,j{\rightarrow}\infty}}1$. The group $G$ is [*Raĭkov complete*]{} if every sequence that is both left and right-Cauchy is convergent. This is equivalent to $G$ being [*completely metrisable*]{}, i.e., that the topology on $G$ can be induced by a complete metric. Also, $G$ is [*Weil complete*]{} if every left-Cauchy sequence in $G$ is convergent. This, in turn, is equivalent to the existence of a compatible complete left-invariant metric on $G$ (such groups are sometime denoted [*CLI*]{} for [*complete left-invariant*]{}). Moreover, in this case, every compatible left-invariant metric is complete.
\[weil\] If $G$ is locally SIN and completely metrisable, then $G$ is Weil complete. In particular, every minimal metric on a completely metrisable group is necessarily complete.
Indeed, suppose that ${\mathcal}O\ni 1$ is an open set witnessing that $G$ is locally SIN and that $(f_i)$ is left-Cauchy. To see that $(f_i)$ is right-Cauchy, fix a neighbourhood $U$ of $1$ and pick some $i_0$ so that $f_i{^{-1}}f_j\in {\mathcal}O$ for all $i,j\geqslant i_0$. Now choose some open $W\ni1 $ so that $f_{i_0}Wf_{i_0}{^{-1}}\subseteq U$ and some open $V\ni 1$ so that $V^{\mathcal}O\subseteq W$. Finally, let $i_1\geqslant i_0$ be so that $f_i{^{-1}}f_j\in V$ whenever $i,j\geqslant i_1$. Then $i,j\geqslant i_1$ implies that $f_{i_0}{^{-1}}f_i\in{\mathcal}O$ and so $f_i=f_{i_0}g$ for some $g\in {\mathcal}O$, whence $$f_jf_i{^{-1}}= f_i\cdot f_i{^{-1}}f_j\cdot f_i{^{-1}}= f_{i_0}g \cdot f_i{^{-1}}f_j\cdot g{^{-1}}f_{i_0}{^{-1}}\in f_{i_0}g Vg{^{-1}}f_{i_0}{^{-1}}\subseteq f_{i_0}W f_{i_0}{^{-1}}\subseteq U.$$ Thus $(f_i)$ is also right-Cauchy and therefore convergent in $G$.
The next result has a long history and many variations. The first occurrence seems to be the paper by A. Gleason [@gleason2] in which it is proved that, in a locally euclidean NSS group, there is an identity neighbourhood in which square roots, whenever they exist, are necessarily unique. We shall need a stronger version of this, namely that in a uniformly NSS group the extraction of square roots, whenever they exist, is left-uniformly continuous. A result of this form, under additional hypotheses, is also proved in Enflo’s paper [@enflo].
\[square roots\] Suppose $G$ is uniformly NSS. Then there is an open set $V\ni1$ so that, for every open $U\ni1$, there is an open $W\ni 1$ so that $$g,f\in V\;\;\;\&\;\;\;g^{-2}f^2\in W\;{\Rightarrow}\; g{^{-1}}f\in U.$$ In particular, the map $g\mapsto g^2$ is injective on $V$.
Since uniformly NSS groups are also locally SIN, we fix a symmetric open set ${\mathcal}O\ni1 $ witnessing both that $G$ is uniformly NSS and locally SIN. Let also $V\ni 1$ be symmetric open so that $\big(V^{\mathcal}OVV\big)^2\subseteq {\mathcal}O$.
To see that the lemma holds for $V$, suppose $U$ is given and pick some $n$ so that $\big(y, y^2, \ldots, y^n\in {\mathcal}O\;{\Rightarrow}\; y\in U\big)$. Let also $W\ni 1$ be an open set with $WW^{\mathcal}O W^{{\mathcal}O^2}\cdots W^{{\mathcal}O^{n-1}}\subseteq V$. Then, for $x,y\in {\mathcal}O$, $$x^{-i}y^{i}=\big[x{^{-1}}y\big]\cdot \big[y{^{-1}}(x{^{-1}}y)y\big]\cdot \big[y^{-2}(x{^{-1}}y)y^2\big]\cdots
\big[y^{-(i-1)}(x{^{-1}}y)y^{i-1}\big]$$ and so, if also $x{^{-1}}y\in W$ and $i\leqslant n$, then $x^{-i}y^i\in WW^{\mathcal}O W^{{\mathcal}O^2}\cdots W^{{\mathcal}O^{i-1}}\subseteq V$. In other words, $$x,y\in {\mathcal}O\;\;\;\&\;\;\; x{^{-1}}y\in W\;\;{\Rightarrow}\;\; {\forall}i\leqslant n\colon\; x^{-i}y^i\in V.$$
Suppose that $g,f\in V$ satisfy $g^{-2}f^2\in W$ and set $y=g{^{-1}}f$ and $x=g{^{-1}}y{^{-1}}g$. Then $x{^{-1}}y=g^{-2}f^2\in W$ and thus $g{^{-1}}y^igy^i =x^{-i}y^i\in V$ for $i\leqslant n$. Note now that $$y^i\in {\mathcal}O\;\;{\Rightarrow}\;\; y^{2i}=(y^ig{^{-1}}y^{-i})\cdot g\cdot (g{^{-1}}y^igy^i)\in V^{\mathcal}OVV$$ for all $i\leqslant n$. We claim that $y^i\in V^{\mathcal}OVV$ for even $i\leqslant n$ and $y^i\in (V^{\mathcal}OVV)^2\subseteq {\mathcal}O$ for odd $i\leqslant n$. This is clear for $i=0,1$, so suppose the result holds for all $i\leqslant j<n$ and consider $i=j+1$. If $i$ is odd, then $j$ is even and thus $y^j\in V^{\mathcal}OVV$ and $y^i=yy^j\in V^2\cdot V^{\mathcal}OVV\subseteq (V^{\mathcal}OVV)^2$. On the other hand, if $i$ is even, then $l=\frac i2\leqslant j$ and so $y^l\in (V^{\mathcal}OVV)^2\subseteq {\mathcal}O$, whence $y^i=y^{2l}\in V^{\mathcal}OVV$.
Thus, $y,y^2,\ldots, y^n\in {\mathcal}O$, whereby $g{^{-1}}f=y\in U$ as claimed.
Our final result originates in work of C. Chevalley [@chevalley], who showed how to construct one-parameter subgroups in locally euclidean NSS groups. Again, a generalisation to the non-locally compact setting was obtained by Enflo in [@enflo], in which he proved the result below under the assumption that the group is [*uniformly dissipative*]{}. However, this assumption excludes, for example, compact Lie groups and therefore does not generalise the classical setting of locally compact Lie groups. Our result below includes this latter setting.
\[one-parameter\] Suppose $G$ is a completely metrisable topological group admitting a minimal metric and that, for every open $W\ni1$, the set $\{g^2{ \; \big| \;}g\in W\}$ is dense in a neighbourhood of $1$. Then there are open sets ${\mathcal}U\supseteq {\mathcal}O\ni1$ so that, for every $f\in {\mathcal}O$, there is a unique one-parameter subgroup $(h^\alpha)_{\alpha\in {\mathbb R}}$ with $h^1=f$ and $h^\alpha\in {\mathcal}U$ for all $\alpha\in [-1,1]$.
Let $d$ be a minimal metric and observe that $d$ is complete by Lemma \[weil\]. We claim that, for all sufficiently small identity neighbourhoods $V$, there is an identity neighbourhood $W$ so that every element of $W$ has a square root in $V$.
Note first that it suffices to show this for closed $V$. Since $G$ has a minimal metric, it is uniformly NSS and hence, by Lemma \[square roots\] it follows that, if $V$ is a sufficiently small neighbourhood of $1$, then, for every open $U\ni 1$, there is an open $\tilde U\ni1$ satisfying $$\label{uniformity}
g,f\in V\;\;\&\;\; g^{-2}f^2\in\tilde U\;{\Rightarrow}\; g{^{-1}}f\in U.$$
Given such a closed neighbourhood $V$, set $W={\overline}{\{g^2{ \; \big| \;}g\in V\}}$, which by assumption is a neighbourhood of $1$, and assume that $f\in W$. We pick $g_n\in V$ so that $g_n^2{\mathop{\longrightarrow}\limits_{n}} f$, which means that $(g_n^2)$ is left-Cauchy and thus, by the above assumption (\[uniformity\]) on $V$, also $(g_n)$ is left-Cauchy. Since $d$ is a complete left-invariant metric and $V$ is closed, it follows that $(g_n)$ is convergent to some $g\in V$, whence $g^2=\lim_{n}g_n^2=f$. In other words, every element of $W$ has a square root in $V$, which proves our claim.
Since $d$ is minimal, there are an identity neighbourhood $V_0$ and a $k\geqslant 1$ so that, for all $g\in G$ and $n\geqslant 1$, $$g, g^2, g^4, g^8, \ldots, g^{2^n}\in V_0\;{\Rightarrow}\; d(g,1)\leqslant 2^{k-n-1}\cdot d(g^{2^n},1).$$ Define inductively identity neighbourhoods $V_0\supseteq V_1\supseteq \ldots\supseteq V_k$ so that every element of $V_{i+1}$ is the square of some element in $V_i$. By shrinking $V_k$, we may assume that $V_k=B_d({\epsilon})$ for some ${\epsilon}>0$.
Now suppose that $h_0\in V_k$ and choose inductively $h_i\in V_{k-i}$ so that $h_{i+1}^2=h_i$. Then $h_k, h_k^2=h_{k-1}, h_k^4=h_{k-2}, \ldots, h_k^{2^k}=h_0$ all belong to $V_0$, whence $$d(h_k,1)\leqslant 2^{k-k-1}\cdot d(h_k^{2^k}, 1)=1/2 \cdot d(h_0, 1)\leqslant {\epsilon}/2.$$ This shows that every $h\in V_k$ has a $2^k$-th root $f \in V_k$ so that $d(f, 1)\leqslant 1/2 \cdot d(h,1)$.
Therefore, if $f\in V_k$ is given, we can choose an infinite sequence $h_0,h_1, h_2,\ldots\in V_k$ beginning at $h_0=f$ so that $h_{i+1}^{2^k}=h_i$ and $d(h_{i+1},1)\leqslant 1/2\cdot d(h_i,1)$ for all $i$. In particular, $h_i^{2^{k\cdot j}}=h_{i-j}$ for all $j\leqslant i$. For every dyadic rational number $\alpha=\frac m{2^{k\cdot i}}$ with $m\in {\mathbb Z}$ and $i\in {\mathbb N}$, we can then unambigously define $h^\alpha=h_i^m$ and see that $h^\alpha\cdot h^\beta=h^{\alpha+\beta}$ for all dyadic rationals $\alpha$ and $\beta$.
Now, if $\alpha\in [0,1[$ is a dyadic rational, write $$\alpha=\frac {a_1}{2^k}+\frac{a_2}{2^{k\cdot 2}}+\ldots+\frac{a_p}{2^{k\cdot p}}$$ with $0\leqslant a_i<2^k$. Then $$h^\alpha= h_1^{a_1}\cdot h_2^{a_2}\cdots h_p^{a_p},$$ whereby $$\begin{split}
d(h^\alpha,1)
&\leqslant a_1\cdot d(h_1,1)+a_2\cdot d(h_2,1)+\ldots+a_{p}\cdot d(h_{p},1)\\
&\leqslant \Big(\frac{a_1}2+\frac{a_2}{2^2}+\ldots+\frac{a_{p}}{2^{p}}\Big)\cdot {\epsilon}\\
&<2^k\cdot {\epsilon}.
\end{split}$$ Moreover, if $\alpha<\frac1{2^{ki}}$, then $a_1=\ldots=a_i=0$, whence $d(h^\alpha,1)<2^{k-i}\cdot {\epsilon}$.
It follows that the mapping $\alpha\mapsto h^\alpha$ is a continuous homomorphism from the additive group of dyadic rationals with the topology induced by ${\mathbb R}$ into $G$. Since $d$ is a complete metric on $G$, it follows that this extends to a continuous one-parameter subgroup $(h^\alpha)_{\alpha\in {\mathbb R}}$ with $h^1=f$ and so that $d(h^\alpha,1)\leqslant 2^k\cdot {\epsilon}$ for all $\alpha\in [-1,1]$.
Now, suppose that $(h^\alpha)_{\alpha\in {\mathbb R}}$ and $(g^\alpha)_{\alpha\in {\mathbb R}}$ are distinct one-parameter subgroups in $G$ with $h^1=g^1$. Then, by the density of the dyadic rationals in ${\mathbb R}$, there must be some dyadic rational $\alpha=\frac 1{2^{n}}$ so that $h^\alpha\neq g^\alpha$. However, as $h^1=g^1$, it follows that there is an $\ell\geqslant 0$ so that $h^{\frac1{2^{\ell+1}}}\neq g^{\frac1{2^{\ell+1}}}$, while $h^{\frac1{2^{\ell}}}=g^{\frac1{2^{\ell}}}$. So the squaring map $f\mapsto f^2$ fails to be injective on any set containing $\{h^\beta\}_{\beta\in [-1,1]}$ and $\{g^\beta\}_{\beta\in [-1,1]}$.
Therefore, if we choose $V_k=B_d({\epsilon})$ small enough so that the squaring map is injective on $B_d(2^k\cdot {\epsilon})$, which is possible by Lemma \[square roots\], then, for every $f\in V_k$, there is a unique one-parameter subgroup $(h^\alpha)_{\alpha\in {\mathbb R}}\subseteq G$ so that $h^1=f$ and $h^\alpha\in B_d(2^k\cdot {\epsilon})$ for all $\alpha\in [-1,1]$. Setting ${\mathcal}U= B_d(2^k\cdot {\epsilon})$ and ${\mathcal}O=B_d({\epsilon})$, the theorem follows.
[99]{}
Garrett Birkhoff, [*A note on topological groups*]{}, Compositio Math. 3 (1936), 427–430.
C. Chevalley, [*On a Theorem of Gleason*]{}, Proc. Amer. Math. Soc. 2 (1951), 122–125.
J. Diestel and A. Spalsbury, [*The joys of Haar measure*]{}, American Mathematical Society 2014.
P. Enflo, [*Uniform structures and square roots in topological groups I*]{}, Israel J. Math. 8 (1970) no. 3, 230–252.
A. M. Gleason, [*Groups without small subgroups*]{}, Ann. of Math. (2) 56, (1952), 193–212.
A. M. Gleason, [*Square roots in locally euclidean groups*]{}, Bull. Amer. Math. Soc. 55 (1949), 446-449.
Shizuo Kakutani, [*Über die Metrisation der topologischen Gruppen*]{}, Proc. Imp. Acad. Tokyo 12 (1936), 82–84. Selected Papers, Vol. 1, Robert R. Kallman Editor, Birkhäuser (1986), 60–62.
V. L. Klee Jr., [*Invariant metrics in groups (solution of a problem of Banach)*]{} Proc. Amer. Math. Soc. 3 (1952) 484–487.
D. Montgomery and L. Zippin, [*Topological transformation groups*]{}, Reprint of the 1955 original. Robert E. Krieger Publishing Co., Huntington, N.Y., 1974.
S. A. Morris and V. Pestov, [*On Lie groups in varieties of topological groups*]{}, Coll. Math. 78 (1998), no. 1, 39–47.
C. Rosendal, [*A topological version of the Bergman property*]{}, Forum Math. 21 (2009), no. 2, 299–332.
C. Rosendal, [*Coarse geometry of topological groups*]{}, preprint.
D. Sullivan, [*Hyperbolic geometry and homeomorphisms*]{}, in Geometric Topology: Proc. Topology Conf. at Athens, Ga., 1977, J.C. Cantrell, ed., Academic Press, New York, London (1979), 543–555.
Terence Tao, [*Hilbert’s fifth problem and related topics*]{}, Graduate Studies in Mathematics, 153. American Mathematical Society, Providence, RI, 2014.
[^1]: The author was partially supported by a Simons Foundation Fellowship (Grant \#229959) and also recognises support from the NSF (DMS 1201295 & DMS 1464974)
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Ira Wolfson,'
- 'Azadeh Maleknejad,'
- and Eiichiro Komatsu
title: 'How attractive is the isotropic attractor solution of axion-SU(2) inflation?'
---
Introduction
============
Cosmic inflation [@Guth:1980zm; @Sato:1980yn] based on a single slowly rolling scalar field [@Linde:1981mu; @Albrecht:1982wi; @Linde:1983gd] yields primordial scalar perturbations which are nearly scale-invariant and nearly Gaussian [@Mukhanov:1981xt; @Starobinsky:1982ee; @Hawking:1982cz; @Bardeen:1983qw]. It can also yield primordial tensor perturbations that are nearly scale invariant and nearly Gaussian [@Starobinsky:1979ty].
The presence of gauge fields during inflation changes this picture, as they can produce non-scale invariant, non-Gaussian scalar and tensor perturbations. For example, a U(1) gauge field coupled to an axion sources scalar and tensor perturbations at quadratic order, resulting in highly non-Gaussian perturbations [@Anber:2009ua; @Barnaby:2010vf; @Barnaby:2011vw; @Anber:2012du]. The lack of detection of scalar non-Gaussianities [@Akrami:2019izv] rules out a large parameter space of such models. Nevertheless one new phenomenology predicted by these models is that tensor perturbations become chiral [@Barnaby:2012xt; @Namba:2015gja; @Anber:2012du].
U(1) gauge field models face an additional challenge as they produce a preferred direction in the universe, breaking the statistical isotropy of perturbations (see [@Maleknejad:2012fw] for a review), that is not detected [@Kim:2013gka; @Akrami:2018odb].
Models with SU(2) gauge fields coupled with an axion avoid both challenges. Since the SU(2) gauge algebra is isomorphic to the SO(3) algebra, the rotational invariance is protected and the statistical isotropy remains unbroken [@Maleknejad:2011jw; @Maleknejad:2011sq]. As for perturbations there is a large parameter space in which SU(2) gauge fields produce chiral tensor perturbations with negligible contributions to scalar perturbations at the linear order [@Dimastrogiovanni:2012ew; @Adshead:2013qp; @Maleknejad:2012fw]. See [@Papageorgiou:2018rfx; @Papageorgiou:2019ecb] for study of non-linear perturbations.
The homogeneous background of gauge fields leads to particle production that may produce significant back-reaction on the dynamics of the homogeneous gauge and scalar fields. The back-reaction due to charged scalar fields and fermions coupled to SU(2) gauge fields and spin-2 particles of SU(2) further restricts the phenomenologically viable parameter space of these models [@Lozanov:2018kpk; @Mirzagholi:2019jeb; @Maleknejad:2018nxz; @Maleknejad:2019hdr; @Domcke:2018eki; @Domcke:2018gfr; @Domcke:2019qmm].
The key to the phenomenological success of these models is the isotropic background of the SU(2) gauge fields. Given an anisotropic background initial geometry, how did the isotropic background emerge? Previous studies have shown that the isotropic background of the SU(2) gauge field is an attractor solution [@Maleknejad:2013npa; @Adshead:2012kp]. However not all possible initial configurations lead to an isotropic end state. As we show in this paper, some lead to violation of slow-roll conditions terminating inflation prematurely. We also show that the attractor’s basin of attraction increases when we introduce another scalar field acting as inflaton and demote the axion-SU(2) system to a spectator sector. Thus the axion-SU(2) spectator model [@Dimastrogiovanni:2016fuu] is both phenomenologically rich and observationally viable.
This paper is organized as follows. In section \[sec:Isotropic Background with SU(2)\] we review SU(2) gauge fields in an isotropic background Friedmann-Robertson-Walker (FRW) metric, coupled to axion fields via a Chern-Simons (CS) topological term. Section \[sec:Anisotropic\] introduces the chromo-natural (CN) system embedded in an anisotropic Bianchi type I metric. We go through some of the theoretic consequences and present new results of numerical analysis. In section \[sec:Anisotropic-Spectator\] the CN system is demoted to a spectator status, and we analytically and numerically study this scalar-driven inflation with a spectator axion-SU(2) component. Our numerical study is focused on the initial condition basin of the isotropic attractor solution. Finally we conclude in section \[sec:Conclusion\].\
\
We work in the mostly positive signature ($-,+,+,+$), and in natural units where $c=\hbar=1$. The reduced Planck mass $M_{Pl}=\sqrt{\frac{1}{8\pi G}}$ is set to $1$.
Isotropic Background of the SU(2) gauge field {#sec:Isotropic Background with SU(2)}
=============================================
Axion-SU(2) inflation
---------------------
The axion-SU(2) inflation is a class of inflationary models where SU(2) gauge fields contribute to the physics of inflation. Within this family of models we look at the chromo-natural (CN) model in which we have an axion field with a cosine potential. The action is given by [@Adshead:2012kp]: $$\begin{aligned}
S=\int d^4 x\sqrt{-g}\left[-\frac{R}{2} -\frac{1}{4}F^a_{\mu\nu}F_a^{\mu\nu} -\frac{1}{2}\left(\partial_{\mu}\chi\right)^2 -\mu^4\left(1+\cos{\frac{\chi}{f}}\right)+\frac{\lambda \chi}{8f}\chi \Tilde{F}^a_{\mu\nu}F_a^{\mu\nu}\right],\label{eq:CN_action}\end{aligned}$$ where $\chi$ is the axion field with the potential $V\left(\chi\right)=\mu^4\left(1+\cos \frac{\chi}{f}\right)$, $f$ is the axion decay constant, $F^a_{\mu\nu}$ is the SU(2) field strength tensor: $$\begin{aligned}
F^a_{\mu\nu}=\partial_{\mu}A^a_{\nu}-\partial_{\nu}A^a_{\mu}-g_{\scaleto{A}{4pt}}\epsilon^a_{bc}A^b_{\mu}A^c_{\nu}\; ,\end{aligned}$$ $\Tilde{F}^a_{\mu\nu}$ is its Hodge dual, $g_{\scaleto{A}{4pt}}$ is the gauge coupling, and $\epsilon^{a}_{bc}$ is the non-Abelian algebra’s set of structure constants. The CN model can be embedded in a simple FRW metric, as was done in [@Maleknejad:2011sq; @Adshead:2012kp; @Obata:2014loa; @Adshead:2016omu] or in some other metric (see [@Maleknejad:2013npa] for Bianchi Type I case). The coupling of the axion and gauge fields through the CS term enables the kinetic gauge-field term to exchange energy with the axion potential, effectively augmenting the conventional drag term for the axion and prolonging slow-roll.\
\
When embedded in an FRW metric, the gauge fields contribute to the energy density and supports the isotropic inflation of space. This can be seen in the Friedmann equations: $$\begin{aligned}
\left\{\begin{array}{cc}
H^2=&\frac{\rho_{\scaleto{A}{4pt}} +\rho_{\chi}}{3}\\
&\\
\frac{\ddot{a}}{a}=&-\frac{\rho_{\scaleto{A}{4pt}} +\rho_{\chi}+3P_{\scaleto{A}{4pt}} +3P_{\chi}}{6}
\end{array}\right.\; ,\end{aligned}$$ with the energy densities given by: $$\begin{aligned}
\rho_{\chi}=\frac{\dot{\chi}^2}{2}+\mu^4\left(1+\cos\frac{\chi}{f}\right),\\
\rho_{\scaleto{A}{4pt}}=2\frac{\delta \mathcal{L_{\scaleto{A}{4pt}}}}{\delta F^{a\; 0}_{\;\sigma}} F^a_{\sigma 0}-\mathcal{L_{\scaleto{A}{4pt}}},\end{aligned}$$ where we call the term: $$\begin{aligned}
\mathcal{L_{\scaleto{A}{4pt}}}= -\frac{1}{4}F^a_{\mu\nu}F_{a}^{\mu\nu} +\frac{1}{4} \left(\frac{\lambda\chi}{f}\right)\Tilde{F}^a_{\mu\nu}F_a^{\mu\nu}\end{aligned}$$ the gauge fields ($A$) Lagrangian. The pressure contributions are given by: $$\begin{aligned}
P_{\chi}=\dot{\chi}^2-\rho_{\chi}\;,\\
\nonumber \\
P_{\scaleto{A}{4pt}}=\frac{\rho_{\scaleto{A}{4pt}}}{3} \; .\end{aligned}$$\
In the FRW spacetime $$\begin{aligned}
ds^2=-dt^2 +a(t)^2\left(d{\bf x}^2\right)\;,\end{aligned}$$ the CN system has an isotropic and homogeneous background solution. In particular, the SU(2) gauge field in the temporal gauge has the following solution [@Maleknejad:2011jw; @Maleknejad:2011sq]: $$\begin{aligned}
A^a_0=0,\hspace{15pt} A^a_i=\psi(t)a(t)\delta^a_i \; , \label{eq:Ansatz-FRW}\end{aligned}$$ where $\psi(t)$ is a pseudo-scalar that parametrizes the effective field value of the gauge field. The energy density for the gauge fields then become: $$\begin{aligned}
\rho_{\scaleto{A}{4pt}}=\frac{3}{2}\left(\dot{\psi}+H\psi\right)^2+\frac{3g_{\scaleto{A}{4pt}}^2\psi^4}{2},\end{aligned}$$ and equations of motion are given by [@Adshead:2012kp]: $$\begin{aligned}
\ddot{\chi}+3H\dot{\chi}=\frac{\mu^4}{f}\sin{\frac{\chi}{f}}-3g_{\scaleto{A}{4pt}}\frac{\lambda}{f}\psi^2\left(\dot{\psi}+H\psi\right)\; , \\
\nonumber \\
\ddot{\psi}+3H\dot{\psi}=-(2H^2+\dot{H})\psi-2g_{\scaleto{A}{4pt}}^2\psi^3 +g_{\scaleto{A}{4pt}}\frac{\lambda}{f}\psi^2 \dot{\chi}\; .\end{aligned}$$ Asserting slow-roll of all fields, we obtain a set of master equations: $$\begin{aligned}
\dot{\chi}\simeq\frac{\frac{\mu^4}{f}\sin{\frac{\chi}{f}}-3g_{\scaleto{A}{4pt}}\frac{\lambda}{f}\psi^2\left(\dot{\psi}+H\psi\right) }{3H}\; , \label{eq:chi_master}\\
\nonumber \\
\dot{\psi}\simeq\frac{-2H^2\psi-2g_{\scaleto{A}{4pt}}^2\psi^3 +g_{\scaleto{A}{4pt}}\frac{\lambda}{f}\psi^2 \dot{\chi}}{3H}\label{eq:psi_master}\; .\end{aligned}$$ Furthermore, discarding $\dot{\chi}$ and $\dot{\psi}$ as minuscule gives [@Adshead:2012kp]: $$\begin{aligned}
\psi\big|_{FRW}\simeq \left( \frac{\mu^4\sin{\frac{\chi}{f}}}{3g_{\scaleto{A}{4pt}}H\lambda}\right)^{\frac{1}{3}}\; .\label{eq:psi_FRW}\end{aligned}$$ As the axion kinetic term is never important in this setup, The slow-roll parameters are approximated by: $$\begin{aligned}
\varepsilon_{\scaleto{H}{4pt}}\equiv -\frac{\dot{H}}{H^2}\simeq \frac{3g_{\scaleto{A}{4pt}}^2\psi^4}{H^2}+\psi^2 \simeq \frac{3g_{\scaleto{A}{4pt}}^2\psi^4}{\mu^4\left(1+\cos{\frac{\chi}{f}}\right)}+\psi^2,\label{eps_adshead}\end{aligned}$$ which holds when the energy density is dominated by the axion potential (i.e. slow-roll), and $\eta_{\scaleto{H}{4pt}}$ is given by:[^1] $$\begin{aligned}
\eta_{\scaleto{H}{4pt}}\equiv \frac{\ddot{H}}{H\dot{H}} \simeq 2g_{\scaleto{A}{4pt}}^2\frac{\psi^4}{H^2}+\frac{\dot{\psi}}{H\varepsilon_{\scaleto{H}{4pt}}}\left(12g_{\scaleto{A}{4pt}}^2\frac{\psi^3}{H^2}+2\psi\right)-2\varepsilon_{\scaleto{H}{4pt}}.\label{eta_adshead}\end{aligned}$$ In [@Adshead:2012kp] the authors point out that the second term in \[eta\_adshead\] has to be of $\mathcal{O}(\varepsilon^2)$ to facilitate slow-roll inflation. Additionally, \[eps\_adshead\] and \[eta\_adshead\] suggest (for $\varepsilon_H\ll 1$ and $\eta_H=\mathcal{O}(\varepsilon_H^2)$) $$\begin{aligned}
\psi^2\ll\frac{H}{\sqrt{2}g_{\scaleto{A}{4pt}}},\hspace{15pt} |\psi|\ll \frac{\mu}{\sqrt{g_{\scaleto{A}{4pt}}}},\end{aligned}$$ which place strict conditions on $\psi$ given the typical value of $\mu$. Embedded in an FRW metric, the gauge fields are isotropic, and to leading order homogeneous.
Spectator Axion-SU(2) Sector {#sec:Spectator_FRW}
----------------------------
Relegating the CN system to a spectator status by adding a dominant minimally coupled scalar inflaton component, the action becomes [@Dimastrogiovanni:2016fuu]: $$\begin{aligned}
S=\int d^4 x \sqrt{-g}\left[-\frac{R}{2} -\frac{1}{4}F^a_{\mu\nu}F_a^{\mu\nu} -\frac{\left(\partial_{\mu}\chi\right)^2}{2} -\mu^4\left(1+\cos{\frac{\chi}{f}}\right)+\frac{\lambda \chi}{8f} \Tilde{F}^a_{\mu\nu}F_a^{\mu\nu}-\frac{\left(\partial_{\mu}\phi\right)^2}{2}-V(\phi)\right],\label{eq:CNspect_action}\end{aligned}$$ which is shorthanded as: $$\begin{aligned}
S=S_{CN}+S_{\phi},\end{aligned}$$ where the subscript $CN$ stands for the chromo-natural action given in Eq. \[eq:CN\_action\].\
Due to the minimal coupling between the axion-SU(2) sector and the inflaton, the equations of motion for the axion and SU(2) fields are the same and using the same ansatz as in \[eq:Ansatz-FRW\] the matter Lagrangian can be read as: $$\begin{aligned}
\mathcal{L}_{m}=a^3\left[\frac{3}{2}\left(\frac{1}{a^2}\left(\frac{\partial(a\psi)}{\partial t}\right)^2 -g_{\scaleto{A}{4pt}}^2\psi^4\right)+\frac{\dot{\chi}^2}{2} -\mu^4\left(1+\cos{\frac{\chi}{f}}\right) -3g_{\scaleto{A}{4pt}}\frac{\lambda\chi}{f}\frac{\psi^2}{a}\frac{\partial(a\psi)}{\partial t}+\frac{\dot{\phi}^2}{2}-V(\phi) \right].\end{aligned}$$ Since the metric is still FRW, the Friedmann equations remain the same: $$\begin{aligned}
H^2=&\frac{\rho}{3}\\ \nonumber
\\
\frac{\ddot{a}}{a}=&-\frac{\rho+3P}{6},\end{aligned}$$ and the total energy density is given by $$\begin{aligned}
\rho=\rho_{\chi}+\rho_{\scaleto{A}{4pt}}+\rho_{\phi},\end{aligned}$$ in which $\rho_{\chi}$ and $\rho_{\scaleto{A}{4pt}}$ are the same as before, and $\rho_{\phi}$ is: $$\begin{aligned}
\rho_{\phi}=\frac{\dot{\phi}^2}{2}+V(\phi).\end{aligned}$$ The pressure is also linearly additive, so the total pressure is given by: $$\begin{aligned}
P=P_{\chi}+P_{\scaleto{A}{4pt}}+P_{\phi},\label{eq:Spectator pressure}\end{aligned}$$ with $$\begin{aligned}
P_{\phi}=\frac{1}{2}\dot{\phi}^2 -V(\phi).\end{aligned}$$ The slow-roll parameters are given by: $$\begin{aligned}
\varepsilon_H\equiv\frac{-\dot{H}}{H^2}=\frac{3}{2}\left(\frac{P+\rho}{\rho}\right),\end{aligned}$$ and $$\begin{aligned}
\eta_H\equiv\frac{\ddot{H}}{H\dot{H}}=\sqrt{\frac{3}{\rho}}\frac{\partial}{\partial t}\ln{\left(|\rho+P|\right)}.\end{aligned}$$ Taking the inflaton pressure and energy as dominant we can write these as: $$\begin{aligned}
P=P_{\phi}+\Delta P;\hspace{15pt} \rho=\rho_{\phi}+\Delta\rho,\end{aligned}$$ where we imply $\frac{\Delta \rho}{\rho_{\phi}},\frac{\Delta P}{P_{\phi}}\ll 1$. In this approximation the quantity $\frac{P+\rho}{\rho}$ becomes: $$\begin{aligned}
\frac{P+\rho}{\rho}\simeq\frac{P_{\phi}+\rho_{\phi}}{\rho_{\phi}}+\frac{1}{\rho_{\phi}}\left[\Delta\rho +\Delta P\right] -\frac{\Delta \rho}{\rho_{\phi}}\left[\frac{P_{\phi}+\rho_{\phi}}{\rho_{\phi}}+\frac{1}{\rho_{\phi}}\left[\Delta\rho +\Delta P\right]\right],\end{aligned}$$ in which by virtue of the SU(2)-axion being a spectator sector, the third term is suppressed.\
This yields: $$\begin{aligned}
\varepsilon\simeq\varepsilon_{\phi}+\frac{3}{2\rho_{\phi}}\left[P_{\chi}+P_{\scaleto{A}{4pt}}+\rho_{\chi}+\rho_{\scaleto{A}{4pt}}\right]=\varepsilon_{\phi}+\varepsilon_{\scaleto{CN}{4pt}}\;,\end{aligned}$$ where $\varepsilon_{\phi}\equiv\frac{3(P_{\phi}+\rho_{\phi})}{2\rho_{\phi}}$ and $\varepsilon_{\scaleto{CN}{4pt}}\equiv\frac{3(P_{\chi}+P_{\scaleto{A}{4pt}}+\rho_{\chi}+\rho_{\scaleto{A}{4pt}})}{2\rho_{\phi}}$. From here on we dispose of the subscript $H$ on $\varepsilon$ and $\eta$. If we want to compare the slow-roll parameter of a spectator system and the CN system this can be written as: $$\begin{aligned}
\varepsilon_{\text{spectator}}\simeq\varepsilon_{\phi}+\frac{\rho_{\scaleto{CN_0}{4pt}}}{\rho_\phi}\varepsilon_{\scaleto{CN_0}{4pt}}, \end{aligned}$$ where the subscript $CN_0$ denotes the original CN system, without the inflaton potential, as in Eq. .\
\
The second slow-roll parameter expansion has two limits. In the limit of $\varepsilon_{\phi}\ll\varepsilon_{\scaleto{CN}{4pt}}$ we have: $$\begin{aligned}
\eta_{\text{spectator}}=\eta_{\phi}+\sqrt{\frac{3}{\rho_{\phi}}}\frac{\partial}{\partial t}\left(\frac{\varepsilon_{\scaleto{CN}{4pt}}}{\varepsilon_{\phi}}\right),\label{eq:FRW_CN_eta}\end{aligned}$$ where $\eta_{\phi}\equiv \sqrt{\frac{3}{\rho_{\phi}}}\frac{\partial}{\partial t}\ln(|\rho_{\phi}+P_{\phi}|)$, the second term is much smaller than the first. Thus if we assume a well behaved Hubble parameter over the inflationary period, we can safely ignore that term. Therefore the second slow-roll parameter is also dominated by the scalar-field-associated parameter. The only caveat for this assumption is that we have to make sure $\varepsilon_{\scaleto{CN}{4pt}}$ does not oscillate too much; otherwise, the temporal derivative might dominate Eq. \[eq:FRW\_CN\_eta\].\
In the limit of $\varepsilon_{\scaleto{CN}{4pt}}\ll\varepsilon_{\phi}$ the term for $\eta_{\text{spectator}}$ becomes: $$\begin{aligned}
\eta_{\text{spectator}}\simeq \eta_{\scaleto{CN}{4pt}}+\sqrt{\frac{3}{\rho_{\phi}}}\frac{\partial}{\partial t}\left(\frac{\varepsilon_{\phi}}{\varepsilon_{\scaleto{CN}{4pt}}}\right),\label{eq:eta_BIG_CN}\end{aligned}$$ where $\eta_{\scaleto{CN}{4pt}}\equiv \sqrt{\frac{3}{\rho_{\phi}}}\partial_t\ln\left(|\rho_{\scaleto{CN}{4pt}}+P_{\scaleto{CN}{4pt}}|\right)$. However, we can construct $\dot{\varepsilon}_{\phi}$ to be sufficiently small such that: $$\begin{aligned}
\eta_{\text{spectator}}\simeq \eta_{\scaleto{CN}{4pt}}.\end{aligned}$$ These expressions for $\eta_{\text{spectator}}$ suggest that in areas where the CN system is unstable, care should be taken choosing the inflaton potential.
Axion-SU(2) system in a Bianchi Type I spacetime {#sec:Anisotropic}
================================================
So far, we have discussed axion-SU(2) systems embedded in an FRW metric. We now turn to consider a homogeneous but anisotropic background geometry. Here we assume an axisymmetric Bianchi type I metric: $$\begin{aligned}
ds^2=-dt^2 +e^{2\alpha(t)}\left[e^{-4\sigma(t)}dx^2 +e^{2\sigma(t)}\left(dy^2+dz^2\right)\right],\label{geometry:BtI}\end{aligned}$$ in which $e^{\alpha(t)}$ is the isotropic scale factor, $a(t)$, and the metric anisotropy is represented by $e^{\sigma(t)}$. As with the case of isotropic background, we define the isotropic Hubble parameter as: $$\begin{aligned}
H\equiv\dot{\alpha}.\end{aligned}$$ When $\sigma(t)=C$ for some constant $C$, the axes can always be rescaled such that the metric reduces to the FRW one. In other words, only the time varying part of $\sigma$ is a physical quantity.\
The spatial triads now have the following form: $$\begin{aligned}
e^a_1=e^{\alpha-2\sigma}\delta^a_1,\quad e^a_2=e^{\alpha+\sigma}\delta^a_2,\quad \text{and}\quad e^a_3=e^{\alpha+\sigma}\delta^a_3\;.\end{aligned}$$ In this geometry the energy-momentum tensor has the diagonal form of: $$\begin{aligned}
T^{\mu}_{\nu}=\left(\begin{array}{cccc}
-\rho(t) & & & \\
&P(t)-2\Tilde{P}(t) & & \\
& & P(t)+\Tilde{P}(t)& \\
& & & P(t)+\Tilde{P}(t)\\
\end{array}\right),\end{aligned}$$ where $\rho$ is the energy density, $P$ is the isotropic pressure and $\Tilde{P}$ is the anisotropic pressure which parametrizes the amount of anisotropy in the energy-momentum tensor. The dynamics in Bianchi cosmology is specified by the following field equations: $$\begin{aligned}
\dot{\alpha}^2 -\dot{\sigma}^2=\frac{\rho}{3}, \label{Friedmann}\end{aligned}$$ $$\begin{aligned}
\ddot{\sigma}+3\dot{\alpha}\dot{\sigma}=\frac{\Tilde{P}}{3},\label{Anisotropic_PART}\end{aligned}$$ $$\begin{aligned}
\ddot{\alpha}+3\dot{\sigma}^2=-\frac{\rho +P}{2}.\label{2nd_friedmann}\end{aligned}$$ From \[Anisotropic\_PART\] we realize that $\Tilde{P}$ is the source of the anisotropy $\dot{\sigma}$. In its absence the initial anisotropy is exponentially suppressed, at a time scale $\dot{\alpha}^{-1}$. However, a non-zero anisotropic pressure can lead to non-trivial dynamics of spatial anisotropies in the metric. When $\dot{\alpha}\equiv H(t)$ and $\sigma=C$, the equations above revert to the usual FRW Friedmann equations.\
We follow the formalism of [@Maleknejad:2013npa] to simplify analytical treatment of the model. Choosing the temporal gauge the consistent truncation for the gauge fields gives: $$\begin{aligned}
A^{a}_{\mu}=\left\{\begin{array}{lr}
0&\mu=0 \\
\psi_i e^{a}_i& \mu=i
\end{array}\right.\; .\end{aligned}$$ The axial symmetry of the metric in the $y-z$ plane yields $\psi_2=\psi_3$; thus, the explicit form of our ansatz is given by: $$\begin{aligned}
A^{a}_i=diag(e^{\alpha-2\sigma}\psi_1,e^{\alpha+\sigma}\psi_2,e^{\alpha+\sigma}\psi_3).\end{aligned}$$ We now introduce the following field re-definitions: $$\begin{aligned}
\psi_1(t) =\frac{\psi(t)}{\beta^2(t)};\quad \psi_2(t)=\psi_3(t)=\beta(t)\psi(t),\end{aligned}$$ in which $\psi(t)$ represents the [*isotropic*]{} field and $\beta(t)$ parametrizes [*anisotropy*]{} in the gauge fields, with $\beta=\pm 1$ being the isotropic gauge field configuration. Moreover, as the point $\beta=0$ is a singularity, $\beta$ cannot change sign.\
With this reformulation the CN matter Lagrangian from the action \[eq:CN\_action\] now reads: $$\begin{aligned}
\mathcal{L}_m&=\beta^2 \left(\psi \left(\dot{\sigma}+\dot{\alpha}+\frac{\dot{\beta}}{\beta}\right)+\dot{\psi}\right)^2+\frac{1}{2 \beta^4}\left(\psi\left(-2 \dot{\sigma}+\dot{\alpha}-\frac{2 \dot{\beta}}{\beta}\right)+\dot{\psi }\right)^2\\ \nonumber
&-\frac{3 g_{\scaleto{A}{4pt}} \lambda \chi \psi^2 \left(\psi\dot{\alpha}+\dot{\psi}\right)}{f}-\mu^4 \left(1+\cos\frac{\chi }{f}\right)-\frac{1}{2} g_{\scaleto{A}{4pt}}^2 \beta^4 \psi^4-\frac{g_{\scaleto{A}{4pt}}^2 \psi^4}{\beta^2}+\frac{1}{2} \dot{\chi}^2.\end{aligned}$$ The field equation of the axion is only coupled to the isotropic part of the gauge field, $\psi$. Therefore, the axion field equation remains unchanged and similar to Eq. \[eq:chi\_master\]. It implies that only the isotropic gauge configuration is sourced by the axion. Later we see that it leads to the exponential decay of the anisotropic part by inflation.\
The energy density $\rho$ can be written as: $$\begin{aligned}
\rho=\rho_{\chi}+\rho_{\scaleto{A}{4pt}}\; , \label{eq:rho_CN}\end{aligned}$$ where $$\begin{aligned}
\rho_{\chi}=\frac{\dot{\chi}^2}{2}+\mu^4\left(1+\cos\frac{\chi}{f}\right)\;,\label{eq:rho_Chi}\end{aligned}$$ and $$\begin{aligned}
\rho_{\scaleto{A}{4pt}}=\frac{1}{2\beta^4}\left(\dot{\psi}+\dot{\alpha}\psi -2\left(\dot{\sigma}+\frac{\dot{\beta}}{\beta}\right)\psi\right)^2 +\beta^2\left(\dot{\psi}+\dot{\alpha}\psi +\left(\dot{\sigma}+\frac{\dot{\beta}}{\beta}\right)\psi\right)^2 +g_{\scaleto{A}{4pt}}^2\frac{(2+\beta^4)\psi^4}{2\beta^2}\;.\label{eq:rho_A}\end{aligned}$$ The isotropic pressure is then given by $$\begin{aligned}
P=\dot{\chi}^2 -\rho_{\chi}+\frac{\rho_{\scaleto{A}{4pt}}}{3}\;,\label{eq:P_isotropic}\end{aligned}$$ while the anisotropic part is $$\begin{aligned}
\Tilde{P}=&\frac{1-\beta^6}{3}\left(\frac{1}{\beta^4}\left(\dot{\psi}+\dot{\alpha}\psi -2\left(\dot{\sigma}+\frac{\dot{\beta}}{\beta}\right)\psi\right)^2 -\frac{1}{\beta^2}g_{\scaleto{A}{4pt}}^2\psi^4\right)\label{eq:P_Anisotropic}\\ \nonumber
& -\beta^2\left(\frac{\dot{\beta}}{\beta}+\dot{\sigma}\right)\left(2\dot{\psi}+2\dot{\alpha}\psi -\left(\dot{\sigma}+\frac{\dot{\beta}}{\beta}\right)\psi\right)\psi.\end{aligned}$$ One constant of motion arising from this effective Lagrangian is the $\sigma$ associated momentum: $$\begin{aligned}
\Pi_{\sigma}=2 \beta^2 \psi \left[\psi\left(\dot{\sigma}+\dot{\alpha}+\frac{\dot{\beta}}{\beta}\right)+\dot{\psi}\right]-\frac{2 \psi}{\beta^4} \left[\psi \left(-2\left( \dot{\sigma}+\frac{\dot{\beta}}{\beta}\right)+\dot{\alpha}\right)+\dot{\psi }\right]+6 \dot{\sigma}\; ,\end{aligned}$$ leading to: $$\begin{aligned}
\dot{\sigma}=\frac{\beta^4 De^{-3\alpha} - \left(\frac{\dot{\beta}}{\beta} \left(\beta^6+2\right)\psi^2 +\left(\dot{\psi}+\dot{\alpha}\psi\right)\left(\beta^6-1\right)\psi\right)}{3\beta^4+\left(\beta^6+2\right)\psi^2},\label{SIGMA}\end{aligned}$$ where $D$ is some arbitrary constant, and represents the initial anisotropy conjugate momentum. In a previous study [@Maleknejad:2013npa] this constant was set to $D=0$, effectively reducing the parameter space. As $D$ is exponentially suppressed by the number of efolds $\alpha$, it could be argued that the system should be indifferent to this initial parameter. However, the understanding of this parameter as akin to an initial anisotropy ‘push’ restricts the initial value to be of order $D\sim H$. To evaluate the restrictions on $D$ we first assume inflation takes place. Formally, in a Bianchi Type I geometry this means that $\ddot{\alpha}>0$ and $\dot{\alpha}^2>\dot{\sigma}^2$. Examining the first slow-roll parameter we have: $$\begin{aligned}
\varepsilon\equiv -\frac{\ddot{\alpha}}{\dot{\alpha}^2}=\frac{\tfrac{\rho+P}{2}+3\dot{\sigma}}{\dot{\alpha}^2}\label{eq:epsilon_Axion_SU(2)}.\end{aligned}$$ Defining $\varepsilon_{iso}\equiv \frac{3\left(\rho+P\right)}{2\rho}$, we rewrite this as: $$\begin{aligned}
\varepsilon=\frac{\rho\varepsilon_{iso}}{3\dot{\alpha}^2}+3\left(\frac{\dot{\sigma}}{\dot{\alpha}}\right)^2.\end{aligned}$$ Taking the upper limit on $\dot{\sigma}$ by setting $\varepsilon_{iso}\gtrsim 0$, we have: $$\begin{aligned}
1>\varepsilon\gtrsim 3\left(\frac{\dot{\sigma}}{\dot{\alpha}}\right)^2.\end{aligned}$$ Since all other constituents of the term \[SIGMA\] have to be small during slow-roll as argued in [@Maleknejad:2013npa], and assuming we start at an initial efolding number of $\alpha=0$, if $D$ dominates the term \[SIGMA\], we have $\dot{\sigma}\sim D/3$. This yields $$\begin{aligned}
\varepsilon\gtrsim\frac{D^2}{3\dot{\alpha}^2}.\end{aligned}$$ Thus it is formally possible for $\dot{\sigma}$ to momentarily achieve values comparable to $\dot{\alpha}$. However, this high value should be attenuated almost immediately to a value more suited for a slow-roll inflation. Specifically in these systems it was shown that $\varepsilon\simeq (1-n_s)^2$ [@Adshead:2013qp; @Dimastrogiovanni:2012ew], with $n_s$ being the scalar spectral index. This should push $\varepsilon$ to be of the order of $10^{-3}$, so we would expect $|D| \ll \dot{\alpha}$. Since $D$ is multiplied by a factor of $e^{-3\alpha}$, it was thought that whatever initial value is assigned, $D$ is irrelevant. However, when $D>$ several $\times\;\dot{\alpha}$, the system becomes unstable and the inflationary solution becomes non-viable. This represents a level of tuning previously overlooked [@Maleknejad:2013npa].\
\
Numerical analysis results {#sec:Numerical-Pure_CN}
--------------------------
Our numerical analysis uses the full equations of motion arising from the action \[eq:CN\_action\]. As mentioned before, the axion field is only coupled to the isotropic part of the gauge field, $\psi$. Therefore, the slow-roll relation \[eq:psi\_FRW\] is still valid in the Bianchi spacetime. We can thus use it to simplify the parameter space. We first specify at the onset of inflation $H_0$, $\chi_0,g_{\scaleto{A}{4pt}},\lambda$ and $f$. We derive $\mu$ using the slow-roll condition: $$\begin{aligned}
\mu=\left(\frac{3H_0}{1+\cos{\frac{\chi_0}{f}}}\right)^{1/4},\label{eq:Extract_mu}\end{aligned}$$ and derive $\psi_0$ by: $$\begin{aligned}
\psi_0=\left(\frac{\mu^4\sin{\frac{\chi}{f}}}{3g_{\scaleto{A}{4pt}}\lambda H_0}\right)^{1/3}\; .\label{eq:Extract_psi}\end{aligned}$$ Specifically, we set $H_0$ to the GUT scale $H_0=10^{-6}M_{Pl}$. We assign $\dot{\psi}_0=\dot{\chi}_0=0$ for simplicity. The baseline parameters we use are specified in Table \[tab:Initial\_parmeters\].
---------------- -------------------- ------------------------ ------------------
$H_0$ $10^{-6}$ $f$ 0.1
$\lambda$ 2000 $g_{\scaleto{A}{4pt}}$ $2\cdot 10^{-6}$
$\chi_0$ $\pi\cdot 10^{-3}$ $\dot{\chi}_0$ 0
$\dot{\psi}_0$ 0
$\mu$ $\sim 10^{-3}$ $\psi_0$ $\sim\pi/200$
---------------- -------------------- ------------------------ ------------------
: Baseline parameters for the numerical analysis of the CN system. The upper parameters: $H_0, f, \lambda, g_{\scaleto{A}{4pt}}$, and $\chi_0$ are specified. The parameters at the bottom are derived from the slow-roll conditions. Unless specifically studied, $D=0$. The slow-roll conditions are used only in the initial parameters derivation stage.[]{data-label="tab:Initial_parmeters"}
\
![Trajectories (blue dots) that start with various initial conditions (red,yellow,green X’s) such that $(\beta_0,\frac{\dot{\beta}_0}{H_0})\in(\pm 10,\pm 10)$. Convergence to the attractors at $\beta=\pm 1$ (black X’s) is apparent. While some trajectories initially turn away from the fixed points and anisotropy increases, they all eventually sufficiently isotropize. The embedded figure shows a single trajectory starting at $\beta_0=2,\frac{\dot{\beta}_0}{H_0}=10$.\[fig:10x10original\]](10x10_phase_space_clatter_new.eps){width="85.00000%"}
![A closer look at trajectories (upper right panel) in a smaller area of phase-space around $\beta=0$ reveals a ‘no-go’ area (black striped area) where convergence is not achieved. This signifies that the system exits slow-roll, and enters a ‘runaway’ dynamics where the system cannot support inflation. Further zoom into the area around $(0,0)$ reveals an internal structure of this ‘no-go’ area. The color shows the amount of efolds the system requires to isotropize. The system does not isotropize in the deep blue regions. \[fig:Zoom\_in\]](zoomINfigures_newer.eps){width="85.00000%"}
The fixed point solution is given by $\beta=\pm 1$; however, it is hard to achieve this value numerically. Thus we employ effective convergence thresholds such that if $|\beta|=1\pm \Delta$, for a sufficiently small $\Delta$, with a vanishingly small $\dot{\beta}$ we are in the convergence target area. After enough consecutive integration steps inside the convergence area, we numerically set $\beta=\pm 1$ (corresponding to the correct attractor) and $\dot{\beta}=0$.\
When calculating the number of efolds it takes the system to isotropize we use an additional exit condition of $$\begin{aligned}
\dot{\sigma}\ll \dot{\alpha}.\end{aligned}$$ In practice, numerically we demand $$\begin{aligned}
|\dot{\sigma}(t)|<\dot{\alpha}(t)\cdot 10^{-3}.\end{aligned}$$ We find that the isotropic attractor solutions at $\beta=\pm 1$ apparently hold for the full numerical analysis. Fig \[fig:10x10original\] shows a set of evenly spaced initial conditions at $(\beta,\dot{\beta}/H_0)\in\left(\pm 10, \pm 10\right)$. Trajectories starting at $\beta>0$ $\left(\beta<0\right)$ converge to $\beta=1$ $\left(\beta=-1\right)$, seemingly regardless of initial conditions. However, upon closer inspection we find some voids in the convergence pattern as seen in Fig. \[fig:Zoom\_in\]. By gradually zooming into the area around the unstable point $(\beta,\dot{\beta}/H_0)=(0,0)$ we gain better resolution of these areas of no convergence. We call these ‘no-go’ areas.\
The existence of these areas can be understood by the kinetic terms of the system. We would expect that when $\frac{\dot{\beta}}{\beta}>\dot{\alpha}$, the kinetic term dominates the dynamics, and we exit the slow-roll regime. In this case, $\big|\dot{\beta}\big|,\big|\beta\big|,\big|\dot{\beta}/\beta\big|\rightarrow \infty$ and looking at the expression for the anisotropy source term \[SIGMA\], we see that the $\frac{\dot{\beta}}{\beta}$ term now dominates and we have $$\begin{aligned}
\dot{\sigma}\propto -\dot{\beta}\beta\psi^2 \;.\end{aligned}$$ Therefore, unless $\psi\rightarrow 0$ the system will tend towards further anisotropization. Fig. \[fig:Zoom\_in\] shows the areas of non-convergence around $\beta=0$. Surprisingly, some areas we should expect to fully anisotropize actually do converge to the isotropic attractor.\
Next, we study the effects of initial anisotropy momentum $D$ on the trajectory of the system. For $D$ on the order of the initial $\dot{\alpha}$, the system behaves as if the initial value of $\dot{\beta}$ is changed. When $D$ is positive we see a behaviour consistent with a larger $\dot{\beta}$, and vice-versa. We show this in Fig. \[fig:D\_behaviour\].\
![Trajectories of the CN system in response to introducing a non-zero initial anisotropy source term $D$. We see that as long as $D\sim H_0$ or smaller the response is similar to a larger or smaller initial $\dot{\beta}$ value.\[fig:D\_behaviour\]](D_dynamics.eps){width="85.00000%"}
\
Spectator Axion-SU(2) Sector in Bianchi Type I Spacetime {#sec:Anisotropic-Spectator}
========================================================
We now relegate the axion-SU(2) system to a spectator status, albeit embedded in a Bianchi type I geometry. Again due to the minimal coupling of the inflaton sector, the changes to the energy density and pressure are straightforward, as the inflaton field does not generate anisotropic pressure. $$\begin{aligned}
\rho=\rho_{\scaleto{CN}{4pt}}+\frac{\dot{\phi}^2}{2}+V(\phi)\; ,\\ \nonumber \\
P= P_{\scaleto{CN}{4pt}}+\frac{\dot{\phi}^2}{2}-V(\phi)\; ,\\ \nonumber \\
\Tilde{P}=\Tilde{P}_{\scaleto{A}{4pt}}\; ,\end{aligned}$$ where $\rho_{\scaleto{CN}{4pt}},P_{\scaleto{CN}{4pt}},\Tilde{P}_{\scaleto{A}{4pt}}$ are given in ,, respectively. As in section \[sec:Spectator\_FRW\], we now denote the energy density and pressure as: $$\begin{aligned}
\rho=\rho_{\phi}+\Delta\rho;\hspace{15pt}P=P_{\phi}+\Delta P.\end{aligned}$$ With this notation we can evaluate the slow-roll parameter $\varepsilon$ in this system as: $$\begin{aligned}
\varepsilon=\varepsilon_{\phi}+\varepsilon_{\scaleto{CN}{4pt}}\end{aligned}$$ where $\varepsilon_{\phi}=\frac{3}{2}\left(\frac{P_{\phi}+\rho_{\phi}}{\rho_{\phi}}\right)$, and $\varepsilon_{\scaleto{CN}{4pt}}=\left(\tfrac{\rho_{\scaleto{CN}{4pt}}+P_{\scaleto{CN}{4pt}}}{2}+3\dot{\sigma}^2\right)/\left(\tfrac{\rho_{\phi}}{3}+\dot{\sigma}^2\right)$. This evaluation is valid if, in addition to $\frac{\Delta \rho}{\rho_{\phi}},\frac{\Delta P}{P_{\phi}}\ll 1$ we also have $\frac{\dot{\sigma}}{\rho_{\phi}}\ll 1$. The second slow-roll parameter is given by: $$\begin{aligned}
\eta=\frac{1}{\dot{\alpha}}\frac{\dddot{\alpha}}{\ddot{\alpha}}=\frac{1}{\dot{\alpha}}\frac{\partial}{\partial t}\ln{\ddot{\alpha}}.\end{aligned}$$ Thus as in and we have two regions: $$\begin{aligned}
\left\{\begin{array}{llr}
\eta\simeq&\eta_{\phi}+\sqrt{\frac{3}{\rho_{\phi}}}\frac{\partial}{\partial t}\left(\frac{\varepsilon_{\scaleto{CN}{4pt}}}{\varepsilon_{\phi}}\right)&\qquad \varepsilon_{\phi}\ll\varepsilon_{\scaleto{CN}{4pt}}, \\
\eta\simeq&\eta_{\scaleto{CN}{4pt}}+\sqrt{\frac{3}{\rho_{\phi}}}\frac{\partial}{\partial t}\left(\frac{\varepsilon_{\phi}}{\varepsilon_{\scaleto{CN}{4pt}}}\right)&\qquad\varepsilon_{\scaleto{CN}{4pt}}\ll\varepsilon_{\phi},
\end{array}\right.\end{aligned}$$ where this time $\eta_{\scaleto{CN}{4pt}}\equiv\sqrt{\frac{3}{\rho_{\phi}}}\frac{\partial}{\partial t}\ln{(|\rho_{\scaleto{CN}{4pt}}+P_{\scaleto{CN}{4pt}}+6\dot{\sigma}^2|)}$, due to the different geometry.\
We thus inherit the same analysis from the spectator system in FRW geometry regarding the stability conditions.
Numerical results
-----------------
![ Three trajectories with different inflaton-to-axion potential ratios $R$ and the identical initial conditions. We see that the system isotopizes more quickly, and stabilizes around a slower-rolling dynamics when $V(\phi)$ is more dominant. \[fig:R-trajectories\]](Trajectories_withR.eps){width="85.00000%"}
We now use the action in \[eq:CNspect\_action\], with a sufficiently flat inflaton potential that supports the required number of efolds. We use the CN realization previously discussed, and ensure we have an overall GUT scale inflation by fixing $H_0=10^{-6}M_{Pl}$. We first define $R$ as the ratio between the inflationary and the axion potentials at the onset of inflation: $$\begin{aligned}
R\equiv\frac{V(\phi_0)}{\mu^4\left(1+\cos\frac{\chi_0}{f}\right)}.\end{aligned}$$ Thus with $R,H_0=10^{-6},\chi_0=\pi\cdot 10^{-3},\phi_0=0$ and $f=0.1$, we can extract $\mu$ and $V(\phi_0)$ by : $$\begin{aligned}
\mu=\left(\frac{3H_0^2}{(1+R)\left(1+\cos\frac{\chi_0}{f}\right)}\right)^{1/4}\; , \\ \nonumber \\
V(\phi_0)=R\mu^4\left(1+\cos\frac{\chi_0}{f}\right)\; .\end{aligned}$$ Additionally, we extract $\psi_0$ from \[eq:Extract\_psi\], with $g_{\scaleto{A}{4pt}}=2\cdot 10^{-6}$ and $\lambda=2000$. The remaining kinetic terms are set to $\dot{\chi}_0=\dot{\psi}_0=0$.\
Every simulated evolution starts with an initial $(\beta,\dot{\beta})$ pair. The overall Hubble parameter is set to $H_0=10^{-6}M_{Pl}$, but it is done in the absence of a kinetic term for $\beta$. Thus, we re-calculate the actual initial Hubble parameter $\dot{\alpha}_0$ including the $\beta,\dot{\beta}$ terms. We then solve the equations of motion numerically for each integration step, while taking care to use the Friedmann equation to check our result. This effectively is a consistency check on our calculations, and typically we have $\dot{\alpha}^2 -\dot{\sigma}^2 -\frac{\rho}{3}=\mathcal{O}(10^{-27})$, which is at the level of machine precision, for the smallest fixed quantity in the system - $\mu^4$.\
![Gradually increasing the inflaton energy density with respect to the spectator axion energy density. The parameter $R$ encodes the domination of the inflaton component over the axion component. We probe the larger phase-space of $(\beta,\dot{\beta}/H_0)\in(10,100)$ to look at the limits of the basin of attractor. The phase-space convergence plots show that as the inflaton becomes more dominant, the system becomes indifferent to initial conditions, and readily converges to the isotropic attractor solution. The color bar shows the number of efolds it takes the system to isotropize.[]{data-label="fig:TurningOnInflaton"}](R_convergence_white.eps){width="85.00000%"}
\
We employ the same numerical methods and exit conditions as in section \[sec:Numerical-Pure\_CN\]. However, when one wishes to compare systems with different $R$ values, a fixed threshold is needed; thus we use $$\begin{aligned}
|\dot{\sigma}(t)|<H_0\cdot 10^{-9}=10^{-15}, \end{aligned}$$ where $H_0$ is fixed regardless of $R$, and $\dot{\alpha}(t)$ is time-dependent and sensitive to the system’s specific setup. We first look at a few trajectories, studying the effect of an increase of $R$. In figure \[fig:R-trajectories\], we observe that as $R$ increases, trajectories tend to stabilize (top left panel). This is due to the dominance of $V(\phi)$ over the axion potential scale $\mu$. We also observe that the system isotropizes more quickly (top right panel) with the increase of $R$, and the inflationary dynamics of $\dot{\alpha}$ more closely resembles a slow-roll inflation (bottom left panel).\
![Anisotropy phase space for $R=1000$. While we see that all regions that converge to the attractor solution do so in $\sim 6$ efolds, we still observe the persistence of the ‘no-go’ areas. This is due to the kinetic term ‘hand-off’ mechanism.[]{data-label="fig:R1000"}](R1000Con.eps){width="85.00000%"}
\
We looked at a larger phase-space than previously studied to probe the limits of the basin of attraction. By simulating anisotropy phase-spaces with different $R$ values, we find that the system gradually stabilizes into an inflaton dominated one ,i.e., the usual scalar-driven slow-roll inflation. Fig. \[fig:TurningOnInflaton\] shows this, as even the larger phase-space, where $(\beta_0,\frac{\dot{\beta}_0}{H_0})\in (\pm 10,\pm 100)$ supports the attractor isotropic solution.\
Even so, there are still locations in the anisotropy phase-space where initial conditions violate slow-roll inflation. This is due to the gauge field kinetic energy term dependence on $\beta$. Recall that $\beta=0$ is the singularity point for the $\psi$ and $\beta$ field equations. Sufficiently close to $\beta=0$ the kinetic term part that is proportional to $\frac{\dot{\beta}^2}{\beta^6}$ becomes unbounded, violates slow-roll, and drives the system to large values of $\beta$. At that point there are two options. Either the system stabilizes and converges to $\beta=\pm 1$ albeit slowly, or the large value of $\dot{\beta}$ is ‘handed-off’ to the kinetic term that is proportional to $\dot{\beta}^2$. In turn this feeds the anisotropy momentum, such that in the limit of large $\beta,\dot{\beta}$ we have: $$\begin{aligned}
\left\{\begin{array}{llc}
\frac{\dot{\sigma}}{\dot{\alpha}}\simeq&-\frac{1}{3}\left(1+\tfrac{\dot{\beta}}{\beta\dot{\alpha}}\right)\left(\beta\psi\right)^2&\qquad \beta\gg1 \\
\frac{\dot{\sigma}}{\dot{\alpha}}\simeq&\frac{1}{3}\left(1-\tfrac{2\dot{\beta}}{\beta\dot{\alpha}}\right)\left(\frac{\psi}{\beta^2}\right)^2&\qquad \beta\ll1
\end{array}\right.\end{aligned}$$ The energy density in the gauge-field sector goes like $$\begin{aligned}
\rho_{\scaleto{A}{4pt}}\sim \left( \frac{\dot{\beta}}{\beta}\right)^2\left(\frac{2}{\beta^4}+\beta^2\right)\psi^2.\end{aligned}$$ Therefore, in the regions where either $\beta\rightarrow 0$, or $\dot{\beta}\gtrsim\beta\gg\dot{\alpha}$ the system becomes mathematically unbounded. Adding the inflaton potential may go a long way to stabilize it, but ultimately, these ‘no-go’ areas should not be expected to vanish altogether. We see that even at $R=1000$, i.e., the system is highly dominated by the inflationary potential, these ‘no-go’ areas still persist when $\dot{\beta}/H_0\gg 1$. This area has not been previously studied. This is shown in figure \[fig:R1000\].\
\
One may suspect these areas coincide with areas where the energy density in the CN sector, including the kinetic terms, initially dominates the inflaton sector. We define $\Tilde{q}$ as the ratio between the energy densities such that: $$\begin{aligned}
\Tilde{q}=\frac{\rho_{\phi}}{\rho_{\scaleto{CN}{4pt}}},\end{aligned}$$ where the energy densities now contain all relevant kinetic terms. In figure \[fig:Panels\] we show that the ‘no-go’ areas do not fully correlate to areas where the CN energy density dominates, even in the presence of kinetic terms. In other words, the ‘no-go’ area is a dynamical effect due to the instability of the system around $\beta\sim 0$ and $\dot{\beta}/H_0\gg1$.
{width="85.00000%"}
(black). We see no correlation between the ‘no-go’ area and areas of initial high $\varepsilon$ or low $\Tilde{q}$.\[fig:Panels\]
\
Moreover, except along a narrow strip along the $\beta\sim 0$ region, the system starts at slow-roll but fails to isotropize, even when the energy in the inflaton sector greatly dominates that of the CN sector.
Conclusion {#sec:Conclusion}
==========
We have studied an axion-SU(2) gauge field system embedded in a homogeneous but anisotropic Bianchi type I geometry. In both of our realizations of the CN system [@Adshead:2012kp; @Dimastrogiovanni:2016fuu] we parametrize the gauge field deviation from isotropy by the dimensionless parameter $\beta(t)$ [@Maleknejad:2013npa]. In this parametrization the prefect isotropic state is given by $\beta=\pm 1$, and a singularity exists at $\beta=0$. We have solved these systems numerically, using the full equations of motion and an energy consistency condition. We have investigated the attractor’s basin of attraction to the homogeneous isotropic solution and confirmed that the system has a single attractor solution in each half-plane (corresponding to $\beta=\pm 1$).\
\
We have found that a non-zero initial anisotropy source term $D$ effectively shifts the initial conditions of the system. Thus, contrary to previous claims based on its exponential suppression, the initial value for $D$ cannot exceed several times the initial Hubble parameter $\dot{\alpha}$. When $D>$ a few $\dot{\alpha}$ the system exits slow-roll either directly, or through the ‘hand-off’ mechanism.\
\
We additionally explored a part of the $\dot{\beta}-\beta$ phase space which has not been previously studied. This regime, e.g. the vicinity of $\beta=0$ with a large kinetic term ($\dot{\beta}/H \gg 1$), contains a no-go area in which the anisotropies do not isotropize and the large kinetic energy in the anisotropies terminates inflation after a few efolds at the most. Thus these areas fail to support the necessary 60 or so efolds of inflation required to account for observations.\
\
In the case of the spectator axion-SU(2) system [@Dimastrogiovanni:2016fuu], stability is significantly increased. The anisotropies decay faster and the system settles into the attractor solution in fewer efolds. The basin of attraction can additionally support larger kinetic terms, and the no-go areas substantially shrink.\
Therfore, we conclude that the spectator axion-SU(2) model is phenomenologically more attractive.
Acknowledgements {#acknowledgements .unnumbered}
================
This reasearch was supported in part by the Excellence Cluster ORIGINS which is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC-2094 – 390783311.\
EK thanks Shinji Mukohyama for the question he asked during the ‘General Relativity - The Next Generation’ conference held at Yukawa Institute for Theoretical Physics from February 19th to 23rd, 2018, which led to this project.\
IW thanks the Minerva Foundation and the Max Planck Society, for supporting his research.
[9]{} A. H. Guth, “The Inflationary Universe: A Possible Solution to the Horizon and Flatness Problems,” Phys. Rev. D [**23**]{} (1981) 347 \[Adv. Ser. Astrophys. Cosmol. [**3**]{} (1987) 139\]. K. Sato, “First Order Phase Transition of a Vacuum and Expansion of the Universe,” Mon. Not. Roy. Astron. Soc. [**195**]{} (1981) 467. A. D. Linde, “A New Inflationary Universe Scenario: A Possible Solution of the Horizon, Flatness, Homogeneity, Isotropy and Primordial Monopole Problems,” Phys. Lett. [**108B**]{} (1982) 389 \[Adv. Ser. Astrophys. Cosmol. [**3**]{} (1987) 149\]. A. Albrecht and P. J. Steinhardt, “Cosmology for Grand Unified Theories with Radiatively Induced Symmetry Breaking,” Phys. Rev. Lett. [**48**]{} (1982) 1220 \[Adv. Ser. Astrophys. Cosmol. [**3**]{} (1987) 158\]. A. D. Linde, “Chaotic Inflation,” Phys. Lett. [**129B**]{} (1983) 177. V. F. Mukhanov and G. V. Chibisov, “Quantum Fluctuations and a Nonsingular Universe,” JETP Lett. [**33**]{} (1981) 532 \[Pisma Zh. Eksp. Teor. Fiz. [**33**]{} (1981) 549\]. A. A. Starobinsky, “Dynamics of Phase Transition in the New Inflationary Universe Scenario and Generation of Perturbations,” Phys. Lett. [**117B**]{} (1982) 175. S. W. Hawking, “The Development of Irregularities in a Single Bubble Inflationary Universe,” Phys. Lett. [**115B**]{} (1982) 295. J. M. Bardeen, P. J. Steinhardt and M. S. Turner, “Spontaneous Creation of Almost Scale - Free Density Perturbations in an Inflationary Universe,” Phys. Rev. D [**28**]{} (1983) 679. A. A. Starobinsky, “Spectrum of relict gravitational radiation and the early state of the universe,” JETP Lett. [**30**]{} (1979) 682 \[Pisma Zh. Eksp. Teor. Fiz. [**30**]{} (1979) 719\]. M. M. Anber and L. Sorbo, “Naturally inflating on steep potentials through electromagnetic dissipation,” Phys. Rev. D [**81**]{} (2010) 043534 \[arXiv:0908.4089 \[hep-th\]\]. N. Barnaby and M. Peloso, “Large Nongaussianity in Axion Inflation,” Phys. Rev. Lett. [**106**]{} (2011) 181301 \[arXiv:1011.1500 \[hep-ph\]\]. N. Barnaby, R. Namba and M. Peloso, “Phenomenology of a Pseudo-Scalar Inflaton: Naturally Large Nongaussianity,” JCAP [**1104**]{} (2011) 009 \[arXiv:1102.4333 \[astro-ph.CO\]\]. M. M. Anber and L. Sorbo, “Non-Gaussianities and chiral gravitational waves in natural steep inflation,” Phys. Rev. D [**85**]{} (2012) 123537 \[arXiv:1203.5849 \[astro-ph.CO\]\]. Y. Akrami [*et al.*]{} \[Planck Collaboration\], “Planck 2018 results. IX. Constraints on primordial non-Gaussianity,” arXiv:1905.05697 \[astro-ph.CO\]. N. Barnaby, J. Moxon, R. Namba, M. Peloso, G. Shiu and P. Zhou, “Gravity waves and non-Gaussian features from particle production in a sector gravitationally coupled to the inflaton,” Phys. Rev. D [**86**]{} (2012) 103508 \[arXiv:1206.6117 \[astro-ph.CO\]\]. R. Namba, M. Peloso, M. Shiraishi, L. Sorbo and C. Unal, “Scale-dependent gravitational waves from a rolling axion,” JCAP [**1601**]{} (2016) 041 \[arXiv:1509.07521 \[astro-ph.CO\]\]. A. Maleknejad, M. M. Sheikh-Jabbari and J. Soda, “Gauge Fields and Inflation,” Phys. Rept. [**528**]{} (2013) 161 \[arXiv:1212.2921 \[hep-th\]\]. Y. Akrami [*et al.*]{} \[Planck Collaboration\], “Planck 2018 results. X. Constraints on inflation,” arXiv:1807.06211 \[astro-ph.CO\]. J. Kim and E. Komatsu, “Limits on anisotropic inflation from the Planck data,” Phys. Rev. D [**88**]{} (2013) 101301 \[arXiv:1310.1605 \[astro-ph.CO\]\].
A. Maleknejad and M. M. Sheikh-Jabbari, “Gauge-flation: Inflation From Non-Abelian Gauge Fields,” Phys. Lett. B [**723**]{} (2013) 224 \[arXiv:1102.1513 \[hep-ph\]\]. A. Maleknejad and M. M. Sheikh-Jabbari, “Non-Abelian Gauge Field Inflation,” Phys. Rev. D [**84**]{} (2011) 043515 \[arXiv:1102.1932 \[hep-ph\]\].
E. Dimastrogiovanni and M. Peloso, “Stability analysis of chromo-natural inflation and possible evasion of Lyth’s bound,” Phys. Rev. D [**87**]{} (2013) no.10, 103501 \[arXiv:1212.5184 \[astro-ph.CO\]\]. P. Adshead, E. Martinec and M. Wyman, “Gauge fields and inflation: Chiral gravitational waves, fluctuations, and the Lyth bound,” Phys. Rev. D [**88**]{} (2013) no.2, 021302 \[arXiv:1301.2598 \[hep-th\]\]. A. Papageorgiou, M. Peloso and C. Unal, “Nonlinear perturbations from the coupling of the inflaton to a non-Abelian gauge field, with a focus on Chromo-Natural Inflation,” JCAP [**1809**]{} (2018) 030 \[arXiv:1806.08313 \[astro-ph.CO\]\]. A. Papageorgiou, M. Peloso and C. Unal, “Nonlinear perturbations from axion-gauge fields dynamics during inflation,” JCAP [**1907**]{} (2019) 004 doi:10.1088/1475-7516/2019/07/004 \[arXiv:1904.01488 \[astro-ph.CO\]\]. K. D. Lozanov, A. Maleknejad and E. Komatsu, “Schwinger Effect by an $SU(2)$ Gauge Field during Inflation,” JHEP [**1902**]{} (2019) 041 \[arXiv:1805.09318 \[hep-th\]\]. L. Mirzagholi, A. Maleknejad and K. D. Lozanov, “Production and Backreaction of Fermions from Axion-$SU(2)$ Gauge Fields during Inflation,” arXiv:1905.09258 \[hep-th\]. A. Maleknejad and E. Komatsu, “Production and Backreaction of Spin-2 Particles of $SU(2)$ Gauge Field during Inflation,” JHEP [**1905**]{} (2019) 174 \[arXiv:1808.09076 \[hep-ph\]\]. A. Maleknejad, “Dark Fermions and Spontaneous $CP$ violation in $SU(2)$-axion Inflation,” arXiv:1909.11545 \[hep-th\]. V. Domcke and K. Mukaida, “Gauge Field and Fermion Production during Axion Inflation,” JCAP [**1811**]{} (2018) 020 \[arXiv:1806.08769 \[hep-ph\]\]. V. Domcke, Y. Ema, K. Mukaida and R. Sato, “Chiral Anomaly and Schwinger Effect in Non-Abelian Gauge Theories,” JHEP [**1903**]{} (2019) 111 \[arXiv:1812.08021 \[hep-ph\]\]. V. Domcke, Y. Ema and K. Mukaida, “Chiral Anomaly, Schwinger Effect, Euler-Heisenberg Lagrangian, and application to axion inflation,” arXiv:1910.01205 \[hep-ph\]. A. Maleknejad and E. Erfani, “Chromo-Natural Model in Anisotropic Background,” JCAP [**1403**]{} (2014) 016 \[arXiv:1311.3361 \[hep-th\]\]. P. Adshead and M. Wyman, “Chromo-Natural Inflation: Natural inflation on a steep potential with classical non-Abelian gauge fields,” Phys. Rev. Lett. [**108**]{} (2012) 261302 \[arXiv:1202.2366 \[hep-th\]\]. E. Dimastrogiovanni, M. Fasiello and T. Fujita, “Primordial Gravitational Waves from Axion-Gauge Fields Dynamics,” JCAP [**1701**]{} (2017) 019 \[arXiv:1608.04216 \[astro-ph.CO\]\]. I. Obata, T. Miura and J. Soda, “Chromo-Natural Inflation in the Axiverse,” Phys. Rev. D [**92**]{} (2015) no.6, 063516 Addendum: \[Phys. Rev. D [**95**]{} (2017) no.10, 109902\] \[arXiv:1412.7620 \[hep-ph\]\]. P. Adshead, E. Martinec, E. I. Sfakianakis and M. Wyman, “Higgsed Chromo-Natural Inflation,” JHEP [**1612**]{} (2016) 137 \[arXiv:1609.04025 \[hep-th\]\].
[^1]: Note that in [@Adshead:2012kp] $\eta_{\scaleto{H}{2pt}}\equiv\frac{d\ln \varepsilon_{\scaleto{H}{2pt}}}{dN}$ whereas we define $\eta_{\scaleto{H}{2pt}}\equiv \frac{\ddot{H}}{H\dot{H}}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Christopher G. Lester'
- Matthias Schott
bibliography:
- 'paper.bib'
title: 'Search for Non-Standard Sources of Parity Violation in Jets at $\sqrt{s}=8$ TeV with CMS Open Data'
---
\[Sec:Intro\]Introduction
=========================
It is rightly beyond doubt that the laws of physics violate parity. The elegant experiments of the 1950s [@Wu:1957my; @Garwin:1957hc; @PhysRev.109.1015] unambiguously showed that the weak interaction of the Standard Model can tell the difference between our universe and its mirror image. However, no attempts have yet been made to identify whether there are non-standard parity violation mechanisms which operate only at high energies probed by the Large Hadron Collider (LHC).
The absence of tests of parity is largely a pragmatic response to the obstacles presented by the LHC: its beams are not polarised, its detectors are not sensitive to polarizations, and it is mathematically impossible to construct a parity violating spin-averaged matrix element within any $CP$-conserving Locally Lorentz-Invariant quantum field theory (LLIQFT), effective or otherwise.[^1] Given that the only route to probing parity-violation in the Standard Model at the LHC would be from within the $C$-even part of its (very small) $CP$-violating sector,[^2] it is not surprising that no such analyses have yet been performed.[^3]
It would be wrong, however, to conclude that genuine tests of parity are therefore of limited value. On the contrary, we will see that genuine tests of parity-invariance are very straightforward to make and can (in principle) provide arbitrarily large signatures for models not fitting into the class of LLIQFTs.
While it is true that there is an overwhelming theoretical preference for LLIQFTs (the Standard Model itself is one, as are most popular extensions including those featuring supersymmetry, leptoquarks, technicolor, axions, additional gauge interactions, [*etc*]{}.) there is no law of nature which [*demands*]{} that new physics be describable only by such theories.
Moreover, given the lack of evidence for new physics found at LHC thus far, the need for the community to search in all possible hiding places is surely greater than ever. In particular, it is hard to imagine any reason why every possible attempt should not be made to test and re-test the fundamental symmetries of nature every time a door opens onto a new energy range.
We use therefore CMS Open Data to make the first LHC search for unequivocal evidence of parity violation. Clearly, the possible physics reach of our study is limited - once by the data statistics used, secondly by the assumed systematic uncertainties; its primary goal is to evaluate potential experimental challenges, while the actual constrains on parity violation are of secondary importance. The Standard Model itself should provide no appreciable signal on account of its approximate $CP$-symmetry. While sensitivity might exist to LLIQFT extensions of the Standard Model having larger sources of $CP$-violation, these are probably constrained by data from rare decays. The primary sensitivity, thus, is to extensions which are outside the LLIQFT class altogether.
The paper is structured as follows: Section \[Sec:Intro\] (this introduction) has explained why genuine tests of parity violation in LHC data are important, and discusses the limitations which the LHC imposes on which sources of parity violation are observable; Section \[Sec:ParityViolation\] discusses the properties of the specific parity-odd event-variable, $\alpha$, used in this paper; Section \[Sec:CMS\] summarises the CMS detector, its relevant reconstruction objects, and the data-sets used in this analysis; Section \[Sec:Calibration\] discusses the calibration of the relevant physics objects used (in particular jets) used in this analysis; Section \[sec:Analysis\] details the actual analysis strategy, signal selection and background estimation; Section \[sec:ResultsAndDatasetAsymmetry\] both defines the test statistic $A_\alpha$ which is used to provide the a measure of non-standard parity violation in the dataset as a whole, and then shows the asymmetries seen in CMS Open Data; Section \[sec:Correction\] unfolds the preceding results for detector effects. The paper concludes in Section \[Sec:Conclusion\].
\[Sec:ParityViolation\]Jet-based Signatures for Non-Standard Parity Violation
=============================================================================
One of the evident disadvantages of searching for the breaking of a particular symmetry without a particular model in mind, is that there the number of potential signatures is almost unbounded. If testing for genuine parity violation, any parity-odd variable which is invariant under the action of relevant non-parity symmetries[^4] has the capacity (assuming a parity-even event selection) to be sensitive to *some* parity-violating model, albeit not necessarily one with a strong theoretical motivation.
Lacking clear guidance from any well motivated non LLIQFT models, we make no claims of generality or optimality concerning our choice to concentrate on a signal region containing three jets. We argue only that the investigation is interesting for the reasons already given in the introduction.
A simple way of constructing a parity-odd variable from jets which is insensitive to rotations is to use a pseudoscalar. This simplest of these is arguably the scalar triple product $[{{\vec p^{{j_1}}}},{{\vec p^{{j_2}}}},{{\vec p^{{j_3}}}}] \equiv ({{\vec p^{{j_1}}}}\times {{\vec p^{{j_2}}}}) \cdot {{\vec p^{{j_3}}}}$, computed from the three jet three-momenta ${{\vec p^{{j_1}}}}$, ${{\vec p^{{j_2}}}}$ and ${{\vec p^{{j_3}}}}$ ordered by the requirement $$\begin{aligned}
p_T({j_1})\ge p_T({j_2}) \ge p_T({j_3})\label{eq:ptordering}.\end{aligned}$$Though the scalar triple product has the beneficial property of changing sign under parity, $({{\vec p^{{j_1}}}}, {{\vec p^{{j_2}}}}, {{\vec p^{{j_3}}}}) \rightarrow (-{{\vec p^{{j_1}}}},- {{\vec p^{{j_2}}}},-{{\vec p^{{j_3}}}})$, it has the disbenefit that it varies not only with the directions but also linearly with the magnitudes of each of the momenta. We therefore choose instead to use a normalised variant whose dependence is primarily on the directions of the momenta:[^5] $$\begin{aligned}
\frac{{{\vec p^{{j_1}}}}\times {{\vec p^{{j_2}}}}}{|{{\vec p^{{j_1}}}}\times {{\vec p^{{j_2}}}}|}
\cdot \frac{{{\vec p^{{j_3}}}}}{ |{{\vec p^{{j_3}}}}|} = \sin\alpha.\end{aligned}$$ We have denoted this quantity ‘$\sin \alpha$’ since it may be interpreted as the sine of the angle $\alpha$ which momentum ${{\vec p^{{j_3}}}}$ makes with the plane spanned by ${{\vec p^{{j_1}}}}$ and ${{\vec p^{{j_2}}}}$, as shown in Figure \[fig:Alpha\]. Positive angles $\alpha$ occur if and only if ${{\vec p^{{j_3}}}}$ is on the same side of that plane as ${{\vec p^{{j_1}}}}\times {{\vec p^{{j_2}}}}$.
![Dependence of $\alpha$ on the longitudinal velocity $\beta$ of the three-jet system for the example configuration described in the text. \[fig:BetaDependence\] ](Figures/Plot_Plane.pdf){width="1.0\linewidth"}
![Dependence of $\alpha$ on the longitudinal velocity $\beta$ of the three-jet system for the example configuration described in the text. \[fig:BetaDependence\] ](Figures/Plot_BetaStarDependence){width="0.93\linewidth"}
Note that magnitude of $\sin\alpha$ is not invariant under longitudinal boosts of the system of three jets. Even the sign of $\alpha$ can be altered by longitudinal boosts for certain jet momenta. Consider, for example, a co-planar set of jet momenta such as ${{\vec p^{{j_1}}}}=(3,0,0)$, ${{\vec p^{{j_2}}}}=(0,2,0)$ and ${{\vec p^{{j_3}}}}=(1,1,0)$. As these are co-planar they have $\alpha=0$. If those momenta are then given a longitudinal boost, the value of $\sin\alpha$ ceases to be zero, but boosts in one direction generate equal and opposite values of $\alpha$ to boosts in the other direction.[^6] Consequently, any search for a new physics process which generates an $\alpha$-asymmetry in the frame of the hard process must contend with the possibility that such an asymmetry might be partly ‘washed away’ by the longitudinal boosts which relate the hard-process frame (HPF) to the lab frame (LF). Evidently, the sort of processes most easily washed away would be those which have three co-planar (or almost co-planar) final-state momenta in the HPF and are generated from initial states (such as gluon-gluon or quark-quark) whose hard processes have symmetrically distributed longitudinal boosts.
Given that a large fraction of the total LHC cross section comes from gluon-gluon and given that conservation of momentum would require co-planarity in the HPF of the outgoing partonic momenta in any $2\rightarrow 2$ or $2\rightarrow 3$ parton-level process, one might be concerned that $\alpha$ appears to lack sensitivity to parity-violating effects in some important places!
Such concern is, however, premature. The *only* way in which a set of co-planar momenta can have a notion of handedness akin to parity is if the plane in which they live is orientable.[^7] It is for the same reason that the (planar) letter **R** can be distinguished from its mirror image only if *one* side of the paper on which both are printed is deemed ‘special’. The plane separating two *identical* interacting initial-state partons in the HPF is necessarily *non*-orientable.[^8] Therefore, although $2\rightarrow 2$ and $2\rightarrow 3$ processes using gluon-gluon and quark-quark initial states are those in which one might have feared the aforementioned ‘wash out’, these same initial states are necessarily those which can carry through no notion of handedness from the intial to the final state in the HPF anyway, let alone one that could be washed away by symmetric boosts from that frame to the LF [^9].
In contrast, asymmetric initial states (such as quark-gluon) *are* capable of distinguishing forward and backward directions in the HPF, and so *are* (at least in principle) able to generate detectable lab-frame asymmetries in $\alpha$. It is important to realise that such asymmetries could be seen even for Lorentz- and rotationally-invariant $2\rightarrow 3$ new-physics processes having final state momenta which are co-planar in the HPF. It is differences between the forward and backward boosts from parton distribution functions which would drive such an asymmetry, *exploiting* the dependence of $\alpha$ on longitudinal boosts. Given the asymmetry is generated by this mechanism, there is some motivation for biasing the signal region to events in which the three-jet system has a high absolute rapidity.
In summary: the three-jet variable which we have called $\alpha$ can (in principle) detect parity-violating effects in a broad class of new-physics models at the LHC. As the particular variable we used to illustrate our search is just one of many that could have been chosen, we make no claims of optimality – indeed we have highlighted many situations in which it has no sensitivity to anomalous parity violation whatsoever. Nonetheless, we have: (i) demonstrated that parity violating models to which our variable *is* sensitive do not need to abandon the equivalence principle, Lorentz invariance or isotropy, and (ii) noted that since $\alpha$ is defined in terms of only three objects the parity violating processes which $\alpha$ is sensitive to would probably include only those which couple to asymmetric initial states, such as quark-gluon.
In future work, it may be interesting to consider other event variables which depend on *two* or *four* (rather than three) final state momenta. A simple variable of the first sort might be the product $\Delta(\phi) \Delta(\eta)$ of the azimuthal and pseudorapidity differences between any two final state objects.[^10] Variables acting on four objects could show sensitivity to non-standard parity violating effects which do not need the initial state to provide an reference direction. One example would be the four-jet (longitudinal boost invariant) event-variable $\delta$ defined by: $$\begin{aligned}
\delta \equiv \frac {
p^\mu({j_1})
p^\nu({j_2})
p^\sigma({j_3})
p^\tau({j_4})
\epsilon_{\mu\nu\sigma\tau}
} {
p_T({j_1})
p_T({j_2})
p_T({j_3})
p_T({j_4})
}.
\label{eq:delta}\end{aligned}$$ The numerator of $\delta$ is the archetypical Lorentz-invariant pseudoscalar, so is certainly invariant with respect to longitudinal boosts and capable of demonstrating non-standard parity violating effects. The numerator makes $\delta$ dimensionless, and is also invariant under longitudinal boosts because it is built exclusively from transverse quantities. A disadvantage of $\delta$, however, is that it requires the inputs to have been ordered in some way, as in (\[eq:ptordering\]). It is by no means a requirement that non-standard parity violating effects should come together with an ordering requirement on momenta. It may also be useful, therefore, to consider completely symmetric pseudoscalars. These are pseudoscalars which (unlike $\delta$) do not change their value if their input momenta are re-ordered. Examples include the four completely symmetric pseudoscalars $S_1$, $S_2$, $S_3$ and $S_4$ defined by $$\begin{aligned}
S_1(p(1),p(2),p(3),p(4)) &=
\sum_{i,j,k,l=1}^3
[i,j,k,l]
(i \cdot j)
(i \cdot j)
(i \cdot k)
,
\\
S_2(p(1),p(2),p(3),p(4)) &=
\sum_{i,j,k,l=1}^3
[i,j,k,l]
(i\cdot j)
(i \cdot k)
(j \cdot l)
,
\\
S_3(p(1),p(2),p(3),p(4)) &=
\sum_{i,j,k,l=1}^3
[i,j,k,l]
(i \cdot i)
(j \cdot j)
(i \cdot k)
\qquad\text{and}
\\
S_4(p(1),p(2),p(3),p(4)) &=
\sum_{i,j,k,l=1}^3
[i,j,k,l]
(i \cdot i)
(j \cdot k)
(i \cdot j),
\label{eq:sfour}\end{aligned}$$ in which $(a\cdot b)$ stands for $p^\mu(a)p_\mu(b)$ and $[a,b,c,d]$ stands for $\epsilon_{\mu\nu\sigma\tau} p^\mu(a) p^\nu(b) p^\sigma(c) p^\tau(d)$. $S_1$ to $S_4$ are not dimensionless, but they could easily be made so in the same way as $\delta$ in (\[eq:delta\]).[^11]
The CMS Detector and Reconstructed Objects,\[Sec:CMS\]
======================================================
The data used in this analysis have been recorded with the CMS detector at the LHC in the year 2012. CMS is a high-energy physics experiment which uses a superconducting solenoid, of 6 m internal diameter, with a magnetic field of 3.8 Tesla. The inner detector (ID) of CMS can reconstruct trajectories of charged particles using a silicon pixel and strip tracker. Electrons and photons are identified and measured in a crystal electromagnetic calorimeter (ECAL), while energies of hadrons or hadronic particle jets are determined in a brass/scintillator hadron calorimeter (HCAL). Muons are identified and measured in the muon system (MS), based on gaseous detectors, which surround the hadronic calorimeter and are embedded in the steel flux-return yoke of the magnet system. CMS uses a right-handed coordinate system. Its origin is defined at the interaction point of the proton collisions, the $x$-axis is pointing towards the center of the LHC, the $y$-axis pointing upwards and the $z$-axis along the counterclockwise-beam direction. The polar angle $\theta$ is measured from the positive $z$-axis, however, mostly expressed in terms of the pseudorapidity $\eta$, defined by $\eta = - \ln(\tan(\theta/2))$. The azimuthal angle $\phi$ is measured in the $x$-$y$ plane. We refer to [@Chatrchyan:2008aa] for a detailed description of the CMS experiment.
The main objects used in this analysis are muons and particle jets, which are briefly discussed in the following. CMS employs a particle-flow algorithm that provides a complete description of the event and identifies electrons, muons, photons, charged hadrons, and neutral hadrons. Muons are reconstructed from a global fit of hits in the MS and the ID, seeded by tracks in the muon system [@CMS-PAS-MUO-10-002]. In this analysis, we require each muon to have a minimal transverse momentum of $p_T>25$ GeV within a pseudo-rapidity range of $|\eta|<2.1$, corresponding to the single muon trigger coverage. In addition, standard quality cuts on the number of hits in the ID and the MS, the $\chi^2$ of the fit as well as on the impact parameters are applied, also following previous CMS measurements [@Chatrchyan:2014mua].
Hadronic jets are reconstructed using an anti-$k_t$ algorithm [@Cacciari:2008gp; @Cacciari:2011ma] with a radius parameter of 0.5 based on particle-flow objects [@CMS-PAS-PFT-10-001; @CMS-PAS-PFT-09-001; @Sirunyan:2017ulk], where the clustering algorithms rejects objects that are coming from a pile-up vertex. A jet area method is used to correct for the remaining pile-up contributions [@Cacciari:2007fd]. Since the four-momenta of particle-flow objects is summed, the jets can be massive, In this analyses, we only study jets with a minimal transverse momentum of $p_T>30$ GeV and a jet rapidity of $|y|<2.4$, since this region allows for the best jet resolution and pile-up rejection. In addition, certain quality criteria on the reconstructed jet properties, such as energy fraction in the ECAL and HCAL or the number of particle-flow objects is applied, following the standard CMS recommendations. Moreover, jets are required to have a minimal distance of $\Delta R>0.5$ to all reconstructed electrons, muons and photons candidates. The reconstructed transverse momenta of jets is used to order them as the $1^{st}$, $2^{nd}$ and $3^{rd}$ jet according to their $p_T$ and denoted as $j_1$, $j_2$ and $j_3$, respectively.
This analysis is built around the CMS Open Data Software Framework available at [@CMS:OpenData]. Version `CMSSW_5_3_32` [@CMS:5:3:32] was used. A dedicated open-source framework, <span style="font-variant:small-caps;">Bacon</span> [@cit:bacon], which was used for several published studies of the CMS Collaboration, e.g. [@Chatrchyan:2014mua], is used to read the Analysis Object Data (AOD), extracting information on reconstructed objects as well as generator level data, if available. The <span style="font-variant:small-caps;">Bacon</span> software framework is also used to apply a GoodRun-List selection provided by the CMS Open Data project [@CMS:GoodRunList], as well as calibration constants for particle jets as well as observables, leading to a separate output-format based on <span style="font-variant:small-caps;">Root</span>-tree objects. The typical event size of one simulated top-quark pair event in the <span style="font-variant:small-caps;">Bacon</span>-output format amounts to 5 kB. For this work, we used an additional software package, which reduces the output files of <span style="font-variant:small-caps;">Bacon</span> further and transforms them into a plain <span style="font-variant:small-caps;">Root</span>-NTuple, denoted as <span style="font-variant:small-caps;">ODNTuple</span> in the following with an average event size of 0.8 kB. Our analysis is based on these <span style="font-variant:small-caps;">ODNTuple</span> data. The full software, which has been used for this work, was previously used and validated for [@Apyan:2019ybx].
Selected Open Data and Simulated Monte Carlo Samples
====================================================
The data acquisition system of CMS records only the event information of collisions with dedicated signatures due to the high-collision rate and the limited bandwidth for data-processing. The data used in this analysis has been collected when significant hadronic activity was present in the event, in particular if the `HLT_PFJet320` trigger had fired, targeted to select events with at least one reconstructed jet with a transverse energy of 320 GeV. In total, files corresponding to 289 from the CMS Open Data [@CMS:QCDData] have been processed, of which 166 fired the `HLT_PFJet320` trigger (Table \[tab:DataSamples\]). The study of a larger data-set was beyond the available computing resources. The integrated luminosity has been calculated using the public available GoodRun-List as well as the corresponding tools. In total, 831,662 triggered events, which pass the GoodRun-List requirement, are used in this analysis.
[l | l | c]{} Data stream / Trigger name & Dataset Name & $\int L dt $ $[\ipb]$\
Incl. Jet-Triggers / `Trigger_HLT_PFJet320` & JetHT\_Run2012C [@CMS:QCDData] & 289 / 166\
An overview of the various signal and background Monte Carlo (MC) samples used in this analysis is given in Table \[tab:MCSamples\], indicating the underlying physics process, the dataset name and the corresponding inclusive cross-section. The inclusive QCD MC was generated with <span style="font-variant:small-caps;">Pythia6</span> [@Sjostrand:2006za] and covers a jet transverse momentum range between 50 and 3000 GeV. In addition, a 4-jet sample is used as alternative QCD sample, which is based on the <span style="font-variant:small-caps;">Alpgen</span> generator [@Mangano:2002ea]. The Drell-Yan processes (W/Z) in the muon decay channel were generated using with the <span style="font-variant:small-caps;">PowhegBox</span> Monte Carlo program [@Alioli:2008gx; @Alioli:2010xd] interfaced to the <span style="font-variant:small-caps;">Pythia</span> v.6.4.26 [@Sjostrand:2006za] parton shower model. All other processes are modeled with the tree-level matrix element event generator <span style="font-variant:small-caps;">MadGraph</span> v5.1.3.30 [@Alwall:2011uj] interfaced with <span style="font-variant:small-caps;">Pythia</span> v6.4.26. In all samples the CT10 PDF set [@Gao:2013xoa] and the Z2\* <span style="font-variant:small-caps;">Pythia</span>6 tune [@Chatrchyan:2013gfi; @Khachatryan:2015pea] are used. The decay of tau-leptons is modeled using the <span style="font-variant:small-caps;">Tauola</span> program [@Davidson:2010rw], while the emission of photons from finale state leptons uses <span style="font-variant:small-caps;">Pythia</span>6. The strong coupling constant $\alpha_s$ has been set to $0.130$ at the $Z$-boson mass scale for all matrix element calculations. The effect of multiple interactions per bunch crossing (pile-up) has been simulated by overlaying MC-generated minimum bias events. The <span style="font-variant:small-caps;">Geant4</span> program was used to simulate the passage of particles through the CMS detector [@Agostinelli:2002hh].
The simulated event samples are reweighted to describe the distribution of the number of pile-up conditions in the data by reweighting the $\rho$ parameter distribution, where $\rho$ denotes the diffuse offset energy density [@Khachatryan:2016kdb]. Moreover, a reweighting of the longitudinal position of the primary $pp$ collision vertex of the MC samples to data has been performed.
[l | l | l]{} Process & Dataset Name & $\sigma$ \[pb\]\
QCD (Inclusive, <span style="font-variant:small-caps;">Pythia6</span>) & QCD\_Pt-15to3000\_TuneZ2star [@CMS:QCD1] & $2\cdot10^{10}$\
QCD (4-Jets , <span style="font-variant:small-caps;">Alpgen</span>) & QCD4Jets\_Pt-400to5600\_TuneZ2Star [@CMS:QCD2] & -\
$pp \rightarrow t\bar{t}+X \rightarrow 2l2\nu2b+ X$ & TTJets\_FullLeptMGDecays\_TuneP11TeV\_8TeV [@CMS:MCTTLep] & 112.3\
$pp \rightarrow t\bar{t}+X \rightarrow 1l1\nu2q2b+ X$ & TTJets\_SemiLeptMGDecays\_8TeV [@CMS:MCTTSemi] & 107.2\
$pp \rightarrow t\bar{t}+X \rightarrow 4q2b+ X$ & TTJets\_HadronicMGDecays\_TuneP11mpiHi\_8TeV [@CMS:MCTTHad] & 25.8\
$pp \rightarrow WW+X \rightarrow 2l2\nu+ X$ & WWJetsTo2L2Nu\_TuneZ2star\_8TeV [@CMS:MCWW] & 5.8\
$pp \rightarrow WZ+X \rightarrow 3l1\nu+ X$ & WZJetsTo3LNu\_8TeV\_TuneZ2Star [@CMS:MCWZ] & 1.1\
$pp \rightarrow Z/\gamma^*+X \rightarrow \mu^+\mu^- + X$ & DYToMuMu\_M-20\_CT10\_TuneZ2star\_v2\_8TeV [@CMS:MCZmumu] & 1931\
\[Sec:Calibration\]Calibration
==============================
Even though the full detector simulation of CMS provides a very good description of the expected event signatures, some remaining differences in reconstruction, trigger and isolation efficiencies as well as in the momentum and energy scales and resolutions between MC simulation and data are present. Dedicated corrections are applied to minimize these differences for the relevant objects in this analysis and are discussed in the following.
The official CMS calibration and corrections for particles jets, in particular the jet energy scale (JES) and the jet energy resolution (JER), has been applied within the <span style="font-variant:small-caps;">Bacon</span> framework. These jet corrections and uncertainties were derived from the simulation, and are confirmed with in situ measurements using the energy balance of dijet and photon+jet events [@Khachatryan:2016kdb]. A reduced set of systematic variations is used to estimate JES and JER uncertainties on the final measurement. In particular, the JES is variated by 2% for $y^{jet}<1.3$ and by 3% for $y^{jet}>1.3$, following [@Khachatryan:2016kdb]. The JER is varied by 20% for $30<{\ensuremath{E_{\mathrm{T}}}}^{jet}<100$ GeV, by 10% for $100<{\ensuremath{E_{\mathrm{T}}}}^{jet}<1$ TeV and by 5% above. The absolute angular $\phi$- and $\eta$-resolution for reconstructed jets with a transverse momentum of 50 GeV is between 0.01 and 0.02 and improves for larger [$p_{\mathrm{T}}$]{} values. We assume a relative uncertainty of 50% on the both angular resolutions.
Events containing exactly two jets with ${\ensuremath{p_{\mathrm{T}}}}>50$ GeV within $y^{jet}<2.4$ are used to test the calibration as well as the assigned uncertainties. A comparison of the ${{\Delta^{\text{rel}} _{{\ensuremath{p_{\mathrm{T}}}}}}}= ( {\ensuremath{p_{\mathrm{T}}}}^{j_{1}}-{\ensuremath{p_{\mathrm{T}}}}^{j_{2}})/( {\ensuremath{p_{\mathrm{T}}}}^{j_{1}}+{\ensuremath{p_{\mathrm{T}}}}^{j_{2}})$ distributions for data and MC, including the assigned systematic uncertainties is shown in Figure \[fig:JetPerformance\], where agreement between data and MC can be seen.
![\[fig:JetPerformance\] Left: Distribution of the ${{\Delta^{\text{rel}} _{{\ensuremath{p_{\mathrm{T}}}}}}}= ({\ensuremath{p_{\mathrm{T}}}}^{j_{1}}-{\ensuremath{p_{\mathrm{T}}}}^{j_{2}})/( {\ensuremath{p_{\mathrm{T}}}}^{j_{1}}+{\ensuremath{p_{\mathrm{T}}}}^{j_{2}})$ in Data and MC. Right: Comparison of the ratio of measured ${\ensuremath{p_{\mathrm{T}}}}(Z)$ and the measured jet energy ${\ensuremath{E_{\mathrm{T}}}}$ for $Z$-boson events in the muon decay channel with exactly one jet with ${\ensuremath{E_{\mathrm{T}}}}>30$ GeV and $20<{\ensuremath{p_{\mathrm{T}}}}(Z)<50$ GeV for Data and MC. All MC-corrections have been applied. The grey band indicates the corresponding systematic uncertainties.](Figures/Plot_QCD_JetStudy_DiJet_Resolution.pdf "fig:"){width="7.3cm"} ![\[fig:JetPerformance\] Left: Distribution of the ${{\Delta^{\text{rel}} _{{\ensuremath{p_{\mathrm{T}}}}}}}= ({\ensuremath{p_{\mathrm{T}}}}^{j_{1}}-{\ensuremath{p_{\mathrm{T}}}}^{j_{2}})/( {\ensuremath{p_{\mathrm{T}}}}^{j_{1}}+{\ensuremath{p_{\mathrm{T}}}}^{j_{2}})$ in Data and MC. Right: Comparison of the ratio of measured ${\ensuremath{p_{\mathrm{T}}}}(Z)$ and the measured jet energy ${\ensuremath{E_{\mathrm{T}}}}$ for $Z$-boson events in the muon decay channel with exactly one jet with ${\ensuremath{E_{\mathrm{T}}}}>30$ GeV and $20<{\ensuremath{p_{\mathrm{T}}}}(Z)<50$ GeV for Data and MC. All MC-corrections have been applied. The grey band indicates the corresponding systematic uncertainties.](Figures/Plot_Zmumu_JetCalib_PTZoOneJet_20_50.pdf "fig:"){width="7.3cm"}
To further validate the applied corrections, $Z$-boson candidate events with an associated production of at least one jet have been selected in data and compared to full simulated Monte Carlo samples. $Z\rightarrow \mu^+\mu^-$ candidate events are selected by requiring events with exactly two oppositely charged, isolated muons with a minimal $p_T$ of 25 GeV within $|\eta|<2.1$. Jets are required to have a minimal transverse momentum of $p_T>30$ GeV within a rapidity range of $|y|<2.4$. The invariant mass of these two lepton candidates has to be between 66 and 116 GeV. This selection ensures a nearly background free selection of $Z$-boson candidates. In order to test the jet calibration, $Z$-boson events with exactly one reconstructed jet are further divided into evens with $20<{\ensuremath{p_{\mathrm{T}}}}(Z)<50$ GeV and $50<{\ensuremath{p_{\mathrm{T}}}}(Z)<100$ GeV. The transverse momentum of the $Z$-boson, precisely measured by its decay leptons, should be balanced in a first approximation by the transverse energy of the select jet, hence the ratio of ${\ensuremath{p_{\mathrm{T}}}}(Z)/{\ensuremath{p_{\mathrm{T}}}}^{jet}$ should peak around 1. The comparison of data and MC of this ratio is also shown in Fig. \[fig:JetPerformance\], where a good agreement within the assigned systematic uncertainties can be seen. This study has been repeated in [@Apyan:2019ybx] with higher jet multiplicities and higher values of ${\ensuremath{p_{\mathrm{T}}}}(Z)$ in the final state, all indicating good closure.
Analysis Strategy, Signal Selection and Background Estimation\[sec:Analysis\]
=============================================================================
It is the goal of this analysis to probe a possible asymmetry between positive and negative values of $\alpha$ (defined in Section \[Sec:ParityViolation\]) in events containing at least three jets in proton-proton collisions at a center of mass energy of $\sqrt{s}=8$ TeV. The dependence of this asymmtry will be investigated in bins of the invariant di-jet mass of the leading and sub-leading jet, $m_{12}$. In addition, the normalized $\alpha$ distribution in data is measured and corrected for detector effects.
Events with at least three jets with a minimal transverse momentum of $p_T>$ 50 GeV within a rapidity range of $|y|<2.4$ are selected for this analysis. Events with a fourth jet with $p_T>30$ GeV within $|y|<3.0$ are vetoed. Standard noise cleaning cuts as well as cuts to reject jets from pile-up collisions are applied. In addition, it is required that the minimal difference between the transverse momentum of the leading and the sub-leading jet be at least 30 GeV, i.e. $\Delta p_T^{j_{1},j_{2}}=p_T^{j_{1}}-p_T^{j_{2}}>$ 30 GeV, and that the minimal transverse momentum difference between the sub-leading and sub-sub-leading be at least 50 GeV, i.e. $\Delta p_T^{j_{2},j_{3}}=p_T^{j_{2}}-p_T^{j_{3}}>$ 50 GeV. These requirements ensure that small changes to jet energies (e.g. from reconstruction effects) do not lead to changes in the $p_T$ ordering of the jets, and so do not result in discrete changes in the orientation of the reference plane. Moreover, we veto fully back-to-back events by requiring that the $\Delta\Phi$ value between $j_1$ and $j_2$ is smaller than 3.0. This selection reduces uncertainty in the normal of the reference plane, since that normal becomes undefined when $j_1$ and $j_2$ are parallel. In total 29,063 events pass this selection. The selected events are further divided along the invariant mass between the leading and sub-leading jet within the ranges $500<m_{12}<700$ GeV, $700<m_{12}<900$ GeV and $900<m_{12}<2000$ GeV as well as for two rapidity regions of the three jet system, $y_{jjj}$, with $|y_{jjj}|<0.7$ and $0.7<|y_{jjj}|$.
Possible contributions to the selected final states involving electroweak decay processes, such as the production of $W+\text{jets}$ and $Z+\text{jets}$ in the hadronic decay mode or the production of $t \bar t$, are estimated by MC simulations and found to be below 0.5 %, limited by the available MC statistics. As an additional cross-check, events with two jets with $p^{j_1}_T>350$ GeV and $p^{j_2}_T>100$ GeV and one additional reconstructed isolated muon with a transverse momentum of at least 50 GeV have been selected. Those events are expected to come from $W/Z$+jets, $t\bar t$ or $WW$ decays, with at least one vector boson decaying leptonically. In total, 31 events pass this selection, of which no events with two muons were found. Assuming that these muons all stem from leptonic $W$-boson decays, we expect $\approx$ 310 electroweak background events, i.e. a background contribution of 1%, which we take as an estimate in the following.
Both QCD MC samples (Table \[tab:MCSamples\]) have been used to study migration effects, i.e. a change in the jet ordering or a wrong charge assignment of $\alpha$ using MC truth information. The minimal [$p_{\mathrm{T}}$]{} difference between the jets as well as the $\Delta\Phi(j_1,j_2)$ requirement results in only 1.5% of events, in which the jet-${\ensuremath{p_{\mathrm{T}}}}$ ordering is different between reconstruction and generator level and 1% of the cases, which lead to wrong sign-assignment of $\alpha$.
In order to improve the description of the MC samples with data, the transverse momenta distribution of the leading and sub-leading jets have been reweighted to data by a two-dimensional approach on an event-by-event basis using all events containing three jets with $p_T>$ 50 GeV within a rapidity range of $|y|<2.4$. The reweighting is based on two dimensional histograms with the ${\ensuremath{p_{\mathrm{T}}}}$ of the leading and sub-leading jet on the axis, which is filled for data and MC. The actual event weights are derived by the ratio of the normalized data histogram over the corresponding normalized histogram of the MC set. It should be noted, that the fiducial volume which is used for the reweighting is significantly larger than the actual fiducial volume of the analysis. Figure \[fig:DataMCReco\] shows the reconstructed $p_T$ distributions after the reweighting procedure of all three jets in all selected events as well as the inclusive distribution of $\alpha$ in data as well as MC simulations, where the distributions are normalized. Data and MC agree well within their uncertainties.
![\[fig:DataMCReco\] Comparison of the [$p_{\mathrm{T}}$]{}distributions for the leading (upper left), sub-leading (upper right) and sub-sub-leading jets (lower left) for data and MC, as well as the inclusive $\alpha$ distribution (lower right) for data and MC. The grey band indicates the corresponding systematic uncertainties.](Figures/Plot_QCD_JetPT1.pdf "fig:"){width="7.3cm"} ![\[fig:DataMCReco\] Comparison of the [$p_{\mathrm{T}}$]{}distributions for the leading (upper left), sub-leading (upper right) and sub-sub-leading jets (lower left) for data and MC, as well as the inclusive $\alpha$ distribution (lower right) for data and MC. The grey band indicates the corresponding systematic uncertainties.](Figures/Plot_QCD_JetPT2.pdf "fig:"){width="7.3cm"} ![\[fig:DataMCReco\] Comparison of the [$p_{\mathrm{T}}$]{}distributions for the leading (upper left), sub-leading (upper right) and sub-sub-leading jets (lower left) for data and MC, as well as the inclusive $\alpha$ distribution (lower right) for data and MC. The grey band indicates the corresponding systematic uncertainties.](Figures/Plot_QCD_JetPT3.pdf "fig:"){width="7.3cm"} ![\[fig:DataMCReco\] Comparison of the [$p_{\mathrm{T}}$]{}distributions for the leading (upper left), sub-leading (upper right) and sub-sub-leading jets (lower left) for data and MC, as well as the inclusive $\alpha$ distribution (lower right) for data and MC. The grey band indicates the corresponding systematic uncertainties.](Figures/Plot_QCD_Reco_Alpha_Full.pdf "fig:"){width="7.3cm"}
Raw Asymmetry Measurements\[sec:ResultsAndDatasetAsymmetry\]
============================================================
The $\alpha$ asymmetry, $A_\alpha$, can be defined as the difference between the sum over the fiducial cross-section in each measurement bin $i$ of $\alpha$ for positive and negative regions of $\alpha$, normalized by the overall cross section, i.e.
$$A_\alpha = \frac{\sum_{i, \alpha>0} d\sigma_i/\sigma^{fid} - \sum_{i, \alpha>0} d\sigma_i/\sigma^{fid}}{\sum_{i, \alpha>0} d\sigma_i/\sigma^{fid} + \sum_{i, \alpha>0} d\sigma_i/\sigma^{fid}} = \frac{\sum_{i, \alpha>0} d\sigma_i/\sigma^{fid} - \sum_{i, \alpha>0} d\sigma_i/\sigma^{fid}}{\sigma}$$
The asymmetry $A_\alpha$ can be approximated by a pure data driven quantity $A^{data}_\alpha$, that reduces to a pure counting exercise: $$A^{data}_\alpha = \frac{N_{\alpha<0} - N_{\alpha>0}}{N_{\alpha<0} + N_{\alpha>0}},$$ where $N$ is the number of events in a particular region of $\alpha$. In the following we assume that $A^{data}_\alpha$ is, to a good approximation, equal to $A_\alpha$, since detector related resolution and efficiency effects are expected to be symmetric in $\alpha$, i.e. affect events with positive and negative values of $\alpha$ in the same way and hence cancel in the ratio. This was tested and confirmed using MC simulations. The resulting asymmetries $A_\alpha$ for the full inclusive selection, as well as for the three invariant mass regions of the leading and sub-leading jet, are shown in Table \[tab:AlphaResults\]. The separation in different invariant di-jet masses allows to probe energy dependencies of parity violating effects. All asymmetries are compatible with the Standard Model expectation of zero, within the statistical uncertainties.
[l |c |c |c ]{} Selection & Number of Events ($\alpha<0$) & Number of Events ($\alpha>0$) & Asymmetry $A_\alpha$\
$m_{jj}$ inclusive & 14661 & 14402 & $0.009 \pm 0.006$\
500<$m_{jj}$<700 GeV & 3211 & 3189 & $0.003 \pm 0.012$\
700<$m_{jj}$<900 GeV & 5615 & 5483 & $0.011 \pm 0.009$\
900<$m_{jj}$<2000 GeV & 5646 & 5551 & $0.008 \pm 0.009$\
$|y_{jjj}|$<0.7 & 10192 & 9993 & $0.010 \pm 0.007$\
0.7<$|y_{jjj}|$ & 4469 & 4409 & $0.007 \pm 0.011$\
A partial asymmetry can be calculated for different regions of $\alpha$, for example by comparing the number of events in $[ -0.96, -0.64]$ to the number of events in $[0.64, 0.96]$. This is graphically illustrated for the inclusive selection as well as the three $m_{jj}$ mass-ranges and two $y_{jjj}$ rapidity ranges in Figure \[fig:AsymmetryReco\] and Figure \[fig:AsymmetryRecoY\], respectively. While the upper part of each plot shows the actual number of reconstructed events with positive and negative values of $\alpha$, the lower part shows the resulting partial asymmetry including the statistical uncertainty. In general a good agreement with the SM expectation is seen and no deviation above $2.5\sigma$ is observed.
![\[fig:AsymmetryReco\] Distribution of reconstructed events with $\alpha<0$ and $\alpha>0$ for all events (upper right), as well as three regions in the invariant mass $m_{12}$; upper right: $0<m_{12}<700$ GeV, lower left: $700<m_{12}<900$ GeV, lower right: $900<m_{12}<2000$ GeV. The ratio in each plots shows the asymmetry value including statistical uncertainties.](Figures/Plot_QCD_Asymmetry_M_All.pdf "fig:"){width="7.3cm"} ![\[fig:AsymmetryReco\] Distribution of reconstructed events with $\alpha<0$ and $\alpha>0$ for all events (upper right), as well as three regions in the invariant mass $m_{12}$; upper right: $0<m_{12}<700$ GeV, lower left: $700<m_{12}<900$ GeV, lower right: $900<m_{12}<2000$ GeV. The ratio in each plots shows the asymmetry value including statistical uncertainties.](Figures/Plot_QCD_Asymmetry_M_0_700.pdf "fig:"){width="7.3cm"} ![\[fig:AsymmetryReco\] Distribution of reconstructed events with $\alpha<0$ and $\alpha>0$ for all events (upper right), as well as three regions in the invariant mass $m_{12}$; upper right: $0<m_{12}<700$ GeV, lower left: $700<m_{12}<900$ GeV, lower right: $900<m_{12}<2000$ GeV. The ratio in each plots shows the asymmetry value including statistical uncertainties.](Figures/Plot_QCD_Asymmetry_M_700_900.pdf "fig:"){width="7.3cm"} ![\[fig:AsymmetryReco\] Distribution of reconstructed events with $\alpha<0$ and $\alpha>0$ for all events (upper right), as well as three regions in the invariant mass $m_{12}$; upper right: $0<m_{12}<700$ GeV, lower left: $700<m_{12}<900$ GeV, lower right: $900<m_{12}<2000$ GeV. The ratio in each plots shows the asymmetry value including statistical uncertainties.](Figures/Plot_QCD_Asymmetry_M_900_2000.pdf "fig:"){width="7.3cm"}
![\[fig:AsymmetryRecoY\] Distribution of reconstructed events with $\alpha<0$ and $\alpha>0$ for events with $|y_{jjj}|<0.7$ (left) and $0.7<|y_{jjj}|$ (right). The ratio in each plots shows the asymmetry value including statistical uncertainties.](Figures/Plot_QCD_Asymmetry_Y_00_07.pdf "fig:"){width="7.3cm"} ![\[fig:AsymmetryRecoY\] Distribution of reconstructed events with $\alpha<0$ and $\alpha>0$ for events with $|y_{jjj}|<0.7$ (left) and $0.7<|y_{jjj}|$ (right). The ratio in each plots shows the asymmetry value including statistical uncertainties.](Figures/Plot_QCD_Asymmetry_Y_07_20.pdf "fig:"){width="7.3cm"}
Results Corrected \[sec:Correction\] for Detector Effects
=========================================================
Jets have also been defined at the generator level, by clustering final-state particles with decay length $c\tau$ > 10 mm, using the anti-$k_t$ algorithm [@Cacciari:2008gp] with radius parameter R = 0.5. The generator-level jets are used to define the fiducial volume of this measurement, by selecting events fulfilling the same kinematic requirements at the generator level as at the reconstruction level.
The normalized three-jet production cross-section, is measured as a function of $\alpha$ in the binning $[-1.60, -1.28, -0.96, -0.64, -0.32, 0.00, 0.32, 0.64, 0.96,1.28, 1.60]$.[^12]
The normalized differential distribution within the fiducial volume is corrected for detector effects and bin-to-bin migrations using an iterative Bayesian unfolding method, with two iterations [@DAgostini:1994fjx; @DAgostini:2010hil]. First, the data are corrected for background contributions as well as events that pass the detector-level selection but not the particle-level selection. Then, the iterative Bayesian unfolding technique is used as a regularized matrix inversion to correct for the detector resolution in events that pass both the detector-level and particle-level selections. The response matrix (Figure \[fig:UnfoldingMatrix\]) which connects the distribution at reconstruction and particle level is estimated using the inclusive QCD MC sample. After the application of the response matrix, a final correction is applied to account for events that pass the particle-level but not detector-level selection. The chosen binning of $\alpha$ ensures a purity of over 90% in each bin.
Statistical uncertainties are estimated by toy variations of the input data statistics, systematic uncertainties are estimated by varying each source of uncertainty and repeating the full unfolding procedure. For the studies of the systematic effects of JES and JER related uncertainties, several scenarios are considered: in the full-correlation scenario, we apply the systematic uncertainties equally to all reconstructed jet candidates of each event. In the no-correlation scenario, we apply the systematic variations to each jet individually, but leave the other jet candidates of the event unaffected[^13]. The raw-calibration scenario is based on jet kinematics, without any applied calibrations. The difference to the nominal results is calculated for each scenario and taken as systematic. The systematic uncertainties due to the angular resolutions are estimated by adding an additional smearing on the reconstructed $\eta$ and $\phi$ values, following Section \[Sec:Calibration\], and then repeating the unfolding procedure. The modeling bias of the unfolding procedure has been tested in two different ways. First, the MC truth level distribution of $\alpha$ was reweighted to the data distributions and then the unfolding procedure repeated. Second, the 4-jet QCD sample was used as a MC prior. The differences between the nominal and the alternative result was taken as systematic uncertainty and added in quadrature.
The normalized unfolded distributions of $\alpha$ are shown in Figure \[fig:UnfoldedResults\] together their overall uncertainties. The numerical values, including a detailed split-up of the associated uncertainties is given in Table \[tab:AlphaResultsUnfolded\]. The total uncertainty ranges between 3% for the central bins to 13% for the bins corresponding to large absolute values of $\alpha$. While the central bins are dominated by uncertainties of the jet energy scale systematics, the bins for $|\alpha|>1$ are dominated by the limited statistics of data and MC as well as modeling uncertainties. It is interesting to note, that these modeling uncertainties are not due from migration effects of $\alpha$, rather than from efficiency corrections that have to be applied between the reconstruction and truth fiducial volumes.
![Unfolded, detector-corrected $\alpha$ distribution with statistical and systematic uncertainties. For a conservative interpretation, the uncertainties should treated as uncorrelated.\[fig:UnfoldedResults\]](Figures/Plot_QCD_UnfoldingMatrix.pdf){width="1.0\linewidth"}
![Unfolded, detector-corrected $\alpha$ distribution with statistical and systematic uncertainties. For a conservative interpretation, the uncertainties should treated as uncorrelated.\[fig:UnfoldedResults\]](Figures/Plot_QCD_Unfolding.pdf){width="1.0\linewidth"}
[l |c |c |c |c |c |c |c |c |c]{} Bin & Value & Stat. & MC-Stat. & Model & Jet & Jet & Jet $\eta$ & Jet $\phi$ & Total\
& & & & & (Uncor.) & (Cor.) & Res. & Res. &\
(-1.60,-1.28) & $0.010\pm0.001$ & 4.3% & 4.9% & 7.1% & 7.2% & 5.3% & 2.5% & 0.8% & 13.5%\
(-1.28,-0.96) & $0.043\pm0.002$ & 2.2% & 2.2% & 0.5% & 2.2% & 2.8% & 0.4% & 1.1% & 5.0%\
(-0.96,-0.64) & $0.097\pm0.005$ & 1.5% & 1.4% & 3.3% & 2.8% & 0.0% & 0.5% & 0.6% & 4.9%\
(-0.64,-0.32) & $0.185\pm0.005$ & 1.1% & 1.0% & 0.7% & 1.0% & 1.7% & 0.3% & 0.1% & 2.6%\
(-0.32,0.00) & $0.165\pm0.005$ & 1.1% & 1.2% & 1.7% & 1.9% & 0.8% & 0.3% & 0.1% & 3.2%\
(0.00,0.32) & $0.164\pm0.005$ & 1.1% & 1.1% & 1.7% & 1.9% & 0.8% & 0.3% & 0.1% & 3.2%\
(0.32,0.64) & $0.179\pm0.005$ & 1.1% & 1.0% & 0.7% & 1.0% & 1.7% & 0.3% & 0.1% & 2.7%\
(0.64,0.96) & $0.100\pm0.005$ & 1.4% & 1.4% & 3.3% & 2.8% & 0.0% & 0.5% & 0.6% & 4.9%\
(0.96,1.28) & $0.044\pm0.002$ & 2.2% & 2.1% & 0.4% & 2.1% & 2.8% & 0.5% & 1.1% & 4.9%\
(1.28,1.60) & $0.012\pm0.001$ & 4.1% & 4.6% & 6.8% & 7.1% & 5.2% & 2.3% & 0.8% & 13.0%\
\[Sec:Conclusion\]Conclusion
============================
We have proposed a way of constraining non-standard sources of parity violation through measurements of the angular distribution, $\alpha$, of the radiation angle of the lowest energetic jet in three-jet events, and have trialled proposal with using CMS open data proton-proton collisions at a center of mass energy of 8 TeV. Potential parity violating effects beyond the Standard Model could lead to an asymmetry in the $\alpha$ distribution, which can be quantified by an asymmetry parameter, $A_\alpha$, which is expected to the zero in the SM. $A_\alpha$ was measured on data in three different kinematic regions, yielding to values which are consistent with zero. No obvious experimental challenges or limitations have been encountered, hence the sensitivity of this measurement could be dramatically improved with higher statistics on data and a better estimate of the corresponding experimental uncertainties. The models to which it is sensitive to could be changed by moving to other variables which probe the final state only, such as the variables $\delta$ and $S_1$ to $S_4$ defined in equations (\[eq:delta\]) to (\[eq:sfour\]). The limitations of the unfolded distribution can be overcome by larger MC samples as well as a more careful evaluation of the associated model uncertainties. We therefore encourage LHC collaborations to perform similar studies in upcoming full Run-2 analysis a center of mass energy of 13 TeV.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank the CMS collaboration for providing the full 2012 data-set as well as for the documentation on the CMS detector performance. This work would have not been possible without the excellent performance of the LHC as well as the existing computing infrastructure and the support from CERN. M.S. would like to thank, in addition, the Fulbright commission as well as the Volkswagen Foundation for the support of this work. Moreover, he would like to thank his previous colleagues at MIT, in particular Aram Apyan, Philip Harris and his host Markus Klute for answering all questions regarding the treatment of the CMS Open Data for this project as well as the pleasant environment during the Fulbright research scholarship. C.G.L. acknowledges fruitful and related discussions with members of the Cambridge Supersymmetry Working Group: notably with Sophie Renner, Thibaut Mueller, Ben Nachman and Thomas Gillam in 2013, and more recently with Zachary Hulcher and Rupert Tombs.
[^1]: To reveal the existence of parity-violation a model must possess at least one matrix element having both a parity-even and a parity-odd part. After trace identities have removed spinor sums, the only parity-odd expressions which can remain in a Lorentz-invariant $|M|^2$ are contractions of the totally antisymmetric alternating tensor with groups of four linearly independent four-momenta: $\epsilon_{\mu\nu\sigma\tau}a^\mu b^\nu c^\sigma d^\tau$. \[fn:epsilon\] While such terms are parity-odd, they are also time-odd. Assuming $CPT$-symmetry, such a matrix element therefore also violates $CP$. $CP$-conserving local Lorentz-invariant quantum field theories therefore cannot generate parity violating differential cross sections.
[^2]: One could, in principle, demonstrate parity-violation unambiguously by using a ‘genuine $CP$-odd’ observable (such as one of those described in [@Han:2009ra]) on a $C$-even final state.
[^3]: This limitation does not prevented the LHC from making measurements of parity-violating parameters within models [*in which a particular mechanism of parity-violation is present by assumption*]{}. For example, the differences between the axial and vector couplings of the $Z$-boson in the Standard Model violate parity and were measured in [@Aad:2015uau; @Khachatryan:2016yte]. However, neither of these papers incontrovertibly demonstrates that nature violates parity
The reason is simple: the angles from which forward-backward asymmetries are calculated are even under parity, unlike primary observables from the experiments of the 1950s. The very same forward backward asymmetries could therefore also be explained, at least in principle, by some alternative parity [*conserving*]{} theory.
[^4]: For example: an ideal parity-testing variable should be invariant under rotation of an LHC event by any angle about the beam pipe axis, or by 180 degrees about any axis at right angles to the beam pipe, unless sensitivity to Lorentz violating effects is desired.
[^5]: The quantity $\sin\alpha$ does not depend [*exclusively*]{} on the directions of ${{\vec p^{{j_1}}}}$, ${{\vec p^{{j_2}}}}$ and ${{\vec p^{{j_3}}}}$ because (\[eq:ptordering\]) uses other properties of ${{\vec p^{{j_1}}}}$, ${{\vec p^{{j_2}}}}$ and ${{\vec p^{{j_3}}}}$ (specifically $p_T$ ordering) to define which of them is which. Nonetheless, the normalisation choice makes $\alpha$ completely independent of measurement errors on momentum magnitudes *until* the point at which any measurement errors become large enough to change the $p_T$ ordering. Had the normalisation been achieved by dividing by $|{{\vec p^{{j_1}}}}||{{\vec p^{{j_2}}}}||{{\vec p^{{j_3}}}}|$ instead of by $|{{\vec p^{{j_1}}}}\times {{\vec p^{{j_2}}}}||{{\vec p^{{j_3}}}}|$, the same insensitivity to small changes in momenta magnitudes would still be present, but the resulting variable would no longer have a simple geometric interpretation as the sine of an angle between two planes. We opt for the former to preserve a geometric interpretation of the variable. We re-emphasise that there is nonetheless considerable arbitrariness in our choice of $\alpha$ – which itself is attributable to the lacking concrete models predicting visible parity violation at the LHC. Were such models to exist, we could speak quantitatively of ‘better’ or ‘worse’ choices of variable, and make optimisations. At the present time, however, we can only described the consequences and limitations of the choice we have made. The body text provides a number of examples of alternative choices which, if made, would result in variables with very different regions of sensitivity. We hope these ideas and others will be the subject of future investigations.
[^6]: Concretely, longitudinal boosts having velocities $\beta_z=\pm 1 / {\sqrt 2}$ change the momenta in the supplied example to ${{\vec p^{{j_1}}}}=(3,0,\pm 3)$, ${{\vec p^{{j_2}}}}=(0,2,\pm 2)$ and ${{\vec p^{{j_3}}}}=(1,1,\pm \sqrt 2)$ which have values of $\sin \alpha$ of $\pm(- 1 /{\sqrt 3} + 1/ {\sqrt 6})$ respectively. A plot showing how $\alpha$ in the above example varies for any $\beta$ is shown in Figure \[fig:BetaDependence\].
[^7]: By orientable, we mean that it has a ‘front’ which is distinguishable from its ‘back’.
[^8]: … assuming isotropy of the laws of physics.
[^9]: Another way of explaining the same point is to consider the following question: “Can a $g g \rightarrow j j j$ new-physics processes result at the LHC in more ‘right-handed’ events of the form ‘$\alpha_+$’ with momenta $
\{
{{\vec p^{{j_1}}}}=
(3,0,3), {{\vec p^{{j_2}}}}=
(0,2,2), {{\vec p^{{j_3}}}}=
(-1,-1,\sqrt 2) \}$ than ‘left-handed’ events of the form ‘$\alpha_-$’ with momenta $
\{ {{\vec p^{{j_1}}}}=
(0,3,3), {{\vec p^{{j_2}}}}=
(2,0,2), {{\vec p^{{j_3}}}}=
(-1,-1,\sqrt 2) \}$ ?” \[The sets of momenta $\alpha_+$ and $\alpha_-$ map into each other under the exchange $x\leftrightarrow y$.\] In answering this question, we would first note that both of these jet configurations are co-planar in the HPF as they have $z$-components which are zero there. Secondly, we could observe that in the HPF an event of type $\alpha_+$ can always be mapped into an event of type $\alpha_-$ by a 180 degree rotation which leaves the gluon-gluon initial state invariant. Therefore, if the equivalence principle holds and the laws of physics are isotropic, then events of type $\alpha_+$ should occur as frequently as events of type $\alpha_-$ in gluon-gluon interactions, and so the answer to the posed question is ‘No!’. If instead the laws of physics are not Lorentz invariant, or not isotropic, then the answer could instead be ‘Possibly!’. A similar argument would fail if the initial state were asymmetric (such as quark-gluon) since then the 180-degree rotation would not leave the initial state invariant and it would not be possible to conclude that the number of $\alpha_+$ and $\alpha_-$ events need be identical.
[^10]: This amazing variable is invariant under longitudinal boosts and does not need the final state particles to be distinguishable or carry an ordering!
[^11]: Note that $S_3$ and $S_4$ are identically zero if evaluated on four-momenta belonging to massless particles. $S_1$ and $S_2$ are therefore more relevance to jet-based variables than $S_3$ and $S_4$.
[^12]: The bin boundaries are at exactly at the decimal values listed. Since $\alpha$ cannot exceed $\pi/2\approx 1.57$ there are no overflow bins.
[^13]: Hence three variations of the results are derived, corresponding to the individual changes to each of the three jets
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this paper, we show the global well-posedness for periodic gKdV equations in the space $H^s(\mathbb{T})$, $s\ge \frac12$ for quartic case, and $s> \frac59$ for quintic case. These improve the previous results of Colliander et al in 2004. In particular, the result is sharp for quintic case. The main approaches are the I-method combining with the resonance decomposition, and a bilinear Strichartz estimate in periodic setting.'
address: 'School of Mathematical Sciences, Beijing Normal University, Laboratory of Mathematics and Complex Systems, Ministry of Education, Beijing 100875, P.R.China'
author:
- Jiguang Bao
- Yifei Wu
title: |
Global well-posedness for periodic generalized\
Korteweg-de Vries equation
---
[^1]
[^2]
Introduction
============
In this paper, we consider the global well-posedness of the Cauchy problem for the periodic generalized Korteweg-de Vries equations (gKdV): $$\label{gkdv}
\left\{ \aligned
&\partial_t u + \partial^3_{x} u =F(u)_x, \quad (t,x)\in [0,T]\times \T,\\
&u(0,x)=\phi(x), \quad x\in \T,
\endaligned
\right.$$ where $u$ is an unknown real function defined on $[0,T]\times \T$, $\phi$ is a given real-valued function, $F$ is a polynomial of degree $k+1$, and $\T=\R/\Z$ is the circle. For simplicity, we may assume that $F(u)=\mu u^{k+1}$. When $\mu =1$, the equation in (\[gkdv\]) is referred to “defocusing", while when $\mu=-1$ it is referred to “focusing". For $k=1$ and $k=2$, they are called by the KdV and modified KdV equations, respectively. These two equations are completely integrable. For $k\ge 3$, they are classified as the generalized KdV equations, which are not completely integrable in general. In particular, the quartic case $k=3$ and the quintic case $k=4$ are of special interest, which are regarded as the mass-subcritical and mass-critical equations.
The Cauchy problem (\[gkdv\]) has been widely studied. The periodic KdV and periodic modified KdV equations are well-posedness in $H^s(\T)$ for any $s\geq -\frac{1}{2}$ and $s\ge \frac12$ respectively. See Kenig, Ponce and Vega [@KPV-96] (also [@Bourgain; @Bourgain-97; @KPV-93-kdv] and [@KwOh-IMRN] for unconditional well-posedness of modified KdV equation) for local results and Colliander, Keel, Staffilani, Takaoka and Tao (I-team) [@CKSTT-03-KDV] for the global results. These above ranges of $s$ are *sharp* in the sense of the uniformly continuous dependence of the solution on the data, see [@KPV-01]. One can also see Kappeler and Topalov [@KaTo-CPDE; @KaTo-Duke] and the reference therein, for global $C^0$-well-posedness for rougher data. When $F$ is a general polynomial of degree $k+1\geq 4$, it was shown by I-team [@CKSTT-04-gkdvT] the local well-posedness in $H^s(\T)$ for any $s\geq \frac{1}{2}$. The authors [@CKSTT-04-gkdvT] also showed the analytic ill-posedness in $H^s(\T)$ for $s<\frac12$. So in this sense, the index $\frac12$ is sharp for local well-posedness in Sobolev space $H^s(\T)$. Moreover, they established the global well-posedness results in $H^s(\T)$ for $s>\frac{13}{14}-\frac{2}{7k}$ (in the defocusing case when $k>4$, which is mass-supercritical). In particular, they proved that the quartic and quintic gKdV equations are global well-posedness in $H^s(\T)$ whenever $s>\frac{5}{6}$ and $s>\frac{6}{7}$ respectively. But there exist some gaps to the local threshold $s=\frac{1}{2}$. In the present paper, we improve the indices and obtain the optimal one for $k=3$ as expected in [@CKSTT-04-gkdvT], while for $k=4$ there is still room to improve to a sharp result. For the related results in real line case, we just refer to [@CKSTT-03-KDV; @Guo; @KPV-96; @K; @Miao-Shao-Wu-Xu:2009:gKdV; @Tao-07] for a few of them. Now our main result can be stated as follows.
\[thm:main\] The Cauchy problems of defocusing generalized KdV equations $$\label{gkdv-4}
\left\{ \aligned
&\partial_t u + \partial^3_{x} u =\partial_x(u^{k+1}), \quad (t,x)\in [0,T]\times \T,\\
&u(0,x)=\phi(x), \quad x\in \T
\endaligned
\right.$$ are globally well-posed in $H^s(\T)$ with $s\ge \frac12$ for $k=3$, and $s>\frac59$ for $k=4$.
Similar results as Theorem \[thm:main\] also hold for the focusing equations with the suitable small initial data which guarantees the positivity of the energy.
Moreover, for general nonlinearity, our method here is also available. However, compared with the local theory, the global result for $k>4$ falls far short of expectations. Even for the quintic case, it still has gap from the sharp local result.
The main approach used here is I-method introduced by I-team, see [@CKSTT-02-DNLS; @CKSTT-03-KDV; @CKSTT-04-gkdvT] for examples. Also, we shall use the resonant decomposition argument given in [@Bourgain-04; @CKSTT-08], see also [@Miao-Shao-Wu-Xu:2009:gKdV; @MWX-DNLS; @LWX-NLS; @LWX-NLS-T] for more related argument. It is known that the problem (\[gkdv-4\]) obeys the conserved Hamiltonian $$E(u):=\int_\T \Big(\frac{1}{2}u_x^2+\frac{1}{k+2}u^{k+2}\Big)\,dx.$$ The scheme of the proof in [@CKSTT-04-gkdvT] is to construct the “almost conservation law" of the “first" modified energy $E(Iu)$ by using the I-operator. Then the global result can be obtained by iteration. Moreover, some suitable “correction-term" may be added to the first modified energy $E(Iu)$. If this is done, one may define the “second" modified energy. Then the better energy increment and global result could be gotten. See [@CKSTT-03-KDV] for a classical application. However, for the gKdV equation, one may note that it is hard to define the second modified energy in a naive way, via adding a “correction-term" to $E(Iu)$ directly. The reason is that the multiplier, introduced to obtain the second modified energy, is singular in the sense that its $L^\infty$-norm is infinity in a nontrivial set. The same thing happens in the mass-critical gKdV equation in the real line case, which was considered in [@Miao-Shao-Wu-Xu:2009:gKdV]. To get around the difficulty, we employ the resonant decomposition method. More precisely, we will split the multiplier into “resonant piece" and “non-resonant piece", and then treat them separately. For “non-resonant piece", we add a “correction-term" to define the second modified energy, while for “resonant piece", we prove that it is relatively small. However in periodic case, compared with the real line case in [@Miao-Shao-Wu-Xu:2009:gKdV], the decomposition should be finer. In fact, we need smaller control in “resonant piece", because of the weaker Strichartz estimate in the periodic setting. Moreover, the result in periodic setting is weaker than the analogous in the real-line setting, and the argument above only is also not enough to obtain the sharp global theory. A main difficulty for the periodic problem is the absence of the bilinear Strichartz estimate. More precisely, in real line setting, one has $$\left\|\int\!\!\!\! \int_{\xi=\xi_1+\xi_2}
e^{ix\xi_1+it\xi_1^3}e^{ix\xi_2+it\xi_2^3}
|\xi_1^2-\xi_2^2|^{\frac{1}{2}}\widehat{\phi_1}(\xi_1)\widehat{\phi_2}(\xi_2)\,d
\xi_1 d \xi\right\|_{L^2_{xt}}\lesssim
\|\phi_1\|_{L^2}\|\phi_2\|_{L^2}.$$ But it doesn’t work in periodic setting. So a novelty in this paper is a variant bilinear Strichartz estimate in the periodic case.
In [@CKSTT-03-KDV], the authors established a type of Strichartz estimate on the long period linear-flow, which is available to the rescaled problem. The kind of the bilinear Strichartz estimate for the periodic nonlinear Schrödinger equation was established simultaneously in [@BuGeTz; @DPST-DCDS-07], and see also [@LWX-NLS-T] for an improvement version to the one in [@DPST-DCDS-07]. Inspired from these papers, we obtain the following bilinear Strichartz estimate on Airy equation here, $$\left\|\int\!\!\!\! \int_{\xi=\xi_1+\xi_2}
e^{ix\xi_1+it\xi_1^3}e^{ix\xi_2+it\xi_2^3}
C(\xi_1,\xi_2,\lambda)\widehat{\phi_1}(\xi_1)\widehat{\phi_2}(\xi_2)\,
\big(d\xi_1\big)_\lambda \big(d\xi\big)_\lambda\right\|_{L^2_{xt}}\lesssim
\|\phi_1\|_{L^2}\|\phi_2\|_{L^2},$$ where $(d\xi)_\lambda$ denotes the normalized counting measure defined in Section 2, and the function $$C(\xi_1,\xi_2,\lambda)\rightarrow |\xi_1^2-\xi_2^2|^{\frac{1}{2}}, \quad
\mbox{as }\lambda \rightarrow \infty.$$ It matches the bilinear Strichartz estimate in real line case when the period (or the scaling parameter) tends to infinity. However, for the rescaled problems under study, the efficacy of the bilinear Strichartz estimate in the periodic case is exactly weaker than the one in the real line case. See Remark \[rem:Bi-Str\] below. This is different from the one on the periodic Schrödinger equation (see [@DPST-DCDS-07; @LWX-NLS-T]).
Outline of the proof
--------------------
### Working space
First, we use the gauge transformation introduced in [@Bourgain; @Staffilani]. Let $$\mathcal G u(t,x)=u\Big(t,x+\int_0^t\!\!\!\int_\T u^k\ dxds\Big).$$ Then we denote functional space $X^s$ as our working space, which is equipped by the norm, $$\label{workingspace}
\|u\|_{X^s}:=\|\mathcal G u\|_{Y^s},$$ where $Y^s$ is the standard (but slightly modified) Bourgain space defined in Section 2. Moreover, we denote $X^s(I)$ to be its restricted space on time interval $I$.
In [@CKSTT-04-gkdvT], I-team employed this gauge transformation to avoid a nontrivial resonance in the original equation. Under this transform, the function $\mathcal G u$ satisfies the equation $$\label{eqs:v}
\partial_t v+\partial_{x}^3v=\mathbb{P}\big[\mathbb{P}(v^k)v_x\big],$$ where $\mathbb P$ denotes the orthogonal projection onto mean zero functions, $$\mathbb{P}f=f-\int_\T f\,dx,$$ that is, $\widehat{\mathbb{P}f}(0)=0$. Then the authors considered $\mathcal G u$ instead to prove the sharp local well-posedness via multilinear estimates.
However, to study the global theory, we can not employ the forms of because it breaks the symmetries, which gives the bad form of the modified energies and thus against finer multiplier estimates, see Step 3 below. So we still consider the original equation, but use the gauged norm which is dependent upon the local theory. This causes difficulties in the multiplier estimates. Fortunately, this difficulty can be overcome by using some good properties of the gauge transformation. For this reason, one shall be careful in the usage of the Bourgain norm.
### $I$-operator
Let $N\gg 1$ be fixed, and the Fourier multiplier operator $I_{N,s}$ be defined as $$\widehat{I_{N,s}f}(\xi)=m_{N,s}(\xi)\hat{f}(\xi).\label{I}$$ Here the multiplier $m_{N,s}(\xi)$ is a smooth, monotone function satisfying $0<m_{N,s}(\xi)\leq 1$ and $$m_{N,s}(\xi)=\biggl\{
\begin{array}{ll}
1,&| \xi|\leq N,\\
N^{1-s}| \xi|^{s-1},&| \xi|>2N.\label{m}
\end{array}$$ Usually, we denote $I_{N,s}$ and $m_{N,s}$ as $I$ and $m$ respectively for short if there is no confusion. Then $$\|f\|_{H^s}\lesssim \|I_{N,s}f\|_{H^1}
\lesssim
N^{1-s}\|f\|_{H^s}.\label{II}$$
### Sketch the proofs
Now we sketch the proof of Theorem \[thm:main\] in the following steps.
`Step 1: Rescaling.`
We rescale the problem by writing $$u_\lambda(t,x)=\lambda^{-\frac{2}{k}}u(t/\lambda^3,x/\lambda);\quad
\phi_{\lambda}(x)=\lambda^{-\frac{2}{k}}\phi(x/\lambda),$$ then $u_\lambda$ satisfies that $$\label{gkdvR}
\left\{ \aligned
&\partial_t u_\lambda + \partial^3_{x} u_\lambda =(u_\lambda^{k+1})_x, \quad (t,x)\in [0,\lambda^3T]\times [0,\lambda],\\
&u_\lambda(0,x)=\phi_\lambda(x),\quad x\in [0,\lambda].
\endaligned
\right.$$ Moreover, the solution of (\[gkdv\]) $u$ exists on $[0,T]$ if and only if $u_\lambda$ exists on $[0,\lambda^3T]$. On the other hand, we get that for any $q\geq 1$ and $s\geq 0$, $$\begin{aligned}
\label{9.51}
\|\phi_\lambda\|_{L^q_x}
=\lambda^{\frac{1}{q}-\frac{2}{k}}\|\phi\|_{L^q_x}; \quad
\|\phi_{\lambda}\|_{\dot{H}^s} = \lambda^{\frac{1}{2}-\frac{2}{k}-s}
\|\phi\|_{\dot{H}^s}.\end{aligned}$$ Hence, by (\[II\]) and $m(\xi)\leq 1$, $$\aligned
E(I\phi_\lambda)=&\; \frac{1}{2}\|\partial_x
I\phi_\lambda\|_{L^2}^2+\frac{1}{k+2}\|I\phi_\lambda\|_{L^{k+2}}^{k+2}\\
\lesssim&\;
N^{2-2s}\|\phi_{\lambda}\|_{\dot{H}^s}^2+\|\phi_\lambda\|_{L^{k+2}}^{k+2}\\
\lesssim&\; N^{2-2s}/\lambda^{\frac{4}{k}+2s-1}\cdot
\|\phi\|_{H^s}+\lambda^{-1-\frac{4}{k}}\|\phi\|_{L^{k+2}}^{k+2}.
\endaligned$$ To normalize the rescaled initial data, we choose $$\label{lambda}
\lambda\sim N^{\frac{1-s}{\frac{2}{k}+s-\frac{1}{2}}}.$$ Then, $$\label{B of Iu}
\|I\phi_\lambda\|_{H^1},\,\,E(I\phi_\lambda) \lesssim 1.$$
`Step 2: Local theory for rescaled solutions`. We need the following local theory,
\[lem:modified-local\] Let $s\geq \frac{1}{2}$ and $\phi$ satisfy $\|I\phi_\lambda\|_{H^1}\lesssim 1$, then Cauchy problem (\[gkdvR\]) is locally well-posed on the interval $[0,\delta]$ with the lifetime $$\delta\sim \lambda^{-\epsilon}\label{delta}$$ for some small $\epsilon>0$. Furthermore, the solution satisfies the estimate $$\|Iu_\lambda\|_{X^1([0,\delta])}
\lesssim
\|I\phi_\lambda\|_{H^1}.\label{LSE}$$
`Step 3: Definition of modified energies`. It will be convenient to define $$f_\lambda(t):= e^{t\partial_x^3}u_\lambda(t),$$ then one may find that $$\partial_t f_\lambda=e^{t\partial_x^3}\partial_x\big(u_\lambda^{k+1}\big).$$ Therefore, we have $$\label{eqs:1719.47}
\partial_t \widehat{f_\lambda}(\xi)=i\xi\int_{\xi_1+\cdots+\xi_{k+1}=\xi}e^{i(-\xi^3+\xi_1^3+\cdots+\xi_{k+1}^3)t}\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+1})
(d\xi_1)_\lambda\cdots (d\xi_{k})_\lambda.$$
We denote $m_j=m( \xi_j)$, $\alpha_{k+2}= \xi_1^3+\cdots+ \xi_{k+2}^3$, $\Gamma_{n}$ to be the hyperplane $$\label{Gamma_n}
\Gamma_n=\left\{( \xi_1,\cdots, \xi_n)\in \left(\frac\Z\lambda\right)^n: \xi_1+\cdots+ \xi_n=0\right\}.$$ From Plancherel’s identity (see below), it follows that $$\begin{aligned}
E(Iu_\lambda)=&\frac{1}{2}\int_0^\lambda |\partial_xIu_\lambda(t,x)|^2\,dx+\frac{1}{k+2}\int_0^\lambda |Iu_\lambda(t,x)|^{k+2}\,dx\\
=&\frac12\int_{\Gamma_2} m_1^2 \xi_1^2\> \widehat{f_\lambda}(t,\xi_1)\widehat{f_\lambda}(t,\xi_2)(d\xi_1)_\lambda\\
&+\frac{1}{k+2}\int_{\Gamma_{k+2}}m_1\cdots m_{k+2}\,e^{i\alpha_{k+2}t}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
(d\xi_1)_\lambda\cdots (d\xi_{k+1})_\lambda.\end{aligned}$$ By , the symmetries of the variables $\xi_j$ in the integration and a direct computation, we have $$\begin{aligned}
&\dfrac{d}{dt}E(Iu_\lambda(t))\notag\\
=&\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\big(M_{k+2}+i\sigma_{k+2}\>\alpha_{k+2}\big)
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
(d\xi_1)_\lambda\cdots (d\xi_{k+1})_\lambda\label{dE1}\\
&+\int_{\Gamma_{2k+2}} e^{i\alpha_{2k+2}t}M_{2k+2}\>
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{2k+2})
(d\xi_1)_\lambda\cdots (d\xi_{2k+1})_\lambda,\label{dE2}\end{aligned}$$ where $$\aligned
M_{k+2}( \xi_1,\cdots, \xi_{k+2}) :=&i\big(m_1^2\xi_1^3+\cdots+m_{k+2}^2\xi_{k+2}^3\big);
\quad \sigma_{k+2}:=\frac1{k+2} m_1\cdots m_{k+2};\\
M_{2k+2}( \xi_1,\cdots, \xi_{2k+2}) :=&\; i(k+2)[\sigma_{k+2}( \xi_1,\cdots, \xi_{k+1},
\xi_{k+2}+\cdots+ \xi_{2k+2}) ( \xi_{k+2}+\cdots+ \xi_{2k+2})]_{sym},
\endaligned$$ and $[m]_{sym}$ denotes the symmetrization of a multiplier $m$ (see [@CKSTT-03-KDV]). So far, this precess is rather standard, and it is the same as what in real line case, see [@F; @Miao-Shao-Wu-Xu:2009:gKdV], etc..
Now we focus our attention on the term , and consider the quantity $$\label{M-phi}
\frac{M_{k+2}}{\alpha_{k+2}}.$$ If it makes sense, then one may use the identity $$\label{identity}
e^{i\alpha_{k+2}s}=\frac{1}{i\alpha_{k+2}} \partial_s\big(e^{i\alpha_{k+2}s}\big),$$ and take the derivative in $s$. This way gives the definition of the second modified energy, and may improve the tiny increment estimate of $E(Iu)$.
One may find is bounded when $k=1,2$. But unfortunately, is singular and thus does not make sense in general when $k\ge 3$. So it fails to define the second modified energy in this way. Here our argument is the resonance decomposition developed in [@Miao-Shao-Wu-Xu:2009:gKdV].
To do this, we first make a convenient reduction. Denote $\xi_1^*,\cdots,\xi_{k+2}^*,\cdots, \xi_{2k+2}^*$ to be the rearrangement of $\xi_1,\cdots,\xi_{k+2}, \cdots,\xi_{2k+2}$, with $
|\xi_1^*|\ge \cdots \ge |\xi_{k+2}^*|\ge \cdots\ge |\xi_{2k+2}^*|.
$
\[rem:reduction\] If $|\xi_1^*|\ll N$, then $M_{k+2}, M_{2k+2}=0$ which gives the conservation of $E(Iu_\lambda)$. Hence one may restrict $|\xi_1^*|\gtrsim N$ in the support of $\Gamma_{k+2}$ and $\Gamma_{2k+2}$.
Now we define the “non-resonance” set using the spirit in [@Miao-Shao-Wu-Xu:2009:gKdV], let $$\Omega=\Omega_1\cup \Omega_2\cup \Omega_3\cup \Omega_4,$$ where $$\begin{aligned}
\Omega_1=&\big\{(\xi_1,\cdots,\xi_{k+2})\in \Gamma_{k+2}: |\xi_3^*|\gg |\xi_4^*|\big\};\\
\Omega_2=&\big\{(\xi_1,\cdots,\xi_{k+2})\in \Gamma_{k+2}: |\xi_1^*|\sim |\xi_2^*|\gtrsim N\gg |\xi_3^*|\sim |\xi_4^*|,
\big|{\xi_1^*}^3+{\xi_2^*}^3\big|\gg \big|{\xi_3^*}^3+\cdots+{\xi_{k+2}^*}^3\big|\big\};\\
\Omega_3=&\big\{( \xi_1,\cdots, \xi_{k+2})\in \Gamma_{k+2}:
|\xi_1^{*}|\gg |\xi_3^{*}|,|\xi_1^{*}+\xi_2^{*}||\xi_1^*|\gg |\xi_3^*|^2\big\};\\
\Omega_4=&\big\{(\xi_1,\cdots,\xi_{k+2})\in \Gamma_{k+2}: |\xi_4^*|\gg|\xi_5^*|,\big|\xi_1^*+\xi_2^*\big|\big|\xi_1^*+\xi_3^*\big|\big|\xi_1^*+\xi_4^*\big|\gg |\xi_5^*||\xi_1^*|^2\\
&\qquad \big|m(\xi_1^*)^2{\xi_1^*}^3+\cdots+m(\xi_4^*)^2{\xi_4^*}^3\big|\gg \big|m(\xi_5^*)^2{\xi_5^*}^3+m(\xi_6^*)^2{\xi_6^*}^3\big|\big\},\end{aligned}$$ and $\xi_6^*=0$ if $k=3$.
Compared with the “non-resonance” sets defined in [@Miao-Shao-Wu-Xu:2009:gKdV], we add the set $\Omega_3$ and slightly change the definition on $\Omega_4$. They are employed to overcome the trouble from the weak Strichartz estimates in the periodic setting. Firstly, is bounded in “non-resonance” set, that is,
\[non-res\] $$\begin{aligned}
\label{eqs:non-res}
|M_{k+2}|\lesssim |\alpha_{k+2}|,\quad \mbox{ in } \Omega.\end{aligned}$$
Secondly, we have
\[es:Mk+2\] In $\Gamma_{k+2}\backslash\Omega$,
- It holds that $$\begin{aligned}
\label{lem:Mk+2-1}
|M_{k+2}|\lesssim m(\xi_1^*)^2|\xi_1^*||\xi_3^*|^2.\end{aligned}$$
- If $|\xi_1^*|\sim |\xi_2^*|\gtrsim N\gg |\xi_3^*|\sim |\xi_4^*|$, then $$\begin{aligned}
\label{lem:Mk+2-2}
|M_{k+2}|\lesssim |\xi_3^*||\xi_4^*||\xi_5^*|.\end{aligned}$$
- If $|\xi_4^*|\gg|\xi_5^*|$, then $$\begin{aligned}
\label{lem:Mk+2-3}
|M_{k+2}|\lesssim m(\xi_1^*)^2|\xi_1^*|^2|\xi_5^*|.\end{aligned}$$
Lemma \[es:Mk+2\] implies that the bound of $M_{k+2}$ in “resonance” set is less than its nature bound (which is $m(\xi_1^*)^2|\xi_1^*|^2|\xi_3^*|$). Here we give a slightly finer bound than the one obtained in [@Miao-Shao-Wu-Xu:2009:gKdV], according to the problem under study.
Based on these two lemmas, we rewrite the term as $$\begin{aligned}
\eqref{dE1}=&\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\Big(\chi_{\Omega}M_{k+2}+i\sigma_{k+2}\>\alpha_{k+2}\Big)
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
(d\xi_1)_\lambda\cdots (d\xi_{k+1})_\lambda\\
&\quad+\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})(d\xi_1)_\lambda\cdots (d\xi_{k+1})_\lambda.\end{aligned}$$ From Lemma \[non-res\], $\frac{1}{i\alpha_{k+2}} \big(\chi_{\Omega}M_{k+2}+i\sigma_{k+2}\>\alpha_{k+2}\big)$ is bounded. Thus by , integration by parts in time, , and combining with , we have $$\begin{aligned}
\dfrac{d}{dt}E(Iu_\lambda(t))=&\int_{\Gamma_{k+2}} \partial_t\big(e^{i\alpha_{k+2}t}\big)\frac{1}{i\alpha_{k+2}} \Big(\chi_{\Omega}M_{k+2}+i\sigma_{k+2}\>\alpha_{k+2}\Big)
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
\notag\\
&+\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
\notag\\
&+\int_{\Gamma_{2k+2}} e^{i\alpha_{2k+2}t}M_{2k+2}\>
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{2k+2})
\notag\\
=&\dfrac{d}{dt}\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\frac{\chi_{\Omega}M_{k+2}+i\sigma_{k+2}\>\alpha_{k+2}}{i\alpha_{k+2}}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
\notag\\
&+\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
\notag\\
&+\int_{\Gamma_{2k+2}} e^{i\alpha_{2k+2}t}\overline{M_{2k+2}}\>
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{2k+2}),\label{3.10}\end{aligned}$$ where $$\aligned
\overline{M_{2k+2}}:=
&i(k+2)\big[\tilde{\sigma}_{k+2}( \xi_1,\cdots, \xi_{k+1}, \xi_{{k+2}}+\cdots+
\xi_{2k+2}) ( \xi_{{k+2}}+\cdots+ \xi_{2k+2}\big)\big]_{sym},\label{M2k+2}
\endaligned$$ and $\tilde{\sigma}_{k+2}=-{\chi_{\Omega}M_{k+2}}/{\alpha_{k+2}}$. In particular, we have the bound of $\overline{M_{2k+2}}$.
\[es:M2k+2\] In $\Gamma_{2k+2}$,
- It holds that $$\begin{aligned}
\label{M2k+2-1}
|\overline{M_{2k+2}}|\lesssim |\xi_1^*|.\end{aligned}$$
- If $|\xi_1^*|\sim |\xi_2^*|\gtrsim N\gg |\xi_3^*|\sim |\xi_4^*|$, then $$\begin{aligned}
\label{M2k+2-2}
|\overline{M_{2k+2}}|\lesssim |\xi_3^*|.\end{aligned}$$
According to , we define $$\begin{aligned}
\label{3.13}
E_I^2(u_\lambda(t)):= E(Iu_\lambda(t))-\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\frac{\chi_{\Omega}M_{k+2}+\sigma_{k+2}\>\alpha_{k+2}}{i\alpha_{k+2}}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2}),\end{aligned}$$ and get $$\begin{aligned}
\dfrac{d}{dt}E_I^2(u_\lambda(t))=&\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})
\\
&+\int_{\Gamma_{2k+2}} e^{i\alpha_{2k+2}}\overline{M_{2k+2}}\>
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{2k+2}).\end{aligned}$$ This gives that $$\begin{aligned}
E_I^2(u_\lambda(t))=&E_I^2(u_\lambda(0))+\int_0^t\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}s}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(s,\xi_1)\cdots\widehat{f_\lambda}(s,\xi_{k+2})
\notag\\
&+\int_0^t\int_{\Gamma_{2k+2}} e^{i\alpha_{2k+2}s}\overline{M_{2k+2}}\>
\widehat{f_\lambda}(s,\xi_1)\cdots\widehat{f_\lambda}(s,\xi_{2k+2}).\label{eqs:EI2}\end{aligned}$$
`Step 4: Energy increment estimates`.
By the preparation in Step 1–Step 3, we can derive the following proposition, which is sufficient to prove Theorem \[thm:main\].
\[thm:main-2\] For a solution $u_\lambda$ to which is smooth-in-time, Schwarz-in-space on the time interval $[0,\delta]$ with $\delta$ satisfying , we have
- (Fixed-time bound) $$\label{fixed-time bound}
\big|E_I^2(u_\lambda(t))-E(Iu_\lambda(t))\big|
\lesssim
N^{-2+}\|Iu_\lambda(t)\|^{k+2}_{H^1_x}.$$
- (Almost conservation law) Let $\|Iu_\lambda\|_{X^1([0,\delta])}\lesssim 1$, then $$\label{Almost conserved}
\left|E^2_I(u_\lambda(\delta))-E^2_I(u_\lambda(0))\right|\le K:= N^{-3+}+N^{-2+}\lambda^{-\frac12}.$$
Now the paper is organized as follows. In Section 2, we introduce some notations and state some preliminary estimates that will be used throughout this paper. In particular, we establish the bilinear Strichartz estimates in this section. In Section 3, we prove Lemma \[non-res\], Lemma \[es:Mk+2\] and Lemma \[es:M2k+2\]. In Section 4, we give the proof of Proposition \[thm:main-2\]. In Section 5, we show that Proposition \[thm:main-2\] implies the global well-posedness stated in Theorem 1.1.
Notations and Preliminary Estimates
===================================
Basic notations and definitions
-------------------------------
We use $A\lesssim B$, $B\gtrsim A$, or sometimes $A=O(B)$ to denote the statement that $A\leq CB$ for some large constant $C$ which may vary from line to line, and may depend on the data and the index $s$. When it is necessary, we will write the constants by $C_1,C_2,\cdots$ to see the dependency relationship. We use $A\sim B$ to mean $A\lesssim B\lesssim A$. We use $A\ll B$, or sometimes $A=o(B)$ to denote the statement $A\leq C^{-1}B$. The notation $a+$ denotes $a+\epsilon$ for any small $\epsilon$, and $a-$ for $a-\epsilon$. $\langle\cdot\rangle=(1+|\cdot|^2)^{1/2}$, $D_x^\alpha=(-\partial^2_x)^{\alpha/2}$ and $J_x^\alpha=(1-\partial^2_x)^{\alpha/2}$. We use $\|f\|_{L^p_xL^q_t}$ to denote the mixed norm $\Big(\displaystyle\int\|f(\cdot,x)\|_{L^q}^p\
dx\Big)^{\frac{1}{p}}$, and $\|f\|_{L^p_{xt}}:=\|f\|_{L^p_xL^p_t}$.
Throughout this paper, we use $\eta$ to denote a [*smooth*]{} cut-off function such that $$\begin{aligned}
\begin{cases}
\eta(x)=1, \ & |x|\le 1,\\
\eta(x)=0, \ &|x|\ge 2.
\end{cases}\end{aligned}$$ For an interval $I\subset \R$, we denote $\chi_I$ as its characteristic function $$\begin{aligned}
\begin{cases}
\chi_I(x)=1, \ & x\in I,\\
\chi_I(x)=0, \ &x\not\in I.
\end{cases}\end{aligned}$$
Now we introduce some other notations and definitions, some of which are employed from [@CKSTT-03-KDV]. We define $(d\xi)_\lambda$ to be the normalized counting measure on $\Z/\lambda$ such that $$\displaystyle\int
a(\xi)\,(d\xi)_\lambda=\frac{1}{\lambda}\sum\limits_{\xi\in
\frac\Z{\lambda}} a(\xi).$$
The Fourier transform of a function $f$ on $\T_{\lambda}=\R/\lambda\Z$ is defined by $$\hat{f}(\xi)=\displaystyle\int_0^\lambda e^{-2\pi i x\xi}f(x)\,dx,$$ and thus the Fourier inversion formula $$f(x)=\displaystyle\int e^{2\pi i x\xi} \hat{f}(\xi)\,(d\xi)_\lambda.$$ Then the following usual properties of the Fourier transform hold, $$\begin{aligned}
&\|f\|_{L^2([0,\lambda])}
= \big\|\hat{f}\big\|_{L^2((d \xi)_{\lambda})} \quad \mbox{(Plancherel)};\label{Plancherel}\\
&\displaystyle\int_0^\lambda f(x)\overline{g(x)}\,dx
= \displaystyle\int \hat{f}(\xi)\overline{\hat{g}(\xi)}\,(d\xi)_\lambda\quad \mbox{(Parseval)}\label{Parseval}; \\
& \widehat{fg}(\xi)=\displaystyle\int
\hat{f}(\xi-\xi_1)\hat{g}(\xi_1) \,(d\xi_1)_\lambda \quad \mbox{(Convolution)}. \label{(Convolution)}\end{aligned}$$ We define the Sobolev space $H^s([0,\lambda])$ with the norm, $$\|f\|_{H^s([0,\lambda])}=\left\|\langle
\xi\rangle^s\hat{f}(\xi)\right\|_{L^2((d\xi)_\lambda)}.$$
For $s,b\in \R $, define the Bourgain space $X_{s,b}$ to be the closure of the Schwartz class under the norm $$\|f\|_{X_{s,b}}:= \left(\int\!\!\!\!\int \langle
\xi\rangle^{2s}\langle\tau- \xi^3\rangle^{2b}|\hat{f}(
\tau,\xi)|^2\,(d\xi)_\lambda d\tau\right)^{\frac{1}{2}},\label{X}$$ for any $\lambda$-periodic function $f$. The space $X_{s,\frac{1}{2}}$ barely fails to control the $L^\infty([0,T], H^s(\R))$ norm. To rectify this we define the slightly stronger space $Y^s$ under the norm $$\label{working space}
\|f\|_{Y^s} := \|f\|_{X_{s,\frac{1}{2}}}+\left\|\langle
\xi\rangle^s\hat{f}\right\|_{L^2((d\xi)_{\lambda}) L^1(d\tau)}.$$ Moreover, we define the restricted space $Y^s(I)$ as $$\|f\|_{Y^s(I)}:= \inf\{\|\tilde f\|_{Y^s}:\tilde f=f, \mbox{ on } I\},$$ and as , $$\|f\|_{X^s(I)}:=\|\mathcal Gf\|_{Y^s(I)}.$$ If there is no confusion, we will not mention the restriction.
Some linear estimates
---------------------
\[lem:non-smooth\] Let $0<\delta <1$, and $f\in X_{0,\frac12}$. Then $\chi_{[0,\delta]}(t)f\in X_{0,b}$ for any $b<\frac12$, and $$\big\|\chi_{[0,\delta]}(t) f\big\|_{X_{0,b}}\lesssim \big\| f\big\|_{X_{0,\frac12}}.$$
Note that $$\big\|J^s_t\chi_{[0,\delta]}(t)\big\|_{L^p_t}+\big\|\chi_{[0,\delta]}(t)\big\|_{L^\infty_t}\lesssim 1$$ for any $s<\frac1p$. Then by the fractional product role, Hölder’s and Sobolev’s inequality, we have $$\begin{aligned}
\big\|\chi_{[0,\delta]}(t)e^{t\partial_{x}^3}f\big\|_{H^b_t}
&\lesssim
\big\|J_t^b\chi_{[0,\delta]}(t)\cdot e^{t\partial_{x}^3}f\big\|_{L^2_t}+
\big\|\chi_{[0,\delta]}(t)\cdot J_t^be^{t\partial_{x}^3}f\big\|_{L^2_t}\\
&\lesssim
\big\|J_t^b\chi_{[0,\delta]}(t)\big\|_{L^p_t} \big\|e^{t\partial_{x}^3}f\big\|_{L^q_t}
+\big\|\chi_{[0,\delta]}(t)\big\|_{L^\infty_t} \big\|J_t^be^{t\partial_{x}^3}f\big\|_{L^2_t}\\
&\lesssim
\big\| f\big\|_{H^{\frac12}_t},\end{aligned}$$ where $
p=2+, b<\frac1p, \frac1p+\frac1q=\frac12.
$ By using this estimate, we have $$\begin{aligned}
\big\|\chi_{[0,\delta]}(t) f\big\|_{X_{0,b}}&=\big\|\chi_{[0,\delta]}(t)e^{t\partial_{x}^3}f(t,x)\big\|_{L^2_xH^b_t}\\
&\lesssim
\big\| e^{t\partial_{x}^3} f\big\|_{L^2_xH^{\frac12}_t}
=
\big\| f\big\|_{X_{0,\frac12}}.\end{aligned}$$ This proves the lemma.
Now we state some preliminary estimates which will be used in the following sections. First we recall some well-known Strichartz estimates (see [@Bourgain; @CKSTT-04-gkdvT], for examples): $$\|f\|_{L^4_{xt}}
\lesssim
\|f\|_{X_{0,\frac{1}{3}}},\label{XE1}$$ and $$\|f\|_{L^6_{xt}}
\lesssim
\lambda^{0+}\|f\|_{X_{0+,\frac{1}{2}+}}.\label{XE2}$$ It follows from the interpolation between (\[XE1\]) and (\[XE2\]) that $$\|f\|_{L^q_{xt}}
\lesssim
\lambda^{0+}\|f\|_{X_{0+,\frac{1}{2}-\sigma(q)}},\label{XE3}$$ for all $4<q<6$ and $\sigma(q)<2(\frac{1}{q}-\frac{1}{6})$.
Since the $L^q_x$-norm is invariant under the gauge transformation, we have almost the same Strichartz estimates between $X^s$ and $Y^s$. In particular, we have the following two estimates.
\[lem:XE6-12\]
- Let $s>\frac12$ and $f\in X^s$, then $$\|f\|_{L^\infty_{xt}} \lesssim
\|f\|_{X^s}.\label{XE4}$$
- Let $s>0$ and $f\in X^s$, then $$\|f\|_{L^6_{xt}}
\lesssim
\lambda^{0+}\|f\|_{X^{s}}.\label{XE5}$$
Let $g=\mathcal G f$. For , by Young’s and Cauchy-Schwartz’s inequalities, we have $$\|f\|_{L^\infty_{xt}}=\|g\|_{L^\infty_{xt}}\leq \left\|\hat{g}\right\|_{L^1((d\xi)_\lambda
d\tau)}\lesssim \left\|\langle
\xi\rangle^{\frac{1}{2}+}\hat g\right\|_{L^2((d\xi)_{\lambda})
L^1(d\tau)}.$$
For , by dyadic decomposition, we write $f=\sum_{j=0}^\infty f_j$, for each dyadic constituents $f_j$ with frequency support $\langle \xi \rangle\sim 2^j$. Then, by (\[XE3\]) and (\[XE4\]), $$\aligned
\|f\|_{L^6_{xt}}
& \leq \sum\limits_{j=0}^\infty
\|f_j\|_{L^6_{xt}}
\lesssim \sum\limits_{j=0}^\infty
\|f_j\|_{L^q_{xt}}^\theta\|f_j\|_{L^\infty_{xt}}^{1-\theta}\\
&\lesssim \lambda^{0+} \sum\limits_{j=0}^\infty
\|f_j\|_{X_{\epsilon,\frac{1}{2}}}^\theta\|f_j\|_{Y^\rho}^{1-\theta}
\lesssim \lambda^{0+}\sum\limits_{j=0}^\infty
2^{[\theta\epsilon+\rho(1-\theta)]j} \|f_j\|_{X_{0,\frac{1}{2}}}^\theta\|f_j\|_{Y^\rho}^{1-\theta}\\
&\lesssim \lambda^{0+}\sum\limits_{j=0}^\infty
2^{[\theta\epsilon+\rho(1-\theta)]j} \|f_j\|_{Y^0},
\endaligned$$ where $\rho>\frac{1}{2}$, and we choose $q=6-$ such that $\epsilon=0+,\theta=1-$. Choosing $q$ close enough to 6 such that $s>\theta\epsilon+\rho(1-\theta)$, then we have the claim by Cauchy-Schwartz’s inequality.
By interpolating between (\[XE4\]) and (\[XE5\]), we have $$\|f\|_{L^q_{xt}}
\lesssim
\lambda^{0+}\|f\|_{Y^{\beta(q)}},\label{XE6}$$ for all $6<q<\infty$ and $\beta(q)>(\frac{1}{2}-\frac{3}{q})$.
Bilinear Strichartz estimate
----------------------------
Now we present the bilinear Strichartz estimate of the periodic version. Let $S_\lambda(t)$ be the solution map to the free KdV equation $$\partial_t u + \partial^3_{x} u=0, \quad \mbox{ in } [0,\lambda^3T]\times [0,\lambda],$$ and the bilinear operator $I_M(f,g)$ satisfy $$\label{Bi operator}
\widehat{I_M(f,g)}(\xi)=\displaystyle\int_{\xi=\xi_1+\xi_2}
\chi_{\left\{|\xi_1^2-\xi_2^2|\gtrsim M\right\}}
\hat{f}(\xi_1)\hat{g}(\xi_2)\,(d \xi_1)_\lambda.$$ First, we recall the following result obtained in (7.29) in [@CKSTT-03-KDV],
Let $\phi_1,\phi_2$ be $\lambda-$periodic functions with both the frequencies supported on $\{\xi:|\xi|\sim N\}$, then $$\label{Bi Str1}
\left\|\eta^2(t)S_\lambda\phi_1 S_\lambda\phi_2\right\|_{L^2_{xt}}\leq
\tilde{C}(N,\lambda)\|\phi_1\|_{L^2_x}\|\phi_2\|_{L^2_x},$$ where $$\label{CNlambda1}
\tilde{C}(N,\lambda)=\biggl\{
\begin{array}{ll}
1,&N\leq 1,\\
(\frac{1}{\sqrt{N}}+\frac{1}{\lambda})^{\frac{1}{2}},&N>1.
\end{array}$$
The results above match Kato’s smoothing effect in the real line case. As a refinement, we give the following bilinear Strichartz estimates.
\[TBi Str\] Let $\phi_1,\phi_2$ be $\lambda-$periodic functions, and the operator $I_M$ be defined in (\[Bi operator\]), then $$\label{Bi Str}
\left\|\eta^2(t)I_M(S_\lambda\phi_1,S_\lambda\phi_2)\right\|_{L^2_{xt}}\leq
C(M,\lambda)\|\phi_1\|_{L^2_x}\|\phi_2\|_{L^2_x},$$ where $$\label{CNlambda}
C(M,\lambda)=\biggl\{
\begin{array}{ll}
1,&M\leq 1,\\
(\frac{1}{M}+\frac{1}{\lambda})^{\frac{1}{2}},&M>1.
\end{array}$$
When $M\lesssim 1$, it easily follows by $L^4_{xt}L^4_{xt}$-Hölder and (\[XE1\]). So we only consider the case $M\gg1$. Then, by Plancherel’s identity, the left-hand side (which denotes simply by LHS) of equals to $$\begin{aligned}
&\left\|\displaystyle\int_{\stackrel{\xi_1+\xi_2=\xi,} {
\tau_1+\tau_2=\tau}}\chi_{\{|\xi_1^2-\xi_2^2|\gtrsim
M\}}\hat{\eta}(\tau_1-\xi_1^3)\hat{\eta}(\tau_2-\xi_2^3)
\phi_1(\xi_1)\phi_2(\xi_2)\,(d\xi_1)_\lambda\,d\tau_1\right\|_{L^2((d\xi)_{\lambda}d\tau)}\\
=&
\left\|\displaystyle\int_{\xi_1+\xi_2=\xi}\chi_{\{|\xi_1^2-\xi_2^2|\gtrsim
M\}}\psi(\tau-\xi_1^3-\xi_2^3)
\phi_1(\xi_1)\phi_2(\xi_2)\,(d\xi_1)_\lambda\right\|_{L^2((d\xi)_{\lambda}d\tau)},\end{aligned}$$ where $\psi=\hat{\eta}\ast\hat{\eta}$. Then by Hölder’s inequality, $$\begin{aligned}
\mbox{LHS of }\eqref{Bi Str}\lesssim &
\Big\|\Big(\int_{\xi_1+\xi_2=\xi}\chi_{\{|\xi_1^2-\xi_2^2|\gtrsim
M\}}^2\psi(\tau-\xi_1^3-\xi_2^3) \,(d\xi_1)_\lambda\Big)^{\frac12}\\
&\quad \cdot
\Big(\int_{\xi_1+\xi_2=\xi} \psi(\tau-\xi_1^3-\xi_2^3)\phi_1(\xi_1)^2\phi_2(\xi_2)^2\,(d\xi_1)_\lambda\Big)^{\frac12}\Big\|_{L^2((d\xi)_{\lambda}d\tau)}\\
\lesssim &\left\|\displaystyle\int_{\xi_1+\xi_2=\xi}\chi_{\{|\xi_1^2-\xi_2^2|\gtrsim
M\}}\psi(\tau-\xi_1^3-\xi_2^3)\,(d\xi_1)_\lambda
\right\|_{L^\infty((d\xi)_{\lambda}d\tau)}^{\frac{1}{2}}\\
&\quad \cdot
\left\|\int_{\xi_1+\xi_2=\xi} \psi(\tau-\xi_1^3-\xi_2^3)\phi_1(\xi_1)^2\phi_2(\xi_2)^2\,(d\xi_1)_\lambda\right\|_{L^1((d\xi)_{\lambda}d\tau)}^{\frac12}\\
\lesssim &\left\|\displaystyle\int_{\xi_1+\xi_2=\xi}\chi_{\{|\xi_1^2-\xi_2^2|\gtrsim
M\}}\psi(\tau-\xi_1^3-\xi_2^3)\,(d\xi_1)_\lambda
\right\|_{L^\infty((d\xi)_{\lambda}d\tau)}^{\frac{1}{2}}\|\phi_1\|_{L^2}\|\phi_2\|_{L^2}.\end{aligned}$$ Therefore, we only need to show $$\label{2.17}
B:=\left\|\displaystyle\int_{\xi_1+\xi_2=\xi}\chi_{\{|\xi_1^2-\xi_2^2|\gtrsim
M\}}\psi(\tau-\xi_1^3-\xi_2^3)\,(d\xi_1)_\lambda
\right\|_{L^\infty((d\xi)_{\lambda}d\tau)} \lesssim C(M,\lambda)^2.$$ Indeed, let the set $$\begin{aligned}
A_{\xi,\tau}=\{\xi_1\in \frac{1}{\lambda}\Z: \xi_2=\xi-\xi_1, |\xi_1^2-\xi_2^2|\gtrsim
M, \tau-\xi_1^3-\xi_2^3=O(1)\},\end{aligned}$$ then we have $$\label{Bound}
B\lesssim \frac{1}{\lambda}\sup\limits_{\xi\in \frac{1}{\lambda}\Z,\tau\in\R}\#A_{\xi,\tau}.$$ So it reduces to estimate $\sup\limits_{\xi\in \frac{1}{\lambda}\Z,\tau\in\R}\#A_{\xi,\tau}$. To this end, we rewrite $A_{\xi,\tau}$ as $$\begin{aligned}
A_{\xi,\tau} &=& \{\xi_1\in \frac{1}{\lambda}\Z: |\xi||2\xi_1-\xi|\gtrsim M,
(\xi_1-\frac{\xi}{2})^2=a+O(1/|\xi|)\}
\label{2.25}\\
&=& \{\xi_1\in \frac{1}{\lambda}\Z: \xi_2=\xi-\xi_1, |\xi_1^2-\xi_2^2|\gtrsim
M, \xi_1=\frac{1}{2}\xi\pm\sqrt{a+O(1/|\xi|)}\},\label{2.26}\end{aligned}$$ where $a=\frac{\tau-\frac{1}{4}\xi^3}{3\xi}$. Here, we only consider the case $''+''$ in , that is, $\xi_1> \frac12\xi$, and denote the corresponding set as $A_{\xi,\tau}^+$. The other case $''-''$ is symmetrical. Now we show that $A_{\xi,\tau}^+$ belongs to a set of length $\frac{1}{M}$. To this end, we consider the following two cases separately: $$\textbf{Case 1}: |a| \lesssim 1/|\xi|; \qquad \textbf{Case 2}: |a| \gg 1/|\xi|.$$ **Case 1:** $ |a| \lesssim 1/|\xi|$. By (\[2.25\]), we have $(2\xi_1-\xi)^2\lesssim
1/|\xi|$ and thus $$\label{2.27}
M^2\lesssim (2\xi_1-\xi)^2\xi^2\lesssim |\xi|.$$ Thus we get $(2\xi_1-\xi)^2\lesssim
1/M^2$, that is, $$\xi_1=\frac{1}{2}\xi+O(\frac 1M).$$ This implies that $A_{\xi,\tau}$ belongs to a set of length $\frac{1}{M}$.
**Case 2:**$|a| \gg 1/|\xi|.$ By (\[2.25\]) again one has $(2\xi_1-\xi)^2\sim a$ and thus $$a\gtrsim \frac{M^2}{\xi^2}.$$ For any $x_1,x_2\in A_{\xi,\tau}^+$, by (\[2.26\]) we find $$\begin{aligned}
|x_1-x_2|&=&\left|\sqrt{a+\varepsilon_1}-\sqrt{a+\varepsilon_2}\right|\\
&=&\frac{|\varepsilon_1-\varepsilon_2|}{\sqrt{a+\varepsilon_1}+\sqrt{a+\varepsilon_2}}\lesssim
\frac{1/|\xi|}{M/|\xi|}=
\frac{1}{M},\end{aligned}$$ where $\varepsilon_1,\varepsilon_2=O(1/|\xi|)$ and we have used the restriction $|a| \gg 1/|\xi|$. This also implies that $A_{\xi,\tau}^+$ belongs to a set of length $\frac{1}{M}$. Therefore, no matter what case, we have $$\#A_{\xi,\tau}\lesssim \frac{\lambda}{M}+1.$$ Then by (\[2.17\]) and (\[Bound\]), we have the claim.
This proposition implies
Let $u=u(t,x),v=v(t,x)$ be the $\lambda-$periodic functions of $x$, then $$\label{EIN}
\left\|\eta^2(t)I_M(u,v)\right\|_{L^2_{xt}}
\lesssim
C(M,\lambda)\|u\|_{X_{0,\frac{1}{2}+}}\|v\|_{X_{0,\frac{1}{2}+}}.$$
\[rem:Bi-Str\] In particular, we set $\lambda$ to be the number in . Then we see that $C(N^2,\lambda)$, for which bound we use in this paper, has the similar size of $\lambda^{-\frac12}$ rather than $N^{-1}$. Indeed, when $s\ge \frac12$, $k=3,4$, $$(1-s)/(\frac{2}{k}+s-\frac{1}{2})<2,$$ thus $\lambda^{-\frac12}>N^{-1}$. This means that the efficacy of the bilinear Strichartz estimate in the periodic case is exactly weaker than the one in the real line case.
\[cor:2.2\] Let $u,v,I_M$ be as Corollary 2.1, and let $\lambda$ be the number in , then for $N\gg 1$, $$\label{CEIN}
\left\|\eta^2(t)I_{N^2}(u,v)\right\|_{L^2_{xt}}
\lesssim
\lambda^{-\frac12+}\|u\|_{X_{0,\frac{1}{2}-}}\|v\|_{X_{0,\frac{1}{2}-}}.$$
First, by interpolating between (\[EIN\]) and the following estimate $$\left\|\eta^2(t)I_M(u,v)\right\|_{L^2_{xt}}
\lesssim \|u\|_{L^4_{xt}}\|v\|_{L^4_{xt}}
\lesssim \|u\|_{X_{0,\frac{1}{3}}}\|v\|_{X_{0,\frac{1}{3}}},$$ we have $$\label{15.05}
\left\|\eta^2(t)I_M(u,v)\right\|_{L^2_{xt}}
\lesssim
C(M,\lambda)^{1-}\|u\|_{X_{0,\frac{1}{2}-}}\|v\|_{X_{0,\frac{1}{2}-}}.$$ In particular, when $M=N^2$, by Remark \[rem:Bi-Str\], $$C(N^2,\lambda)=\lambda^{-\frac12},\quad \mbox{whenever } N\gg 1.$$ This proves the corollary.
Proof of Lemmas \[non-res\]–\[es:M2k+2\]
========================================
Proof of Lemma \[non-res\]
--------------------------
Note that $$\Omega=\Omega_1\cup \Omega_2\cup \Omega_3\cup \Omega_4,$$ so we need to prove that in every $\Omega_j, j=1,2,3,4$, $$|M_{k+2}|\lesssim |\alpha_{k+2}|.$$ The estimates in $\Omega_1, \Omega_2$ and $\Omega_4$ are almost the same as Lemma 4.2 in [@Miao-Shao-Wu-Xu:2009:gKdV], however, as one of the key lemmas in this paper, we still give a detail proof here for the sake of completeness. To simplify the notations, we set $\xi_j^*=\xi_j, j=1,\cdots, k+2$.
In $\Omega_1$, we note that $\xi_1\cdot\xi_2<0$, thus, $$\begin{aligned}
|\alpha_{k+2}|=&\big|\xi_1^3+\xi_2^3+\xi_3^3\big|+o(|\xi_3^3|)\\
= & \big|(\xi_1+\xi_2)(\xi_1^2-\xi_1\xi_2+\xi_2^2)+\xi_3^3\big|+o(|\xi_3^3|)\\
= & \big|\xi_3(\xi_1^2-\xi_1\xi_2+\xi_2^2-\xi_3^2)\big|+o(|\xi_3^3|)
\ge \big|\xi_3\xi_1^2\big|+o(|\xi_3^3|)\sim |\xi_3||\xi_1|^2.\end{aligned}$$ Moreover, by the mean value theorem, $$\begin{aligned}
\big|M_{k+2}\big|
\lesssim &
\big|m_1^2\xi_1^3+m_2^2\xi_2^3\big|+\big|m_3^2\xi_3^3\big|+\cdots+\big|m_{k+2}^2\xi_{k+2}^3\big|\\
\lesssim &
m_1^2\big|\xi_1+\xi_2\big|\xi_1^2+\big|\xi_3^3\big|\\
\lesssim &
\big|\xi_3\big|\xi_1^2.\end{aligned}$$ Thus we obtain the desirable estimates in $\Omega_1$.
In $\Omega_2$, we have $$|\alpha_{k+2}|\sim \big|\xi_1^3+\xi_2^3\big|.$$ Moreover, $$\begin{aligned}
\big|M_{k+2}\big|
\le &
\big|m_1^2\xi_1^3+m_2^2\xi_2^3\big|+\big|\xi_3^3+\cdots+m_{k+2}^2\xi_{k+2}^3\big|\\
\lesssim &
m_1^2\big|\xi_1^3+\xi_2^3\big|+\big|\xi_3^3+\cdots+m_{k+2}^2\xi_{k+2}^3\big|
\lesssim
\big|\xi_1^3+\xi_2^3\big|.\end{aligned}$$ So these give the desirable estimates in $\Omega_2$.
In $\Omega_3$, on one hand, since $$\label{1216.53}
\big|\xi_1^3+\xi_2^3\big|\sim \xi_1^2|\xi_1+\xi_2|\gg |\xi_1||\xi_3|^2\gg |\xi_3|^3,$$ thus, $$|\alpha_{k+2}|= \big|(\xi_1^3+\xi_2^3)+(\xi_3^3+\cdots+\xi_{k+2}^3)\big|\sim \big|\xi_1^3+\xi_2^3\big|.$$ On the other hand, by the mean value theorem and , $$\begin{aligned}
\big|M_{k+2}\big|
\lesssim &
\big|m_1^2\xi_1^3+m_2^2\xi_2^3\big|+\big|m_3^2\xi_3^3\big|+\cdots+\big|m_{k+2}^2\xi_{k+2}^3\big|\\
\lesssim &
m_1^2\big|\xi_1^3+\xi_2^3\big|+\big|\xi_3^3\big|\\
\lesssim &
\big|\xi_1^3+\xi_2^3\big|.\end{aligned}$$ Combining these two estimates, gives the desirable estimates in $\Omega_3$.
For $\Omega_4$, we set $\overline{\xi_4}=\xi_4+\xi_5+\xi_6$, then $\xi_1+\xi_2+\xi_3+\overline{\xi_4}=0$. Therefore, $$\begin{aligned}
\alpha_{k+2}=&\xi_1^3+\xi_2^3+\xi_3^3+\overline{\xi_4}^3+\big(\xi_4^3-\overline{\xi_4}^3\big)+\xi_5^3+\xi_6^3\\
=&3(\xi_1+\xi_2)(\xi_1+\xi_3)(\xi_1+\overline{\xi_4})+\big(\xi_4^3-\overline{\xi_4}^3\big)+\xi_5^3+\xi_6^3\\
= &3(\xi_1+\xi_2)(\xi_1+\xi_3)(\xi_1+\xi_4)+O(|\xi_5+\xi_6|\xi_1^2)\\
= &3(\xi_1+\xi_2)(\xi_1+\xi_3)(\xi_1+\xi_4)+O(|\xi_5|\xi_1^2).\end{aligned}$$ By the definition of $\Omega_4$, $ |\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|\gg|\xi_5|\xi_1^2$. Thus we have $$\begin{aligned}
\label{17.26}
|\alpha_{k+2}|\sim |\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|.\end{aligned}$$ By the similar way and the mean value theorem, we have $$\begin{aligned}
\label{17.27}
M_{k+2}=&m(\xi_1)^2\xi_1^3+m(\xi_2)^2\xi_2^3+m(\xi_3)^2\xi_3^3+m(\overline{\xi_4})^2\overline{\xi_4}^3
+\big(m(\xi_4)\xi_4^3-m(\overline{\xi_4})\overline{\xi_4}^3\big)+\xi_5^3+\xi_6^3\notag\\
=&m(\xi_1)^2\xi_1^3+m(\xi_2)^2\xi_2^3+m(\xi_3)^2\xi_3^3+m(\overline{\xi_4})^2\overline{\xi_4}^3+O(|\xi_5|\xi_1^2).\end{aligned}$$ Now we claim that $$\begin{aligned}
\label{17.28}
\big|m(\xi_1)^2\xi_1^3+m(\xi_2)^2\xi_2^3+m(\xi_3)^2\xi_3^3+m(\overline{\xi_4})^2\overline{\xi_4}^3\big|
\lesssim
|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|.\end{aligned}$$ To prove this, we split it into two cases: either $|\xi_1|-|\xi_4|=o(|\xi_1|)$, or $|\xi_1|-|\xi_4|\sim |\xi_1|$. The first case follows from the following double mean value theorem.
\[lem:DMVT\] Let $f(\xi)=m(\xi)^2\xi^3$, then for $|\eta|,|\lambda|\ll |\xi|$, $$\label{DMVT}
\left| f(\xi+\eta+\lambda)-f(\xi+\eta)-f(\xi+\lambda)+f(\xi)\right|
\lesssim \big|f''(\xi)\big||\eta||\lambda| .$$
Indeed, using , we have $$\begin{aligned}
\big|m(\xi_1)^2\xi_1^3+m(\xi_2)^2\xi_2^3&+m(\xi_3)^2\xi_3^3+m(\overline{\xi_4})^2\overline{\xi_4}^3\big|
\lesssim
m_1^2|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\overline{\xi_4}|\\
\lesssim&
|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|+O(|\xi_5|\xi_1^2)
\sim
|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|.\end{aligned}$$ In the second case, we also have $|\xi_1|-|\xi_3|\sim |\xi_1|$. Thus, it gives that $$|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|\sim\xi_1^2|\xi_1+\xi_2|.$$ Therefore, $$\begin{aligned}
\big|m(\xi_1)^2\xi_1^3+m(\xi_2)^2\xi_2^3&+m(\xi_3)^2\xi_3^3+m(\overline{\xi_4})^2\overline{\xi_4}^3\big|\\
\lesssim&
\big|m(\xi_1)^2\xi_1^3+m(\xi_2)^2\xi_2^3\big|+\big|m(\xi_3)^2\xi_3^3+m(\overline{\xi_4})^2\overline{\xi_4}^3\big|\\
\lesssim&
m_1^2\xi_1^2|\xi_1+\xi_2|+m_3^2\xi_3^2|\xi_3+\xi_4|
\lesssim
\xi_1^2|\xi_1+\xi_2|+\xi_1^2|\xi_5|\\
\sim&
|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|.\end{aligned}$$ This proves . Now combining with , we have $$\begin{aligned}
|M_{k+2}|\lesssim |\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|+O(|\xi_5|\xi_1^2)
\sim |\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|.\end{aligned}$$ Together with , we find $|M_{k+2}|\lesssim |\alpha_{k+2}|$, which is the desirable estimate in $\Omega_4$. This completes the proof of the lemma.
Proof of Lemma \[es:Mk+2\]
--------------------------
We may assume that $|\xi_1|\ge \cdots \ge |\xi_{k+2}|$ by symmetries, and set $\xi_6=0$ if $k=3$. Recall that $$\Gamma_{k+2}\backslash \Omega=\big(\Gamma_{k+2}\backslash \Omega_1\big)\cap \big(\Gamma_{k+2}\backslash \Omega_2\big)\cap \big(\Gamma_{k+2}\backslash \Omega_3\big)\cap \big(\Gamma_{k+2}\backslash \Omega_4\big).$$
First, we consider . If $|\xi_1|\sim |\xi_2|\sim |\xi_3|$, then $$\begin{aligned}
\big|M_{k+2}\big|
\lesssim
\big|m_1^2\xi_1^3\big|
\sim
m_1^2|\xi_1||\xi_3|^2.\end{aligned}$$ If $|\xi_1|\sim |\xi_2|\gg |\xi_3|$, then by the definition of $\Omega_1$ and $\Omega_3$, we have in $\Gamma_{k+2}\backslash \Omega$, $$|\xi_1|\sim |\xi_2|\gg |\xi_3|\sim |\xi_4|,\quad \mbox{ and } \quad |\xi_1|\big|\xi_1+\xi_2|\big|\lesssim |\xi_3|^2.$$ Then by the mean value theorem and the inequality $m(\xi)^2|\xi|\le m(\eta)^2|\eta|$ if $|\xi|\le |\eta|$, we have $$\begin{aligned}
\big|M_{k+2}\big|
\le &
\big|m_1^2\xi_1^3+m_2^2\xi_2^3\big|+\big|m_3^2\xi_3^3+\cdots+m_{k+2}^2\xi_{k+2}^3\big|\\
\lesssim &
m_1^2\xi_1^2\big|\xi_1+\xi_2\big|+m_3^2|\xi_3|^3\\
\lesssim &
m_1^2|\xi_1||\xi_3|^2+m_3^2|\xi_3|^3\\
\lesssim &
m_1^2|\xi_1||\xi_3|^2.\end{aligned}$$ This proves .
Now we consider . By the definition of $\Omega_2$, we have $\big|\xi_1^3+\xi_2^3\big|\lesssim \big|\xi_3^3+\cdots+\xi_{k+2}^3\big|$ in $\Gamma_{k+2}\backslash \Omega_2$. Then by the mean value theorem, we have $$\begin{aligned}
\big|M_{k+2}\big|
\le &
\big|m_1^2\xi_1^3+m_2^2\xi_2^3\big|+\big|\xi_3^3+\cdots+\xi_{k+2}^3\big|\\
\lesssim &
m_1^2\big|\xi_1^3+\xi_2^3\big|+\big|\xi_3^3+\cdots+\xi_{k+2}^3\big|\\
\lesssim &
\big|\xi_3^3+\cdots+\xi_{k+2}^3\big|\\
\lesssim &
\big|\xi_3\big|\big|\xi_4\big|\big|\xi_5\big|.\end{aligned}$$
We turn to consider . According to the definition of $\Omega_3$, we split it into the following two subsets, $$\aligned
A_{1} =&\; \{(\xi_1,\cdots,\xi_6)\in \Gamma_{k+2}\backslash \Omega: |\xi_4|\gg|\xi_5|,
|m^2_1\xi_1^3+\cdots+m^2_4 \xi_4^3|\lesssim\big|m_5^2\xi_5^3+m_6^2\xi_6^3\big|\}; \\
A_{2} =&\; \{(\xi_1,\cdots,\xi_6)\in \Gamma_{k+2}\backslash \Omega: |\xi_4|\gg|\xi_5|,
|( \xi_1+ \xi_2)( \xi_1+ \xi_3)( \xi_1+ \xi_4)|\lesssim
| \xi_1|^2| \xi_5|\}.
\endaligned$$
In $A_{1}$, we have $$\begin{aligned}
\big|M_{k+2}\big|
\le &
\big|m^2_1\xi_1^3+\cdots+m^2_4 \xi_4^3\big|+\big|m_5^2\xi_5^3+m_6^2\xi_6^3\big|\\
\lesssim &
m_5^2|\xi_5|^3
\lesssim
m_1^2\xi_1^2|\xi_5|.\end{aligned}$$
In $A_{2}$, we may assume that $|\xi_1+\xi_3|\ll |\xi_1|$ or $|\xi_1+\xi_4|\ll |\xi_1|$. Otherwise, if $|\xi_1+\xi_3|\gtrsim |\xi_1|$ and $|\xi_1+\xi_4|\gtrsim |\xi_1|$, then from the relation in $A_{3}$, we have $| \xi_1+ \xi_2|\lesssim |\xi_5|$, which is included in $A_{1}$. Therefore, by the definition of $\Omega_1$, we have $|\xi_1|\sim |\xi_2|\sim |\xi_3|\sim |\xi_4|$ in $A_3$. Further, we set $\xi_1>0$ by symmetries, and then have three cases as follows, $$\begin{aligned}
(1), \xi_1>0, \xi_2<0,&\,\xi_3<0,\xi_4>0;\quad
(2), \xi_1>0, \xi_2<0,\xi_3>0,\xi_4<0;\\
&(3), \xi_1>0, \xi_2>0,\xi_3<0,\xi_4<0.\end{aligned}$$
For (1), we take $\xi=\xi_1,\eta=-(\xi_1+\xi_2), \lambda=-(\xi_1+\xi_3)$, then $|\eta|\lesssim |\lambda|\ll |\xi|$. Using Lemma \[lem:DMVT\], we have $$\begin{aligned}
|m^2_1\xi_1^3+\cdots+m^2_4 \xi_4^3|\lesssim m_1^2|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|
\lesssim m_1^2|\xi_1|^2|\xi_5|,\end{aligned}$$ where we have used $ \big|f''(\xi_1)\big|\sim m^2_1|\xi_1| $. Thus, $$\begin{aligned}
\big|M_{k+2}\big|
\le &
\big|m^2_1\xi_1^3+\cdots+m^2_4 \xi_4^3\big|+\big|m_5^2\xi_5^3+m_6^2\xi_6^3\big|\\
\lesssim &
m_1^2|\xi_1+\xi_2||\xi_1+\xi_3||\xi_1+\xi_4|+\big|m_5^2\xi_5^3+m_6^2\xi_6^3\big|\\
\lesssim &
m_1^2|\xi_1|^2|\xi_5|+m_5^2|\xi_5|^3
\lesssim
m_1^2\xi_1^2|\xi_5|.\end{aligned}$$
For (2), we take $\xi=\xi_1,\eta=-(\xi_1+\xi_2), \lambda=-(\xi_1+\xi_4)$; For (3), we take $\xi=\xi_1,\eta=-(\xi_1+\xi_3), \lambda=-(\xi_1+\xi_4)$. Then by the same argument, we get the desired estimates. This proves the lemma.
Proof of Lemma \[es:M2k+2\]
---------------------------
The proof of this lemma is essentially presented in Lemma 4.4 in [@Miao-Shao-Wu-Xu:2009:gKdV], so we only present the sketch of the proof here. Again, we assume that $|\xi_1|\ge \cdots \ge |\xi_{k+2}|$ by symmetries. First, from Lemma \[non-res\], we have $|\tilde{\sigma}_{k+2}|\lesssim 1$. Therefore, $|M_{2k+2}|\lesssim |\xi_1|$. Now we prove . From , we rewrite $M_{2k+2}$ as $$\begin{aligned}
M_{2k+2}=& i\tilde{\sigma}_{k+2}( \overline{\xi_1},\xi_{k+2}, \xi_{k+3},\cdots, \xi_{2k+2}) \overline{\xi_1}
+i\tilde{\sigma}_{k+2}(\xi_1, \overline{\xi_2}, \xi_{k+3}, \xi_{k+4},\cdots, \xi_{2k+2}) \overline{\xi_2}\\
&+i\tilde{\sigma}_{k+2}(\xi_1, \xi_2, \overline{\xi_3}, \xi_{k+4}, \xi_{k+5},\cdots, \xi_{2k+2}) \overline{\xi_3} +\cdots+i\tilde{\sigma}_{k+2}(\xi_1, \xi_2, \xi_3, \cdots, \xi_{k+1}, \overline{\xi_{k+2}}) \overline{\xi_{k+2}},\end{aligned}$$ where $\overline{\xi_j}=\xi_j+\cdots+\xi_{k+j}$. Since $|\overline{\xi_j}|\lesssim |\xi_j|\le |\xi_3|$ for any $j=3,\cdots, k+2$, and $|\tilde{\sigma}_{k+2}|\lesssim 1$, we have $$\begin{aligned}
&\big|i\tilde{\sigma}_{k+2}(\xi_1, \xi_2, \overline{\xi_3}, \xi_{k+4}, \xi_{k+5},\cdots, \xi_{2k+2}) \overline{\xi_3} +\cdots+i\tilde{\sigma}_{k+2}(\xi_1, \xi_2, \xi_3, \cdots, \xi_{k+1}, \overline{\xi_{k+2}}) \overline{\xi_{k+2}}\big|\\
\lesssim & |\overline{\xi_3}|+\cdots+|\overline{\xi_{k+2}}|\lesssim |\xi_3|.\end{aligned}$$ Furthermore, by the mean value theorem, $$\begin{aligned}
&\big|i\tilde{\sigma}_{k+2}( \overline{\xi_1},\xi_{k+2}, \xi_{k+3},\cdots, \xi_{2k+2}) \overline{\xi_1}
+i\tilde{\sigma}_{k+2}(\xi_1, \overline{\xi_2}, \xi_{k+3}, \xi_{k+4},\cdots, \xi_{2k+2}) \overline{\xi_2}\big|\\
\lesssim & |\overline{\xi_1}+\overline{\xi_{2}}|\lesssim |\xi_3|.\end{aligned}$$ This proves the lemma.
Proof of Proposition \[thm:main-2\]
===================================
We first give the fixed-time bound in Proposition \[thm:main-2\]. By , it reduces to the following lemma.
\[lem:fixed-time\] For any $1/2\leq s<1$, $$\left|\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\frac{\chi_{\Omega}M_{k+2}+\sigma_{k+2}\>\alpha_{k+2}}{i\alpha_{k+2}}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})\right| \lesssim
N^{-2+}\|Iu_\lambda(t)\|^{k+2}_{H^1_x}.$$
First, we assume that $\widehat {u_\lambda}$ is positive, otherwise one may replace it by $|\widehat{u_\lambda}|$. Second, we also assume that $|\xi_1|\ge \cdots \ge |\xi_{k+2}|$ by symmetries. Moreover, by the reduction in Remark \[rem:reduction\], we further assume $|\xi_1|\sim |\xi_2|\gtrsim N$. Now by Lemma \[non-res\], we have $$|\frac{\chi_{\Omega}M_{k+2}+\sigma_{k+2}\>\alpha_{k+2}}{i\alpha_{k+2}}|\lesssim 1.$$ Therefore, by Hölder’s inequality and Sobolev’s inequality, we have $$\begin{aligned}
&\left|\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}t}\frac{\chi_{\Omega}M_{k+2}+\sigma_{k+2}\>\alpha_{k+2}}{i\alpha_{k+2}}
\widehat{f_\lambda}(t,\xi_1)\cdots\widehat{f_\lambda}(t,\xi_{k+2})\right| \\
=&
\left|\int_{\Gamma_{k+2}} \frac{\chi_{\Omega}M_{k+2}+\sigma_{k+2}\>\alpha_{k+2}}{i\alpha_{k+2}}
\widehat{u_\lambda}(t,\xi_1)\cdots\widehat{u_\lambda}(t,\xi_{k+2})\right| \\
\lesssim &
\left|\int_{\Gamma_{k+2}}
\widehat{u_\lambda}(t,\xi_1)\cdots\widehat{u_\lambda}(t,\xi_{k+2})\right| \\
=&
N^{2(s-1)}\Big|\int_{\Gamma_{k+2}}|\xi_1|^{-s+k\epsilon}|\xi_2|^{-s}
\widehat{|\nabla| Iu_\lambda}(t,\xi_1)\widehat{|\nabla| Iu_\lambda}(t,\xi_{2})\widehat{|\nabla|^{-\epsilon} u_\lambda}(t,\xi_{3})\cdots
\widehat{|\nabla|^{-\epsilon} u_\lambda}(t,\xi_{k+2})
\Big|\\
\lesssim &
N^{-2+k\epsilon}\Big|\int
{|\nabla| Iu_\lambda}(t,x){|\nabla| Iu_\lambda}(t,x)
\big({|\nabla|^{-\epsilon} Iu_\lambda}(t,x)\big)^{k}
\Big|\\
\lesssim &
N^{-2+k\epsilon}\big\|Iu_\lambda\big\|_{L^\infty_tH^1_x}^2\big\||\nabla|^{-\epsilon} u_\lambda\big\|_{L^\infty_{xt}}^k\\
\lesssim &
N^{-2+k\epsilon}\big\|Iu_\lambda\big\|_{L^\infty_tH^1_x}^{k+2},\end{aligned}$$ where we have used that for any $\frac12\le s \le 1$, $$\big\||\nabla|^{-\epsilon} u_\lambda\big\|_{L^\infty_{xt}}\lesssim
\big\| u_\lambda\big\|_{L^\infty_{t}H^\frac12_x}
\lesssim
\big\| Iu_\lambda\big\|_{L^\infty_{t}H^1_x}.$$ This proves the lemma.
Lemma \[lem:fixed-time\] proves . To prove , by , we need to estimate $$\begin{aligned}
\int_0^\delta\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}s}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(s,\xi_1)\cdots\widehat{f_\lambda}(s,\xi_{k+2}),\end{aligned}$$ and $$\begin{aligned}
&\int_0^\delta\int_{\Gamma_{2k+2}} e^{i\alpha_{2k+2}s}\overline{M_{2k+2}}\>
\widehat{f_\lambda}(s,\xi_1)\cdots\widehat{f_\lambda}(s,\xi_{2k+2}).\end{aligned}$$ These are included in the following two lemmas.
\[p+2-linear\] Let $s\geq\frac{1}{2}$, and $ \|Iu_\lambda\|_{X^1([0,\delta])}\lesssim 1$, then $$\left|\int_0^\delta\!\!\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}s}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(s,\xi_1)\cdots\widehat{f_\lambda}(s,\xi_{k+2})\right|
\lesssim
K,
\label{Lambda6}$$ where $K= N^{-3+}+N^{-2+}\lambda^{-\frac12}$.
Before estimation, we give several reductions. First, let $v=\mathcal G u$, then $$\hat v(t,\xi)=e^{2\pi i \xi\int_0^t \!\!\int_\T u^k\,dxds} \hat u(t,\xi).$$ So for $\xi_1+\cdots+\xi_{k+2}=0$, $$\hat u(t,\xi_1)\cdots \hat u(t,\xi_{k+2})=\hat v(t,\xi_1)\cdots \hat v(t,\xi_{k+2}).$$ After rescaling, this gives that $$\hat u_\lambda(t,\xi_1)\cdots \hat u_\lambda(t,\xi_{k+2})=\hat v_\lambda(t,\xi_1)\cdots \hat v_\lambda(t,\xi_{k+2}).$$ Thus, $$\begin{aligned}
&\int_0^\delta\!\!\int_{\Gamma_{k+2}} e^{i\alpha_{k+2}s}\big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{f_\lambda}(s,\xi_1)\cdots\widehat{f_\lambda}(s,\xi_{k+2})\\
=&
\int_0^\delta\!\!\int_{\Gamma_{k+2}} \big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{u_\lambda}(s,\xi_1)\cdots\widehat{u_\lambda}(s,\xi_{k+2})\\
=&
\int_0^\delta\!\!\int_{\Gamma_{k+2}} \big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{v_\lambda}(s,\xi_1)\cdots\widehat{v_\lambda}(s,\xi_{k+2}).\end{aligned}$$
Second, to extend the integration domain from $[0,\delta]$ to $\R$, we insert the non-smooth cutoff function $\chi_{[0,\delta]}(t)$ into one of $v_\lambda$ and use the estimate in Lemma \[lem:non-smooth\]. This allows us to turn to show $$\left|\int_\R \!\int_{\Gamma_{k+2}} \big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{v_\lambda}(s,\xi_1)\cdots\widehat{v_\lambda}(s,\xi_{k+2})\right|
\lesssim
K\>
\|Iv_\lambda\|_{X_{1,\frac{1}{2}-}}\|Iv_\lambda\|_{Y^1}^{k+1},$$ where $v_\lambda$ is time supported on $[0,\delta]$. But the $0+$ loss is not essential and will be recorded by $N^{0+}$, thus it will not be mentioned. Then by Plancherel’s identity, it turns to show $$\label{3.11}
\left|\iint_{\Gamma_{k+2}\times \Gamma_{k+2}} \big(1-\chi_{\Omega}\big)M_{k+2}
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{k+2},\xi_{k+2})\right|
\lesssim K\>\|Iv_\lambda\|_{Y^1}^{k+2},$$ where the set $\Gamma_{k+2}\times
\Gamma_{k+2}=\{( \xi,\tau): \xi_1+\cdots+ \xi_{k+2}=0,
\tau_1+\cdots+\tau_{k+2}=0\}$ and we write $ \xi=( \xi_1,\cdots, \xi_{k+2})$, $\tau=(\tau_1,\cdots,\tau_{k+2})$ for short.
Third, by symmetry we may assume that $$\aligned \big| \xi_{1}\big|\ge |\xi_2|\ge \cdots \geq
\big| \xi_{{k+2}} \big|.
\endaligned$$ Also, by dyadic decomposition, we may write $$|\xi_j|\sim N_j, \quad \mbox{ for } j=1,\cdots, k+2.$$ According to the reduction in Remark \[rem:reduction\], we further assume $|N_1|\sim |N_2|\gtrsim N$. The last reduction is that after replacing $\widehat{v_\lambda}(\tau,\xi)$ by $|\widehat{v_\lambda}(\tau,\xi)|$ if necessary, we assume that $\widehat{v_\lambda}(\tau,\xi)$ is positive.
Now we divide it into four regions: $$\aligned
A_1 =& \{( \xi,\tau)\in (\Gamma_{k+2}\backslash\Omega)\times \Gamma_{k+2}:
| \xi_2|\gtrsim N\gg | \xi_3|\}; \\
A_2 =& \{( \xi,\tau)\in (\Gamma_{k+2}\backslash\Omega)\times \Gamma_{k+2}:
| \xi_3|\gtrsim N\gg | \xi_4|\}; \\
A_3 =& \{( \xi,\tau)\in (\Gamma_{k+2}\backslash\Omega)\times \Gamma_{k+2}:
| \xi_4|\gtrsim N\gg | \xi_5|\}; \\
A_4 =& \{( \xi,\tau)\in (\Gamma_{k+2}\backslash\Omega)\times \Gamma_{k+2}:
| \xi_5|\gtrsim N\}.
\endaligned$$
**Estimate in $A_1$.** By Lemma \[es:Mk+2\] (2), we have $$\big|M_{k+2}\big|\lesssim |\xi_3||\xi_4||\xi_5|.$$ Therefore, by Lemma \[lem:XE6-12\] and Corollary \[cor:2.2\], we have $$\begin{aligned}
LHS \mbox{ of } (\ref{3.11})
\lesssim&
\iint_{A_1} |\xi_3||\xi_4||\xi_5|
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{k+2},\xi_{k+2})\\
\lesssim&
N^{-2s+}\iint_{A_1} |\xi_1|^s|\xi_2|^{s-}|\xi_3||\xi_4|^{1-}|\xi_5|^{1-}
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{k+2},\xi_{k+2})\\
\lesssim&
N^{-2+}\iint
\big(|\nabla|P_{N_1}Iv_\lambda\big)(t,x)\big(|\nabla|^{1-}P_{N_2}Iv_\lambda\big)(t,x)\big(|\nabla|P_{N_3}Iv_\lambda\big)(t,x)\\
&\qquad\cdot\big(|\nabla|^{1-}P_{\ll N }Iv_\lambda\big)^{2}(t,x)\big(P_{\ll N }v_\lambda\big)^{k-3}(t,x)\,dxdt \\
\lesssim&
N^{-2+}
\big\|\eta(t)^2I_{N^2}\big(|\nabla|P_{N_1}Iv_\lambda, |\nabla|P_{N_3}Iv_\lambda\big)\big\|_{L^2_{xt}}\big\||\nabla|^{1-}Iv_\lambda\big\|_{L^6_{xt}}^3\big\|v_\lambda\big\|_{L^\infty_{xt}}^{k-3} \\
\lesssim& \;
N^{-2+}\lambda^{-\frac12}.\end{aligned}$$
**Estimate in $A_2$.** By the definition of $\Omega_1$, $A_2=\emptyset$.
**Estimate in $A_3$.** We split it into two parts again, and define $$\aligned
A_{31} =& \{( \xi,\tau)\in A_3:
| \xi_1|\sim | \xi_2|\sim| \xi_3|\sim| \xi_4|\}; \\
A_{32} =& \{( \xi,\tau)\in A_3:
| \xi_1|\sim | \xi_2|\gg| \xi_3|\sim| \xi_4|\}; \\
\endaligned$$
[Estimate in $A_{31}$.]{} By Lemma \[es:Mk+2\] (3), we have $$|\chi_{A_{31}}\,{M}_{k+2}|\lesssim m( \xi_1 )^2| \xi_1 |^{2}|\xi_5 |\sim m( \xi_1 )m( \xi_2)| \xi_1 ||\xi_2|^{1-}|\xi_3|^{0+}|\xi_5|.$$ Therefore, by Lemma \[lem:XE6-12\] and Corollary \[cor:2.2\], $$\aligned
LHS \mbox{ of } (\ref{3.11})
\lesssim&
\iint_{A_3} m( \xi_1 )m( \xi_2)| \xi_1 ||\xi_2|^{1-}|\xi_3|^{0+}|\xi_5|
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{k+2},\xi_{k+2})\\
\lesssim&
\iint
\big(|\nabla|P_{N_1}Iv_\lambda\big)(t,x)\big(|\nabla|^{1-}P_{N_2}Iv_\lambda\big)(t,x)\big(|\nabla|^{0+}P_{N_3}Iv_\lambda\big)(t,x)\\
&\qquad\qquad\cdot
\big(P_{N_4}v_\lambda\big)(t,x)\big(|\nabla|P_{N_5}Iv_\lambda\big)(t,x)\big(P_{\ll N }v_\lambda\big)^{k-3}(t,x)\,dxdt \\
\lesssim&
\big\|\eta(t)^2I_{N^2}\big(|\nabla|P_{N_1}Iv_\lambda, |\nabla|P_{N_5}Iv_\lambda\big)\big\|_{L^2_{xt}}\big\||\nabla|^{1-}P_{N_2}Iv_\lambda\big\|_{L^{6}_{xt}}\\
&\qquad\qquad\cdot
\big\||\nabla|^{0+}P_{N_3}v_\lambda\big\|_{L^{6}_{xt}}\big\|P_{N_4}v_\lambda\big\|_{L^{6}_{xt}}\big\|P_{\ll N}v_\lambda\big\|_{L^\infty_{xt}}^{k-3} \\
\lesssim&
N^{-2+}
\big\|\eta(t)^2I_{N^2}\big(|\nabla|P_{N_1}Iv_\lambda, |\nabla|P_{N_5}Iv_\lambda\big)\big\|_{L^2_{xt}}
\big\||\nabla|^{1-}Iv_\lambda\big\|_{L^{6}_{xt}}^3\big\|v_\lambda\big\|_{L^\infty_{xt}}^{k-3} \\
\lesssim& \;
N^{-2+}\lambda^{-\frac12}.
\endaligned$$
[Estimate in $A_{32}$.]{} Note that both the estimates in Lemma \[es:Mk+2\] (1) and (3) hold in $A_{32}$, so for any $\epsilon>0$, $$\begin{aligned}
|\chi_{A_{32}}\>{M}_{k+2}|\lesssim &
\big[m( \xi_1 )^2| \xi_1 |^{2}|\xi_5 |\big]^{1-\epsilon}\big[m( \xi_1 )^2| \xi_1 ||\xi_3|^{2}\big]^\epsilon\\
= &
m( \xi_1 )^2| \xi_1 |^{2-\epsilon}|\xi_3|^{2\epsilon}|\xi_5|^{1-\epsilon}\\
\lesssim &
m( \xi_1 )^2| \xi_1 ||\xi_2|^{1-\epsilon}|\xi_3|^{\epsilon}|\xi_4|^{\epsilon}\langle \xi_5\rangle.\end{aligned}$$ Therefore, by Lemma \[lem:XE6-12\] and Corollary \[cor:2.2\], $$\aligned
LHS \mbox{ of } (\ref{3.11})
\lesssim&
\iint_{A_3} m( \xi_1 )^2| \xi_1 ||\xi_2|^{1-\epsilon}|\xi_3|^{\epsilon}|\xi_4|^{\epsilon}\langle \xi_5\rangle
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{k+2},\xi_{k+2})\\
\lesssim&
\iint
\big(|\nabla|P_{N_1}Iv_\lambda\big)(t,x)\big(|\nabla|^{1-\epsilon}P_{N_2}Iv_\lambda\big)(t,x)\big(|\nabla|^\epsilon P_{N_3}v_\lambda\big)^2(t,x)\\
&\qquad\qquad\cdot
\big(\langle\nabla\rangle P_{N_5}Iv_\lambda\big)(t,x)\big(P_{\ll N }v_\lambda\big)^{k-3}(t,x)\,dxdt \\
\lesssim&
\big\|\eta(t)^2I_{N^2}\big(|\nabla|P_{N_1}Iv_\lambda, \langle\nabla\rangle P_{N_5}Iv_\lambda\big)\big\|_{L^2_{xt}}\big\||\nabla|^{1-}P_{N_2}Iv_\lambda\big\|_{L^{6}_{xt}}\\
&\qquad\qquad\cdot
\big\||\nabla|^{0+}P_{N_3}v_\lambda\big\|_{L^{6}_{xt}}^2\big\|P_{\ll N}v_\lambda\big\|_{L^\infty_{xt}}^{k-3} \\
\lesssim&
N^{-2+}
\big\|\eta(t)^2I_{N^2}\big(|\nabla|P_{N_1}Iv_\lambda, \langle\nabla\rangle P_{N_5}Iv_\lambda\big)\big\|_{L^2_{xt}}
\big\||\nabla|^{1-}Iv_\lambda\big\|_{L^{6}_{xt}}^3\big\|v_\lambda\big\|_{L^\infty_{xt}}^{k-3} \\
\lesssim& \;
N^{-2+}\lambda^{-\frac12}.
\endaligned$$
**Estimate in $A_4$.** Moreover, we split $A_4$ into two subregions: $$\aligned
A_{41} =& \{( \xi,\tau)\in A_5:
| \xi_4|\gg | \xi_5|\}; \\
A_{42} =& \{( \xi,\tau) \in A_5:
| \xi_4|\sim | \xi_5|\}.
\endaligned$$ The estimate in $A_{41}$ can be treated as the estimate in $A_{3}$, since they have the same bound on $M_{k+2}$. So we omit the details. Now we consider the estimate in $A_{42}$. In this part, by Lemma \[es:Mk+2\] (1) and the relationship $|\xi_3|\sim |\xi_4|\sim |\xi_5|$, we have $$\begin{aligned}
|M_{k+2}|\lesssim m(\xi_1)^2|\xi_1||\xi_3|^2
\lesssim &
m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|^{\frac12+3\epsilon}|\xi_3|^{\frac12-\epsilon}|\xi_4|^{\frac12-\epsilon}|\xi_5|^{\frac12-\epsilon}\\
\lesssim &
N^{-\frac32-}m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|^{\frac12+}\cdot m(\xi_3)|\xi_3|^{1-}\cdot m(\xi_4)|\xi_4|^{1-}\cdot m(\xi_5)|\xi_5|\\
\lesssim &
N^{-2+}m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|^{1-}\cdot m(\xi_3)|\xi_3|^{1-}\cdot m(\xi_4)|\xi_4|^{1-}\cdot m(\xi_5)|\xi_5|.\end{aligned}$$ Further, we claim that $$\label{7.25}
| \xi_1|-| \xi_5|\gtrsim | \xi_1|\gtrsim N.$$ Indeed, if $| \xi_j|=| \xi_1|+o(| \xi_1|)$, for all $j=1,\cdots, 5$, then there exist $\mu_j\in \{-1,1\}$ such that $$\xi_j= \mu_j \xi_1+o(| \xi_1|).$$ Therefore, $$|\xi_6|=\big|\xi_1+\cdots+ \xi_5\big|=|\mu_1+\cdots+\mu_5||\xi_1|+o(|\xi_1|).$$ Note that $|\mu_1+\cdots+\mu_5|\ge 1$, we have $|\xi_6|\sim |\xi_1|$, but this is not the case in $A_4$. So we have $| \xi_1|-| \xi_5|\gtrsim | \xi_1|$ and thus $$\big|\xi_1^2-\xi_5^2\big|\gtrsim N^2.$$ Therefore, we have $$\aligned
LHS \mbox{ of } (\ref{3.11})
\lesssim&
N^{-2+}\iint_{A_4} m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|^{1-}\cdot m(\xi_3)|\xi_3|^{1-}\cdot m(\xi_4)|\xi_4|^{1-}\cdot m(\xi_5)|\xi_5|\\
&\qquad \qquad\cdot
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{k+2},\xi_{k+2})\\
\lesssim&
N^{-2+}\iint \eta^2(t)
I_{N^2}\big(|\nabla|Iv_\lambda, |\nabla|Iv_\lambda\big)(t,x)\\
&\qquad \qquad\cdot
\big(|\nabla|^{1-}P_{\gtrsim N}Iv_\lambda\big)^3(t,x)
\big(P_{\ll N }v_\lambda\big)^{k-3}(t,x)\,dxdt \\
\lesssim&
N^{-2+}\big\|\eta(t)^2I_{N^2}\big(|\nabla|Iv_\lambda, |\nabla|Iv_\lambda\big)\big\|_{L^2_{xt}}
\big\||\nabla|^{1-}Iv_\lambda\big\|_{L^{6}_{xt}}^3\big\|v_\lambda\big\|_{L^\infty_{xt}}^{k-3} \\
\lesssim& \;
N^{-2+}\lambda^{-\frac12}.
\endaligned$$
**Estimate in $A_5$ (if $k=4$).** It also can be treated as the estimate in $A_4$, and thus we obtain the same conclusion as what in $A_4$.
\[2p+2-linear\] Let $s\geq\frac{1}{2}$, and $ \|Iu\|_{X^1([0,\delta])}\lesssim 1$, then $$\left|\int_0^t\!\!\int_{\Gamma_{2k+2}}\!\! e^{i\alpha_{2k+2}s}\overline{M_{2k+2}}\>
\widehat{f_\lambda}(s,\xi_1)\cdots\widehat{f_\lambda}(s,\xi_{2k+2})\right|
\lesssim
K'\>
\|Iu_\lambda\|_{Y^1}^{2k+2}.
\label{Lambda6}$$ where $K'=N^{-3+}+N^{-2+}\lambda^{-1}$.
By the reductions at the beginning of the proof of Proposition \[p+2-linear\], it suffices to show $$\label{3.11}
\left|\iint_{\Gamma_{2k+2}\times \Gamma_{2k+2}} \overline {M_{2k+2}}
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{2k+2},\xi_{2k+2})\right|
\lesssim K'\>\|Iv_\lambda\|_{Y^1}^{2k+2},$$ where the hyperplane $\Gamma_{2k+2}^2
=\{( \xi_1,\cdots, \xi_{2k+2},\tau_1,\cdots,\tau_{2k+2}): \xi_1+\cdots+ \xi_{2k+2}=0,
\tau_1+\cdots+\tau_{2k+2}=0\}$. Also, we may assume that $$\aligned | \xi_{1}|\ge |\xi_2|\ge \cdots \geq
| \xi_{{2k+2}} |,\quad |\xi_j|\sim N_j, \mbox{ for } j=1,\cdots, 2k+2,
\endaligned$$ and $\widehat{v_\lambda}(\tau,\xi)$ is positive.
Now we consider the following three subregions separately: $$\aligned
B_1 =& \{( \xi_1,\cdots, \xi_{2k+2},\tau_1,\cdots,\tau_{2k+2})\in \Gamma_{2k+2}^2:
| \xi_1|\sim| \xi_2|\gtrsim N\gg | \xi_3|\}; \\
B_2 =& \{( \xi_1,\cdots, \xi_{2k+2},\tau_1,\cdots,\tau_{2k+2})\in \Gamma_{2k+2}^2:
| \xi_1|\sim| \xi_2|\geq | \xi_3|\gtrsim N \gg |\xi_4|\};\\
B_3 =& \{( \xi_1,\cdots, \xi_{2k+2},\tau_1,\cdots,\tau_{2k+2})\in \Gamma_{2k+2}^2:
| \xi_4|\gtrsim N\}.
\endaligned$$
**Estimate in $B_1$.** By Lemma \[es:M2k+2\] (2), we have $|\overline{M_{2k+2}}|\lesssim|\xi_3|$. Then, by (\[CEIN\]), we have $$\aligned
\mbox{LHS of (\ref{3.11})}\lesssim &\; \int_{B_1} |\xi_3|
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{2k+2},\xi_{2k+2})\\
\lesssim &\; N^{-2}\int_{B_1} m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|\cdot m(\xi_3)|\xi_3|\cdot m(\xi_4)\langle\xi_4\rangle
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{2k+2},\xi_{2k+2})\\
\lesssim&\; N^{-2}\big\|\eta(t)^2I_{N^2}\big(|\nabla|P_{N_1}Iv_\lambda, |\nabla| P_{N_3}Iv_\lambda\big)\big\|_{L^2_{xt}}\\
&\quad\cdot
\big\|\eta(t)^2I_{N^2}\big(|\nabla|P_{N_2}Iv_\lambda, \langle\nabla\rangle P_{N_4}Iv_\lambda\big)\big\|_{L^2_{xt}}\big\|v_\lambda\big\|_{L^\infty_{xt}}^{2k-2}\\
\lesssim&\; N^{-2+}\lambda^{-1}. \endaligned$$
**Estimate in $B_2$.** Similar to the proof of , we have $$\label{7.26}
| \xi_1|-| \xi_3|\gtrsim | \xi_1|\gtrsim N.$$ Moreover, from , we have $|M_{2k+2}|\lesssim |\xi_1|$. Then $$\begin{aligned}
|M_{2k+2}|\lesssim |\xi_1|
\lesssim
N^{-2}m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|\cdot m(\xi_3)|\xi_3|\cdot m(\xi_4)\langle\xi_4\rangle.\end{aligned}$$ Therefore, we have the same estimate as what in $B_1$, and get also $$\aligned
\mbox{LHS of (\ref{3.11})}
\lesssim N^{-2+}\lambda^{-1}.
\endaligned$$
**Estimate in $B_3$.** In this part, $$\begin{aligned}
|M_{2k+2}|\lesssim |\xi_1|
\lesssim
N^{-3+s}N_4^{-s}m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|\cdot m(\xi_3)|\xi_3|\cdot m(\xi_4)|\xi_4|.\end{aligned}$$ Therefore, we have $$\aligned
\mbox{LHS of (\ref{3.11})}
\lesssim &\; N^{-3+s}N_4^{-s}\int_{B_3} m(\xi_1)|\xi_1|\cdot m(\xi_2)|\xi_2|\cdot m(\xi_3)|\xi_3|\cdot m(\xi_4)|\xi_4|\\
&\qquad \qquad\cdot
\widehat{v_\lambda}(\tau_1,\xi_1)\cdots\widehat{v_\lambda}(\tau_{2k+2},\xi_{2k+2})\\
\lesssim&\; N^{-3+s}N_4^{-s}\big\||\nabla|P_{N_1}Iv_\lambda\big\|_{L^4_{xt}}
\cdots
\big\||\nabla|P_{N_4}Iv_\lambda\big\|_{L^4_{xt}}
\|v_\lambda\|_{L^\infty_{xt}}^{2k-2}\\
\lesssim&\; N^{-3+s}N_4^{-s+}\\
\lesssim&\; N^{-3+}.
\endaligned$$ This gives the proof of the lemma.
Since $K'\le K$, combining with the results on Lemma \[lem:fixed-time\]–Lemma \[2p+2-linear\], we prove Proposition \[thm:main-2\].
Proposition \[thm:main-2\] implies Theorem \[thm:main\]
=======================================================
Suppose that $$\begin{aligned}
\label{2.288}
\sup\limits_{t\in [0,(j-1)\delta]}E(Iu_\lambda(t))\le 2E(I\phi_\lambda)\quad \mbox{for some } j\in \mathbb N.\end{aligned}$$ Then by local theory in Lemma \[lem:modified-local\], we have $$\|Iu_\lambda\|_{X^1([(i-1)\delta,i\delta])}
\lesssim
1, \quad \mbox{for any } 1\le i \le j.$$ So by in Proposition \[thm:main-2\] and a simple iteration, we have $$\begin{aligned}
\label{2.16}
\left|E^2_I(t)-E^2_I(0)\right|&\le j K\notag\\
&\le C_0:=\frac12E(I\phi_\lambda)
\end{aligned}$$ for any $ t\le j\delta, j\le C_0 K^{-1}$. By , we have $$\big|E(Iu(t))-E_I^2(t)\big|\lesssim N^{-2+}\|Iu(t)\|^{k+2}_{H^1_x}\lesssim N^{-2+}E(Iu(t))^\frac{k+2}2.$$ This combining with , gives us that for any $ t\le j\delta$, $j$ satisfying and $\le C_0 K^{-1}$, $$\begin{aligned}
E(Iu(t))&\le E_I^2(t)+CN^{-2+}E(Iu(t))^\frac{k+2}2\\
&\le
E_I^2(0)+\frac12E(I\phi_\lambda)+CN^{-2+}E(Iu(t))^\frac{k+2}2\\
&\le
\frac32E(I\phi_\lambda)+CN^{-2+}\big(E(Iu(t))^\frac{k+2}2+E(I\phi_\lambda)^\frac{k+2}2\big).\end{aligned}$$ So by continuity argument, we have for any $ t\le j\delta$, $$E(Iu(t))\le 2E(I\phi_\lambda).$$ This extends to $[0,j\delta]$. Thus by finite induction, we obtain that $$\sup\limits_{t\in [0,C_0\delta K^{-1}]}E(Iu_\lambda(t))\le 2E(I\phi_\lambda).$$ This proves that $u_\lambda$ exists on $[0,C_0\delta K^{-1}]$, which implies that $u$ exists on $[0,C_0\delta \lambda^{-3} K^{-1}]$. Suppose that $$\delta\lambda^{-3} K^{-1}\ge N^{0+},$$ then $u$ exists for arbitrary time by choosing large $N$.
Since $$\delta\sim \lambda^{0+},\quad \lambda\sim N^{\frac{1-s}{\frac{2}{k}+s-\frac{1}{2}}}, \quad
K= N^{-3+}+N^{-2+}\lambda^{-\frac12},$$ we have $\delta\lambda^{-3} K^{-1}\ge N^{0+}$ as long as $$\begin{aligned}
2>\frac52\cdot\frac{1-s}{\frac{2}{k}+s-\frac{1}{2}};\quad
3>\frac{3(1-s)}{\frac{2}{k}+s-\frac{1}{2}}.\end{aligned}$$ Particularly, when $k=3$, it holds for any $s\ge \frac12$; when $k=4$, it holds for any $s> \frac59$. This completes the proof of Theorem 1.1.
[99]{}
Bourgain, J.: [*Fourier transform restriction phenomena for certain lattice subsets and applications to nonlinear evolution equations. II: the KdV-Equation.*]{} Geom. Funct. Anal., [**3**]{}, 209–262, (1993).
Bourgain, J.: Remark on normal forms and the “I-method” for periodic NLS. J. Anal. Math., 94, 127–157, (2004).
Bourgain, J.: [*Periodic Korteweg-de Vries equation with measures as initial data*]{}, Selecta Math. (N.S.) [**3**]{}, 115–159, 209–262, (1997).
Burq, N.; Gerard, P.; Tzvetkov, N.: Bilinear eigenfunction estimates and the Nonlinear Schrödinger equations on surfaces, Inventiones Math., 159, 187–223, (2005).
Colliander, J.; Keel, M.; Staffilani, G.; Takaoka, H.; and Tao, T.: [*A refined global well-posedness result for Schrödinger equations with derivatives.*]{} SIAM J. Math. Anal., [**34**]{}, 64-86, (2002).
Colliander, J.; Keel, M.; Staffilani, G.; Takaoka, H.; and Tao, T.: [*Sharp global well-posedness for KdV and modified Kdv on $\R$ and $\T$*]{}. J. Amer. Math. Soc., [**16**]{}, 705–749, (2003).
Colliander, J.; Keel, M.; Staffilani, G.; Takaoka, H.; and Tao, T.: [*Multilinear estimates for periodic [K]{}d[V]{} equations, and applications.*]{} J. Funct. Anal., [**211**]{} (1), 173–218, (2004).
Colliander, J.; Keel, M.; Staffilani, G.; Takaoka, H.; and Tao, T.: [*Resonant decompositions and the I-method for cubic nonlinear Schrödinger on $\R^2$*]{}. Discrete and Contin. Dyn. Syst., [**21**]{} (3), 665–686, (2008).
De Silva, D.; Pavlovic, N.; Staffilani, G.; and Tzirakis, N.: Global well-posedness for a periodic nonlinear Schroödinger equation in 1D and 2D. Discrete Contin. Dyn. Syst., **19** (1), 37–65, (2007).
Farah, L. G.: [*Global rough solutions to the critical generalized KdV equation*]{}. J. Differ. Eqns., **249**, 1968–1985, (2010).
Guo, Z.: [*Global well-posedness of Korteweg-de Vries equation in $H^{-3/4}(\R)$*]{}. J. Math. Pures Appl., [**91**]{}, 583–597, (2009).
Kappeler, T. and P. Topalov. [*Global well-posedness of mKdV in $L^2(\T, \R)$.*]{} Comm. Partial Differ. Eqs, **30** (13), 435–49, (2005).
Kappeler, T. and P. Topalov. [*Global well-posedness of KdV in $H^{-1}(\T, \R)$.*]{} Duke Math. J., [**135**]{} (2), 327–360, (2006).
Kenig, C. E.; Ponce, G.; and Vega, L.: [*The Cauchy problem for the Korteweg-de Vries equation in Sobolev spaces of negative indics*]{}. Duke Math. J., [**71**]{} (1), 1–21, (1993).
Kenig, C. E., Ponce G. and Vega, L.: [*A bilinear estimate with applications to the KdV equation*]{}. J. Amer. Math. Soc., [**9**]{} (2), 573–603, (1996).
Kenig, C. E., Ponce G. and Vega, L.: [*On the ill-posedness of some canonial dispersive equations*]{}. Duke Math. J., [**106**]{} (2), 617–633, (2001).
Kishimoto, N.: [*Well-posedness for the Cauchy problem for the Korteweg-de Vries equation at the critical regularity*]{}. Diff. Integr. Eqs., [**22**]{}, 447-464, (2009).
Kwon S.; and Oh T.: [*On unconditional well-posedness of modified KdV*]{}. Int. Math. Res. Not., Vol., **2012**, No. 15, 3509–3534.
Miao, C.X.; Shao, S.L.; Wu Y.F.; and Xu, G.X.: [*The low regularity global solutions for the critical generalized KdV equation*]{}. Dynamics of PDE, **7**, No.3, 265–288, (2010).
Miao, C.X.; Wu Y.F.; and Xu, G.X.: *Global well-posedness for Schrodinger equation with derivative in $H^{\frac12}(\R)$*. J. Differ. Eqs., **251**, 2164–2195, (2011).
Li, Y.S.; Wu Y.F.; and Xu G.X.: *Global well-posedness for the mass-critical nonlinear Schrödinger equation*. SIAM J. Math. Anal., **43**, 322–340, (2011).
Li, Y.S.; Wu Y.F.; and Xu G.X.: *Global well-posedness for the periodic mass-critical nonlinear Schrödinger equation*. J. Differ. Eqs., **250**, 2715–2736, (2011).
Staffilani G.: [*On solutions for periodic generalized KdV equations.*]{} Int. Math. Res. Not. [**18**]{}, 899–917, (1997).
Tao, T.: [*Scattering for the quartic generalised Korteweg-de Vries equation*]{}. J. Differ. Eqns., [**232**]{} (2), 623–651, (2007).
[^1]: The first author was partially supported by the NSF of China (No.11371060).
[^2]: The second author was partially supported by the NSF of China (No. 11101042) and the China Postdoctoral Science Foundation (No.2012T50068).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We present a new sorting algorithm, called , that exploits the existence of monotonic runs for sorting efficiently partially sorted data. This algorithm is a variant of the well-known algorithm , which is the sorting algorithm used in standard libraries of programming languages such as Python or Java. More precisely, is a so-called $k$-aware merge-based algorithm, a class that was introduced by Buss and Knop that captures “-like” algorithms.
In this article, we prove that, although is simple to implement, and differs only slightly from , its computational cost, in number of comparisons performed, is optimal up to a small additive linear term: this makes the first $k$-aware algorithm to benefit from this property, which is also a 33% improvement over ’s worst-case. This suggests that could be a strong contender for being used instead of .
Then, we investigate the optimality of $k$-aware algorithms: we give lower and upper bounds on the best approximation ratios of such algorithms, compared to optimal sorting algorithms. In particular, we design generalisations of whose computational costs are optimal up to arbitrarily small multiplicative factors.
author:
- Vincent Jugé
date: 'Université Paris-Est, LIGM (UMR 8049), CNRS, ENPC, ESIEE Paris, UPEM'
title: 'Adaptive Shivers Sort: An Alternative Sorting Algorithm'
---
This is the final draft of the article that was submitted for publication in the proceedings of the 31^th^ Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2020). This version contains all the proofs of the results mentioned in the proceedings version of the paper. Its presentation may differ slightly from that of the proceedings version.
Introduction {#sec:intro}
============
The problem of sorting data has been one of the first and most extensively studied problems in computer science, and sorting is ubiquitous, due to its use as a sub-routine in a wealth of various algorithms. Hence, as early as the 1940’s, sorting algorithms were invented, which enjoyed many optimality properties regarding their complexity in time (and, more precisely, in number of comparisons or element moves required) as well as in memory. Every decade or so, a new major sorting algorithm was invented, either using a different approach to sorting or adapting specifically tuned data structures to improve previous algorithms: [@goldstine1947planning], [@hoare1961algorithm], [@williams1964algorithm], [@dijkstra1982smoothsort], [@moffat1996splaysort], …
In 2002, Tim Peters, a software engineer, created a new sorting algorithm, which he called [@Peters2015]. This algorithm immediately demonstrated its efficiency for sorting actual data, and was adopted as the standard sorting algorithm in core libraries of wide-spread programming languages such as Python and Java. Hence, the prominence of such a custom-made algorithm over previously preferred *optimal* algorithms contributed to the regain of interest for the study of sorting algorithms.
Understanding the reasons behind the success of is still an ongoing task. These reasons include the fact that is well adapted to the architecture of computers (e.g., for dealing with cache issues) and to realistic distributions of data. In particular, a model that successfully explains why is adapted to sorting realistic data involves *run decompositions*, as illustrated in Figure \[fig:runs\]. Such decompositions were already used by Knuth [@Knuth98], which predated , and adapted the traditional algorithm as follows: is based on splitting arrays into monotonic subsequences, also called *runs*, and on merging these runs together. Thus, all algorithms sharing this feature of are also called *natural* merge sorts.
$
S=(~\underbrace{12,7,6,5}_{\text{first run}},
~\underbrace{5,7,14,36}_{\text{second run}},
~\underbrace{3,3,5,21,21}_{\text{third run}},
~\underbrace{20,8,5,1}_{\text{fourth run}}~)
$
In addition to being a natural merge sort, also includes many optimisations, which were carefully engineered, through extensive testing, to offer the best complexity on *real-life* data and *real-life* computers. As a result, the general structure of can be split into three main components: (i) a simple policy for choosing which runs to merge, (ii) a more complex policy for choosing which *small* runs to merge, (e.g., runs of length less than 32), and (iii) an equally complex sub-routine for merging runs. The second and third components were those which were the most finely tuned, hence understanding the subtleties of why they are efficient and how they could be improved seems difficult. The first component, however, is quite simple, and therefore it offers the best opportunities for modifying and improving .
#### Context and related work
The success of has nurtured the interest in the quest for sorting algorithms that would be adapted to arrays with few runs. However, the *ad hoc* conception of made its complexity analysis less easy than what one might have hoped, and it is only in 2015, a decade after had been largely deployed, that Auger et al. proved that required ${\mathcal{O}}(n \log(n))$ comparisons for sorting arrays of length $n$ [@AuNiPi15].
Even worse, because of the lack of a systematic and theoretical analysis of this algorithm, several bugs were discovered only recently in both Python and Java implementations of [@GoRoBoBuHa15; @auger2018worst]. Meanwhile, since the inception of , several natural merge sorts have been proposed, all of which were meant to offer easy-to-prove complexity guarantees. Such algorithms include , introduced by Shivers in [@shivers02], as well as Takaoka’s [@Ta09] (equivalent constructions of this algorithm were also obtained in [@BaNa13]), Buss and Knop’s [@BuKno18], and the most recent algorithm , due to Munroe and Wild [@munro2018nearly].
These algorithms share most of the nice properties of , as summarised in Figure \[fig:prop:sort\] (columns 1 to 4). For instance, except , these are *stable* algorithms, which means that they sort repeated elements in the same order that these elements appear in the input. This is very important for merge sorts, because only adjacent runs will be merged, which allows merging directly arrays instead of having to use linked lists. Similarly, and are *online* algorithms. This means that they can sort (possibly infinite) streams of data, with the property that, at any moment after having been given access to the first $n$ items of the stream, they will already have presorted these items, which can then be actually sorted in time ${\mathcal{O}}(n)$. [^1] Finally, all these algorithms sort arrays of length $n$ in time ${\mathcal{O}}(n\log n)$, and, for all of them except , they even do it in time ${\mathcal{O}}(n+n\log\rho)$, where $\rho$ is the number of runs of the array. This is optimal in the model of sorting by comparisons [@Mannila1985], using the classical counting argument for lower bounds.
Some of these algorithms even adapt to the run lengths, and not only to the number of runs: if the array consists of $\rho$ runs of lengths $r_1,\ldots,r_\rho$, then they run in ${\mathcal{O}}(n + n {\mathcal{H}})$, where ${\mathcal{H}}$ is defined as ${\mathcal{H}}= H(r_1/n,\ldots,r_\rho/n)$ and $H(x_1,\ldots,x_\rho) = -\sum_{i=1}^\rho x_i \log_2(x_i)$ is the binary Shannon entropy. Since ${\mathcal{H}}\leqslant \log_2(\rho)$, this finer upper bound is once again optimal.
Algorithm Time complexity Stable Online Worst-case merge cost
----------- -------------------------------------- -------- -------- --------------------------------------------------------------------------------------
${\mathcal{O}}(n + n \log(\rho))$ ${\text{\rlap{$n \log_2(\rho)$}}\phantom{n \log_2(n)}} + {\mathcal{O}}(n)$
${\mathcal{O}}(n + n {\mathcal{H}})$ ${\text{\rlap{$3/2 \, n {\mathcal{H}}$}}\phantom{n \log_2(n)}} + {\mathcal{O}}(n)$
${\mathcal{O}}(n \log n)$ $n \log_2(n) + {\mathcal{O}}(n)$
${\mathcal{O}}(n + n {\mathcal{H}})$ ${\text{\rlap{$n {\mathcal{H}}$}}\phantom{n \log_2(n)}} + {\mathcal{O}}(n)$
${\mathcal{O}}(n + n {\mathcal{H}})$ ${\text{\rlap{$c_\alpha n {\mathcal{H}}$}}\phantom{n \log_2(n)}} + {\mathcal{O}}(n)$
${\mathcal{O}}(n + n {\mathcal{H}})$ ${\text{\rlap{$n {\mathcal{H}}$}}\phantom{n \log_2(n)}} + {\mathcal{O}}(n)$
${\mathcal{O}}(n + n {\mathcal{H}})$ ${\text{\rlap{$n {\mathcal{H}}$}}\phantom{n \log_2(n)}} + {\mathcal{O}}(n)$
Focusing only on the time complexity, four algorithms seem on par with each other, and finer complexity evaluations are required to separate them. Except , it turns out that these algorithms are, in fact, described only as policies for merging runs, the actual sub-routine used for merging runs being left implicit. Therefore, we settle for the following cost model.
Since naive merging algorithms approximately require $m+n$ element comparisons and element moves for merging two arrays of lengths $m$ and $n$, and since $m+n$ element moves may be needed in the worst case (for any values of $m$ and $n$), we measure below the complexity in terms of *merge cost* [@AuNiPi15; @BuKno18; @golin1993queue; @munro2018nearly]: the cost of merging two runs of lengths $m$ and $n$ is defined as $m+n$, and we identify the complexity of an algorithm with the sum of the costs of the merges processed while applying this policy.
In this new model, one can prove that the average merge cost of any natural merge sort must be at least $n {\mathcal{H}}+ {\mathcal{O}}(n)$. This makes and the only sorting algorithms with an optimal merge cost, as shown in the last column of Figure \[fig:prop:sort\]. Since is stable, it should be a natural option for succeeding as standard sorting algorithm in Python or Java. Nevertheless, and although its has implementations similar to that of , its merge policy is arguably more complicated, as outlined by the fact that it is not online anymore (and that making online would also make it even more complicated).
Therefore, there is yet to find an natural merge sort whose structure would be extremely close to that of , thus being online, and that would have an optimal merge cost (up to an additive term ${\mathcal{O}}(n)$).
In order to achieve this goal, a first direction to look at is that of $k$-aware algorithms. The class of $k$-aware algorithms was invented by Buss and Knop [@BuKno18], and characterises those algorithms whose merge policy is similar to that of . More precisely, is based on discovering runs on the fly, and storing them into a stack; then, it merges only runs that lie on the top of the stack, and such decisions are based only on the lengths of those top runs.
The rationale behind this process is that processing runs in such a way should be adapted to the architecture of computers, for instance by avoiding cache misses. Then, one says that a natural merge sort is $k$-aware if deciding which runs should be merged is based only on the lengths of the top $k$ runs of the stack, and if the runs merged belong themselves to these top $k$ runs.
Focusing on $k$-aware algorithms seems all the more relevant because some of the nice features of were also due to the high degree of tuning of the components (ii) and (iii). Hence, if one does not modify these components, and if one follows a merge policy that behaves in a way similar to that of , one may reasonably hope that those nice features of would be kept intact, even though their causes are not exactly understood. This suggests identifying natural merge sorts with their merge policy, and integrating the components (ii) and (iii) later.
#### Contributions
We propose a new natural merge sort, which we call . As advertised above, we will identify this algorithm with its run merge policy. is a blend between the algorithms and ; the purpose being to borrow nice properties from both algorithms. As a result, the merge policy of is extremely similar to that of , which means that switching from one algorithm to the other should be essentially costless: it would require changing only a dozen lines in the code of Java, and no unforeseen decrease in performance.
is a $3$-aware algorithm, which is stable and enjoys an optimal $n{\mathcal{H}}+ {\mathcal{O}}(n)$ upper bound on its merge cost. In particular, this algorithm is also online, like every $k$-aware algorithm, and it runs in time ${\mathcal{O}}(n + n{\mathcal{H}})$. Hence, appears as optimal with respect to all the criteria mentioned in Figure \[fig:prop:sort\]; it is the first known $k$-aware algorithm with a merge cost of $n{\mathcal{H}}+ {\mathcal{O}}(n)$, thereby answering a question left open by Buss and Knop in [@BuKno18]. Moreover, and due to its simple policy, the complexity proof of is very simple as well; below, we propose a short, self-contained version of this proof.
Then, still aiming to compare and , we investigate their *best-case* merge costs. It turns out that the merge cost of , in every case, is bounded between $n {\mathcal{H}}$ and $n ({\mathcal{H}}+2)$, these bounds being both close and optimal for any stable merge sort; at the same time, the merge cost of is only bounded above by $n({\mathcal{H}}+ \Delta)$, where $\Delta = 24/5 - \log_2(5) \approx 2.478$, which is slightly suboptimal. Hence, we design a variant of , called , which loses the online property, but whose merge cost also enjoys an $n ({\mathcal{H}}+2)$ upper bound.
Finally, we further explore the question, raised by Buss and Knop in [@BuKno18], of the optimality of $k$-aware algorithms on *all* arrays. More precisely, this question can be stated as follows: for a given integer $k$ and a real number $\varepsilon > 0$, does there exist a $k$-aware algorithm whose merge cost is at most $1+\varepsilon$ times the merge cost of any stable merge cost on any array? We prove that the answer is always negative when $k = 2$; for all $k \geqslant 3$, the set of numbers $\varepsilon$ for which the answer is positive forms an interval with no upper bound, and we prove that the lower bound of that interval is a positive real number, which tends to $0$ when $k$ grows arbitrarily.
and related algorithms {#sec:description}
======================
In this section, we describe the run merge policy of the algorithm and of related algorithms. The merge policy of algorithm is depicted in Algorithm \[alg:ASS\]. For the ease of subsequent sections, and because it does not make any proof harder, we shall consider as a parameterised algorithm: in addition to the array to sort, this algorithm also requires a positive integer $c$ as parameter.
In what follows, the value of the parameter $c$ will be considered as fixed once and for all. In subsequent sections, we will consider most specifically two choices for the parameter $c$, by either setting $c = 1$ or by setting $c = n+1$, where $n$ is the length of the array to be sorted. We respectively call and the algorithms obtained with these two parameter choices. At any moment except when we will specifically refer to the algorithm , the reader may safely imagine that we set $c = 1$.
$\operatorname{\texttt{runs}}{\ensuremath{\leftarrow}}$ the run decomposition of $A$ $\operatorname{{\mathcal{S}}}{\ensuremath{\leftarrow}}$ an empty stack ()[$\operatorname{\texttt{runs}}\neq \emptyset$]{}[ remove a run $R$ from $\operatorname{\texttt{runs}}$ and push $R$ onto $\operatorname{{\mathcal{S}}}$ ]{}
This algorithm is based on discovering monotonic runs and on maintaining a stack of such runs, which may be merged or pushed onto the stack according to whether the conditions of cases \#1 to \#4 apply. In particular, since these conditions only refer to the values of $\ell_1$, $\ell_2$ and $\ell_3$, and since only the runs $R_1$, $R_2$ and $R_3$ may be merged, this algorithm falls within the class of $3$-aware stable sorting algorithms such as described by Buss and Knop [@BuKno18] as soon as the choice of $c$ is independent of the input. This is the case of , but not of other variants such as .
Let us also present briefly some related algorithms. Like , all of these algorithms rely on discovering and maintaining runs in stack, although their merge policies follows different rules. More precisely, each of them can be obtained from by modifying the **while** loop of lines \[algline:inner\_while\_translated\] to \[algline:end\_inner\_loop\].
The first algorithm we present is , and is due to Peters [@Peters2015]. It is obtained by using the following **while** loop instead of that of :
As mentioned in the introduction, this algorithm enjoys a ${\mathcal{O}}(n + n {\mathcal{H}})$ worst-case merge cost. Two crucial elements in achieving this result consist in showing that some invariant (the fact that $r_i + r_{i+1} \leqslant r_{i+2}$ for all $i \geqslant 3$) is maintained, and in avoiding merging newly pushed runs until their length is comparable with that of the runs stored in the top of the stack. Original versions of missed the case $\heartsuit$, which made the invariant invalid and caused several implementation bugs [@GoRoBoBuHa15; @auger2018worst]. Yet, even this flawed version of the algorithm managed to enjoy a ${\mathcal{O}}(n + n {\mathcal{H}})$ worst-case merge cost. Unfortunately, the constant hidden in the ${\mathcal{O}}$ is rather high, as outlined by the following result.
\[thm:TS-merge-cost\] The worst-case merge cost of on inputs of length $n$ is bounded above by $3/2 \, n {\mathcal{H}}+ {\mathcal{O}}(n)$ and bounded below by $3/2 \, n \log_2(n) + {\mathcal{O}}(n)$.
The lower bound was proved in [@BuKno18], and the upper bound is proved in an extended version of [@auger2018worst].
Hence, and in order to lower the constant from $3/2$ to $1$, it was important to look for other merge policies. With this idea in mind, one hope comes from the algorithm , which was invented by Shivers [@shivers02]. It is obtained by using the following one-line **while** loop instead of that of :
This sorting algorithm *may* merge a newly pushed run even if its length is much larger than those of the runs stored in the stack, and therefore it is *not* adaptive to the number of runs nor to their lengths. Nevertheless, it still enjoys the nice property of having a worst-case merge cost that is optimal up to an additive linear factor, when the only complexity parameter is $n$.
The worst-case merge cost of on inputs of length $n$ that decompose into $\rho$ monotonic runs is both bounded above by $n \log_2(n) + {\mathcal{O}}(n)$ and bounded below by $\omega(n \log_2(\rho))$.
This result was proved in [@shivers02; @BuKno18]. Its proof, which we omit here, is very similar to our own analysis of in Section \[sec:analysis\] below. A crucial element of both proofs, as will be stated in Lemma \[lem:invariant-li\], is the fact that the sequence $\ell_1,\ell_2,\ldots,\ell_h$ shall always be increasing after some fixed index: that index is $2$ in the proof of [@BuKno18], and it is $3$ in Lemma \[lem:invariant-li\]. In essence, this invariant is similar to that of , but it allows decreasing the associated constant hidden in the ${\mathcal{O}}$ notation from $3/2$ to $1$.
The very idea of integrating features from these two algorithms led Buss and Knop to invent the algorithm [@BuKno18]. This algorithm is obtained by using the following **while** loop instead of that of :
The hope here is that both avoiding to merge newly pushed runs while they are too large and maintaining an invariant (on the integers $\ell_i$) similar to that of would make very efficient. Unfortunately, this algorithm suffers from the same design flaw as the original version of , and the desired invariant is *not* maintained. Even worse, the effects of not maintaining this invariant are much more severe here, as underlined by the following result.
\[thm:complexity-AugSS\] The worst-case merge cost of on inputs of length $n$ is ${\mathcal{O}}(n^2)$.
Consider some integer $k \geqslant 1$, and let $n = 4k$ and $\rho = 2k$. Let also $r_1,\ldots,r_\rho$ be the sequence of run lengths defined by $r_{2i-1} = 3$ and $r_{2i} = 1$ for all $i \leqslant k$. Note that $r_1+\ldots+r_\rho = n$.
Now, let us apply the algorithm on an array of $n$ integers that splits into increasing runs $R_1,\ldots,R_\rho$ of lengths exactly $r_1,\ldots,r_\rho$. One verifies quickly that the algorithm performs successively the following operations:
- push rhe runs $R_1$ and $R_2$ (case \#1);
- for all $i \in \{2,\ldots,k-1\}$, push the run $R_{2i-1}$ (case \#1), then merge the runs $R_{2i-2}$ and $R_{2i-3}$ (case \#2’), and push the run $R_{2i}$ (case \#1 again);
- keep merging the last two runs on the stack (line \[alg:collapse\]): we first merge the runs $R_{\rho-1}$ and $R_\rho$, and then, the $(m+1)$^th^ such merge involves runs of sizes $4$ and $4 m$.
Therefore, the merge cost of on that array is
$\displaystyle
{\mathsf{mc}}= \sum_{i=1}^k (r_{2i-1}+r_{2i}) + \sum_{m=1}^{k-1} 4 (m+1) =
n^2/8 + 3 n/2 - 4.
$
Conversely, in any (natural or not) merge sort, any element can be merged at most $n-1$ times, and therefore the total merge cost of such a sorting algorithm is at most $n(n-1)$.
Hence, it still remained to design an algorithm that would successfully integrate the features of both and . This new algorithm is .
Worst-case merge cost analysis of {#sec:analysis}
==================================
We mentioned in the introduction that enjoys excellent worst-case upper bounds in terms of merge cost. This statement is outlined by the result below.
*Note to the reviewers:* the proof of Theorem \[thm:complexity-nH+D\] is rejected to Appendix \[sec:app:proofs\], and is not meant to be present in the final version of the paper. The ideas of this proof are presented in page , after the proof of Theorem \[thm:complexity-nH+3\], and shall be present in the final version.
[theorem]{}[compNHDthm]{}\[thm:complexity-nH+D\] The merge cost of is bounded above by $n \left({\mathcal{H}}+ \Delta\right)$, where $\Delta = 24/5 - \log_2(5) \approx 2.478$.
The remainder of this section is devoted to proving the following weaker variant of Theorem \[thm:complexity-nH+D\], which already provides us with an $n \left({\mathcal{H}}+ 3\right)$ upper bound and whose proof is far less technical and therefore much more insightful.
\[thm:complexity-nH+3\] For every value of the parameter $c$, the merge cost of is bounded above by $n ({\mathcal{H}}+ 3 - \{\log_2(n/c)\}) - \rho-1$, where $\{x\} = x - \lfloor x \rfloor$ denotes the fractional part of the real number $x$.
In what follows, the value of the parameter $c$ is fixed once and for all. Then, we denote by $r$ the length of a run $R$, by $\ell$ the integer $\lfloor \log_2(r / c) \rfloor$ and by $\lambda$ the real number $\{\log_2(r / c)\} = \log_2(r / c) - \ell$. We adapt readily these notations when the name of the run considered varies, e.g., we note by $r'$ the length of the run $R'$, by $\ell'$ the integer $\lfloor \log_2(r' / c) \rfloor$ and by $\lambda'$ the real number $\{\log_2(r' / c)\}$. In particular, we will commonly denote the stack by $(R_1,\ldots,R_h)$, where $R_k$ is the $k$^th^ top-most run of the stack. The length of $R_k$ is then denoted by $r_k$, and we set $\ell_k = \lfloor \log_2(r_k / c) \rfloor$. With this notation in mind, we first prove two auxiliary results.
\[lem:l-small-increase\] When two runs $R$ and $R'$ are merged into a single run $R''$, we have $\ell'' \leqslant \max\{\ell,\ell'\}+1$.
Without loss of generality, we assume that $r \leqslant r'$. In that case, it comes that
$\displaystyle
2^{\ell''} c \leqslant r'' = r + r' \leqslant 2 r' < 2 \times 2^{\ell'+1} c = 2^{\ell'+2} c,
$
and therefore that $\ell'' \leqslant \ell'+1$.
\[lem:invariant-li\] At any time during the execution of the algorithm, if the stack of runs is ${\mathcal{S}}= (R_1,\ldots,R_h)$, we have:
currentlabel
$\displaystyle
\ell_2 \leqslant \ell_3 < \ell_4 < \ldots < \ell_h.\label{ell-incr}$[()]{.nodecor}
The proof is done by induction. First, if $h \leqslant 2$, there is nothing to prove: this case occurs, in particular, when the algorithm starts.
We prove now that, if the inequalities of hold at some point, they still hold after an update on the stack, triggered by any of the cases \#1 to \#4. This is done by a case analysis, denoting by ${\mathcal{S}}= (R_1,\ldots,R_h)$ the stack before the update and by $\overline{{\mathcal{S}}} = ({\overline{R}}_1,\ldots,{\overline{R}}_{\overline{h}})$ the stack after the update:
- If case \#1 just occured, then a new run ${\overline{R}}_1$ was just pushed onto the stack; we have $\overline{h} = h+1$, and ${\overline{R}}_i = R_{i-1}$ for all $i \geqslant 2$. Since holds in ${\mathcal{S}}$, we already have $\overline{\ell}_4 < \overline{\ell}_5 < \ldots < \overline{\ell}_{\overline{h}}$. Moreover, since case \#1 occured, none of the conditions for triggering the cases \#2 to \#4 holds in ${\mathcal{S}}$. This means that $\ell_1 < \ell_2 < \ell_3$ or, equivalently, that $\overline{\ell}_2 < \overline{\ell}_3 < \overline{\ell}_4$, which shows that holds in $\overline{{\mathcal{S}}}$.
- If one of the cases \#2 to \#4 just occured, then $\overline{h} = h-1$, ${\overline{R}}_i = R_{i+1}$ for all $i \geqslant 3$, and ${\overline{R}}_2$ is either equal to $R_3$ (in case \#4) or to the result of the merge between the runs $R_2$ and $R_3$ (in cases \#2 and \#3). Thanks to Lemma \[lem:l-small-increase\], this means that either $\overline{\ell}_2 = \ell_3$ (in case \#4) or that $\overline{\ell}_2 \leqslant \max\{\ell_2,\ell_3\}+1 = \ell_3+1$. Moreover, since holds in ${\mathcal{S}}$, we already have $\ell_3 < \overline{\ell}_3 < \overline{\ell}_4 < \ldots < \overline{\ell}_{\overline{h}}$. It follows that $\overline{\ell}_2 \leqslant \ell_3+1 \leqslant \overline{\ell}_3$, which shows that also holds in $\overline{{\mathcal{S}}}$.
Before going further in the proof of Theorem \[thm:complexity-nH+3\], we first need to split run merges into several classes. When merging two runs $R$ and $R'$ into one bigger run $R''$, we say that the merge of $R$ is *expanding* if $\ell'' \geqslant \ell + 1$, and is *non-expanding* otherwise. Hence, we refer below to the merge of $R$ with $R'$ and to the merge of $R'$ with $R$ as if these were two separate objects. In particular, the merge cost is itself split in two parts: one part, for a cost of $r$, is assigned to the merge of $R$, and the other part, for a cost of $r'$, is assigned to the merge of $R'$.
Then, note that, if $\ell \leqslant \ell'$, the merge of $R$ with $R'$ is necessarily expanding. Consequently, when two runs $R$ and $R'$ are merged, either the merge of $R$ or of $R'$ is expanding. In particular, if $\ell = \ell'$, then both merges of $R$ and of $R'$ must be expanding. Hence, we say that the merge between $R$ and $R'$ is *intrinsically expanding* if $\ell = \ell'$.
\[lem:expanding-nH\] The total cost of expanding merges is at most $n ({\mathcal{H}}- \{\log_2(n/c)\}) + \Lambda$, where $\Lambda$ is defined as $\Lambda = \sum_{i=1}^\rho r_i \lambda_i$.
While the algorithm is performed, the elements of a run $R$ of initial length $r$ may take part to at most
$\displaystyle \lfloor \log_2(n/c) \rfloor - \ell =
(\log_2(n/c) - \{\log_2(n/c)\}) - (\log_2(r/c) - \lambda) =
\log_2(n/r) + \lambda - \{\log_2(n/c)\}$
expanding merges. Consequently, if the array is initially split into runs of lengths $r_1,\ldots,r_\rho$, the total cost of expanding merges is at most
$\displaystyle \sum_{i=1}^\rho r_i \left(\log_2(n/r_i) + \lambda_i - \{\log_2(n/c)\}\right) =
n ({\mathcal{H}}- \{\log_2(n/c)\}) + \Lambda.
\vspace{-\baselineskip}$
It remains to prove that the total cost of non-expanding merges is at most $3n - \Lambda - \rho$. This requires further splitting sequences of merges based on the case that triggered these merges. Hence, when discussing the various merges that may arise, we also call *$\#k$-merge* every merge triggered by a case $\#k$.
Now, we define the *starting sequence* of a run $R$ as the (possibly empty) maximal sequence of consecutive \#2-merges that follows immediately the push of $R$ onto the stack. We call *middle sequence* of $R$ the maximal sequence of consecutive \#2- and \#3-merges that follows the starting sequence of $R$, *critical merge* of $R$ the \#4-merge that immediately follows the middle sequence of $R$ (if it exists), and *ending sequence* of $R$ the maximal sequence of consecutive merges that follows the critical merge of $R$ and precedes the next run push. If the middle sequence of $R$ is followed by a push run (or by exiting the main loop of ), then the critical merge of $R$ simply does not exist, and the ending sequence of $R$ is then defined as the empty sequence.
Below, we also call *non-expanding cost* of a sequence of merges the total cost of the *non-expanding merges* included in this sequence. Before looking more closely at the non-expanding cost of the starting, middle and ending sequences, we first prove invariants similar to that of Lemma \[lem:invariant-li\].
\[lem:invariant-mid-end\] Each middle sequence consists only of \#3-merges, and each ending sequence consists only of \#4-merges.
Let ${{\mathcal{M}}_{\mathrm{mid}}}$ be a middle sequence. By construction, it can contain \#2-merges and \#3-merges only. Hence, we just need to prove that it cannot contain \#2-merges, i.e., that $\ell_1 < \ell_3$ at any time during the sequence (including just before it starts). This statement is proved by induction.
First, the sequence starts with a \#3-merge, which means precisely that $\ell_1 < \ell_3$ just before it starts. Then, we prove that, if the inequality $\ell_1 < \ell_3$ holds at some point, it still holds after a \#3-merge. Indeed, let us denote by ${\mathcal{S}}= (R_1,\ldots,R_h)$ the stack before the merge and by $\overline{{\mathcal{S}}} = ({\overline{R}}_1,\ldots,{\overline{R}}_{\overline{h}})$ the stack after the merge: since ${\overline{R}}_3 = R_4$ and ${\overline{R}}_1 = R_1$, it follows from Lemma \[lem:invariant-li\] that $\overline{\ell}_1 = \ell_1 < \ell_3 < \ell_4 = \overline{\ell}_3$, which proves that ${{\mathcal{M}}_{\mathrm{mid}}}$ consists only of \#3-merges.
Similarly, let ${{\mathcal{M}}_{\mathrm{end}}}$ be an ending sequence. This time, we need to prove that it cannot contain \#2-merges or \#3-merges, i.e., that $\ell_1 < \ell_3$ and $\ell_2 < \ell_3$ at any time during the sequence. We also proceed by induction. First, if ${{\mathcal{M}}_{\mathrm{end}}}$ is non-empty, then the critical merge of $R$ exists, and it is a \#4-merge. This means that $\ell_1 < \ell_3$ and $\ell_2 < \ell_3$ just before that merge occurs.
Then, we prove that, if the inequalities $\ell_1 < \ell_3$ and $\ell_2 < \ell_3$ hold at some point, they still hold after a \#4-merge. Indeed, let us denote again by ${\mathcal{S}}$ and $\overline{{\mathcal{S}}}$ the stacks before and after the merge. We have ${\overline{R}}_3 = R_4$, ${\overline{R}}_2 = R_3$, and ${\overline{R}}_1$ is the result of the merges of $R_1$ and $R_2$. Hence, and due to Lemmas \[lem:l-small-increase\] and \[lem:invariant-li\], it follows that $\overline{\ell}_1 \leqslant \max\{\ell_1,\ell_2\}+1 \leqslant \ell_3 < \ell_4 = \overline{\ell}_3$ and that $\overline{\ell}_2 = \ell_3 < \ell_4 = \overline{\ell}_3$.
\[lem:invariant-end\] At any time during an ending sequence, including just before it starts, we have $\ell_1 \leqslant \ell_2$.
Lemma \[lem:invariant-mid-end\] states that the ending sequence consists only of \#4-merges, and is preceded by a \#4-merge. Hence, we just need to prove that the inequality $\ell_1 \leqslant \ell_2$ must hold after any \#4-merge.
Indeed, let ${\mathcal{S}}= (R_1,\ldots,R_h)$ the stack before the merge and by $\overline{{\mathcal{S}}} = ({\overline{R}}_1,\ldots,{\overline{R}}_{\overline{h}})$ the stack after the merge: we have ${\overline{R}}_2 = R_3$, and ${\overline{R}}_1$ is the result of the merges of $R_1$ and $R_2$. Moreover, since the conditions for cases \#2 and \#3 are not satisfied by ${\mathcal{S}}$, we have $\ell_1 < \ell_3$ and $\ell_2 < \ell_3$. It follows from Lemma \[lem:l-small-increase\] that $\overline{\ell}_1 \leqslant \max\{\ell_1,\ell_2\}+1 \leqslant \ell_3 = \overline{\ell}_2$.
\[cor:mid-end-intr-expanding\] Each middle or ending sequence contains intrinsically expanding merges only.
Let $\mathbf{m}$ be a merge of a middle or ending sequence, and let ${\mathcal{S}}= (R_1,\ldots,R_h)$ be the stack just before the merge occurs:
- If $\mathbf{m}$ belongs to a middle sequence, then Lemma \[lem:invariant-mid-end\] proves that it is a \#3-merge, between the runs $R_2$ and $R_3$, such that $\ell_2 \leqslant \ell_3$ (because of Lemma \[lem:invariant-li\]) and $\ell_3 \leqslant \ell_2$ (because it is a \#3-merge). Consequently, this merge is intrinsically expanding.
- If $\mathbf{m}$ belongs to an ending sequence, then Lemma \[lem:invariant-mid-end\] proves that it is a \#4-merge, between the runs $R_1$ and $R_2$, such that $\ell_1 \leqslant \ell_2$ (because of Lemma \[lem:invariant-end\]) and $\ell_2 \leqslant \ell_1$ (because it is a \#4-merge). Consequently, this merge is also intrinsically expanding.
\[lem:2pow-x\] For all real numbers $x$ such that $0 \leqslant x \leqslant 1$, we have $2^{1-x} \leqslant 2 - x$.
Every function of the form $x \mapsto \exp(t x)$, where $t$ is a fixed real parameter, is convex. Therefore, the function $f : x \mapsto 2^{1-x} - (2-x)$ is convex too. It follows, for all $x \in [0,1]$, that $f(x) \leqslant \max\{f(0),f(1)\} = 0$, which completes the proof.
With the help of these auxiliary results, we may now prove the following statement, whose proof is also the most technical one in this section.
\[lem:non-expanding-run-2r\] The total non-expanding cost of the critical merge (if it exists) and of the starting, middle and ending sequences of a run $R$ is at most $(2 - \lambda) r - 1$.
Let ${\mathcal{S}}= (R,R_2,\ldots,R_h)$ be the stack just after the run $R$ has been pushed, and let $k$ be the largest integer such that $\ell_k \leqslant \ell$. We will respectively denote by ${{\mathcal{M}}_{\mathrm{start}}}$, ${{\mathcal{M}}_{\mathrm{mid}}}$ and ${{\mathcal{M}}_{\mathrm{end}}}$ the starting, middle and ending sequences of $R$, and by ${\mathbf{M}_{\mathrm{cr}}}$ the critical merge of $R$.
The starting sequence ${{\mathcal{M}}_{\mathrm{start}}}$ consists in merging the run $R_2$ with $R_3$, then merging the resulting run successively with $R_4, R_5, \ldots, R_k$. Due to Lemma \[lem:invariant-li\], the non-expanding merges of ${{\mathcal{M}}_{\mathrm{start}}}$ involve at most once each of the runs $R_3,\ldots,R_k$, when these runs are first merged. Let $k'$ be the largest integer, if any, such that the first merge of $R_{k'}$ is non-expanding, and let $R^\ast$ be the run into which $R_{k'}$ is merged. Then, all runs $R_2,\ldots,R_{k'}$ are merged into that run, and thus the non-expanding cost of ${{\mathcal{M}}_{\mathrm{start}}}$ is at most $r_3+\ldots+r_{k'-1}+r_{k'} = r^\ast - r_2 \leqslant r^\ast - 1$. Since $\lfloor \log_2(r^\ast / c) \rfloor = \ell_{k'} \leqslant \ell$, it even follows that $r^\ast < 2^{\ell+1} c = 2^{1-\lambda} r$, and therefore the non-expanding cost of ${{\mathcal{M}}_{\mathrm{start}}}$ is at most $2^{1-\lambda} r-1$.
Then, Corollary \[cor:mid-end-intr-expanding\] ensures that the sequences ${{\mathcal{M}}_{\mathrm{mid}}}$ and ${{\mathcal{M}}_{\mathrm{end}}}$ contain expanding merges only, hence their non-expanding cost is $0$.
Finally, assume that the critical merge ${\mathbf{M}_{\mathrm{cr}}}$ exists and is non-expanding, and let also $\overline{{\mathcal{S}}} = ({\overline{R}}_1,\ldots,{\overline{R}}_{\overline{h}})$ be the stack just before it occurs. Note that ${\overline{R}}_1 = R$. Then, it must be the case that $\ell = \overline{\ell}_1 > \overline{\ell}_2$. Moreover, if the starting sequence ${{\mathcal{M}}_{\mathrm{start}}}$ had a non-zero non-expanding cost, then the above-defined run $R^\ast$ turns out to be entirely contained in the run ${\overline{R}}_2$. It follows that $r+r^\ast \leqslant r+\overline{r}_2 < 2^{\ell+1} c =
2^{1-\lambda} r$.
Let us gather the above results. First, the starting sequence ${{\mathcal{M}}_{\mathrm{start}}}$ has a non-expanding cost bounded above by $r^\ast - 1\leqslant 2^{1-\lambda} r - 1$. Then, the non-expanding costs of both sequences ${{\mathcal{M}}_{\mathrm{mid}}}$ and ${{\mathcal{M}}_{\mathrm{end}}}$ are equal to $0$. Finally, if the merge ${\mathbf{M}_{\mathrm{cr}}}$ is well-defined and non-expanding, its non-expanding cost is exactly $r$, and we also have $r+r^\ast - 1 \leqslant 2^{1-\lambda} r - 1$.
Hence, in all cases, it the total non-expanding cost of the sequences ${{\mathcal{M}}_{\mathrm{start}}}$, ${{\mathcal{M}}_{\mathrm{mid}}}$ and ${{\mathcal{M}}_{\mathrm{end}}}$ and of ${\mathbf{M}_{\mathrm{cr}}}$ is at most $2^{1-\lambda} r - 1$. Using Lemma \[lem:2pow-x\] completes the proof.
We may now conclude with the proof of Theorem \[thm:complexity-nH+3\].
Lemma \[lem:expanding-nH\] states that the total cost of expanding merges is at most $n {\mathcal{H}}+ \Lambda$. Then, Lemma \[lem:non-expanding-run-2r\] states that the total non-expanding cost of the critical merge and of the starting, middle and ending sequences of a given run $R$ is at most $(2 - \lambda) r - 1$. Hence, and taking all runs into account, the total non-expanding cost of these sequences is at most $2n - \Lambda - \rho$.
It remains to take care of the non-expanding cost of the runs merged in line \[alg:collapse\] of Algorithm \[alg:ASS\]. The merges performed in line \[alg:collapse\] are the same merges as those that would occur in the *starting sequence* of an additional (fictitious) run of length $n$ that we would have appended to our array. Performing the analysis of Lemma \[lem:non-expanding-run-2r\] in that special case proves that the total non-expanding cost of these merges is at most $n - 1$. This completes the proof of Theorem \[thm:complexity-nH+3\].
Now that Theorem \[thm:complexity-nH+3\] has been proved, let us present briefly the ideas that lead to Theorem \[thm:complexity-nH+D\] itself. Although the main interest of Theorem \[thm:complexity-nH+D\] is that the constant $\Delta$ it refers to is optimal, as underlined in Proposition \[pro:ass-not-optimal\], the tools it is based on are quite more complex, and seeing why these tool fit together is less intuitive.
\[proof:thm:complexity-nH+D\]The cost allocation used for proving Theorem \[thm:complexity-nH+3\] is not sufficient to prove the better bound mentioned in Theorem \[thm:complexity-nH+D\]. Therefore, we must define a notion of *potential* of a sequence of runs, which will fit more closely the merge cost needed to merge this sequence, and which is just a sum of potentials of individual runs.
The potential of a run $R$ is defined as the real number
$\displaystyle \mathbf{Pot}(R) = \begin{cases}
2^\ell \, \Phi(r_\bullet) \, c - \ell \, r & \text{if $R$ is in the stack;} \\
2^{\ell+1} \, c - \ell \, r & \text{if $R$ has not yet been pushed onto the stack,}
\end{cases}$
where $\ell \in \mathbb{Z}$ and $r_\bullet \in [0,1)$ are defined by the relation $r = 2^\ell (1+r_\bullet) c$, and $\Phi : [0,1] \mapsto \mathbb{R}$ is the function defined by $\Phi : x \mapsto \max\{(2-5x)/3,1/2-x,0\}$.
The proof itself consists in showing that the merges performed by can be gathered into groups whose merge cost is bounded above by the variation of this potential, which necessitates careful disjuntions of cases (e.g., these cases may depend on which of the starting, middle or ending sequences exist, after a run $R$ has been pushed onto the stack). Theorem \[thm:complexity-nH+D\] then follows from evaluating the global variation in potential during the algorithm.
Best-case and worst-case merge costs {#sec:best-case}
====================================
In the introduction, we mentioned that, unlike , the algorithm is $3$-aware and adapted for sorting streams of data, which suggests that it might be preferred to . However, the worst-case merge cost of is only of $n ({\mathcal{H}}+ 2)$, whereas the above analysis only proves that the merge cost of is bounded above by $n ({\mathcal{H}}+ \Delta)$. Hence, and most notably in cases where ${\mathcal{H}}$ is small, might be the best option. Furthermore, in other cases than the worst case, we do have little information on the relative costs of and of .
We address these problems as follows. First, we derive lower bounds on the best-case merge cost of any merge policy. Then, we prove that the worst-case merge cost of is actually optimal. Finally, we study , which is the special case of obtained by setting $c = n+1$, as mentioned in Section \[sec:description\], and we prove there that also matches the worst-case optimal merge cost of .
Best-case merge cost
--------------------
Given a sequence $\mathbf{r} = (r_1,\ldots,r_\rho)$ of run lengths and an array of length $n = r_1 + \ldots + r_\rho$ that splits into monotonic runs of lengths $r_1,\ldots,r_\rho$, what is the best merge cost of any merge policy? An answer to this question is given by Theorem \[thm:lower-bound-merge-cost\], which is a slight refinement of results from [@BaNa13; @Mannila1985; @munro2018nearly].
This answer comes from the analysis of *merge trees*, which we describe below, and of the algorithm [@BaNa13; @Ta09], which is *not* a stable natural merge sort, and whose merge policy is described in Algorithm \[alg:MinS\]. Note that may require merging *non-adjacent* runs, which might therefore be less easy to implement than just merging adjacent runs. Yet, such merges (between runs of lengths $m$ and $n$) can still be carried in time $m+n$, for instance by using linked lists, and therefore the merge cost remains an adequate measure of complexity for this algorithm.
$\operatorname{\texttt{runs}}{\ensuremath{\leftarrow}}$ the run decomposition of $A$
Given a merge policy and a sequence of runs, we define a binary rooted tree, called *merge tree*, as follows. Every node of the tree is identified with a run, either present in the initial sequence or created by the algorithm. The runs in the sequence are the leaves of the tree, and when two runs $R_1$ and $R_2$ are merged together in a run $R'$, then $R'$ is identified with the internal node whose children are $R_1$ and $R_2$: if the run $R_1$ was placed to the left of $R_2$ in the sequence to be sorted, then $R_1$ is the left sibling of $R_2$.
Any such tree is the tree of a binary prefix encoding on a text on the alphabet $\{1,2,\ldots,\rho\}$, which contains $r_i$ characters $i$ for all $i \leqslant \rho$. Furthermore, denoting by $d_i$ the depth of the leaf $R_i$, both the length of this code and the merge cost of the associated merge policy are equal to $\sum_{i=1}^\rho d_i r_i$. Hence, character encoding algorithms unsurprisingly yield efficient sorting algorithms.
\[thm:lower-bound-merge-cost\] The merge cost of any (stable or not) merge policy on a non-sorted array is minimised by the algorithm , and is at least $\max\{n, n {\mathcal{H}}\}$.
The merge tree of the algorithm is a Huffman tree. This means that is indeed the natural merge sort with the least merge cost, and that this merge cost is the length of a Huffman code for a text containing $r_i$ occurrences of the character $i$. Such a text has entropy ${\mathcal{H}}$, and therefore Shannon theorem proves that the associated Huffman code has length at least $n {\mathcal{H}}$. Furthermore, it is clear that the merge cost of ${\textsf{MinimalSort}\xspace}$ must be at least $n$, which completes the proof.
Optimality of worst-case merge costs
------------------------------------
The lower bound provided by Theorem \[thm:lower-bound-merge-cost\] matches quite well the worst-case merge costs of both and . In particular, and independently of the sequence to be sorted, the merge cost of (respectively, ) lies between $n {\mathcal{H}}$ and $n({\mathcal{H}}+ 2)$ (respectively, $n({\mathcal{H}}+ \Delta)$).
This shows, among others, that both and are very close to optimal when ${\mathcal{H}}$ is large. When ${\mathcal{H}}$ is small, however, the respective performances of and are still worth investigating. In particular, and given the tiny margin of freedom between these lower and upper bounds, it becomes meaningful to check whether our upper bounds are indeed optimal.
We prove below that, unlike or for any fixed value of $c$, the worst-case merge costs of and of are optimal.
\[pro:pos-optimal\] Let ${\mathcal{S}}$ be a stable merge policy. Assume that there exist real constants $\alpha$ and $\beta$ such that the merge cost of ${\mathcal{S}}$ can be bounded above by $n (\alpha {\mathcal{H}}+ \beta)$. Then, we have $\alpha \geqslant 1$ and $\beta \geqslant 2$.
First, Theorem \[thm:lower-bound-merge-cost\] proves, by considering arbitrarily large values of ${\mathcal{H}}$, that $\alpha \geqslant 1$.
Second, let $n \geqslant 6$ be some integer, and let $\mathbf{a}$ be some array of data that splits into runs of lengths $2$, $n-4$ and $2$. We have $n {\mathcal{H}}= 4 \log_2(n/2) + (n-4) \log_2\left(1+4/(n-4)\right) \leqslant 4 \log_2(n) + 2$. However, every stable merge policy has a merge cost $2n-2$ when sorting $\mathbf{a}$. Hence, by considering arbitrarily large values of $n$, it follows that $\beta \geqslant 2$.
Note that Proposition \[pro:pos-optimal\] indeed proves that the worst-case merge cost of is optimal, since it matches the upper bound of $n(\alpha {\mathcal{H}}+ \beta)$ with $\alpha = 1$ and $\beta = 2$. Addressing the optimality of the worst-case merge cost of requires considering another example.
\[pro:ass-not-optimal\] Let $c$ be a fixed parameter value, and let $\beta$ be a real constant such that the merge cost of can be bounded above by $n({\mathcal{H}}+ \beta)$ when $n$ is large enough. Then, we have $\beta \geqslant \Delta$.
Let $k \geqslant 2$ some integer, and let $m = \lfloor 2^k c \rfloor$. Consider the sequence of run lengths $\mathbf{r} = (1,2 m -4,1,m+1,1,2 m+1,1)$. For every sequence $s \in {\mathcal{S}}_{\mathbf{r}}$, we have $n = 5 m+2 = 5 \cdot 2^k c + {\mathcal{O}}(1)$ and
$\displaystyle
n {\mathcal{H}}= {\textstyle \sum_{i=1}^7} \log_2(n/r_i) \, r_i = n \left(\log_2(5) - 4/5 + o(1)\right).$
On the other hand, the merge cost of is equal to ${\mathsf{mc}}= 20m-3 = n \left(4 + o(1)\right)$. Since ${\mathsf{mc}}\leqslant n ({\mathcal{H}}+\beta)$, it follows that $\beta \geqslant 4 - (\log_2(5) - 4/5) = 24/5 - \log_2(5)$.
Hence, when $c$ is fixed a priori, and in particular when we set $c = 1$, the worst-case merge time of is *not* optimal. This is not very surprising since, unlike , the algorithm cannot take into account the total length of the input until that end is indeed reached.
Indeed, optimal worst-case merge times can be recovered by taking into account the length of the input, when setting the value of the parameter $c$. This is how we designed , which is nothing but the algorithm when sorting arrays of length $n$.
\[pro:LASS-worst-case\] The merge cost of is bounded above by $n({\mathcal{H}}+2)$.
Theorem \[thm:complexity-nH+3\] states that has a merge cost
$\displaystyle {\mathsf{mc}}\leqslant n ({\mathcal{H}}+ 3 - \{\log_2(n/(n+1))\}) - \rho-1.$
Since
$\displaystyle \{\log_2(n/(n+1))\} = \log_2(n/(n+1)) - \lfloor \log_2(n/(n+1)) \rfloor =
1 - \log_2(1+1/n) \geqslant 1 - 2/n,$
it follows that ${\mathsf{mc}}\leqslant n ({\mathcal{H}}+ 2) - \rho + 1 \leqslant n ({\mathcal{H}}+2)$.
Approximately optimal sorting algorithms {#sec:optimal}
========================================
Although is optimal up to an additive factor of at most $\Delta n$, this factor may still be of importance when considering arrays of data with small run-length entropy ${\mathcal{H}}$. Note that such arrays may have arbitrarily large lengths, but also arbitrarily many monotonic runs. This is, for instance, the case of arrays whose run lengths form the sequence $\mathbf{r} = (1,1,\ldots,1,k^2)$ whose $k$ first terms are integers $1$: although this sequence contains $k+1$ terms, it is associated with a value of ${\mathcal{H}}\approx 2 \log_2(k) / k$.
Hence, and since the parameters $n$, $\rho$ and ${\mathcal{H}}$ may vary more or less independently from each other (besides the fact that $2^{\mathcal{H}}\leqslant \rho \leqslant n$), we aim for the uniform approximation result captured by the following definition.
Let ${\mathcal{A}}$ be some stable natural merge sort, and let $\varepsilon \geqslant 0$ be a real number. We say that ${\mathcal{A}}$ is *asymptotically $\varepsilon$-optimal* if there exists an integer $\rho \geqslant 1$ such that, for every stable natural merge sort ${\mathcal{B}}$ and every array to be sorted and consisting of at least $\rho$ monotonic runs, the respective merge costs $\mathbf{m}_a$ and $\mathbf{m}_b$ of ${\mathcal{A}}$ and ${\mathcal{B}}$ are such that $\mathbf{m}_a \leqslant (1 + \varepsilon) \mathbf{m}_b$. If, furthermore, we can set $\rho = 1$, then we simply say that ${\mathcal{A}}$ is *$\varepsilon$-optimal*.
Below, we prove two results. The first one is negative, and states that the algorithms and (or even and ) fall into a family of algorithms, called $k$-aware algorithms, that cannot contain any $\varepsilon$-optimal when $\varepsilon$ is sufficiently small. The second one, on the other hand, is positive, and consists in building up, for every $\varepsilon > 0$, a variant of that is both $\varepsilon$-optimal and $k$-aware for some constant $k$.
First, we define more precisely the family of $k$-aware algorithms, to which all these algorithms belong; our notion subsumes and generalises slightly the notion of *awareness* of Buss and Knop [@BuKno18]. In particular, while and are $(4,3)$- and $(3,3)$-aware algorithms in the sense of Buss and Knop, this novel notion also captures and , which are respectively length-$(\infty,3)$- and length-$(3,3)$-aware algorithms.
Let $k$ and $\ell$ be elements of the set $\{0,1,2,\ldots\} \cup \{\infty\}$, with $k \geqslant \ell$. A deterministic sorting algorithm is said *$(k,\ell)$-aware* (or simply *$k$-aware* if $k = \ell$) if it sorts arrays of data by manipulating a stack of runs and operating as follows:
- the algorithm discovers the monotonic runs (in which the data is split) from the left to the right, which it pushes on the stack when discovering them;
- the algorithm is allowed to merge two consecutive runs in the $\ell$ top runs in its stack only, and its decision may be based only on the lengths of the top $k$ runs of the stack; if $\ell = \infty$, the algorithm may merge any two consecutive runs in its stack; if $k = \infty$, it is granted an infinite memory, and thus its decisions may be based all the push or merge operations it performed (and on the lengths of the runs involved in these operations).
If, furthermore, the algorithm is given access to the length of the array and can base its decisions on this information, then we say that it is a *length-$(k,\ell)$-aware* algorithm.
We present now results going in opposite directions, and which, taken together, form a first step towards finding the best approximation factor of $k$-aware algorithms: one direction is supported by Propositions \[pro:no-epsilon-optimal\] and \[pro:never-epsilon-optimal\], and the other direction is supported by Theorem \[thm:eta-optimal\].
\[pro:no-epsilon-optimal\] Let $k \geqslant 2$ be an integer, and let
$\displaystyle \varepsilon_k = 1/(5(k+3)2^{k+3}).$
No length-$(\infty,k)$-aware sorting algorithm is asymptotically $\varepsilon_k$-optimal.
Our proof is structured as follows. Let $\rho = 2^{k+1}$, and let $n \geqslant 2$ be some positive integer. We design two arrays $\mathbf{a}^1$ and $\mathbf{a}^{-1}$ with $\rho+2^n$ runs and with the same length, such that no length-$(\infty,k)$-aware algorithm can approach the merge cost of an optimal stable merge algorithm by a factor $1 + \varepsilon_k$ on both arrays $\mathbf{a}^1$ and $\mathbf{a}^{-1}$.
First, here is how we build these two arrays. Let $r_1 = r_{\rho-1} = 3 \times 2^{2n+1}$, $r_i = 2^{2n+2}$ for all $i \in \{2,\ldots,\rho-2\}$, $r_\rho = 2^{2n+k+4} (k+3)$ and $r_i = 1$ for all $i \in \{\rho+1,\ldots,\rho+2^n\}$. For $\epsilon = \pm 1$, let us denote by $\mathbf{r}^\epsilon$ the sequence $(r_1,\ldots,r_{\rho-2},r_{\rho-1}+\epsilon 2^{2n},r_\rho-\epsilon 2^{2n},r_{\rho+1},\ldots,r_{\rho+2^n})$. Then, let $\mathbf{a}^\epsilon$ be an array of data that belongs to ${\mathcal{S}}_{\mathbf{r}^\epsilon}$, i.e., that splits into monotonic runs $R_1,\ldots,R_{\rho+2^n}$ of lengths $r_1,\ldots,r_{\rho-2},r_{\rho-1}+\epsilon 2^n,r_\rho-\epsilon 2^n,r_{\rho+1},\ldots,r_{\rho+2^n}$. Note that $\mathbf{a}^\epsilon$ has more than $2^n$ runs, and that its length does not depend on $\epsilon$.
Let us split the array $\mathbf{a}^\epsilon$ into the following three *blocks*: the *left block* consists in the runs $R_1,\ldots,R_{\rho-1}$, the *middle block* consists of the run $R_\rho$, and the *right block* consists of the runs $R_{\rho+1},\ldots,R_{\rho+2^n}$.
Then, let ${\mathcal{B}}$ be the algorithm that operates on $\mathbf{a}^\epsilon$ as follows. First, ${\mathcal{B}}$ merges every pair of runs $R_{2i}$ and $R_{2i-\epsilon}$, for $i \leqslant 2^k-1$. After that step has been carried, the number of runs in each block is a power of two: ${\mathcal{B}}$ sorts each block by acting like , i.e., it keeps merging every pair of adjacent runs, thereby reducing by half the number of runs in the block, until every block contains only one run. Finally, ${\mathcal{B}}$ merges the middle and right block, then merges the resulting block with the left block. One checks easily that the merge costs of ${\mathcal{B}}$ on $\mathbf{a}^1$ and $\mathbf{a}^{-1}$ are less than $\mathbf{M}$, where $\mathbf{M} = 2 r_\rho + (k+2) 2^{2n+k+3} + (n+2) 2^n$.
Then, let ${\mathcal{A}}$ be some length-$(\infty,k)$-aware algorithm. Let us assume that, denoting by $\mathbf{m}_a^\epsilon$ and by $\mathbf{m}_b^\epsilon$ the respective merge costs of ${\mathcal{A}}$ and ${\mathcal{B}}$ on the array $\mathbf{a}^\epsilon$, we have both $\mathbf{m}_a^1 \leqslant (1 + \varepsilon_k) \mathbf{m}_b^1$ and $\mathbf{m}_a^{-1} \leqslant (1 + \varepsilon_k) \mathbf{m}_b^{-1}$. We shall prove that this assumption is invalid, thereby proving Proposition \[pro:no-epsilon-optimal\].
Indeed, let us first focus on the array $\mathbf{a}^1$. If, when sorting the array $\mathbf{a}^1$, the algorithm ${\mathcal{A}}$ includes the run $R_\rho$ in three merge operations or more, its merge cost will be at least $3 r_\rho$. On the other hand, since $n \geqslant 2$, we know that
$\displaystyle \mathbf{M} - 2 r_\rho = (k+2) 2^{2n+k+3} + (n+2) 2^n \leqslant
2^{2n+k+3} (k+3) = r_\rho / 2.$
It follows that $\mathbf{m}_a^1 \geqslant 3 r_\rho \geqslant 6 \mathbf{M} / 5 \geqslant 6 \mathbf{m}_b^1 / 5 > (1+\varepsilon_k) \mathbf{m}_b^1$.
Hence, we know that ${\mathcal{A}}$ can merge $R_\rho$ at most twice. This means that ${\mathcal{A}}$ merges $R_\rho$ once with the entire left block, and once with the entire right block. Since the left and right blocks have respective lengths $2^{2n+k+3}$ and $2^n$, these two merges that involve $R_\rho$ have a total merge cost of $2 r_\rho + 2^{n+1} + 2^{2n+k+3}$ at least. It remains to evaluate the cost of those merges that ${\mathcal{A}}$ sorts respectively the left and right blocks.
The merges that ${\mathcal{B}}$ performs while sorting both blocks are the merges of a Huffman tree, hence the merge cost of ${\mathcal{B}}$ on both blocks is optimal among all the (possibly non stable) merge sorts. In particular, if the merge cost of ${\mathcal{A}}$ on the left block is *not* optimal, and since that cost must be a multiple of $2^{2n}$, it means that $\mathbf{m}_a^1 \geqslant \mathbf{m}_b^{1} + 2^{2n}$. Since $\mathbf{m}_b^{1} \leqslant \mathbf{M} \leqslant 5r_\rho/2 = 5 (k+3) 2^{2n+k+3} = 2^{2n}/\varepsilon_k$, it follows that $\mathbf{m}_a^1 \geqslant (1 + 2^{2n}/\mathbf{M}) \mathbf{m}_b^1 \geqslant (1 + \varepsilon_k) \mathbf{m}_b^1$.
Hence, let us assume that the merge cost of ${\mathcal{A}}$ on the left block is optimal, and let $d_i$ the number of merges involving the run $R_i$ while ${\mathcal{A}}$ sorts the left block. Let $\mathbf{D}$ be the largest of the integers $d_i$, and let $\mathbf{I}$ be the number of indices $i$ such that $d_i = \mathbf{D}$.
By optimality of ${\mathcal{A}}$, we know that larger runs $R_i$ must be associated smaller integers $d_i$. Hence, it comes that $\mathbf{D} \geqslant d_i \geqslant d_1 \geqslant d_{\rho-1}$ for all $i \in \{2,\ldots,\rho-2\}$. Then, if $d_{\rho-1} \neq \mathbf{D}$, let $R_i$ and $R_j$ two runs that ${\mathcal{A}}$ merges together, with $d_i =d_j = \mathbf{D}$. The resulting run is longer than $R_{\rho-1}$, which proves that $\mathbf{D} \geqslant d_1 \geqslant d_{\rho-1} \geqslant \mathbf{D}-1$.
Then, Kraft inequality states that $1 = \sum_{i=1}^{\rho-1} 2^{-d_i} = 2^{-\mathbf{D}} \mathbf{I} + 2^{1-\mathbf{D}} (\rho-1-\mathbf{I})$. Since $\rho = 2^k$, it comes that $2^{k+1} = 2^\mathbf{D} + 2 + \mathbf{I}$, with $1 \leqslant \mathbf{I} \leqslant 2^{k-1}$, which means that $\mathbf{D} = k$ and that $\mathbf{I} = 2^k-2 = \rho-2$. This means that $d_1=\ldots=d_{\rho-2} = k$ and that $d_{\rho-1} = k-1$ and, given that ${\mathcal{A}}$ can merge adjacent runs only, it must perform the same merges as ${\mathcal{B}}$.
Therefore, ${\mathcal{A}}$ needs to merge the runs $R_{2^k+1}$ and $R_{2^k+2}$. When it does, the stack of ${\mathcal{A}}$ may contain at most $k-2$ runs on top of $R_{2^k+2}$. However, since ${\mathcal{A}}$ performed the same merges as ${\mathcal{B}}$, the runs $R_{2^k+2^i+1}$ (for $1 \leqslant i \leqslant k-1$) cannot belong to the same merged run unless that run also contains the run $R_{2^k+1}$. Hence, when $R_{2^k+1}$ and $R_{2^k+2}$ are merged, they belong to pairwise distinct merged runs, and therefore they cannot all fit on the stack. In particular, when this merge occurs, the run $R_{\rho-1}$ may not have been pushed onto the stack, i.e., discovered by the algorithm ${\mathcal{A}}$.
One proves similarly that, when sorting the array $\mathbf{a}^{-1}$, the algorithm ${\mathcal{A}}$ must sort the left block by performing the same merges of ${\mathcal{B}}$, sort the right block, and finally merge the middle block with the right block and then with the left block. While doing so, we also prove that ${\mathcal{A}}$ must merge the runs $R_{2^k}$ and $R_{2^k+1}$ before having yet discovered the run $R_{\rho-1}$.
Finally, recall that $\mathbf{a}^1$ and $\mathbf{a}^{-1}$ have the same length. Hence, since ${\mathcal{A}}$ is deterministic, the merges its performs while sorting either $\mathbf{a}^1$ or $\mathbf{a}^{-1}$ and before discovering the run $R_{\rho-1}$ must be the same. However, these families of merges are incompatible, since they require merging $R_{2^k+1}$ either with $R_{2^k+2}$ or with $R_{2^k}$. Hence, our initial assumption on ${\mathcal{A}}$ was false, which completes the proof.
Unsurprisingly, this result can be strengthened dramatically in the case of length-$(\infty,2)$-aware algorithms, which are not asymptotically $\varepsilon$-optimal for any $\varepsilon$.
\[lem:konig\] Consider the following dynamic system. Starting with one empty stack ${\mathcal{S}}$, one successively performs operations of the following type: either (i) one picks some element $k$ of the stack and deletes all the elements above, making $k$ the top element of the stack, or (ii) if one has already pushed the integers $0,1,\ldots,\ell-1$ onto the stack, one pushes the integer $\ell$ on the top of the stack.
Then, for all integers $h$ and all functions $f : \mathbb{Z}_{\geqslant 0} \mapsto \mathbb{Z}_{\geqslant 0}$, there exists an integer $m$ such that, when the integer $m$ is pushed onto the stack, either (i) the stack ${\mathcal{S}}$ has been of height $h$ at some point, or (ii) some integer $k$ has been picked at least $f(k)$ times.
Lemma \[lem:konig\] is nothing but a variant of König’s lemma. Below, we proceed by induction on $h$, proving that, for every function $f$, some integer $m$ satisfying the statement of Lemma \[lem:konig\] exists: we will denote this integer by $m(f,h)$.
First, for $h = 0$, one proves the statement by choosing $m(f,0) = 1$. Now, consider some integer $h \geqslant 0$ such that, for all functions $f$, there exists an integer $m(f,h)$ satisfying the statement of Lemma \[lem:konig\].
Then, consider some finite execution of the dynamic system such that no integer $k$ has been picked $f(k)$ times or more, the stack has never been of height $h$, and $\overline{m}$ integers have been pushed onto the stack. We can split the set $\{1,\ldots,\overline{m}-1\}$ into $f(0)+1$ sub-intervals $I_0,\ldots,I_{f(0)}$, where $I_k$ is the set of those positive integers that were pushed onto the stack before $0$ was picked $k$ times (but after it had been picked $k-1$ times).
Then, we also set $I_{f(0)+1} = \{\overline{m}\}$, and denote by $\mathbf{i}_k$ the smallest element of $I_k \cup \ldots \cup I_{f(0)+1}$ for all $k \leqslant f(0)+1$. Let us also denote by $f_\ell$ the function $x \mapsto f(x+\ell)$. For all $k \leqslant f(0)$, the induction hypothesis proves that $I_k$ must be of cardinality at most $m(f_{\mathbf{i}_k},h-1)$. In particular, an immediate induction proves that $\mathbf{i}_k \leqslant \mathbf{i}'_k$ for all $k \leqslant f(0)+1$, where we set
$\displaystyle
\mathbf{i}'_0 = 1 \text{ and }
\mathbf{i}'_{k+1} = \mathbf{i}'_k + \max_{j \leqslant \mathbf{i}'_k} m(f_j,h-1)$
for all $k \geqslant 0$. Thus, if we set $m(f,h) = 1 + \mathbf{i}'_{f(0)+1}$, we have proved that $\overline{m} < m(f,h)$, which completes the induction and the proof.
\[pro:never-epsilon-optimal\] Let ${\mathcal{A}}$ be a length-$(\infty,2)$-aware sorting algorithm. The worst-case merge cost of ${\mathcal{A}}$ is bounded below by $\omega(n ({\mathcal{H}}+1))$. In particular, ${\mathcal{A}}$ is not asymptotically $\varepsilon$-optimal for any real number $\varepsilon$.
Let us assume in the entire proof, for the sake of contradiction, that there exist an integer $\mathbf{z}$ and a length-$(\infty,2)$-aware algorithm ${\mathcal{A}}$ whose merge cost is bounded above by $\mathbf{z} n ({\mathcal{H}}+1)$. Then, for all integers $k \leqslant \ell$, let $\mathbf{a}_{k,\ell}$ be an array that decomposes into $k+1$ runs of respective lengths $2^{\ell-1},2^{\ell-2},\ldots,2^{\ell-k}$ and $2^\ell+2^{\ell-k}$: the array $\mathbf{a}_{k,\ell}$ has length $2^{\ell+1}$.
The entropy of any array $\mathbf{a}_{k,\ell}$ is defined by
$\displaystyle {\mathcal{H}}_{k,\ell} = \sum_{i=2}^{k+1} \log_2(2^i)/2^i - (1+2^{-k})\log_2((1+2^{-k})/2)/2 < \sum{i \geqslant 2} i/2^i + 1 = 5/2,$
and therefore the cost of those merges used by ${\mathcal{A}}$ for sorting $\mathbf{a}_{k,\ell}$ is smaller than $7 \mathbf{z} 2^\ell$. Moreover, if the stack of ${\mathcal{A}}$ is of height $h$ when the last run of $\mathbf{a}_{k,\ell}$ is discovered, then that run will take part to $h$ merges, for a total cost of $2^\ell h$ at least. It follows that $h < 7 \mathbf{z}$.
However, once the integer $\ell$ is fixed, and provided that it is executed on some array $\mathbf{a}_{k,\ell}$, the algorithm ${\mathcal{A}}$ cannot distinguish between the arrays $\mathbf{a}_{0,\ell},\ldots,\mathbf{a}_{\ell,\ell}$ until it discovers their last run. In particular, if, when treating some array $\mathbf{a}_{k,\ell}$, and just before pushing its $i$^th^ run (with $i \leqslant k$), the stack of ${\mathcal{A}}$ turns out to be of height $h \geqslant 7 \mathbf{z}$, then ${\mathcal{A}}$ might as well discover that it was, in fact, treating the array $\mathbf{a}_{i,\ell}$, contradicting the previous paragraph. Therefore, the stack of ${\mathcal{A}}$ may *never* be of height $7 \mathbf{z}$ or more.
At the same time, if some run of length $2^{\ell-j}$ takes part to $7 \mathbf{z} 2^j$ merges, then of course these merges have a total cost of $7 \mathbf{z} 2^\ell$ at least. Hence, every run of length $2^{\ell-j}$ must take part to at most $f(j)$ merges, where $f(j) = 7 \mathbf{z} 2^j$. Then, observe that our stack follows exactly the dynamics described in Lemma \[lem:konig\], where picking a integer $i$ means that one merges the run whose original length was $2^{\ell-i}$, and which lies in second position at the top of the stack, with the run that lies on the top of the stack. It follows from Lemma \[lem:konig\] that $\ell \leqslant m(f,7 \mathbf{z})$.
Consequently, and when $\ell$ is large enough, we know that there must exist arrays on which the merge cost of ${\mathcal{A}}$ is at least $\mathbf{z} n ({\mathcal{H}}+1)$. In particular, since would have sorted ${\mathcal{A}}$ for a cost of $n({\mathcal{H}}+ 2)$, it follows that ${\mathcal{A}}$ is not $(\mathbf{z}/2-1)$-optimal.
Furthermore, since no algorithm can require a merge cost greater than $\rho n$ on arrays of length $n$ and with $\rho$ runs, every array falsifying the $(\mathbf{z}/2-1)$-optimality of ${\mathcal{A}}$ must have at least $\mathbf{z}/2-2$ runs. Hence, for every $\varepsilon > 0$ and every $\rho \geqslant 0$, by choosing $\mathbf{z} \geqslant \max\{\varepsilon,2\rho+4\}$, one checks that ${\mathcal{A}}$ cannot be asymptotically $\varepsilon$-optimal either.
\[thm:eta-optimal\] Let $k \geqslant 8$ be an integer, and let
$\displaystyle \eta_k = 11 / \log_2((k-3)/4).$
There exists a $k$-aware sorting algorithm that is $\eta_k$-optimal.
Note that, if $k \geqslant 3$, then Theorems \[thm:complexity-nH+D\] and \[thm:lower-bound-merge-cost\] already prove that , which is a $3$-aware (and thus a $k$-aware) algorithm, is also $(\Delta+1)$-optimal. In particular, numerical computations show that $\Delta + 1 \leqslant \eta_k$ when $k \leqslant 22$, which already proves Theorem \[thm:eta-optimal\] in that case.
Furthermore, and although both constants $\varepsilon_k$ and $\eta_k$ become arbitrarily small when $k \to \infty$, there is a double exponential gap between those constants. This mismatch might not be too surprising, given that Proposition \[pro:no-epsilon-optimal\] focuses on all length-$(\infty,k)$-aware algorithms, while Theorem \[thm:eta-optimal\] deals with the much more restricted class of $k$-aware algorithms.
$\operatorname{\texttt{runs}}{\ensuremath{\leftarrow}}$ the run decomposition of $A$\[alg:m:init\] $\operatorname{{\mathcal{S}}}{\ensuremath{\leftarrow}}$ an empty stack (\[alg:m:main\])[$\operatorname{\texttt{runs}}\neq \emptyset$]{}[ remove a run $R$ from $\operatorname{\texttt{runs}}$ and push $R$ onto $\operatorname{{\mathcal{S}}}$ $b {\ensuremath{\leftarrow}}$ [**true**]{} ]{} (\[alg:m:collapse\])[$h \geqslant 2\kappa+2$]{}[ merge the runs $R_{h-1}$ and $R_h$]{} \[alg:m:flS\] apply the optimal merge policy on the remaining $h$ runs (with $h \leqslant 2\kappa+1$) \[alg:m:end\]
(\[alg:lowstack\])[ $r_{{\text{\textsf{large}}}}{\ensuremath{\leftarrow}}\lfloor (r_1+r_2+\ldots+r_{2\kappa}) / \kappa \rfloor$ $\ell_{{\text{\textsf{large}}}}{\ensuremath{\leftarrow}}\lfloor \log_2(r_{{\text{\textsf{large}}}})\rfloor$ $b {\ensuremath{\leftarrow}}$ [**true**]{} $i {\ensuremath{\leftarrow}}3$ ]{} (\[alg:endlowstack\])[ $r_{{\text{\textsf{large}}}}{\ensuremath{\leftarrow}}\lfloor (r_1+r_2+\ldots+r_h) / \kappa \rfloor$ $\ell_{{\text{\textsf{large}}}}{\ensuremath{\leftarrow}}\lfloor \log_2(r_{{\text{\textsf{large}}}})\rfloor$ $i {\ensuremath{\leftarrow}}3$ ]{} (\[alg:testMerge\])[ ]{}
The sequel of this article is devoted to proving Theorem \[thm:eta-optimal\]. The proof is organised as follows. We design, for every integer $\kappa \geqslant 3$, a variant of (i.e., of for the parameter $c = 1$) that is both $(2\kappa+2)$-aware and $\eta_{2\kappa+3}$-optimal. This generalised algorithm is called [$\kappa$-]{}. It is a parameterised algorithm that whose parameter is the integer $\kappa$.
Algorithm \[alg:ASSm\] presents the algorithm [$\kappa$-]{}, which we split into one central part and several auxiliary functions. on which it relies. In what follows, we will consider the parameters $c$ and $\kappa$ as fixed once and for all. The algorithm works in three phases: the phase $1$ covers lines \[alg:m:init\] to \[alg:m:collapse\], the phase $2$ covers the line \[alg:m:flS\] (and, therefore, the lines \[alg:endlowstack\] to \[alg:m:collapse:end\]), and the phase $3$ covers the line \[alg:m:end\]. Moreover, we will say below that the loop that spans the lines \[alg:m:main\] to \[alg:m:cs1\] is the *main loop* of [$\kappa$-]{}, and that the lines \[alg:m:init\] to \[alg:m:flS\] (i.e., the union of phases $1$ and $2$) form the *main body* of [$\kappa$-]{}.
The algorithm [$\kappa$-]{}is clearly $(2\kappa+2)$-aware. Therefore, it remains to prove that it is also $\eta_{2\kappa+3}$-optimal. The main idea is that, when the main body of [$\kappa$-]{}is executed on some array $\mathbf{a}$, it performs all the merges that would have been performed by itself, except those merges that would involve at least one *very large* run (roughly said, a run is called *very large* if its size is at least $n/\kappa$, where $n$ is the length of the array $\mathbf{a}$).
Let us first present the key results that will pave our way towards proving Theorem \[thm:eta-optimal\]. Initial results concern properties of itself. First, in Definition \[def:state\], we introduce the notion of *state*, which is the sequence obtained by pushing each remaining run (yet to be discovered) onto the stack that manipulates. Then, we prove, with Proposition \[pro:first-merge\], that the merges performed by depend only on the current state, and not on the actual height of the stack. This leads to Theorem \[thm:stable-border\], which states the following *stability* property: denoting by $M$ the set of run merges performed by , and if we first perform merges contained in an arbitrary subset $M'$ of $M$, then executing on the resulting state leads to performing all the remaining merges in $M \setminus M'$.
Then, we focus on properties of [$\kappa$-]{}. Like Proposition \[pro:first-merge\], Proposition \[pro:first-merge-2\] states that the merges performed during each of the phases 1 and 2 of [$\kappa$-]{}depend only on the current state. Propositions \[pro:cassm-subset-cass\] and \[pro:flS-study\] further state that these merges also belong to $M$: more precisely, they are exactly those merges that do not involve any very large run.
Finally, investigating properties of optimal stable merge policies, we can prove that both the stack ${\mathcal{S}}$ obtained after the main body of [$\kappa$-]{}and stacks that one might have obtained from an optimal stable merge sort do have some *regularity* properties, as stated in Proposition \[pro:flS-study\] and Corollary \[cor:nice-border\]. Moreover, Lemma \[lem:refinement\] helps us to quantify the overhead due to using a sub-optimal stack ${\mathcal{S}}$, and Proposition \[pro:cassm-best\] quantifies the overhead due to constructing ${\mathcal{S}}$ via merges borrowed from . We complete the proof by observing that, although these overheads grow with the share of the array that is covered by non-very-large runs, the entropy ${\mathcal{H}}$ also grows accordingly, which allows us to get uniform bound on the overhead due to using [$\kappa$-]{}instead of an optimal merge policy.
Now, let us turn to the actual proof of Theorem \[thm:eta-optimal\]. Below, and as suggested by our presentation of Algorithm \[alg:ASSm\], we change the convention used in previous sections for representing stacks, and decide to represent a stack ${\mathcal{S}}$ as a tuple $(R_1,\ldots,R_h)$, where $R_k$ is the $k$^th^ *bottom-most* run of the stack. Indeed, in what follows, we will identify every array with the sequence of monotonic runs it can be split into. Hence, we will often refer to the sequence of runs on which the algorithm or [$\kappa$-]{}are executed, and it will also be convenient to view stacks as sequences of runs on which the algorithm may be executed. In particular, if $\ell_1 > \ell_2 > \ldots > \ell_\rho$, executing or [$\kappa$-]{}on the sequence $(R_1,\ldots,R_\rho)$ leads to producing the stack $(R_1,\ldots,R_\rho)$ itself.
Then, in order to simplify our study of the dynamics of both algorithms and [$\kappa$-]{}, we adopt a slightly different point of view on the objects they handle. More precisely, both algorithms, when sorting a sequence of runs, handle a stack of runs as well as a tail of the sequence to be sorted. It turns out that, for tracking which runs are merged, the stack and the sequence tail can be combined into one single object, which will be a sequence of runs.
As we will show, this object allows us to identify the sequence of run merges preformed by the algorithm, which is why we call it a *state*.
\[def:state\] Consider the execution of the algorithm (or [$\kappa$-]{}) on a sequence of runs to be sorted. At any step of the algorithm, the algorithm handles both a stack ${\mathcal{S}}= (R_1,\ldots,R_h)$ of runs and a sequence ${\mathcal{R}}= (R_{h+1},\ldots,R_\rho)$ of those runs that have yet to be discovered and pushed onto the stack.
We call *state* of the algorithm, at that step, the sequence $(R_1,\ldots,R_\rho)$, i.e., the concatenation of ${\mathcal{S}}$ and ${\mathcal{R}}$. Finally, two steps of the algorithm are said to be *consecutive* if they are separated by a run push or merge operation.
We focus now on describing how and [$\kappa$-]{}transform a state into another one. We first treat the (easiest) case of .
\[def:successor\] Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs of length $\rho \geqslant 2$. Then, let $m = \min\{x \colon 1 \leqslant x \leqslant \rho-1$ and $\ell_x \leqslant \max\{\ell_{x+1},\ell_{x+2}\}\}$, with the convention that $\ell_{\rho+1} = \infty$ (so that $m$ always exists).
We say that $m$ is the *merge point* of the sequence ${\mathcal{R}}$. Furtherore, we call *successor* of ${\mathcal{R}}$, and denote by ${\textsf{succ}}({\mathcal{R}})$, the sequence of runs
$\displaystyle (R_1,\ldots,R_{m-1},R',R_{m+2},\ldots,R_\rho),$
where $R'$ is the run obtained by merging $R_m$ and $R_{m+1}$.
\[pro:first-merge\] Let $\mathbf{s}$ and $\mathbf{s}'$ be two successive steps of an execution of , and let $\mathbf{S}$ and $\mathbf{S}'$ be the associated states. Then, (i) if a run push occurred between $\mathbf{s}$ and $\mathbf{s}'$, we have $\mathbf{S}' = \mathbf{S}$, and (ii) if a run merge occurred between $\mathbf{s}$ and $\mathbf{s}'$, we have $\mathbf{S}' = {\textsf{succ}}(\mathbf{S})$.
First, if a run push occurred between $\mathbf{s}$ and $\mathbf{s}'$, it is clear that $\mathbf{S}' = \mathbf{S}$. Hence, we assume that a run merged occurred between $\mathbf{s}$ and $\mathbf{s}'$. Let ${\mathcal{S}}= (R_1,\ldots,R_h)$ be the stack at step $\mathbf{s}$, and let ${\mathcal{R}}$ be the sequence of those runs that are yet to be pushed onto the stack at step $\mathbf{s}$, so that $\mathbf{S}$ is the concatenation of ${\mathcal{S}}$ and ${\mathcal{R}}$.
Lemma \[lem:invariant-li\], once rephrased with our new convention for representing stacks, states that $\ell_1 > \ell_2 > \ldots > \ell_{h-2} \geqslant \ell_{h-1}$. Then, since a merge occurred, it was either a merge between the runs $R_{h-2}$ and $R_{h-1}$, meaning that $\ell_{h-2} \leqslant \max\{\ell_{h-1},\ell_h\}$, or between the runs $R_{h-1}$ and $R_h$, meaning that $\ell_{h-2} > \ell_h \geqslant \ell_{h-1}$ (or that ${\mathcal{R}}$ is empty and $\ell_{h-2} > \ell_{h-1} > \ell_h$). In both cases, it comes at once that $\mathbf{S}' = {\textsf{succ}}(\mathbf{S})$.
We prove now some stability result fot , which is stated precisely in Theorem \[thm:stable-border\] and can be roughly stated as follows: “Consider the set $M$ of run merges performed by an execution of on a sequence ${\mathcal{R}}$ of runs. Then, let assume that, instead, we started by performing some merges $\mathbf{m}_1,\ldots,\mathbf{m}_k \in M$ on the sequence ${\mathcal{R}}$, obtaining a sequence ${\mathcal{R}}'$ as a result of those merges, and let $M'$ be the sequence of merges performed by an execution of on that sequence ${\mathcal{R}}'$. It holds that $M$ is the disjoint union of the sets $M'$ and $\{\mathbf{m}_1,\ldots,\mathbf{m}_k\}$.”
\[def:border\] Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs, and let ${\mathcal{T}}$ be the merge tree obtained by running the algorithm on ${\mathcal{R}}$. We say that ${\mathcal{T}}$ is the merge tree *induced* by ${\mathcal{R}}$ for (we may omit referring to if the context is clear).
Then, a *border* of ${\mathcal{T}}$ is a maximal set $\mathcal{B}$ of pairwise incomparable (for the strict ancestor relation) nodes of ${\mathcal{T}}$; i.e., no strict ancestor of a node in $\mathcal{B}$ belongs to $\mathcal{B}$. The set of ancestors of nodes in $\mathcal{B}$, including nodes in $\mathcal{B}$ themselves, is called the *ancestor sub-tree* of $\mathcal{B}$ in ${\mathcal{T}}$.
Every set of pairwise incomparable nodes of a merge tree ${\mathcal{T}}$ is naturally endowed with a left-to-right ordering. Hence, we will often identify a border $\mathcal{B}$ with the unique sequence of runs $(R'_1,\ldots,R'_k)$ such that each $R'_i$ lies to the left of $R'_{i+1}$ (i.e., if $R$ is the least common ancestor to $R'_i$ and $R'_{i+1}$, then $R'_i$ descends from the left child of $R$) and such that $\{R'_1,\ldots,R'_k\} = \mathcal{B}$. Building on this identification, we may also say that a border $\mathcal{B}$ induces a merge tree.
\[thm:stable-border\] Let ${\mathcal{T}}$ be the merge tree induced by a sequence of runs ${\mathcal{R}}$, and let ${\mathcal{R}}'$ be a border of ${\mathcal{T}}$. The merge tree induced by the sequence ${\mathcal{R}}'$ coincides with the ancestor sub-tree of ${\mathcal{R}}'$ in ${\mathcal{T}}$.
Let $\rho = |{\mathcal{R}}|$ and $k = |{\mathcal{R}}'|$. Below, we identify the set ${\mathcal{R}}$ with a sequence $(R_1,\ldots,R_\rho)$ of runs. Then, if $\rho \geqslant 2$, let $m$ be the merge point of ${\mathcal{R}}$, and let ${\overline{R}}$ be the parent of the runs $R_m$ and $R_{m+1}$, i.e., the run obtained by merging $R_m$ and $R_{m+1}$; if $\rho = 1$, we just set $m = 0$. Since $k \leqslant \rho$, we prove Theorem \[thm:stable-border\] by induction on the triple $(\rho,\rho - k, m)$.
First, if $\rho = k$, then ${\mathcal{R}}= {\mathcal{R}}'$, and thus the result is immediate.
Second, let us assume that $\rho = k+1$. In that case, there exists an integer $p \leqslant \rho-1$ such that ${\mathcal{R}}' = {\mathcal{R}}\setminus \{R_p,R_{p+1}\} \cup
\{R'\}$, where $R'$ is the parent of the runs $R_p$ and $R_{p+1}$, i.e., the run obtained by merging $R_p$ and $R_{p+1}$. Observe that either $p = m$ or $m \leqslant p-2$:
- If $m = p$, we further note that the tree ${\mathcal{T}}$ can be reconstructed from the sequence of states $\mathbf{S}_0,\ldots,\mathbf{S}_{\rho-1}$ of , with $\mathbf{S}_0 = {\mathcal{R}}$ and $\mathbf{S}_{i+1} = {\textsf{succ}}(\mathbf{S}_i)$ for all $i \geqslant 0$, as follows. First, $\mathbf{S}_1$ is the ordered set of leaves of ${\mathcal{T}}$, and then, if $\mathbf{S}_{i+1}$ is obtained by removing two consecutive runs $R_\ominus$ and $R_\oplus$ from $\mathbf{S}_i$ and replacing them by a new run $R$, the run $R$ is a new internal node of ${\mathcal{T}}$, whose left and right children are $R_\ominus$ and $R_\oplus$.
Here, Proposition \[pro:first-merge\] proves that $\mathcal{B}$ is just the state $\mathbf{S}_1$. It follows that every merge performed while executing on $\mathcal{B}$ is also performed while executing on ${\mathcal{R}}$, which completes the proof.
- If $m \leqslant p-2$, then let also $\overline{{\mathcal{R}}} = {\mathcal{R}}\setminus
\{R_m,R_{m+1}\} \cup \{{\overline{R}}\}$ and $\overline{{\mathcal{R}}}' = {\mathcal{R}}\setminus
\{R_m,R_{m+1},R_p,R_{p+1}\} \cup \{{\overline{R}},R'\}$ be borders of ${\mathcal{T}}$. Let ${\mathcal{T}}'$, $\overline{{\mathcal{T}}}$ and $\overline{{\mathcal{T}}}'$ be the merge trees induced by ${\mathcal{R}}'$, $\overline{{\mathcal{R}}}$ and $\overline{{\mathcal{R}}}'$. We will show that they are all ancestor sub-trees of ${\mathcal{T}}$. The induction hypothesis proves that $\overline{{\mathcal{T}}}$ is the ancestor sub-tree of $\overline{{\mathcal{R}}}$ in ${\mathcal{T}}$. Hence, the runs $R_p$ and $R_{p+1}$ are siblings in both trees ${\mathcal{T}}$ and $\overline{{\mathcal{T}}}$. Consequently, $\overline{{\mathcal{R}}}'$ is also a border of $\overline{{\mathcal{T}}}$, and therefore $\overline{{\mathcal{T}}}'$ is the ancestor sub-tree of $\overline{{\mathcal{R}}}'$ in $\overline{{\mathcal{T}}}$ (and thus in ${\mathcal{T}}$ as well).
Furthermore, Proposition \[pro:first-merge\] proves that the runs $R_m$ and $R_{m+1}$ are siblings in both trees ${\mathcal{T}}$ and ${\mathcal{T}}'$. It follows that $\overline{{\mathcal{R}}}'$ is a border of ${\mathcal{T}}'$ and, by induction hypothesis, that $\overline{{\mathcal{T}}}'$ is the ancestor sub-tree of $\overline{{\mathcal{R}}}'$ in ${\mathcal{T}}'$.
Hence, the tree $\overline{{\mathcal{T}}}'$ is obtained by removing the four leaves $R_m$, $R_{m+1}$, $R_p$ and $R_{p+1}$ from ${\mathcal{T}}$, and is also a strict sub-tree of ${\mathcal{T}}'$, which contains both leaves $R_m$ and $R_{m+1}$. This means that ${\mathcal{T}}'$ is obtained by removing the two leaves $R_p$ and $R_{p+1}$ from ${\mathcal{T}}$, which makes it the ancestor sub-tree of ${\mathcal{R}}'$ in ${\mathcal{T}}$, as desired.
Finally, we assume that $\rho \geqslant k+2$. Let $R$ be an internal node of ${\mathcal{T}}$ that belongs to ${\mathcal{R}}'$, and let $R_\oplus$ and $R_\ominus$ be its children. The set ${\mathcal{R}}'' = {\mathcal{R}}' \setminus \{R\} \cup
\{R_\oplus,R_\oplus\}$ is a border of ${\mathcal{T}}$ of cardinality $k+1$. Let also ${\mathcal{T}}'$ and ${\mathcal{T}}''$ be the merge trees respectively induced by ${\mathcal{R}}'$ and ${\mathcal{R}}''$. Then, ${\mathcal{T}}''$ is the ancestor sub-tree of ${\mathcal{R}}''$ in ${\mathcal{T}}$, and thus ${\mathcal{T}}'$ is the ancestor sub-tree of ${\mathcal{R}}'$ in both ${\mathcal{T}}''$ and ${\mathcal{T}}$.
So far, we have obtained a concise description of the dynamics of , where states play a crucial role, and from which we could derive the stability result given by Theorem \[thm:stable-border\]. In order to analyse the behaviour of [$\kappa$-]{}itself, we further need to define the notion of a (*very*) *large* run in a sequence of runs.
\[def:large\] Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs. We say that a run $R_i$ is a *large* run of ${\mathcal{R}}$ if $\ell_i \geqslant
\ell_{{\text{\textsf{large}}}}+ 1$, where we set
$\displaystyle r_{{\text{\textsf{large}}}}=
\lfloor (r_1+r_2+\ldots+r_{\min\{\rho,2\kappa\}}) / \kappa \rfloor \text{ and }
\ell_{{\text{\textsf{large}}}}=
\lfloor \log_2(r_{{\text{\textsf{large}}}}) \rfloor.$
We also say that $R_i$ is a *very large* run of ${\mathcal{R}}$ if $\ell_i \geqslant \ell^\star_{{\text{\textsf{large}}}}$, where we set
$\displaystyle r^\star_{{\text{\textsf{large}}}}= (r_1+\ldots+r_\rho) / \kappa \text{ and }
\ell^\star_{{\text{\textsf{large}}}}= \lfloor \log_2(r^\star_{{\text{\textsf{large}}}}) \rfloor.$
Below, we may omit any reference to ${\mathcal{R}}$ if it is clear from the context, among others when speaking of a run within a stack (because, during an execution of [$\kappa$-]{}, the stack itself always embeds into one unique state).
\[def:kappa-successor\] Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs. Then, let $m = \min\{x \colon 1 \leqslant x \leqslant \rho-1$, $R_x$ and $R_{x+1}$ are not large, and $\ell_x \leqslant \max\{\ell_{x+1},\ell_{x+2}\}\}$, with the convention that $\ell_{\rho+1} = \infty$; note that $m$ may not necessarily exist.
We say that $m$ is the *$\kappa$-merge point* of the sequence ${\mathcal{R}}$. Then, we call *$\kappa$-successor* of ${\mathcal{R}}$, and denote by ${\textsf{succ$_\kappa$}}({\mathcal{R}})$, the sequence of runs
$\displaystyle (R_1,\ldots,R_{m-1},R',R_{m+2},\ldots,R_\rho),$
where $R'$ is the run obtained by merging $R_m$ and $R_{m+1}$; if $m$ does not exist, we just define ${\textsf{succ$_\kappa$}}({\mathcal{R}})$ as the empty sequence.
Similarly, let $m^\star = \min\{x \colon 1 \leqslant x \leqslant \rho-1$, $R_x$ and $R_{x+1}$ are not very large, and $\ell_x \leqslant \max\{\ell_{x+1},\ell_{x+2}\}\}$, with the convention that $\ell_{\rho+1} = \infty$. We say that $m^\star$ is the *$\star\kappa$-merge point* of ${\mathcal{R}}$. We also call *$\star\kappa$-successor* of ${\mathcal{R}}$, and denote by ${\textsf{succ$_{\star\kappa}$}}({\mathcal{R}})$, the sequence of runs
$\displaystyle (R_1,\ldots,R_{m^\star-1},R',R_{m^\star+2},\ldots,R_\rho),$
where $R'$ is the run obtained by merging $R_{m^\star}$ and $R_{m^\star+1}$; if $m^\star$ does not exist, we define ${\textsf{succ$_{\star\kappa}$}}({\mathcal{R}})$ as the empty sequence.
Since Definition \[def:kappa-successor\] is a variant of Definition \[def:successor\] taking into account the notion of large runs, it may be expected that a variant of Proposition \[pro:first-merge\], involving $\kappa$-successors, should exist for the algorithm [$\kappa$-]{}. This is the object of Proposition \[pro:first-merge-2\] below.
\[lem:m-is-small\] Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs of length $\rho \geqslant 2\kappa+1$, and let $m$ be its $\kappa$-merge point. This $\kappa$-merge point always exists. Furthermore, if $m \geqslant 2\kappa$, then none of the runs $R_{2\kappa-1},R_{2\kappa},\ldots,R_{m+1}$ is large.
We first prove that less than $\kappa$ of the runs $R_1,\ldots,R_{2\kappa}$ are large. Indeed, let $s = r_1+\ldots+s_{2\kappa}$. For every large run $R_i$, we have $\lfloor \log_2(r_i) \rfloor = \ell_i > \ell_{{\text{\textsf{large}}}}=
\lfloor \log_2(r_{{\text{\textsf{large}}}}) \rfloor$, which means that $r_i > r_{{\text{\textsf{large}}}}= \lfloor s / \kappa \rfloor$, and thus that $r_i > s / \kappa$. Hence, if there exist $\kappa$ large runs $R_{i_1},R_{i_2},\ldots,R_{i_\kappa}$, with $1 \leqslant i_1 < i_2 < \ldots < i_\kappa \leqslant 2\kappa$, it comes at once that $s \geqslant r_{i_1}+r_{i_2}+\ldots+r_{i_\kappa} > \kappa s/\kappa$, which is a contradiction.
Thus, by pigeonhole principle, there exists at least one integer $j \leqslant \kappa$ such that neither $R_{2j-1}$ nor $R_{2j}$ is large. Then, given the definition of $\kappa$-merge points, we may, without loss of generality, and up to appending an immensely large run to the sequence ${\mathcal{R}}$, assume that $R_\rho$ is large. Now, let $k$ be the smallest integer such that $k \geqslant 2j+1$ and $R_k$ is large. By construction, neither $k-1$ nor $k-2$ is large, and $\ell_{k-2} \leqslant \ell_k$, proving that $m$ exists and that $m \leqslant k-2$. The result follows.
\[pro:first-merge-2\] Let $\mathbf{s}$ and $\mathbf{s}'$ be two successive steps during an execution of the phase 1 of [$\kappa$-]{}(i.e., lines \[alg:m:init\] to \[alg:m:collapse\]), and let $\mathbf{S}$ and $\mathbf{S}'$ be the associated states. Let also ${\mathcal{S}}= (R_1,\ldots,R_h)$ be the stack at step $\mathbf{s}$, and let $m$ be the $\kappa$-merge point of $\mathbf{S}$. Then, (i) if $h \leqslant 2\kappa$ or $m \geqslant h-1$, a run push occurred between $\mathbf{s}$ and $\mathbf{s}'$, and we have $\mathbf{S}' = \mathbf{S}$; and (ii) if $h \geqslant 2\kappa+1$ and $m \leqslant h-2$, a run merge occurred between $\mathbf{s}$ and $\mathbf{s}'$, and we have $\mathbf{S}' = {\textsf{succ$_\kappa$}}(\mathbf{S})$.
We consider several cases separately. First, if $h \leqslant 2\kappa$, we simply had to push a run onto ${\mathcal{S}}$, so that $\mathbf{S} = \mathbf{S}'$.
Assume now that $h = 2\kappa+1$. Without loss of generality, we may assume that the function (in line \[alg:m:cs1\] of the algorithm) was called on the stack ${\mathcal{S}}$, and followed by a run push if did not trigger any merge. A call to ${\textsf{TestMerge}}(i,\ell_{{\text{\textsf{large}}}})$ triggers a merge if and only if $R_{i-2}$ and $R_{i-1}$ are not large, and $\ell_{i-2} \leqslant \max\{\ell_{i-1},\ell_i\}$. Since consists in calling successively ${\textsf{TestMerge}}(3,\ell_{{\text{\textsf{large}}}}),\ldots,
{\textsf{TestMerge}}(h,\ell_{{\text{\textsf{large}}}})$ untils one such call triggers a merge (if any), it comes at once that (i) if $m \leqslant h-2$, then the runs $R_m$ and $R_{m+1}$ are merged, and that (ii) if $m \geqslant h-1$, then no runs are merged.
Finally, assume that $h \geqslant 2 \kappa+2$. Without loss of generality, we tackle the case of line \[alg:m:collapse\] by appending to our states a fictitious run of arbitrarily large size. This allows us to assume that both states $\mathbf{s}$ and $\mathbf{s}'$ belong to the main loop of . We claim that, in that case, we must have $m \geqslant h-2$. Let $\overline{\mathbf{s}}$ be the step that preceded $\mathbf{s}$ (we readily adopt the notations $\overline{r}_{{\text{\textsf{large}}}}$, $\overline{{\mathcal{S}}}$, $\overline{h}$ and $\overline{m}$) and let us prove the claim by induction on the number of (push or merge) operations already performed.
First, if $h = 2\kappa+2$ and $\mathbf{s}$ was obtained from $\overline{\mathbf{s}}$ by pushing a run onto a stack $\overline{{\mathcal{S}}}$ of height $2\kappa+1$, we just proved above that $m \geqslant 2\kappa$. In all other cases, we already had $\overline{h} \geqslant 2\kappa+2$, and therefore $\overline{m} \geqslant \overline{h}-2$. Hence, we distinguish two sub-cases:
- If $\mathbf{s}$ was obtained from $\overline{\mathbf{s}}$ by pushing a run, then we could not have $m = \overline{h}-2$, unless what the call to $(\overline{h},\infty)$ would have triggered a merge. This means that $\overline{m} \geqslant \overline{h}-1 = h - 2$.
- If $\mathbf{s}$ was obtained from $\overline{\mathbf{s}}$ by merging two runs, observe that $r_{{\text{\textsf{large}}}}= \overline{r}_{{\text{\textsf{large}}}}$. By induction hypothesis, we know that $R_i = {\overline{R}}_i$ for all $i \leqslant \overline{m}-1$, and that $R_{\overline{m}}$ is the result of the merge of ${\overline{R}}_{\overline{m}}$ and ${\overline{R}}_{\overline{m}+1}$. In particular, every run $R_i$ (with $i \leqslant \overline{m}-1$) that was large in $\overline{\mathbf{S}}$ is still large in $\mathbf{S}$. This means, among others, that $\ell_x = \overline{\ell}_x$ for all $x \leqslant \overline{m}-1$, and thus that $m \geqslant \overline{m}-2$.
Let us further assume that $m = \overline{m}-2$. This means that neither $R_m$ nor $R_{m+1}$ is large (in $\mathbf{S}$), and that $\ell_m \leqslant \max\{\ell_{m+1},\ell_{m+2}\}$. Hence, neither $R_m$ nor $R_{m+1}$ was large (in $\overline{\mathbf{S}}$), and ${\overline{R}}_{m+2}$ was not large as well. By definition of $\overline{m}$, it follows that $\ell_m > \max\{\ell_{m+1},\overline{\ell}_{m+2}\}$ and that $\ell_{m+1} > \max\{\overline{\ell}_{m+2},\overline{\ell}_{m+3}\}$. This implies that $\overline{\ell}_{m+2} \leqslant \ell_{m+2} \leqslant \max\{\overline{\ell}_{m+2},\overline{\ell}_{m+3}\} + 1 \leqslant \ell_{m+1}$, and therefore that $\ell_m > \ell_{m+1} \geqslant \ell_{m+2}$, contradicting the fact that $\ell_m \leqslant \max\{\ell_{m+1},\ell_{m+2}\}$.
Consequently, we have $m \geqslant \overline{m}-1 \geqslant \overline{h}-3 = h - 2$, which proves the claim.
Now that the claim is proved, let us come back to the proof of Proposition \[pro:first-merge-2\]. Like when $h = 2\kappa+1$, we may assume that a call to ${\textsf{TestMerge}}(h,\infty)$ was called on the stack ${\mathcal{S}}$: this call either triggered a merge between the runs $R_{h-2}$ and $R_{h-1}$, or it was immediately followed by a run push. Hence, it suffices to prove that $m = h-2$ in the former case, and that $m \geqslant h-1$ in the latter case.
Since $m \geqslant h-2 \geqslant 2\kappa$, Lemma \[lem:m-is-small\] proves that none of the runs $R_{2\kappa-1},\ldots,R_{m+1}$ is large. In particular, neither $R_{h-2}$ nor $R_{h-1}$ is large. Hence, we have $m = h-2$ if and only if $\ell_{h-2} \leqslant \max\{\ell_{h-1},\ell_h\}$, and $m \geqslant h-1$ otherwise. Since the former case coincides with the fact that calling $(h,\infty)$ triggers a run merge, Proposition \[pro:first-merge-2\] follows.
From the similarity between Propositions \[pro:first-merge\] and \[pro:first-merge-2\], we can now derive the following result.
\[pro:cassm-subset-cass\] Let ${\mathcal{R}}$ be a sequence of runs. Let $M$ be the set of merges performed when executing on ${\mathcal{R}}$, and let $M'$ be the set of merges performed when executing on ${\mathcal{R}}$ the phase 1 of [$\kappa$-]{}. The set $M'$ is a subset of $M$.
Let $\mathbf{S}_0,\ldots,\mathbf{S}_{\rho-1}$ be the sequence of states of , with $\mathbf{S}_0 = {\mathcal{R}}$ and $\mathbf{S}_{i+1} = {\textsf{succ}}(\mathbf{S}_i)$ for all $i \geqslant 0$. We denote by $R_{i,1},\ldots,R_{i,\rho_i}$ the runs that form the state $\mathbf{S}_i$. Then, let $m$ be the $\kappa$-merge point of ${\mathcal{R}}$, and let $m_i$ be the merge point of the state $\mathbf{S}_i$, for all $i \geqslant 0$. We first prove that the runs $R_m$ and $R_{m+1}$ are siblings of each other in the tree ${\mathcal{T}}$ induced by ${\mathcal{R}}$.
Let $k$ be the smallest integer such that $m_k \geqslant m-k-1$. For all $i \leqslant k-1$ and all $j \geqslant m$, we have $R_{i,j-i} = R_j$, while $R_{i,m-i-1}$ is obtained by merging runs among $R_1,\ldots,R_{m-1}$, including the run $R_{m-1}$ itself. It follows that $\ell_{k,m-k-1} \geqslant \ell_{m-1}$, while $\ell_{k,j-k} = \ell_j$ for all $j \geqslant k$.
Moreover, by definition of $m$, we have (i) $\ell_{{\text{\textsf{large}}}}\geqslant \max\{\ell_m,\ell_{m+1}\}$, and either (ii) $\ell_{{\text{\textsf{large}}}}< \max\{\ell_{m-1},\ell_m\}$ or (iii) $\ell_{m-1} > \max\{\ell_m,\ell_{m+1}\}$. Due to (i), the inequality (ii) rewrites as $\ell_{{\text{\textsf{large}}}}< \ell_{m-1}$, which implies (iii). Hence, the inequality (iii) is necessarily true. It follows that
$\displaystyle \ell_{k,m-k-1} \geqslant \ell_{m-1} > \max\{\ell_m,\ell_{m+1}\} =
\max\{\ell_{k,m-k},\ell_{k,m+1-k}\},$
which means that $m_k \neq m-k-1$. Since, at the same time, we have
$\displaystyle \ell_{k,m-k} = \ell_m \leqslant \max\{\ell_{m+1},\ell_{m+2}\} =
\max\{\ell_{k,m+1-k},\ell_{k,m+2-k}\},$
we conclude that $m_k = m-k$, which means must merge the runs $R_m = R_{k,m-k}$ and $R_{m+1} = R_{k,m+1-k}$.
Now, let $\mathbf{S}^\kappa_0,\ldots,\mathbf{S}^\kappa_p$ be the sequence of states of the phase 1 of [$\kappa$-]{}, with $\mathbf{S}^\kappa_0 = {\mathcal{R}}$ and $\mathbf{S}^\kappa_{i+1} = {\textsf{succ$_\kappa$}}(\mathbf{S}^\kappa_i)$ for all $i \geqslant 0$. We have just proved that $R_m$ and $R_{m+1}$ are siblings in the tree ${\mathcal{T}}$ induced by ${\mathcal{R}}$, which means that $\mathbf{S}^\kappa_1$ is a border of ${\mathcal{T}}$. Similarly, every state $\mathbf{S}^\kappa_{i+1}$ is a border of the tree induced by $\mathbf{S}^\kappa_i$, and therefore of ${\mathcal{T}}$ as well. In particular, the merge performed by [$\kappa$-]{}when transforming the state $\mathbf{S}^\kappa_i$ into $\mathbf{S}^\kappa_{i+1}$ was also performed by . This completes the proof.
Then, and although the function refers to very large runs instead of large runs, the above results can also be replicated when considering merges that are performed in line \[alg:m:flS\] of the algorithm.
\[pro:flS-study\] Let ${\mathcal{R}}$ be a sequence of runs. Let $M$ be the set of merges performed when executing on ${\mathcal{R}}$, and let $M'$ be the set of merges performed when executing on ${\mathcal{R}}$ the line \[alg:m:flS\] [$\kappa$-]{}. The set $M'$ is a subset of $M$.
Moreover, let ${\mathcal{S}}= (R_1,\ldots,{\mathcal{R}}_h)$ be the stack obtained at the end of line \[alg:m:flS\]. Then, for all $i \leqslant h$, (i) if $\ell_i \geqslant \ell^\star_{{\text{\textsf{large}}}}+2$, the run $R_i$ was already a run of ${\mathcal{R}}$, and (ii) if $i \geqslant 2$ and $\ell_{i-1} \leqslant \ell^\star_{{\text{\textsf{large}}}}$, we have $\ell_i \geqslant \ell^\star_{{\text{\textsf{large}}}}+1$.
We first prove an analog of Proposition \[pro:first-merge-2\]: adapting its proof in the case where we had $h = 2\kappa+1$ (and we now have $h \leqslant 2\kappa+1$ only), we observe that (i) any two distinct successive states $\mathbf{S}$ and $\mathbf{S}'$ obtained while executing the function obey the relation $\mathbf{S}' = {\textsf{succ$_{\star\kappa}$}}(\mathbf{S})$, and that (ii) the final state that we obtain, just before going to line \[alg:m:end\], and which is equal to the stack ${\mathcal{S}}$, does not have any $\star\kappa$-merge point.
Then, exactly like in the proof of Proposition \[pro:cassm-subset-cass\], the merge performed to transform the state $\mathbf{S}$ into $\mathbf{S}'$ would also have been performed by the algorithm , which proves the first part of Proposition \[pro:flS-study\].
Second, let us observe that a very large run is also a large run. Hence, for every run $R$ obtained by merging two runs $R_\ominus$ and $R_\oplus$ during the main body of [$\kappa$-]{}, we had $\ell^\star_{{\text{\textsf{large}}}}\geqslant \ell_\ominus$ and $\ell^\star_{{\text{\textsf{large}}}}\geqslant \ell_\oplus$, so that $\ell^\star_{{\text{\textsf{large}}}}+ 1 \geqslant \ell$. Hence, every run $R_i$ of ${\mathcal{S}}$ such that $\ell_i \geqslant \ell^\star_{{\text{\textsf{large}}}}$ has never been merged, and was indeed a run of the original sequence ${\mathcal{R}}$.
Similarly, and adapting the proof of Lemma \[lem:m-is-small\], we observe that, if ${\mathcal{S}}$ contains two consecutive runs $R_i$ and $R_{i+1}$ that are not very large, then the $\star\kappa$-merge point of ${\mathcal{S}}$ exists. Thus, if ${\mathcal{S}}$ is the final state obtained in line \[alg:m:flS\], and since that state does not have a $\star\kappa$-merge point, it means that, among any two consecutive runs of ${\mathcal{S}}$, at least one must be very large. This completes the proof.
As a nice consequence of these results, and although this will not be directly useful for proving Theorem \[thm:eta-optimal\], we can observe that using [$\kappa$-]{}instead of comes with no overhead in terms of merge cost.
\[cor:cassm-better-than-cass\] The merge cost of [$\kappa$-]{}is never higher than the merge cost of .
Let ${\mathcal{R}}$ be some sequence of runs on which and [$\kappa$-]{}are executed, and let ${\mathcal{T}}$ be the merge tree induced by ${\mathcal{R}}$ (for the algorithm ). Propositions \[pro:cassm-subset-cass\] and \[pro:flS-study\] prove that the stack ${\mathcal{S}}$ obtained at the end of [$\kappa$-]{}’s main body is a border of ${\mathcal{T}}$, and that the merges performed in this main body are those merges of ${\mathcal{T}}$ that lie below the border ${\mathcal{S}}$.
Then, the merge cost of line \[alg:m:end\] cannot be higher than the merge cost used by to merge the runs in ${\mathcal{S}}$, because the former cost is that of an optimal merge policy. The result follows.
Proposition \[pro:flS-study\] provides us with a characterisation of the sequence ${\mathcal{S}}$ on which the optimal merge procedure of line \[alg:m:end\] is executed. Furthemore, and together with Proposition \[pro:cassm-subset-cass\], it proves that every merge performed prior to line \[alg:m:end\] would also have been performed by . This suggests that the cost of the merges performed to transform the original sequence ${\mathcal{R}}$ into ${\mathcal{S}}$ is close to optimal. Now, in order to prove Theorem \[thm:eta-optimal\], we must now compare [$\kappa$-]{}to an optimal stable merge sorting policy. The main obstacle that might prevent us from performing this comparison is that the main body, on which we have focused so far, does not perform all merges: it performs only a part of these merges, reducing the sequence ${\mathcal{R}}$ on which it is executed to a sequence ${\mathcal{S}}$ of length at most $2\kappa+1$, which is the stack obtained after line \[alg:m:flS\].
Hence, we shall prove two different statements. One is that the main body of [$\kappa$-]{}indeed reduces ${\mathcal{R}}$ to the sequence ${\mathcal{S}}$ with an optimal merge cost, up to some error that we shall evaluate. The other one is that the sequence ${\mathcal{S}}$, although it might not be a border of any merge tree induced by ${\mathcal{R}}$ for an optimal stable merge policy, is still good enough that the cost penalty incurred by using ${\mathcal{S}}$ as an intermediate state is not too high.
For the sake of readability, and given a sequence of runs ${\mathcal{R}}= (R_1,\ldots,R_\rho)$, we identify below every sub-interval $X$ of $\{1,\ldots,\rho\}$ with the run obtained by merging those runs $R_k$ such that $k \in X$. Accordingly, we have $R_{\{k\}} = R_k$, and we also set $r_X = \sum_{k \in X} r_k$, and so on. Furthermore, if $X = \{x_1,\ldots,x_k\}$, we denote by ${\mathsf{mc}}_X$ the minimal merge cost of a stable merge policy that would sort the sequence $(R_{x_1},\ldots,R_{x_k})$ and merge these runs into the run $R_X$.
\[def:diff:partition\] Let $\mathbf{I} = (I_1,\ldots,I_n)$ and $\mathbf{J} = (J_1,\ldots,J_m)$ be two partitions of the set $\{1,\ldots,\rho\}$ into intervals, and such that $\mathbf{J}$ refines $\mathbf{I}$, i.e., that every interval $J_j$ is contained in some interval $I_i$.
For all $i \leqslant n$, we set $\mathsf{d}_i = \#\{j \colon J_j \subseteq I_i\}$, i.e., $\mathsf{d}_i$ is the number of intervals $I_i$ in which $J_j$ is subdivided. We say that $\mathbf{J}$ is a *$\mathsf{d}$-refinement* of $\mathbf{I}$ if $\mathsf{d}_i \leqslant \mathsf{d}$ for all $i \leqslant n$.
Finally, we call *distortion* between $\mathbf{I}$ and $\mathbf{J}$ the integer
$\displaystyle \Delta(\mathbf{I},\mathbf{J}) = \sum_{i=1}^n r_{\mathbf{I}_i} \mathbf{1}_{\mathsf{d}_i \neq 1},$
i.e., the cumulative length of all those runs $R_{\mathbf{I}_i}$ such that $\mathbf{I}_i$ does not belong to the partition $\mathbf{J}$.
\[lem:refinement\] Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs to be sorted and let $\mathsf{d}$ be a positive integer. Let also $\mathbf{I} = (I_1,\ldots,I_n)$ and $\mathbf{J} = (J_1,\ldots,J_m)$ be two partitions of the set $\{1,\ldots,\rho\}$ into intervals, such that $\mathbf{J}$ is a $\mathsf{d}$-refinement of $\mathbf{I}$. Finally, let $\mathsf{mc}_{\mathbf{I}}$ be the minimal merge cost of a stable merge policy that would sort the sequence ${\mathcal{R}}_{\mathbf{I}} = (R_{I_1},\ldots,R_{I_n})$, let $\mathsf{mc}'_{\mathbf{I}} = \sum_{i=1}^n \mathsf{mc}_{I_i}$, and let $\mathsf{MC}_{\mathbf{I}} = \mathsf{mc}_{\mathbf{I}} +
\mathsf{mc}'_{\mathbf{I}}$. We define the integer $\mathsf{MC}_{\mathbf{J}}$ similarly. Then, it holds that $|\mathsf{MC}_{\mathbf{I}} -
\mathsf{MC}_{\mathbf{J}}| \leqslant \lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J})$.
First, consider some interval $I_i$ such that $\mathsf{d}_i \geqslant 2$, and let $j$ be the integer such that $I_i$ is the disjoint union of $J_{j+1},J_{j+2},\ldots,J_{j+\mathsf{d}_i}$. We can merge $R_{J_{j+1}},\ldots,R_{J_{j+\mathsf{d}_i}}$ into one unique run $R_{I_i}$ by using a balanced binary merge tree of height $\lceil \log_2(\mathsf{d}_i) \rceil$. The total merge cost of these operations is thus bounded above by $\lceil \log_2(\mathsf{d}) \rceil \, r_{\mathbf{I}_i}$. Proceeding in this way for every interval $I_i$ such that $\mathsf{d}_i \geqslant 2$, we have transformed the sequence ${\mathcal{R}}_{\mathbf{J}} = (R_{J_1},\ldots,R_{J_m})$ into the sequence ${\mathcal{R}}_{\mathbf{I}} = (R_{I_1},\ldots,R_{I_n})$ for a cost of $\lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J})$ at most. It follows both that $\mathsf{mc}_{\mathbf{J}} \leqslant \mathsf{mc}_{\mathbf{I}} +
\lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J})$ and that $\mathsf{mc}'_{\mathbf{I}} \leqslant \mathsf{mc}'_{\mathbf{J}} +
\lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J})$.
Then, let ${\mathcal{T}}$ be the merge tree, for the optimal merge policy, induced by the sequence ${\mathcal{R}}_{\mathbf{J}} = (R_{J_1},\ldots,R_{J_m})$. We transform the tree ${\mathcal{T}}$ as follows. For every interval $I_i$ such that $\mathsf{d}_i \geqslant 2$, let $R_{J_j}$ be a leaf of ${\mathcal{T}}$ of least depth such that $J_j \subseteq I_i$. Then, we replace the leaf $R_{J_j}$ by $R_{I_i}$ and, for all the other leaves $R_{J_k}$ such that $J_k \subseteq I_i$, we remove the leaf $R_{J_k}$ and we replace its father by its sibling. Finally, after these removals and replacements have been performed, we update the value of each internal node of the resulting tree, so that each internal node be the merge of its two children. The tree we obtain after this procedure has been applied is a possible merge tree for the sequence ${\mathcal{R}}_{\mathbf{I}} = (R_{I_1},\ldots,R_{I_n})$, and the merge cost is its associated to is not higher than merge cost of ${\mathcal{T}}$ itself. This means that $\mathsf{mc}_{\mathbf{I}} \leqslant \mathsf{mc}_{\mathbf{J}}$.
Finally, consider once again an interval $I_i$ such that $\mathsf{d}_i
\geqslant 2$, and let us focus at the individual merges that are performed for merging all the runs $R_k$ (with $k \in I_i$) into the one run $R_{I_i}$. Each such merge involves two runs $R_X$ and $R_Y$, where $X$ and $Y$ are adjacent sub-intervals of $I_i$. Let $x = \max X = \min Y - 1$. If there exists an interval $J_j$ such that $\min J_j \leqslant x < \max J_j$, then we shall merge the runs $R_{X \cap J_j}$ and $R_{Y \cap J_j}$ instead of merging $R_X$ and $R_Y$; if $x = \max J_j$, then we just skip performing that merge. Doing so, we have diminished the merge cost of each individual merge, and we have transformed the sequence $(R_k)_{k \in I_i}$ into the sequence $(R_{J_j})_{J_j \subseteq I_i}$ instead of the run $R_{I_i}$ itself. Repeating this procedure for every interval $I_i$ such that $\mathsf{d}_i \geqslant 2$, we have transformed the sequence ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ into the sequence ${\mathcal{R}}_{\mathbf{J}}$ for a cost of $\mathsf{mc}'_{\mathbf{I}}$ at most. It follows that $\mathsf{mc}'_{\mathbf{J}} \leqslant \mathsf{mc}'_{\mathbf{I}}$.
Gathering the above inequalities, we conclude by observing that
$\displaystyle \begin{cases}
\mathsf{MC}_{\mathbf{I}} = \mathsf{mc}_{\mathbf{I}} + \mathsf{mc}'_{\mathbf{I}} \leqslant \mathsf{mc}_{\mathbf{J}} + \mathsf{mc}'_{\mathbf{J}} +
\lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J}) = \mathsf{MC}_{\mathbf{J}} +
\lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J}) \text{ and } \\
\mathsf{MC}_{\mathbf{J}} = \mathsf{mc}_{\mathbf{J}} + \mathsf{mc}'_{\mathbf{J}} \leqslant \mathsf{mc}_{\mathbf{I}} + \mathsf{mc}'_{\mathbf{I}} +
\lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J}) = \mathsf{MC}_{\mathbf{I}} +
\lceil \log_2(\mathsf{d}) \rceil \, \Delta(\mathbf{I},\mathbf{J}).
\end{cases}
\vspace{-1.5\baselineskip}
$
In view of Proposition \[pro:flS-study\] and of Lemma \[lem:refinement\], it would be tempting to prove that the merge tree ${\mathcal{T}}$ induced by the sequence ${\mathcal{R}}$, when following an optimal merge policy, has a border that is a $\mathsf{d}$-refinement of the stack ${\mathcal{S}}$ obtained at the end of the main body. The following results lead us to a variant of this statement.
\[no-small\] Let $R$ be some run obtained while applying an optimal merge policy on a sequence ${\mathcal{R}}$ or runs (i.e., $R$ may be either an original run already contained in ${\mathcal{R}}$, or be obtained by merging two other runs). Then, let $\overline{R}_1,\overline{R}_2,\ldots,\overline{R}_k$ be the runs with which $R$ is successively merged according to that optimal policy. For all $i \geqslant 3$, it holds that $\overline{r}_i \geqslant r$.
Let us first assume that $\overline{r}_3 < r$. Then, the total cost of the merges or $R$ with $\overline{R}_1$, $\overline{R}_2$ and $\overline{R}_3$ is $\mathsf{mc} = 3(r + \overline{r}_1) + 2 \overline{r}_2 + \overline{r}_3$. However, if we had used a balanced binary tree, of height 2, for merging these four runs, (i.e., merging first the two leftmost runs, then the two rightmost runs, and finally the two resulting runs), each run would have participated to $2$ merges only, for a total cost of $\mathsf{mc}' = 2 (r + \overline{r}_1 + \overline{r}_2 + \overline{r}_3) =
\mathsf{mc} - (r + \overline{r}_1 - \overline{r}_3) < \mathsf{mc}$. This contradicts the optimality of our merge policy, which proves that $\overline{r}_3 \geqslant r$.
The same reasoning, applied to the run obtained by merging $R$ and $\overline{R}_1,\ldots,\overline{R}_i$, shows that $\overline{r}_{i+3} \geqslant r + \overline{r}_1 + \ldots + \overline{r}_i \geqslant r$ for all $i \leqslant k-3$, which completes the proof.
\[cor:nice-border\] Let ${\mathcal{R}}$ be a sequence of runs, and let ${\mathcal{T}}$ be the merge tree induced by ${\mathcal{R}}$ for some optimal stable merge policy. There exists a border ${\mathcal{B}}= (R_1,\ldots,R_k)$ of ${\mathcal{T}}$ such that, (i) for all $i \leqslant k$, if $\ell_i \geqslant \ell^\star_{{\text{\textsf{large}}}}+ 2$, then the run $R_i$ was already a run of ${\mathcal{R}}$, and (ii) among any $7$ consecutive runs of ${\mathcal{B}}$, at least one is very large.
Let us first recall how we obtained the stack produced at the end of [$\kappa$-]{}’s main body. Starting from the merge tree ${\mathcal{T}}'$ induced by ${\mathcal{R}}$ (for the merge policy of ), we simply decided to keep removing any two sibling leaves whenever none of them was very large. Now, we use a similar strategy.
Starting from the border ${\mathcal{B}}= {\mathcal{R}}$, we keep modifying the border ${\mathcal{B}}$ as follows: whenever two consecutive runs $R_\ominus$ and $R_\oplus$ of ${\mathcal{B}}$ are siblings in ${\mathcal{T}}$, and if neither of them is very large, then we remove them from ${\mathcal{B}}$, and replace them by their parent $\overline{R}$, which is simply the run obtained by merging $R_\ominus$ and $R_\oplus$. Then, when no such operation is anymore possible, we prove that ${\mathcal{B}}$ is the border mentioned in the statement of Corollary \[cor:nice-border\].
Indeed, if a run $R$ of ${\mathcal{B}}$ was obtained by merging two runs $R_\ominus$ and $R_\oplus$, then we must have $\ell \leqslant \max\{\ell_\ominus,\ell_\oplus\} + 1 \leqslant
\ell^\star_{{\text{\textsf{large}}}}+ 1$, which proves the statement (i).
Then, for every run $R$ of ${\mathcal{B}}$ that is very large, let $\langle R \rangle$ denote the set of not-very-large runs it shall successively be merged to, according to the merge policy, plus $R$ itself. Lemma \[no-small\] proves that $\langle R \rangle$ contains at most $4$ runs. Moreover, if a run $R'$ of ${\mathcal{B}}$ is not very large, it means that its sibling, in ${\mathcal{T}}$, has a descendant $R''$ that belongs to ${\mathcal{B}}$ and is very large. It follows that $R' \in \langle R'' \rangle$. Thus, overall, the border ${\mathcal{B}}$ is the disjoint union of those sets $\langle R \rangle$ such that $R$ is a very large run of ${\mathcal{B}}$. By construction, each such set contains a very large run, and it contains at most $4$ runs, which are consecutive runs of ${\mathcal{B}}$. In particular, if $R_{i-3},\ldots,R_{i+3}$, are consecutive runs, there exists a very large run $R_j$ such that $R_i \in \langle R_j \rangle$, proving that $|i-j| \leqslant j$. The statement (ii) follows.
Then, we can finally derive the proof of Theorem \[thm:eta-optimal\].
\[pro:cassm-best\] Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs, and let $n^\ast$ be the cumulative length of those runs $R_i$ that are not very large in ${\mathcal{R}}$. Then, let ${\mathsf{mc}}_\kappa$ be the merge cost of [$\kappa$-]{}’s main body, and let ${\mathsf{mc}}_{{\mathcal{S}}}$ be the optimal merge cost of a stable merge policy that would transform the sequence ${\mathcal{R}}$ into the stack ${\mathcal{S}}$ obtained after the main body was executed. It holds that ${\mathsf{mc}}_\kappa \leqslant {\mathsf{mc}}_{{\mathcal{S}}} + 4 n^\ast$.
In this proof, let us forget the chronological order in which run merges are performed in the main body of [$\kappa$-]{}, and let us just focus on which merges are performed. Proposition \[pro:cassm-subset-cass\] states that these merges would all be performed by itself, and therefore we can just pretend that we shall count the total merge cost of , but omitting those merges that were not performed in [$\kappa$-]{}’s main body.
Let ${\mathcal{T}}$ be the merge tree induced by ${\mathcal{R}}$, and let $n = r_1+\ldots+r_\rho$. Observe that, if a run $R$ is very large in ${\mathcal{R}}$, then it is also large in every border of ${\mathcal{T}}$ that contains $R$. It follows from Proposition \[pro:first-merge-2\] that $R$ will never be merged during the main body of [$\kappa$-]{}.
Now, we we split the merge cost of the main body into two parts, depending on which merges were expanding or non-expanding. Borrowing the notations of Lemma \[lem:refinement\], let ${\mathcal{S}}= (R_{I_1},\ldots,R_{I_h})$ be the stack obtained at the end of [$\kappa$-]{}’s main body, and let $\mathbf{I} = (I_1,\ldots,I_h)$ be the associated partition of $\{1,\ldots,\rho\}$.
Then, consider some interval $X \in \{I_1,\ldots,I_h\}$. We recall that $\mathsf{mc}_X$ is the optimal cost of a merge policy that merges all the runs $(R_k)_{k \in X}$ into the single run $R_X$. Then, let also $\mathsf{mc}_{\kappa,X}$ be the cost of those *expanding* merges performed by [$\kappa$-]{}on sub-runs of $R_X$, and let ${\mathcal{H}}_X = \sum_{k \in X} r_k/r_X \log_2(r_X/r_k)$. As mentioned in Theorem \[thm:lower-bound-merge-cost\], we know that $\mathsf{mc}_X \geqslant r_X {\mathcal{H}}_X$.
Moreover, for all $k \in X$, the run $R_k$ has taken part to at most $\ell_X - \ell_k$ expanding merges. It follows that
$\displaystyle \mathsf{mc}_{\kappa,X} \leqslant \sum_{k \in X} r_k (\ell_X - \ell_k)
\leqslant \sum_{k \in X} r_k (1 + \log_2(r_X) - \log_2(r_X/r_k)) =
r_X (1 + {\mathcal{H}}_X) \leqslant \mathsf{mc}_X + r_X.$
Second, we can refine Lemma \[lem:non-expanding-run-2r\] in the case of very large runs $R$. Indeed, every such run is also a run $R_{I_i}$ (with $i \leqslant h$). Adapting the (second paragraph of the) proof of Lemma \[lem:non-expanding-run-2r\] proves that the total non-expanding cost of the critical merge and of the starting, middle and ending sequences of $R_{I_i}$ is at most $r_{I_{i-1}}$, or even $0$ if $i=1$ or if $R_{I_{i-1}}$ is also a very large run of ${\mathcal{R}}$.
Accounting for all very large runs of ${\mathcal{R}}$, the total non-expanding cost of those merges is therefore bounded above by $n^\ast$, and even by $n^\ast - R_{I_h}$ if $R_{I_h}$ is *not* a very large run of ${\mathcal{R}}$. Similarly, the total non-expanding cost of those merges that were performed in line \[alg:m:collapse\] or \[alg:m:collapse:end\] is at most $r_{I_h}$, or $0$ if $R_{I_h}$ is a not-very-large run of ${\mathcal{R}}$. Hence, the total non-expanding cost of these merges cannot exceed $n^\ast$.
Then, Lemma \[lem:non-expanding-run-2r\] itself takes care of the non-expanding cost that follows pushing a non-very-large run. Accounting for all such pushes, we obtain a non-expanding cost that cannot exceed $2n^\ast$. Consequently, the total merge cost of [$\kappa$-]{}’s main body is at most
$\displaystyle \sum_{X \colon \ell_X \leqslant \ell^\star_{{\text{\textsf{large}}}}} \mathsf{mc}_{\kappa,X} +
n^\ast + 2n^\ast \leqslant
\sum_{X \colon \ell_X \leqslant \ell^\star_{{\text{\textsf{large}}}}} (\mathsf{mc}_X + r_X) +
3 n^\ast = {\mathsf{mc}}_{{\mathcal{S}}} + 4 n^\ast.
\vspace{-2.5\baselineskip}$
Let ${\mathcal{R}}= (R_1,\ldots,R_\rho)$ be a sequence of runs of cumulative length $n = \sum_{i=1}^\rho r_i$. Then, let ${\mathcal{S}}$ be the stack obtained after executing [$\kappa$-]{}’s main body on ${\mathcal{R}}$, and let ${\mathcal{B}}$ be the border mentioned in Corollary \[cor:nice-border\]. Furthermore, let $\mathbf{I} = (I_1,\ldots,I_h)$ and $\mathbf{I}' = (I'_1,\ldots,I'_{h'})$ be the partitions of $\{1,\ldots,\rho\}$ into intervals such that ${\mathcal{S}}= (R_{I_1},\ldots,R_{I_h})$ and ${\mathcal{B}}= (R_{I'_1},\ldots,R_{I'_{h'}})$. and let also $\mathbf{J} = (J_1,\ldots,J_k)$ be the coarsest partition that refines both $\mathbf{I}$ and $\mathbf{I}'$.
We first prove that $\mathbf{J}$ is a $15$-refinement of both partitions $\mathbf{I}$ and $\mathbf{I}'$. Indeed, let $R_X$ be some run in ${\mathcal{S}}$ or in ${\mathcal{B}}$. If $\ell_X \geqslant \ell^\star_{{\text{\textsf{large}}}}+ 2$, then $R_X$ is also a run of ${\mathcal{R}}$ and of both of ${\mathcal{S}}$ and ${\mathcal{B}}$, hence $X \in \mathbf{J}$.
Then, let $R_X$ be some run in ${\mathcal{S}}$ such that $\ell_X \leqslant \ell^\star_{{\text{\textsf{large}}}}+ 1$, and let $J_a,\ldots,J_{a+b}$ be the intervals in $\mathbf{J}$ contained in $X$. Since $\ell_X \leqslant \ell^\star_{{\text{\textsf{large}}}}+ 1$, we know that $r_X < 2^{\ell^\star_{{\text{\textsf{large}}}}+ 2}$. Moreover, each of the sets $J_{a+1},\ldots,J_{a+b-1}$ must belong to $\mathbf{I}'$, i.e., $R_{J_{a+1}},\ldots,R_{J_{a+b-1}}$ are consecutive runs of ${\mathcal{B}}$. If $b \geqslant 15$, then these runs can be split in $2$ subsequences of $7$ consecutive runs. Then, Corollary \[cor:nice-border\] proves that at least $2$ of these, let us say $R_Y$ and $R_{Y'}$, are very large. It follows that
$\displaystyle 2^{\ell^\star_{{\text{\textsf{large}}}}+ 2} > r_X \geqslant r_{Y} + r_{Y'} \geqslant
2 \times 2^{\ell^\star_{{\text{\textsf{large}}}}+1},$
which is a contradiction. It follows that $b \leqslant 14$, which means that $\mathbf{J}$ is a $15$-refinement of $\mathbf{I}$.
Similarly, Proposition \[pro:flS-study\] proves that, among any two consecutive runs of ${\mathcal{S}}$, at least one must be very large, and it follows that $\mathbf{J}$ is a $5$-refinement of $\mathbf{I}'$.
Now, let us use the notations of Lemma \[lem:refinement\] and of Proposition \[pro:cassm-best\], and denote by $\mathsf{mc}$ and by $\mathsf{mc}_{{\textsf{opt}}}$ the respective costs of [$\kappa$-]{}and of an optimal stable merge policy. Observing that $\mathsf{mc}_{{\textsf{opt}}}= \mathsf{MC}_{\mathbf{I}'}$, that $\mathsf{mc}'_{\mathbf{I}} = \mathsf{mc}_{{\mathcal{S}}}$ and that $\Delta(\mathbf{I},\mathbf{J}) = n^\ast$, it follows that $$\begin{aligned}
\mathsf{mc} & = \mathsf{mc}_{\mathbf{I}} + \mathsf{mc}_\kappa
\leqslant \mathsf{MC}_{\mathbf{I}} + 4n^\ast
\leqslant \mathsf{MC}_{\mathbf{J}} + (4 + \lceil \log_2(15) \rceil) n^\ast \\
& \leqslant \mathsf{MC}_{\mathbf{I}'} + (4 + \lceil \log_2(15) \rceil +
\lceil \log_2(5) \rceil) n^\ast = \mathsf{mc}_{{\textsf{opt}}}+ 11 n^\ast.\end{aligned}$$
Furthermore, if $R_X$ is a run of ${\mathcal{S}}$ that is not very large, then
$\displaystyle \log_2(r_X / 2) \leqslant \lfloor \log_2(r_X) \rfloor =
\ell_X \leqslant \ell^\star_{{\text{\textsf{large}}}}=
\lfloor \log_2(\lfloor n/\kappa \rfloor)\rfloor
\leqslant \log_2(n / \kappa),$
and it follows that $r_X \leqslant 2 n / \kappa$. In particular, for all $k \in X$, we also have $r_k \leqslant 2 n / \kappa$, and thus
$\displaystyle
n {\mathcal{H}}\geqslant \sum_{X \colon \ell_X \leqslant \ell^\star_{{\text{\textsf{large}}}}} \, \sum_{k \in X} r_k \log_2(n / r_k) \geqslant n^\ast \log_2(\kappa / 2).
$
Since Theorem \[thm:lower-bound-merge-cost\] proves that $\mathsf{mc}_{{\textsf{opt}}}\geqslant n {\mathcal{H}}$, it follows that
$\displaystyle \mathsf{mc} \leqslant (1 + 11 n^\ast / (n {\mathcal{H}})) \mathsf{mc}_{{\textsf{opt}}}\leqslant (1 + 11 / \log_2(\kappa / 2)) \mathsf{mc}_{{\textsf{opt}}}=
(1 + \eta_{2\kappa+3}) \mathsf{mc}_{{\textsf{opt}}}.
$
Since $\eta_{2\kappa+3} \leqslant \eta_{2\kappa+2}$, and since [$\kappa$-]{}is both $(2\kappa+2)$-aware and $(2\kappa+3)$-aware, this completes the proof.
[10]{}
Nicolas Auger, Vincent Jug[é]{}, Cyril Nicaud, and Carine Pivoteau. On the worst-case complexity of timsort. In [*26th Annual European Symposium on Algorithms (ESA 2018)*]{}. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, 2018. Extended version available at: <https://arxiv.org/abs/1805.08612>.
Nicolas Auger, Cyril Nicaud, and Carine Pivoteau. Merge strategies: [F]{}rom [M]{}erge [S]{}ort to [T]{}im[S]{}ort. Research Report hal-01212839, hal, 2015. URL: <https://hal-upec-upem.archives-ouvertes.fr/hal-01212839>.
J[é]{}r[é]{}my Barbay and Gonzalo Navarro. On compressing permutations and adaptive sorting. , 513:109–123, 2013.
Sam Buss and Alexander Knop. Strategies for stable merge sorting. In [*Proceedings of the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms*]{}, pages 1272–1290. Society for Industrial and Applied Mathematics, 2019.
Stijn De Gouw, Jurriaan Rot, Frank de Boer, Richard Bubel, and Reiner H[ä]{}hnle. Open[JDK]{}’s [J]{}ava.utils.[C]{}ollection.sort() is broken: [T]{}he good, the bad and the worst case. In [*International Conference on Computer Aided Verification*]{}, pages 273–289. Springer, 2015.
Edsger Dijkstra. Smoothsort, an alternative for sorting in situ. In [*Theoretical Foundations of Programming Methodology*]{}, pages 3–17. Springer, 1982.
Herman Goldstine and John von Neumann. Planning and coding of problems for an electronic computing instrument. 1947.
Mordecai Golin and Robert Sedgewick. Queue-mergesort. , 48(5):253–259, 1993.
Tony Hoare. Algorithm 64: [Q]{}uicksort. , 4(7):321, 1961.
Donald Knuth. . Addison Wesley Longman Publish. Co., Redwood City, CA, USA, 1998.
Heikki Mannila. Measures of presortedness and optimal sorting algorithms. , 34(4):318–325, 1985.
Alistair Moffat, Gary Eddy, and Ola Petersson. Splaysort: Fast, versatile, practical. , 26(7):781–797, 1996.
J. Ian Munro and Sebastian Wild. Nearly-optimal mergesorts: Fast, practical sorting methods that optimally adapt to existing runs. In Yossi Azar Hannah Bast and Grzegorz Herman, editors, [*26th Annual European Symposium on Algorithms (ESA 2018)*]{}, Leibniz International Proceedings in Informatics (LIPIcs), pages 63:1–63:15, 2018.
Tim Peters. Timsort description, accessed june 2015. <http://svn.python.org/projects/python/trunk/Objects/listsort.txt>.
Olin Shivers. A simple and efficient natural merge sort. Technical report, Georgia Institute of Technology, 2002.
Tadao Takaoka. Partial solution and entropy. In Rastislav Kr[á]{}lovi[č]{} and Damian Niwi[ń]{}ski, editors, [*Mathematical Foundations of Computer Science 2009*]{}, pages 700–711, Berlin, Heidelberg, 2009. Springer Berlin Heidelberg.
John Williams. Algorithm 232: [H]{}eapsort. , 7:347–348, 1964.
Appendix: Proof of Theorem \[thm:complexity-AugSS\] {#sec:detailed-analysis}
===================================================
We reproduce here the full statement of Theorem \[thm:complexity-nH+D\], and then we provide a full proof.
Like the proof of Theorem \[thm:complexity-nH+3\], the proof we draw below relies on Lemmas \[lem:l-small-increase\] and on proofs on invariants and on the dynamics of the algorithm, which are provided in Lemmas \[lem:invariant-li\], \[lem:invariant-mid-end\] and \[lem:invariant-end\].
However, since we look for a finer upper bound on the total merge cost of the algorithm, we will forget the distinction between expanding and non-expanding merge costs, although we will reuse the notion of intrinsically expanding merge. Instead, we shall rely on the following notion of *potential* or a run, which will allow a precise evaluation or the merge costs.
Let $\Phi : [0,1] \mapsto \mathbb{R}$ be the function defined by $\Phi : x \mapsto \max\{(2-5x)/3,1/2-x,0\}$. Let also $R$ be a run of length $r$, and let $\ell \in \mathbb{Z}$ and $r_\bullet \in [0,1)$ be defined by the relation $r = 2^\ell (1+r_\bullet) c$. We define the *potential* of the run $R$ as the real number
$\displaystyle \mathbf{Pot}(R) = \begin{cases}
2^\ell \, \Phi(r_\bullet) \, c - \ell \, r & \text{if $R$ is in the stack;} \\
2^{\ell+1} \, c - \ell \, r & \text{if $R$ has not yet been pushed onto the stack.}
\end{cases}$
At each step of the algorithm, some runs belong to the stack, and may even result from the merge of previous runs, while other runs may not have been discovered yet, and therefore have not yet been pushed onto the stack. Hence, we define the *global potential* of that step as the sum of the individual potentials of the runs, which may or may not belong to the stack.
A first case where the variation of global potential fits well merge costs is the case where two runs of similar lengths are merged with each other.
\[lem:merge-cost-equitable\] Consider some intrinsically expanding merge between two runs $R$ and $R'$, during an execution of the algorithm. Let $\mathbf{Pot}_{\mathrm{before}}$ and $\mathbf{Pot}_{\mathrm{after}}$ be the respective global potentials *before* and *after* the merge, and let ${\mathsf{mc}}$ be the cost of that merge. It holds that $\mathbf{Pot}_{\mathrm{before}} \geqslant
\mathbf{Pot}_{\mathrm{after}} + {\mathsf{mc}}$.
Let $R''$ be the run obtained by merging $R$ and $R'$. Since our merge is intrinsically expanding, it means that $\ell = \ell'$, and it then comes that $r'' = r + r' = 2^\ell (2 + r_\bullet
+ r'_\bullet) c = 2^{\ell''} (1 + r''_\bullet) c$, where $\ell'' = \ell + 1$ and $r''_\bullet = (r_\bullet + r'_\bullet) / 2$.
Since $\Phi$ is a maximum of affine functions, it is convex, hence $2 \Phi(r''_\bullet) \leqslant
\Phi(r_\bullet) + \Phi(r'_\bullet)$. It follows that $$\begin{aligned}
\mathbf{Pot}_{\mathrm{before}} - \mathbf{Pot}_{\mathrm{after}} & =
(2^\ell \, \Phi(r_\bullet) \, c - \ell \, r) +
(2^{\ell'} \, \Phi(r'_\bullet) \, c - \ell' \, r') -
(2^{\ell''} \, \Phi(r''_\bullet) \, c - \ell'' \, r'') \\
& =
2^\ell \left(\Phi(r_\bullet) + \Phi(r'_\bullet) - 2 \Phi(r''_\bullet)\right) \, c + (r + r')
\geqslant r + r' = {\mathsf{mc}}.\end{aligned}$$
Unfortunately, not all merges are covered by Lemma \[lem:merge-cost-equitable\]. Yet, Lemmas \[lem:invariant-li\], \[lem:invariant-mid-end\] and \[lem:invariant-end\] provide us with a description of the dynamics of , which will allow us to deal with all other stack modifications due to cases \#1 to \#4.
Indeed, and although variations of global potential might not be a proper estimation of the cost of single merges, we can still cluster merges into groups whose total cost is well approached by variations of global potential. These groups are the following variant of starting, middle and ending sequences.
\[def:cluster\] We call *cluster sequence* of a run $R$ the sequence of merges formed of those merges that belong to the starting sequence of $R$ and of the critical merge of $R$ (if it exists).
Note, although this might not sound a priori obvious, that the cluster sequence of a run $R$ is formed of consecutive merges only. This is a consequence of the following result, which refines Lemma \[lem:invariant-mid-end\].
\[lem:middle=>no-ending\] If the middle sequence of a run $R$ is non-empty, then the critical merge of $R$ does not exist, and the ending sequence of $R$ is empty.
We just need to prove that no \#4-merge can follow a \#3-merge. Indeed, let $\mathbf{m}$ be a \#3-merge. We denote by ${\mathcal{S}}= (R_1,\ldots,R_h)$ the stack before $\mathbf{m}$ occurs and by $\overline{{\mathcal{S}}} = ({\overline{R}}_1,\ldots,{\overline{R}}_{\overline{h}})$ the stack after $\mathbf{m}$ occurred.
Since $\mathbf{m}$ is a \#3-merge, it must be the case that $\ell_1 < \ell_3 \leqslant \ell_2$. Then, we have $\overline{r}_1 = r_1$ and $\overline{r}_2 = r_2 + r_3 \geqslant r_2$, and therefore $\overline{\ell}_2 \geqslant \ell_2 > \ell_1 = \overline{\ell}_1$. Hence, no \#4-merge follows the merge $\mathbf{m}$.
Before going further, we mention one technical lemma that will be useful in subsequent computations.
\[lem:ugly-phi\] For all real numbers $x$ such that $0 \leqslant x \leqslant 1$, we have $1 \geqslant x + 2 \Phi(x/2) - \Phi(x) \geqslant
x + \Phi(x)$.
First, since the function $\Phi$ is non-increasing, the inequality $x + 2 \Phi(x/2) - \Phi(x) \geqslant
x + \Phi(x)$ is immediate. Moreover, $\Phi$ is affine on each of the intervals $[0,1/4]$, $[1/4,1/2]$ and $[1/2,1]$, and therefore so is the function $f : x \mapsto x + 2 \Phi(x/2) - \Phi(x)$. Hence, $f(x) \leqslant \max\{f(0),f(1/4),f(1/2),f(1)\} = 1$ whenever $0 \leqslant x \leqslant 1$, which means that $1 \geqslant x + 2 \Phi(x/2) - \Phi(x)$.
\[lem:merge-cost-cluster\] Consider some run $R$ that is to be pushed onto the stack, and let ${{\mathcal{M}}_{\mathrm{cluster}}}$ be the cluster sequence of $R$. Let $\mathbf{Pot}_{\mathrm{before}}$ and $\mathbf{Pot}_{\mathrm{after}}$ be the respective global potentials just before ${{\mathcal{M}}_{\mathrm{cluster}}}$ starts and just after it ends. Let also ${\mathsf{mc}}$ be the total merge cost of ${{\mathcal{M}}_{\mathrm{cluster}}}$. It holds that $\mathbf{Pot}_{\mathrm{before}} \geqslant
\mathbf{Pot}_{\mathrm{after}} + {\mathsf{mc}}$.
Let ${\mathcal{S}}= (R_1,\ldots,R_h)$ and ${\mathcal{S}}' = (R'_1,\ldots,R'_{h'})$ be the respective stacks obtained just before pushing $R$ and just after the cluster sequence of $R$ ends. Let also $k = \max\{i \mid 2 \leqslant i \leqslant h$ and $\ell_i \leqslant \ell\}$, or $k = 1$ if $\ell < \ell_3$. For the sake of convenience, we further set $\Delta_{\mathbf{Pot}} = \mathbf{Pot}_{\mathrm{before}} -
\mathbf{Pot}_{\mathrm{after}}$.
Note, due to Lemma \[lem:invariant-li\], that $\ell_k \geqslant \ell_i + (k-i)$ for all $i \leqslant k$, and therefore that, at the end of the starting sequence of $R$, all runs $R_1,\ldots,R_k$ were merged into a single run ${\overline{R}}$ such that $\overline{r} = r_1 + \ldots + r_k$ and $\ell_k \leqslant \overline{\ell} \leqslant \ell_k+1$.
Unfortunately, we shall now distinguish *four* cases, depending on whether the critical merge ${\mathbf{M}_{\mathrm{cr}}}$ of $R$ exists and, if it does, whether $\ell = \ell_k$ and whether ${\mathbf{M}_{\mathrm{cr}}}$ is expanding for $R$.
1. If ${\mathbf{M}_{\mathrm{cr}}}$ does not exist, then ${\mathsf{mc}}= \sum_{i=1}^k (k+1-i) r_i - r_1$. Furthermore, we know that $r'_1 = r$, that $r'_2 = \overline{r}$ and that $r'_j = r_{j+k-2}$ for all $j \geqslant 3$. Moreover, in that case, we know that $\ell'_2 > \ell'_1 = \ell \geqslant \ell_k$. If follows that $\ell'_2 = \ell + 1$ and therefore that $\ell'_2 \geqslant \ell_i + (k+1-i)$ for all $i \leqslant k$. Thus, we observe that $$\begin{aligned}
\Delta_{\mathbf{Pot}} & =
2^{\ell+1} \, c -
2^{\ell} \, \Phi(r_\bullet) \, c -
2^{\ell'_2} \, \Phi(r'_{2\bullet}) \, c +
{\textstyle\sum_{i=1}^k} \left((\ell'_2 - \ell_i) r_i + 2^{\ell_i} \, \Phi(r_{i\bullet}) \, c \right) \\
& \geqslant
2^{\ell+1} \, c -
2^{\ell} \, \Phi(r_\bullet) \, c -
2^{\ell+1} \, \Phi(r'_{2\bullet}) \, c
+
{\textstyle\sum_{i=1}^k} (k+1-i) r_i \\
& \geqslant {\mathsf{mc}}+ 2^{\ell} \, \left(2 - \Phi(r_\bullet) - 2 \Phi(r'_{2\bullet})\right) \, c \\
& \geqslant {\mathsf{mc}}\hspace{48mm} \text{since $\Phi(x) \leqslant 2/3$ whenever $0 \leqslant x \leqslant 1$.}
\end{aligned}$$
2. If ${\mathbf{M}_{\mathrm{cr}}}$ exists, then ${\mathsf{mc}}= r + \sum_{i=1}^k (k+2-i) r_i - r_1$, and we know that $r'_1 = r + \overline{r}$ and that $r'_j = r_{j+k-1}$ for all $j \geqslant 2$. If, furthermore, ${\mathbf{M}_{\mathrm{cr}}}$ is non-expanding for $R$, then this means that $\ell'_1 = \ell \geqslant \overline{\ell}+1$. Thus, we also observe that $$\begin{aligned}
\Delta_{\mathbf{Pot}} & =
(\ell'_1 - \ell) r + 2^{\ell+1} \, c -
2^{\ell'_1} \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} \left((\ell'_1 - \ell_i) r_i +
2^{\ell_i} \, \Phi(r_{i\bullet}) \, c\right) \\
& \geqslant
2^{\ell+1} \, c -
2^\ell \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} (k+1-i) r_i \\
& \geqslant {\mathsf{mc}}+ 2^\ell \, \left(1 - \Phi(r'_{1\bullet}) - r'_{1\bullet}\right) \, c \\
& \geqslant {\mathsf{mc}}\hspace{48mm} \text{thanks to Lemma~\ref{lem:ugly-phi}.}
\end{aligned}$$
3. If ${\mathbf{M}_{\mathrm{cr}}}$ exists and $\ell = \ell_k$, then $\overline{\ell} = \ell$ and $\ell'_1 = \ell+1$. In that case, it also holds that $r_{k\bullet} \leqslant \overline{r}_\bullet$ and that $\overline{r}_\bullet \leqslant \overline{r}_\bullet + r_\bullet =
2 r'_{1\bullet}$. Hence, since $\Phi$ is decreasing, we know that $\Phi(r'_{1\bullet}) \leqslant
\Phi(\overline{r}_\bullet/2)$. In that case, it follows that $$\begin{aligned}
\Delta_{\mathbf{Pot}} & =
(\ell'_1 - \ell) r + 2^{\ell+1} \, c -
2^{\ell'_1} \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} \left((\ell'_1 - \ell_i) r_i + 2^{\ell_i} \,
\Phi(r_{i\bullet}) \, c \right) \\
& \geqslant
r + 2^{\ell+1} \, c -
2^{\ell+1} \, \Phi(r'_{1\bullet}) \, c +
2^\ell \, \Phi(r_{k\bullet}) \, c +
{\textstyle\sum_{i=1}^k} (k+1-i) r_i \\
& \geqslant
{\mathsf{mc}}- r'_1 + r + 2^{\ell+1} \, c -
2^{\ell+1} \, \Phi(r'_{1\bullet}) \, c +
2^\ell \, \Phi(r_{k\bullet}) \, c \\
& \geqslant
{\mathsf{mc}}+ 2^\ell \, \left(1 - \overline{r}_\bullet + \Phi(\overline{r}_\bullet) - 2
\Phi(\overline{r}_\bullet/2)\right) \, c
\\
& \geqslant
{\mathsf{mc}}\hspace{48mm} \text{thanks to Lemma~\ref{lem:ugly-phi}.}
\end{aligned}$$
4. If ${\mathbf{M}_{\mathrm{cr}}}$ exists, is expanding for $R$, and $\ell > \ell_k$, then $\ell'_1 \geqslant \ell + 1 \geqslant \ell_k + 2$, hence we observe that $$\begin{aligned}
\Delta_{\mathbf{Pot}} & =
(\ell'_1 - \ell) r + 2^{\ell+1} \, c -
2^{\ell'_1} \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} \left((\ell'_1 - \ell_i) r_i + 2^{\ell_i} \, \Phi(r_{i\bullet}) \, c \right) \\
& \geqslant
r + 2^{\ell+1} \, c -
2^{\ell+1} \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} (k+2-i) r_i \\
& \geqslant
{\mathsf{mc}}+ 2^{\ell+1} \, \left(1 - \Phi(r'_{1\bullet})\right) \, c \\
& \geqslant
{\mathsf{mc}}\hspace{48mm} \text{since $\Phi(x) \leqslant 2/3$ whenever $0 \leqslant x \leqslant 1$.}
\end{aligned}$$
Throughout the main loop of , every merge belongs either to a cluster sequence or to a middle or ending sequence. In the latter cases, that merge is intrinsically expanding. Therefore, Lemmas \[lem:merge-cost-equitable\] and \[lem:merge-cost-cluster\] prove that the total merge cost of the main loop is bounded from above by the variation of global potential between the beginning and the end of that main loop.
Hence, it remains to study merges that occur in line \[alg:collapse\] of Algorithm \[alg:ASS\]. This is the object of the next result.
\[lem:merge-cost-final-collapse\] Let $\mathbf{Pot}_{\mathrm{before}}$ and $\mathbf{Pot}_{\mathrm{after}}$ be the respective global potentials just before line \[alg:collapse\] starts and just after it ends. Let also ${\mathsf{mc}}$ be the total cost of the merges taking place in line \[alg:collapse\]. It holds that $\mathbf{Pot}_{\mathrm{before}} + n \geqslant
\mathbf{Pot}_{\mathrm{after}} + {\mathsf{mc}}$, where $n$ is the length of the array to be sorted.
Let ${\mathcal{S}}= (R_1,\ldots,R_h)$ and ${\mathcal{S}}' = (R'_1)$ be the respective stacks obtained just before and just after line \[alg:collapse\]. Once again, let also $\Delta_{\mathbf{Pot}} = \mathbf{Pot}_{\mathrm{before}} -
\mathbf{Pot}_{\mathrm{after}}$.
Lemma \[lem:invariant-li\] proves that $\ell_h \geqslant \ell_i + (h-i)$ for all $i \leqslant h$, and therefore that $\ell_h+1 \geqslant \ell'_1 \geqslant \ell_h$. Furthermore, we know that $r'_1 = n = r_1+\ldots+r_h$ and that ${\mathsf{mc}}= \sum_{i=1}^h(h+1-i)r_i - r_1$.
This time, we shall distinguish two cases, depending on whether $\ell_h = \ell'_1$.
1. If $\ell'_1 = \ell_h+1$, and since $\Phi(x) \leqslant 2/3$ for all $x$, we observe that $$\begin{aligned}
\Delta_{\mathbf{Pot}} + n & =
r'_1 - 2^{\ell'_1} \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} \left((\ell'_1 - \ell_i) r_i + 2^{\ell_i} \,
\Phi(r_{i\bullet}) \, c \right) \\
& \geqslant
r'_1 - 2^{\ell'_1} \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} (h+1-i) r_i \\
& \geqslant
{\mathsf{mc}}+ 2^{\ell'_1} \, \left(1 + r'_{1\bullet} - \Phi(r'_{1\bullet})\right) \, c \\
& \geqslant {\mathsf{mc}}.
\end{aligned}$$
2. If $\ell'_1 = \ell_h$, then $r_{h\bullet} \leqslant
r'_{1\bullet}$, and therefore $\Phi(r_{h\bullet}) \geqslant
\Phi(r'_{1\bullet})$. Thus, we observe that $$\begin{aligned}
\Delta_{\mathbf{Pot}} + n & =
r'_1 - 2^{\ell'_1} \, \Phi(r'_{1\bullet}) \, c +
{\textstyle\sum_{i=1}^k} \left((\ell'_1 - \ell_i) r_i + 2^{\ell_i} \,
\Phi(r_{i\bullet}) \, c \right) \\
& \geqslant
r'_1 - 2^{\ell'_1} \, \Phi(r'_{1\bullet}) \, c
+ 2^{\ell_h} \, \Phi(r_{h\bullet}) \, c +
{\textstyle\sum_{i=1}^k} (h-i) r_i \\
& \geqslant
{\mathsf{mc}}+ r'_1 - {\textstyle\sum_{i=1}^k} r_i +
2^{\ell'_1} \, \left(\Phi(r_{h\bullet}) - \Phi(r'_{1\bullet})\right) \, c \\
& \geqslant {\mathsf{mc}}.
\end{aligned}$$
Before concluding this section, we shall first prove another technical lemma.
\[lem:log+phi\] For all real numbers $x$ such that $0 \leqslant x \leqslant 1$, we have
$\displaystyle \log_2(1+x) + \Phi(x)/(1+x) \geqslant \log_2(5) - 9/5.$
Let $f(x) = \log_2(1+x) + \Phi(x)/(1+x)$. Since $\Phi$ is affine on each of the intervals $[0,1/4]$, $[1/4,1/2]$ and $[1/2,1]$, we treat these intervals separately.
1. If $0 \leqslant x \leqslant 1/4$, then $\Phi(x) = (2-5x)/3$, hence
$\displaystyle (1+x)^2 f'(x) = (1+x)\log_2(e)-7/3 \leqslant
5/4 \log_2(e) - 7/3 \approx -0.5,$
and therefore $f(x) \geqslant f(1/4) = \log_2(5)-9/5$.
2. If $1/4 \leqslant x \leqslant 1/2$, then $\Phi(x) = 1/2-x$, hence
$\displaystyle (1+x)^2 f'(x) = (1+x)\log_2(e)-3/2 \geqslant
5/4 \log_2(e) - 3/2 \approx 0.3,$
and therefore $f(x) \geqslant f(1/4) = \log_2(5)-9/5$ too.
3. If $1/2 \leqslant x \leqslant 1$, then $\Phi(x) = 0$, hence $f(x) = \log_2(1+x) \geqslant \log_2(3/2) \geqslant \log_2(5)-9/5$.
Like in Section \[sec:analysis\], we may now conclude with the proof of Theorem \[thm:complexity-nH+D\].
Let $\mathbf{Pot}_{\mathrm{start}}$ and $\mathbf{Pot}_{\mathrm{end}}$ be the respective global potentials when the algorithm starts and when it ends, and let ${\mathsf{mc}}$ be the total merge cost of .
The potential of a run $R$ that has not yet been pushed onto the stack is $$\begin{aligned}
\mathbf{Pot}(R) & = 2^{\ell+1} \, c - \ell \, r \\
& = \left( 2 / (1 + r_\bullet) - \log_2(r / c) +
\log_2(1 + r_\bullet) \right) \, r \\
& \leqslant (2 - \log_2(r / c)) \, r \hspace{18mm} \text{by applying Lemma~\ref{lem:2pow-x} to $x = \log_2(1 + r_\bullet)$.}\end{aligned}$$ Hence, it the array to be sorted consists of $\rho$ runs of lengths $r_1,\ldots,r_\rho$, we have $$\begin{aligned}
\mathbf{Pot}_{\mathrm{start}} & \leqslant
{\textstyle\sum_{i=1}^\rho} 2 r_i - r_i \log_2(r_i/c) \\
& \leqslant (2 - \log_2(n/c)) n - {\textstyle\sum_{i=1}^\rho} r_i \log_2(r_i/n) \\
& \leqslant (2 - \log_2(n/c) + {\mathcal{H}}) n.\end{aligned}$$ Moreover, if we let $\hat{\ell} \in \mathbb{Z}$ and $n_\bullet \in [0,1)$ such that $n = 2^{\hat{\ell}}(1+n_\bullet)c$, then
$\displaystyle \mathbf{Pot}_{\mathrm{end}} = 2^{\hat{\ell}} \, \Phi(n_\bullet) \, c - \hat{\ell} n = (\Phi(n_\bullet)/(1+n_\bullet) - \hat{\ell}) \, n.$
Hence, it follows from Lemmas \[lem:merge-cost-equitable\], \[lem:merge-cost-cluster\] and \[lem:merge-cost-final-collapse\] that $$\begin{aligned}
{\mathsf{mc}}& \leqslant \mathbf{Pot}_{\mathrm{start}} - \mathbf{Pot}_{\mathrm{end}} + n \\
& \leqslant
(3 - \log_2(n/c) + {\mathcal{H}}+ \hat{\ell} - \Phi(n_\bullet)/(1+n_\bullet)) \, n \\
& \leqslant
(3 + {\mathcal{H}}- \log_2(1 + n_\bullet) - \Phi(n_\bullet)/(1+n_\bullet)) \, n \\
& \leqslant
({\mathcal{H}}+ 24/5 - \log_2(5)) \, n \hspace{6mm}
\text{thanks to Lemma~\ref{lem:log+phi}.}
\vspace*{-3.5\baselineskip}\end{aligned}$$
[^1]: Although the current implementation of in Python and Java is also used for sorting data streams, this online property is not exploited.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We examine the internal structure of the ground states of a trapped Bose-Einstein condensate in which atoms have three internal hyperfine spins. We determine a set of collective spin states which minimize the interaction energy between condensate atoms. We also examine the internal dynamics of an initially spin polarized condensate. The time scale of spin-mixing is predicted.'
address: |
[$^{1}$Rochester Theory Center for Optical Science and Engineering]{}\
[and Department of Physics and Astronomy,]{} [University of Rochester, Rochester, New York 14627]{}\
[$^2$Laboratory for Laser Energetics, and Department of Physics and Astronomy]{}\
[University of Rochester, Rochester, New York 14627]{}
author:
- |
\
[C.K. Law$^{1}$, [H. Pu]{}$^{2}$ and N.P. Bigelow$^{2}$]{}
title: 'Quantum spins mixing of spinor Bose-Einstein condensates'
---
0.2in
Bose-Einstein condensates (BEC) of atoms with internal degrees of freedom are new forms of macroscopically coherent matter which exhibit rich quantum structures. In the case of BEC with two internal spin states [@Myatt; @Hall], theoretical studies have predicted interesting phenomena such as quantum entanglement of spins [@Cirac], suppression of quantum phase diffusion [@Law] interference effects [@Savage]. Recently, Stamper-Kurn [*et al.*]{} [@Stamper] have realized an optically trapped BEC in which all three hyperfine states in the lowest energy manifold of sodium atoms are involved. Such a three-component condensate raises new questions regarding the more complex ground state structure [@Stenger; @Ho] and internal spin dynamics. One of the key features here is that there are spin exchange interactions which constantly mix different condensate spin components while the system as a whole remains in the ground state. For example two atoms with respective hypefine spins $+1$ and $-1$ interact and become two atoms with hypefine spin 0. Therefore an important problem is to determine how atoms organize their spins in the ground state and how a spin polarized BEC loses its polarization because of spin exchange interactions.
In this paper we approach the questions using an algebraic method found in quantum optics. We identify the fact that the interaction between spin components in a BEC is analogous to 4-wave mixing in nonlinear optics. However, since the trap is like a matter wave cavity, a more appropriate optical analogy is the 4-wave mixing in a high finesse cavity (i.e., a cavity QED system). With the help of the methods developed in a related cavity QED problem [@Puri; @Wu], we are able to study the organization of spins in the condensate ground state. We find that there exists a class of quantum superposition states which minimize the interaction energy. These quantum states are recognized as collective spin states which are characterized by strong correlations among different spin components, and in some cases we find that the number of atoms in individual spin component shows large fluctuations. In this paper we also examine the [*internal dynamics*]{} of the spin-mixing process arising from the nonlinear interactions between condensate atoms [@Goldstein]. For an initially spin polarized BEC, we predict the time scale at which spins become strongly mixed.
To begin we consider a dilute gas of trapped bosonic atoms with hyperfine spin $f=1$. The second quantized Hamiltonian of the system is given by $(\hbar =1)$, $$\begin{aligned}
{\cal H}&=&\sum\limits_\alpha ^{} {\int_{}^{} {d^3x}\hat \Psi _\alpha ^{\dagger}
\left( {-{{\nabla ^2} \over {2M}}+V_T} \right)}\hat \Psi _\alpha \nonumber \\
&+&\sum\limits_{\alpha ,\beta ,\mu ,\nu }^{} \Omega_{\alpha \beta \mu \nu }
{\int_{}^{} {\hat \Psi _\alpha ^{\dagger}
\hat \Psi _\beta ^{\dagger}\hat \Psi _\mu \hat \Psi _\nu d^3x}}\end{aligned}$$ where $\hat \Psi_{\kappa}$ $(\kappa=-1,0,1)$ is the atomic field annihilation operator associated with atoms in the hyperfine spin state $\left| {f=1,m_f=\kappa } \right\rangle $. The summation indices in (1) run through the values $-1,0,1$. The mass of the atom is given by $M$ and the trapping potential $V_T$ is assumed to be the same for all three components. The interactions between atoms are characterized by the coefficients $\Omega_{\alpha \beta \mu \nu }$ which are obtained from the two-body interaction model [@Ho; @Goldstein; @Ohmi; @Walls], $$U(\vec x_1,\vec x_2)=\delta (\vec x_1-\vec x_2)
\sum\limits_{F=0}^2 {g_F\sum\limits_{M_F=-F}^F {\left| {F,M_F} \right\rangle }}
\left\langle {F,M_F} \right|.$$ Here ${\left| {F,M_F} \right\rangle }$ is the total hyperfine spin state formed by two atoms each with spin $f=1$, and $g_F \equiv {{4\pi \hbar ^2a_F} / M}$ with $a_F$ is the $s-$wave scattering length in the $F$ channel. The interaction (2) is based on rather general symmetry assumptions of the system, because it preserves angular momentum and the rotation symmetry in hyperfine spin space [@Ho]. The model also makes use of the $\delta$ potential approach which has been widely used in one-component dilute BEC.
By expanding the total spin state ${\left| {F,M_F} \right\rangle }$ in terms of basis vectors $\left| {f=1,m_f=\alpha } \right\rangle \otimes \left| {f=1,m_f=\beta }
\right\rangle $, we obtain the Hamiltonian in the form ${\cal H}={\cal H}_S+{\cal H}_A$, where $$\begin{aligned}
{\cal H}_S&=&\sum\limits_\alpha ^{} {\int_{}^{} {d^3x}
\hat \Psi _\alpha ^{\dagger}
\left( {-{{\nabla ^2} \over {2M}}+V_T} \right)}\hat \Psi _\alpha \nonumber \\
&+&{\lambda _s \over 2}\sum\limits_{\alpha ,\beta }^{} {\int_{}^{}
{\hat \Psi _\alpha ^{\dagger}\hat \Psi _\beta ^{\dagger}
\hat \Psi _\alpha \hat \Psi _\beta d^3x}}\end{aligned}$$ $$\begin{aligned}
{\cal H}_A&=&{{\lambda _a} \over 2}\int_{}^{}
{\left( {\hat \Psi _1^{\dagger}
\hat \Psi _1^{\dagger}
\hat \Psi _1\hat \Psi _1} \right.}
+\hat \Psi _{-1}^{\dagger}
\hat \Psi _{-1}^{\dagger}\hat \Psi _{-1}\hat \Psi _{-1} \nonumber \\
&+&2\hat \Psi _1^{\dagger}\hat \Psi _0^{\dagger}\hat \Psi _1
\hat \Psi _0+2\hat \Psi _{-1}^{\dagger}
\hat \Psi _0^{\dagger}\hat \Psi _{-1}\hat \Psi _0 \nonumber \\
&-&2\hat \Psi _1^{\dagger}\hat \Psi _{-1}^{\dagger}\hat \Psi _1\hat \Psi _{-1}
+2\hat \Psi _0^{\dagger}\hat \Psi _0^{\dagger}\hat \Psi _1\hat \Psi _{-1} \nonumber\\
&+& \left. {2\hat \Psi _1^{\dagger}
\hat \Psi _1^{\dagger}\hat \Psi _0\hat \Psi _0} \right)\end{aligned}$$ Here $\lambda_s = (g_0 + 2g_2)/3$ and $\lambda_a = (g_2-g_0)/3$ are defined. The Hamiltonian ${\cal H}$ is written as the sum of a symmetric part ${\cal H}_S$ and a non-symmetric part ${\cal H}_A$, where ${\cal H}_S$ remains unchanged for any interchange of the spin component indices.
In this paper we assume that the symmetric interaction ${\cal H}_S$ is strong compared with ${\cal H}_A$. This occurs for atoms whose scattering lengths in different $F$ channels have approximately same values such that $|\lambda_s| \gg |\lambda_a|$. Recent estimations have indicated that sodium and rubidium atoms indeed have such a property. With the symmetric ${\cal H}_S$ being the dominant Hamiltonian, the condensate wavefunctions $\phi_{\kappa}(\vec x)$ $(\kappa =0,\pm 1)$ for each spin component are approximately described by the same wavefunction $\phi (\vec x)$, i.e., $\phi_{\kappa} (\vec x)=\phi (\vec x)$, which is defined by the Gross-Pitaevskii Equation through ${\cal H}_S$ , $$\left( {-{{\nabla ^2} \over {2M}}
+V_T+ \lambda_s N|\phi |^2} \right)\phi =\varepsilon \phi$$ where $\varepsilon$ is the mean field energy or the chemical potential.
Under the condition that atoms in different spin states are described by the same wavefunction, we can approximate field operators at the zero temperature by, $$\hat \Psi _\kappa
\approx \hat a_\kappa \phi (\vec x) \ \ \ \ \ \ \ \kappa=0,\pm 1$$ Here $\hat a_{\kappa}$ is the annihilation operator associated with the condensate mode, and it satisfies the usual commutation relation $\left[ {\hat a_\kappa ,\hat a_\gamma } \right]=0$ and $\left[ {\hat a_\kappa ,\hat a_\gamma ^{\dagger}} \right]
=\delta _{\kappa \gamma }$. Using (5) and (6), ${\cal H}_S$ and ${\cal H}_A$ have leading parts $H_s$ and $H_a$ respectively, $$\begin{aligned}
{\cal H}_S & \approx & \varepsilon \hat N-\lambda _s' \hat N(\hat N-1)
\equiv H_s\\
{\cal H}_A & \approx & \lambda _a'\left( {\hat a_1^{\dagger}\hat a_1^{\dagger}
\hat a_1\hat a_1+\hat a_{-1}^{\dagger}\hat a_{-1}^{\dagger}\hat a_{-1}\hat a_{-1}} \right.
+2\hat a_1^{\dagger}\hat a_0^{\dagger}\hat a_1\hat a_0 \nonumber \\
&+&2\hat a_{-1}^{\dagger}\hat a_0^{\dagger}\hat a_{-1}\hat a_0
-2\hat a_1^{\dagger}\hat a_{-1}^{\dagger}\hat a_1\hat a_{-1}
+2\hat a_0^{\dagger}\hat a_0^{\dagger}\hat a_1\hat a_{-1} \nonumber \\
&+&\left. {2\hat a_1^{\dagger}\hat a_{-1}^{\dagger}\hat a_0\hat a_0}
\right)\equiv H_A\end{aligned}$$ Here $2\lambda _i' \equiv \lambda _i \int_{}^{} {\left| {\phi (\vec x)}
\right|^4d^3 x}$ $(i=s,a)$, and $\hat N \equiv \hat a_1^{\dagger}\hat a_1
+\hat a_0^{\dagger}\hat a_0+\hat a_{-1}^{\dagger}\hat a_{-1}$ is the total number of atoms in the condensate.
Our goal is to find the quantum states that minimize the energy $H_s+H_a$. Since $H_s$ is a function of $\hat N$ only, $H_s$ is a constant operator for a fixed number of atoms. Therefore it is sufficient to look for the ground state of $H_a$. It is quite remarkable that a similar structure of $H_a$ also appeared in nonlinear wave-mixing processes in cavity QED [@Puri]. We follow Refs.[@Puri; @Wu] and identify the algebraic structure of the system. We notice that the operators $\hat L_-\equiv \sqrt 2\left( {\hat a_1^{\dagger}\hat a_0
+\hat a_0^{\dagger}
\hat a_{-1}} \right)$, $\hat L_+\equiv \sqrt 2\left( {\hat a_0^{\dagger}\hat a_1
+\hat a_{-1}^{\dagger}
\hat a_0} \right) $ and $\hat L_z\equiv \left( {\hat a_{-1}^{\dagger}\hat a_{-1}
-\hat a_1^{\dagger}
\hat a_1} \right)$ obey angular momentum commutation relations: $\left[ {\hat L_+,\hat L_-} \right]=2\hat L_z$ and $\left[ {\hat L_z,\hat L_\pm } \right]=\pm \hat L_{\pm}$. In other words, the operators $\hat L_{+}$, $\hat L_{-}$ can be interpreted as raising and lowering operators of a kind of ‘orbital angular momentum’, and $\hat L_z$ is the ‘z-component’ in the standard notations. From the theory of angular momentum, $\hat L^2$ and $\hat L_z$ have a complete set of common eigenvectors $\left| {l,m_l} \right\rangle$ defined by $$\begin{aligned}
&& \hat L^2\left| {l,m_l} \right\rangle=l(l+1)\left| {l,m_l} \right\rangle
\\
&& \hat L_z\left| {l,m_l} \right\rangle=m_l\left| {l,m_l} \right\rangle\end{aligned}$$ where $m_l=0,\pm 1, \pm 2,...,\pm l$. For a given total number of atoms $N$, the allowable values of $l$ are $l=0,2,4,...N$ if $N$ is even, and $l=1,3,5,...,N$ if $N$ is odd.
With the help of the angular momentum operators, $H_a$ takes a very simple form, $$H_a=\lambda _a'\left( {\hat L^2-2\hat N} \right).$$ This is the main result of the paper because the energy spectrum of $H_a$ is now solved. Eq. (11) indicates that $\left| {l,m_l} \right\rangle$ are eigenstates of $H_a$ with the energy $E_{l}^{a}$ $$E_l^{a} = \lambda _a'\left[ {l(l+1)-2N} \right].$$ The lowest energy state of $H_a$ depends on the sign of $\lambda_a'$. In the following we discuss two cases: (I) $\lambda_a' >0 $ and (II) $\lambda_a' < 0 $.
\(I) $\lambda_a' >0 $: In this case $\left| {l=0,m_l=0} \right\rangle $ is the ground state of $H_a$. Using the Fock states $\left| {n_1,n_0,n_{-1}} \right\rangle $ defined by the number operators $\hat n_j \equiv \hat a_{j}^{\dagger} \hat a_{j}$ for the three spin components (i.e., $ \hat n_j \left| {n_1,n_0,n_{-1}} \right\rangle
=n_{j} \left| {n_1,n_0,n_{-1}} \right\rangle$), $\left| {l=0,m_l=0} \right\rangle $ has the form $$\left| {l=0,m_l=0} \right\rangle
=\sum\limits_{k=0}^{[N/ 2]} {A_k\left| {k,N-2k,k} \right\rangle }$$ where the amplitudes $A_k$ obey the recursion relation $$A_k=-\sqrt {{{N-2k+2} \over {N-2k+1}}}A_{k-1}.$$ We see that the state $\left| {l=0,m_l=0} \right\rangle$ is a quantum superposition of a chain of Fock states $\left| {k,N-2k,k} \right\rangle$ in which the numbers of atoms in the spins $1$ and $-1$ are equal. We stress that such a quantum state is a [*collective*]{} spin state which cannot be expressed as product states of individual atoms. The amplitudes $A_k$ are arranged in such a way that the interaction energy $H_a$ is almost completely cancelled. This can be seen from the disappearence of $N^2$ dependence in the energy of $H_a$. It is not difficult to show that for the state (13), the average numbers of atoms in each component are all equal, i.e., $\left\langle {\hat n_0} \right \rangle
= \left\langle {\hat n_1} \right \rangle
=\left\langle {\hat n_{-1}} \right \rangle =N/3$. Since $A_k$ are almost uniformly distributed (see Fig. 1a), there are large fluctuations of particle numbers in individual components although the total particle number $N$ is fixed. More precisely, we find that $\left\langle {\Delta \hat n_0} \right \rangle \approx 2N/ \sqrt{5}$ for $N \gg 1$, i.e., a super-Poisson distribution. Our further calculations indicate that super-Poisson distribution of particle numbers are a common feature for low energy eigenstates of $H_a$ when $\lambda_a' >0 $.
\(II) $\lambda_a' <0 $: In this case $H_a$ has $2N+1$ degenerate ground states given by $\left| {l=N,m_l} \right\rangle $ where $m_l=0,\pm 1, \pm 2,...,\pm N$. The energy (12) of these states is $\lambda _a'N(N-1)$, and the general form of $\left| {l=N,m_l} \right\rangle $ is given by $$\left| {l=N,m_l} \right\rangle
=\sum\limits_k^{} {B_k^{(m_l)}\left| {k,N-2k-m_l,k+m_l}
\right\rangle }.$$ Here the summation index $k$ runs over all physical Fock states $\left| {k,N-2k-m_l,k+m_l} \right\rangle$ (i.e., those with non-negative numbers in each component). The simplest case of (15) is $\left| {l=N,m_l=-N} \right\rangle = \left| {N,0,0} \right\rangle$, and with this we can construct the amplitudes $B_k^{(m_l)}$ by repeatedly applying the raising operator $\hat L_+$. To give an illustration, we plot in Fig. 1b the Fock state amplitudes $B_k^{(m_l)}$ for several $m_l$’s. We see that $B_k^{(m_l)}$ has a narrow distribution which indicates well defined particle numbers in each spin components. It is interesting that all the degenerate states (15) have sub-Poisson number fluctuations in each spin component (see the inset of Fig. 1b). This feature is just the opposite to the previous case $\lambda_a' >0 $.
Finally, let us look at the spin-mixing dynamics of an initially spin-polarized condensate in which all atoms in the condensate are prepared in the spin 0 state at $t=0$, i.e., $
\left| {\psi (0)} \right\rangle =\left| {0,N,0} \right\rangle $. In this case two atoms in the spin 0 state can be converted into one atom in the spin $1$ state and the other in the spin $-1$ state. Using the $H_s+H_a$ as our approximate Hamiltonian, the system at time $t$ is given by, $$\left| {\psi (t)} \right\rangle
=e^{-i\theta_N (t) }\sum\limits_{l=0}^{N}
{C_l}e^{-i\lambda _a'l(l+1)t}\left| {l,m_l=0} \right\rangle$$ where $C_l=\left\langle {{l,m_l=0}}
\mathrel{\left |
{\vphantom {{l,m_l=0} {0,N,0}}} \right.
\kern-\nulldelimiterspace} {{0,N,0}} \right\rangle $ and $\theta_N (t) = (\varepsilon N+\lambda _s'N(N-1))t$. In Fig. 2 we present the time dependence of the particle number in the spin-0 component for $N=10^2,10^3,10^4$ cases. We see that the number of atoms in the spin-0 component becomes steady at $\left\langle {\hat n_0} \right\rangle=N/2$ after a time $t_c$, $$t_c \approx {1 \over {2\lambda_a' \sqrt{N}}}.$$ This is the time scale for the spin-mixing process purely due to the nonlinear interaction between condensate atoms [@remark]. In the Thomas-Fermi (large $N$) limit, we find that for a spherial harmonic trap, $(g_2-g_0)t_c\approx 5.1 N^{1/ 10}
\left( {(g_0+2g_2) / M\omega ^2} \right)^{3/ 5}$ where $\omega$ is trap frequency. Therefore $t_c$ becomes quite insensitive to $N$ in the Thomas-Fermi limit. To give a realistic example, for a sodium condensate with $N= 10^4$ and $\omega =2\pi \times$370Hz, we find that $t_c$ is about $0.4$ seconds, assuming $a_2 \approx 2.6$nm and $a_0 \approx 2.3$nm.
We remark that the spin mixing dynamics can be quite different for different initial conditions. In Fig. 3, we give an example for the case when all three components initially have the same atom numbers, i.e., $\left| {\psi (0)} \right\rangle =\left| {N/3,N/3,N/3} \right\rangle$. It is quite surprising that the particle number executes fast oscillations with a frequency of the order of $\lambda_a' N$, and then the system suddenly becomes steady. This interesting behavior indicates that there are complex quantum dynamics governed by the nonlinear interaction $H_a$. In fact since $H_a$ has a discrete spectrum, quantum recurrence or revival is expected in a much longer time scale (which is typically of the order of $\pi / \lambda_a'$). It is worth further exploring the quantum dynamics in the context of either BEC or cavity QED.
To conclude, we have examined the spin-mixing interaction of a Bose-Einstein condensate with three internal spin components. It is quite remarkable that the model interaction (2) (which is based on general symmetry assumptions) can lead to a simple algebric representation, and from which we can construct the collective spin states which minimize the interaction energy among condensate atoms. These collective states exhibit spin correlations and characteristic particle number fluctuations which depend crucially on the sign of $\lambda_a'$. We have also investigated the spin mixing dynamics due to the nonlinear interaction between condensate atoms. The time scale of mixing for an initially spin polarized system is identified. This study provides a theoretical treatment of the structure and dynamics of spinor BEC. However, our analysis are limited to interaction between condensate atoms, it remains to be answered that how non-condensate atoms will decohere the condensate structure. Further work along this direction would be necessary.
CKL would like to thank Prof. J.H. Eberly for discussions. This research was supported by NSF grants PHY-9415583 and PHY-9457897, and the David and Lucile Packard Foundation.
C.J. Myatt, E.A. Burt, R.W. Ghrist, E.A. Cornell, and C.E. Wieman, Phys. Rev. Lett. [**78**]{}, 586 (1997).
D.S. Hall, M.R. Matthews, J.R. Ensher, C.E. Wieman, E.A. Cornell,cond-mat/9804138
I. Cirac, M. Lewenstein, K. Molmer, and P. Zoller, Phys. Rev. A [**57**]{}, 1208 (1998).
C.K. Law, H. Pu, N.P. Bigelow and J.H. Eberly, Phys. Rev. A [**58**]{} 531 (1998).
C.M. Savage, Janne Ruostekoski and D. F. Walls, Phys. Rev. A [**57**]{}, 3805 (1998).
D.M. Stamper-Kurn, M.R. Andrews, A.P. Chikkatur, H.J. Miesner, J. Stenger, and W. Ketterle, Phys. Rev. Lett. [**80**]{}, 2027 (1998).
J. Stenger, S. Inouye, D.M. Stamper-Kurn, H.J. Miesner, A.P. Chikkatur and W. Ketterle (private communication).
T.L. Ho, Phys. Rev. Lett. (to be published).
Liwei Wang, R.R. Puri and J.H. Eberly, Phys. Rev. A [**46**]{}, 7192 (1992).
Ying Wu, Phys. Rev. A [**54**]{}, 4534 (1996).
A different aspect of spin mixing dynamics in the context of atomic phase conjugation was reported by E.V. Goldstein and P. Meystre, cond-mat/9806165.
T. Ohmi and K. Machida, cond-mat/9803160.
W. Zhang and D.F. Walls, Phys. Rev. A [**57**]{}, 1248 (1998).
Our further calculations indicate that $\left\langle {\Delta \hat n_j } \right\rangle$ also become steady for $t >t_c$. These particle number fluctuations (for $t>t_c$) are found to be quite significant: $\left\langle {\Delta \hat n_1} \right\rangle
= \left\langle {\Delta \hat n_1} \right\rangle
= \left\langle {\Delta \hat n_0} \right\rangle /2
\approx 0.18 N$.
=3.5in
=3.5in
=3.5in
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Davide Pierangeli$^\dag$,$^{1,2}$, Mushegh Rafayelyan$^\dag$,$^{3}$ Claudio Conti$^\ddag$,$^{2,1}$ Sylvain Gigan$^\ddag$,$^{3}$'
title: 'Supplementary Material for *Scalable spin-glass optical simulator* '
---
Supplementary Figures
=====================
![ []{data-label="FigureS1"}](SupplementaryFigure1.pdf){width="1.05\columnwidth"}
![ []{data-label="FigureS2"}](SupplementaryFigure2.pdf){width="1.05\columnwidth"}
![ []{data-label="FigureS3"}](SupplementaryFigure3.pdf){width="1.05\columnwidth"}
[28]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
-
bibliography:
- 'PartialLists.bib'
title: Algebraic Comparison of Partial Lists in Bioinformatics
---
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using the perturbative method for the simulation of charmed particles, the dynamical origin of charm quark elliptic flow is studied in the framework of a multi-phase transport (AMPT) model. Besides the expected ordering relative to that of light quarks according to quark masses, charm quark elliptic flow is seen to be sensitive to the parton scattering cross section. To describe the observed large elliptic flow of electrons from the decay of charmed mesons, a charm quark elastic scattering cross section much larger than that estimated from the perturbative QCD is required.'
address:
- 'Department of Chemistry and Physics, Arkansas State University, P.O. Box 419, State University, AR 72467-0419, USA'
- 'Institute of Theoretical Physics, Shanghai Jiao Tong University, Shanghai 200240, China'
- 'Cyclotron Institute and Physics Department, Texas A&M University, College Station, Texas 77843-3366, USA'
author:
- 'Bin Zhang, Lie-Wen Chen and Che-Ming Ko'
title: 'Charm elliptic flow in Au+Au collisions at RHIC [^1] '
---
INTRODUCTION
============
Because of their large masses, charm quarks are produced very early and propagate through the quark-gluon plasma formed in relativistic heavy ion collisions. Any modifications of charm quark spectrum thus carry information on the properties of the quark-gluon plasma. Although charmed hadrons are at present not directly observable in central nucleus-nucleus collisions at the Relativistic Heavy Ion Collider (RHIC), experimental data on the transverse momentum spectrum of electrons from their decays have already provided useful information on the interaction of charm quarks in the quark-gluon plasma. For example, the transverse momentum spectrum of these electrons is found to be insensitive to the charm final-state interactions as results from both the PYTHIA model and the blastwave hydrodynamic model are consistent with the experimental data [@Batsouli:2002qf]. On the other hand, the large elliptic flow of these electrons is consistent with the prediction of the coalescence model [@Lin:2003jy; @Greco:2003vf] which assumes that charm and light quarks are in thermal equilibrium and have same elliptic flow. In the present talk, we discuss in the framework of the AMPT model [@Zhang:1999bd; @Lin:2000cx; @Lin:2001yd; @Lin:2001zk; @Lin:2003ah; @Lin:2004en] the mechanism for the generation of charm quark elliptic flow and the dependence of its value on the charm scattering cross section in the quark-gluon plasma [@Zhang:2005vq].
THE AMPT MODEL
==============
The AMPT model has four components: initial conditions, parton cascade, hadronization, and hadron cascade. For studying charm elliptic flow, we use the version with string melting, in which hadrons that are generated from the HIJING model [@Wang:1991ht] are converted to partons according to their valence structures with a formation time that is determined by the transverse momentum of the parent hadron, in order to simulate the evolution of the energy stored in initial strings and the effects of particle production from the coherent color field. The space-time evolution of resulting partonic system is modeled by the ZPC model [@Zhang:1997ej], which includes elastic scatterings between partons with a cross section given by the leading pQCD and regulated by a screening mass that is taken as a parameter to adjust the magnitude of the cross section. After the partonic system freezes out, closest quarks and anti-quarks are recombined into hadrons with their subsequent evolution simulated by the ART model [@Li:1995pr]. Using parton cross sections $6$-$10$ mb, the AMPT model with string melting can give a good description of measured low transverse momentum particle spectrum [@Chen:2004vh], elliptic flow [@Lin:2001zk; @Chen:2004vh], higher-order anisotropic flows [@Chen:2004dv], and the pion interferometry [@Lin:2002gc].
SIMULATION OF CHARMED PARTICLES
===============================
Charmed particles are rare particles even at RHIC as only about two pairs are produced in the mid-rapidity region of central Au+Au collisions at available top energies. To simulate charm particles efficiently, we use the perturbative method [@Randrup:1980qd] by introducing an enhancement factor for their production from initial hard scattering, so that each charmed particle carries a probability given by the inverse of corresponding enhancement factor, and neglects the effects due to charmed particle scattering on un-charmed particles. Using the power law parametrization of D meson spectrum measured in d+Au collisions by the STAR collaboration [@Adams:2004fc; @Tai:2004bf; @Ruan:2004if], we first generate the transverse momentum distribution of D mesons between rapidity of -2 and +2 with their distribution in the transverse plane according to the positions of initial nucleon-nucleon collisions. The initial phase-space distribution of charm quarks is then obtained by dissociating D mesons into their valence quarks after a formation time given by inverse of the D meson transverse momentum. The charm scattering cross section with other partons in the quark-gluon plasma is taken to be the same as the cross section for collisions between light quarks. Both cross sections of 3 mb, which is similar to that given by the perturbative QCD, and 10 mb that is needed for describing observables related to light quarks, are used in present study. At the freeze-out of partons, charm quarks are combined with light quarks into D mesons according to the coordinate-space coalescence model used in the AMPT model. For the scattering of charmed mesons with other hadrons, their cross sections are simply taken to be the same as the charm-parton cross section as their effect on the charmed meson spectrum is insignificant.
RESULTS AND DISCUSSIONS
=======================
From their transverse momentum spectra at mid-rapidity, charm quarks are found to approach thermal equilibrium when their scattering cross sections increase. This result is reminiscent of the transition from the charm spectrum obtained from the PYTHIA to that of the blastwave hydrodynamic model. As in Ref.[@Batsouli:2002qf], the spectrum of electrons from the decay of final charmed mesons is not very sensitive to charm final-state interactions and is compatible with experimental data [@Adler:2004ta] for both cross sections.
![Elliptic flows of quarks, D mesons, and electrons from D meson decay in minimum bias Au+Au collisions at $\sqrt{s_{NN}}=200$ GeV from the AMPT model.[]{data-label="fig:v2pt2"}](v2pt2.eps){width="100mm" height="100mm"}
For elliptic flows, results from the AMPT model are shown in Fig. \[fig:v2pt2\]. It is seen from panels (a) and (b) that the magnitude of charm quark elliptic flow increases with increasing parton scattering cross section. For both parton cross sections, there is, however, a strong mass ordering of quark elliptic flows with charm quark elliptic flow saturating to about the same value at a larger transverse momentum compared with that of light quarks. The elliptic flows of D mesons and their decay electrons are shown in panels (c) and (d). Both are seen to follow essentially that of charm quarks as the momentum of a charmed meson is largely given by that of charm quark when light quarks have only bare masses as in the AMPT model. A larger charmed meson elliptic flow of about 10% at $p_t=2$ GeV/c is obtained if an isotropic instead of screened Coulomb cross section is used for charm quark scattering as suggested recently in Ref.[@vanHees:2004gq] that charm quark scattering is dominated by the formation of charmed meson resonances in the quark-gluon plasma. To explain the large charm elliptic flow of more than 10% in the available data [@Adler:2005ab; @Laue:2004tf] thus requires effects beyond the perturbative QCD in the quark-gluon plasma.
We have not included in the present study electrons from the decay of mesons consisting of bottom quark. Their contribution is expected to become important at $p_t>3$ GeV/c. Also, the effect of radiative energy loss on charm quark elliptic flow, which becomes non-negligible as the charm quark transverse momentum increases, is not considered. For charm quarks with high transverse momentum, charm meson production will be mainly from fragmentation instead of recombination. These effects need to be included for a more complete study of heavy quark elliptic flow at high transverse momentum. Our results are, nevertheless, consistent with other recent studies of charm collective flow [@vanHees:2004gq; @Chen:2004cx; @Molnar:2004ph; @Bratkovskaya:2004ec; @Moore:2004tg; @Djordjevic:2005db; @Petreczky:2005nh; @vanHees:2005wb], i.e., charm elliptic flow is sensitive to charm final-state interactions and a large elliptic flow seen in available data requires a charm scattering cross section larger than that given by the perturbative QCD.
[9]{}
S. Batsouli, S. Kelly, M. Gyulassy and J.L. Nagle, Phys. Lett. B 557 (2003) 26.
Z.W. Lin and D. Molnar, Phys. Rev. C 68 (2003) 044901.
V. Greco, C.M. Ko and R. Rapp, Phys. Lett. B 595 (2004) 202.
B. Zhang, C.M. Ko, B.A. Li and Z.W. Lin, Phys. Rev. C 61 (2000) 067901.
Z.W. Lin, S. Pal, C.M. Ko, B.A. Li and B. Zhang, Phys. Rev. C 64 (2001) 011902.
Z.W. Lin, S. Pal, C.M. Ko, B.A. Li and B. Zhang, Nucl. Phys. A 698 (2002) 375.
Z.W. Lin and C.M. Ko, Phys. Rev. C 65 (2002) 034904.
Z.W. Lin and C.M. Ko, Phys. Rev. C 68 (2003) 054904.
Z.W. Lin, C.M. Ko, B.A. Li, B. Zhang and S. Pal, arXiv:nucl-th/0411110.
B. Zhang, L.W. Chen and C.M. Ko, Phys. Rev. C 72 (2005) 024906.
X.N. Wang and M. Gyulassy, Phys. Rev. D 44 (1991) 3501.
B. Zhang, Comput. Phys. Commun. 109 (1998) 193.
B.A. Li and C.M. Ko, Phys. Rev. C 52 (1995) 2037; B.A. Li, A.T. Sustich, B. Zhang, and C.M. Ko, Int. Jour. Phys. E 10, 267-352 (2001).
L.W. Chen, V. Greco, C.M. Ko and P.F. Kolb, Phys. Lett. B 605 (2005) 95.
L.W. Chen, C.M. Ko and Z.W. Lin, Phys. Rev. C 69 (2004) 031901.
Z.W. Lin, C.M. Ko and S. Pal, Phys. Rev. Lett. 89 (2002) 152301.
J. Randrup and C.M. Ko, Nucl. Phys. A 343 (1980) 519.
J. Adams *et al.* \[STAR Collaboration\], Phys. Rev. Lett. 94 (2005) 062301. A. Tai \[STAR Collaboration\], J. Phys. G 30 (2004) S809. L. Ruan \[STAR Collaboration\], J. Phys. G 30 (2004) S1197. S.S. Adler *et al.* \[PHENIX Collaboration\], Phys. Rev. Lett. 94 (2005) 082301. H. van Hees and R. Rapp, Phys. Rev. C 71 (2005) 034907. S.S. Adler \[PHENIX Collaboration\], Phys. Rev. C 72 (2005) 024901. F. Laue \[STAR Collaboration\], J. Phys. G 31 (2005) S27. L.W. Chen and C.M. Ko, J. Phys. G 31 (2005) S49. D. Molnar, J. Phys. G 31 (2005) S421. E.L. Bratkovskaya, W. Cassing, H. Stocker and N. Xu, Phys. Rev. C 71 (2005) 044901. G.D. Moore and D. Teaney, Phys. Rev. C 71 (2005) 064904. M. Djordjevic, M. Gyulassy, R. Vogt and S. Wicks, arXiv:nucl-th/0507019.
P. Petreczky and D. Teaney, arXiv:hep-ph/0507318.
H. van Hees, V. Greco and R. Rapp, arXiv:nucl-th/0508055.
[^1]: Work supported by the U.S. National Science Foundation under Grant No.’s PHY-0140046 (BZ) and PHY-0457265 (CMK), the Welch Foundation under Grant No. A-1358 (CMK), and the National Natural Science Foundation of China under Grant No.’s 10105008 and 10575071(LWC).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a theoretical model for atomic hydrogen ionization by electron impact in the instantaneous approximation and the more accurate non-instantaneous approach using the methods of Quantum Electrodynamics, for the binary coplanar and the coplanar asymmetric geometries. All electrons are described by plane wave functions in the coplanar binary geometry but in the asymmetric geometry the ejected electron is described by a Sommerfeld-Maue wave function. It is shown that the two models give the same results in the non relativistic limit for the binary coplanar geometry where the interactions can be treated as instantaneous. However, this is no longer true for the relativistic case where one has to take into account both the instantaneous interaction and the radiation interaction. These results are obtained in the first order of perturbation theory.'
author:
- 'S. Taj'
- 'B. Manaut'
- 'A. Makhoute'
- 'Y. Attaourti'
title: |
Instantaneous versus non-instantaneous approach to relativistic\
ionization of atomic hydrogen by electron impact
---
Introduction
============
The interaction of electrons with atoms is the field that most deeply probes both the structure and reaction dynamics of a many-body system [@1]. Electron-atom collisions that ionize the target provide a large and interesting diversity of phenomena. The reason for this is that a three-body final state allows a wide range of kinematic regions to be investigated. These different kinetic regions depend sensitively on different aspects of the description of the collision. Up to now, there has been no calculation of differential cross-sections by a method that is generally valid. The understanding of the ionization by electron impact has advanced by an iterative process involving experiments and calculations that emphasize different aspects of the reaction. Kinematic regions have been found that are completely understood in the sense that absolute differential cross-sections in detailed agreement with experiment can be calculated.\
These form the basis of a structure probe, electron momentum spectroscopy, that is extremely sensitive to one-electron and electron-correlation properties of the target ground states of the residual ion in the case of heavy atoms. Other kinematic regions require a complete description of the collision which may be facilitated by including the boundary condition for the three charged particles in the final states. This point is not trivial at all because there is no separation distance at which the Coulomb forces in the three-body system are strictly negligible. The pioneering experiments of Ehrhardt et al [@2] are of this type. Electron-impact ionization has been studied experimentally using relativistic electrons by Dangerfield and Spicer [@3], Hoffman et all (1979) [@4] and Anholt (1979) [@5]. The measurements with relativistic electrons have all been of total cross-sections typically for the $K$ and $L$ shells of heavy atoms. Theoretical models for total ionization cross-sections have been developed by a number of authors including Scoffield (1978) [@6] and Moiseiwitsh and Stockmann (1980) [@7]. This field has been reviewed by Moiseiwitsh (1980) [@8]. Fuss, Mitroy and Spicer (1982) [@9] have developed a theoretical model for the binary (e,2e) reaction. This (e,2e) reaction is that were the outgoing electrons have equal energy. The theoretical model they have developed uses the impulse and relativistic plane wave Born approximations (RPWBA). Nakel and Whelan (1999) [@10] have reviewed the experimental and theoretical developments in the study of relativistic (e,2e). They argued that the fully relativistic distorted wave approximation (RDWBA) proposed by Walters et al [@11] was the simple possible approximation that allows to gain understanding of the relativistic (e,2e) processes.\
As for electron impact ionization for a hydrogen atomic target, the first work that relied on the model developed by Fuss et al [@9], was presented by Attaourti et al [@12] who studied the importance of the relativistic electronic dressing during the process of laser-assisted ionization of atomic hydrogen by electron impact. Taking a zero laser electric field, one recovers the binary (e,2e) process in the absence of the laser field in the instantaneous approximation where a direct non relativistic interaction potential was used. Later on, Attaourti et al (2005) [@13] developed a simple semi-relativistic model using a Sommerfeld-Maue wave function to describe the ejected electron and the Darwin semi-relativistic wave function to describe the hydrogen atomic target in its ground state [@14] also in the instantaneous interaction. This simple model allows to investigate both the relativistic binary (e,2e) reaction process (RPWBA) and the semi-relativistic (e,2e) process in the coplanar asymmetric geometry (SRCBA).\
The purpose of the present work is to go beyond the instantaneous approximation and use the propagator approach of QED to include both the instantaneous and radiation contributions necessary to describe more accurately the (e,2e) processes in the relativistic domain. Indeed, it is well known that non relativistically, interactions are instantaneous but this is no longer true in the relativistic case. This important feature of relativistic collision processes is explicitly contained in the formalism of QED. This important point is clearly explained for example by F. Gross (1999) [@15], W. Greiner (1996) [@16] and many others. To what extent the contribution of both the instantaneous and radiation parts to these (e,2e) processes affect and modify the non polarized differential cross section is the question we want to address and study.\
The organization of this paper is as follows : in section 2, we review the RPWBA formalism in the instantaneous approximation and give the theoretical results for the same model within the framework of QED. In section 3, we review the SRCBA formalism in the instantaneous approximation and give theoretical results using full QED calculations. In section 4, we discuss some relevant results and in section 5, we end by a conclusion. Throughout this work, we use atomic units and the metric tensor $g_{\mu\nu}=diag(1,-1,-1,-1)$.
Binary coplanar geometry
========================
Theoretical model within the frame work of the instantaneous approximation
---------------------------------------------------------------------------
The transition matrix element for the direct channel (we do not consider exchange effects) is given by $$\begin{aligned}
S_{fi}=-i\int dt<\psi _{p_{f}}(x_{1})\phi _{f}(x_{2})\mid
V_{d}\mid \psi _{p_{i}}(x_{1})\phi _{i}(x_{2})>\label{1}\end{aligned}$$ where $V_{d}=1/r_{12}-1/r_{1}$ is the direct interaction potential ($t_{1}=t_{2}=t\Longrightarrow x_{1}^{0}=x_{2}^{0}=x^{0}$) and in the RPWBA, $\psi _{p_{f}}(x_{1})$ is the wave function describing the scattered electron
$$\begin{aligned}
\psi
_{p_{f}}(x_{1})=\frac{u(p_{f},s_{f})}{\sqrt{2E_{f}V}}e^{-ip_{f}.x_{1}}\label{2}\end{aligned}$$
given by a free Dirac solution normalized to the volume $V$. For the incident electron, we use $$\begin{aligned}
\psi
_{p_{i}}(x_{1})=\frac{u(p_{i},s_{i})}{\sqrt{2E_{i}V}}e^{-ip_{i}.x_{1}}\label{3}\end{aligned}$$ For the atomic target, $\phi _{i}(x_{2})=\phi
_{i}(t,\mathbf{r}_{2})$ is the relativistic wave function of atomic hydrogen in its ground state. For the ejected electron, we use again a free Dirac solution normalized to the volume $V$ and $\phi _{f}(x_{2})$ is given by : $$\begin{aligned}
\phi _{f}(x_{2})=\psi
_{p_{B}}(x_{2})=\frac{u(p_{B},s_{B})}{\sqrt{2E_{B}V}}
e^{-ip_{B}.x_{2}}\label{4}\end{aligned}$$
The unpolarized triple differential cross section (TDCS) [@16] is then given by : $$\begin{aligned}
\frac{d\overline{\sigma }}{dE_{B}d\Omega _{B}d\Omega _{f}}
&=&\frac{|\mathbf{p}_f||\mathbf{p}_B|}{|\mathbf{p}_i| c^{4}}
\frac{(2E_{i}E_{f}/c^{2}-p_{i}.p_{f}+c^{2})}{\mid
\mathbf{p}_{f}\mathbf{-p}
_{i}\mid ^{4}} \nonumber\\
&&\times 4E_B\Big| \Phi _{1,1/2,1/2}(\mathbf{q}=\mathbf{\Delta
-p}_{B})\nonumber\\&&-\Phi
_{1,1/2,1/2}(\mathbf{q}=\mathbf{-p}_{B})\Big| ^{2}\label{5}\end{aligned}$$
The functions $\Phi _{1,1/2,1/2}(\mathbf{q})$ are the Fourier transforms of the relativistic atomic hydrogen wave functions [@12], and the quantity $\mathbf{\Delta
=p}_{i}-\mathbf{p}_{f}\ $ is the momentum transfer.
Theoretical model with the inclusion of the photon propagator
-------------------------------------------------------------
The transition matrix element for the direct channel in the non-instantaneous approximation is given by : $$S_{fi}=-i\int_{-\infty }^{+\infty }dt<\phi _{f}(x)|{A\!\!\!/}(x)| \phi
_{i}(x)>\label{8}$$ where ${A\!\!\!/}(x)=\gamma_{\mu}A^{\mu}=
A_{0}(x)\gamma_{0}-\mathbf{A(x).\gamma}$. Contravariant four vectors are written $x^\mu=(t,\mathbf{x})$. $\phi _{f}(x)=\psi
_{p_b}(x)$ is defined in Eq. (\[4\]), and $\phi
_{i}(t,\mathbf{x})=\psi _{t}(x)$ is the relativistic wave function of atomic hydrogen in its ground state. The electromagnetic potential $A^{\mu}$ due to the scattered electron is given by $$A^{\mu}(x)=-4\pi\int G(x-y)J^{\mu}(y)d^4y\label{9}$$ where $J^{\mu}(y)$ is the four-vector current for the electron $$\begin{aligned}
J^{\mu}(y)=\overline{\psi}_{p_f}(y)\gamma^{\mu}\psi_{p_i}(y)\label{10},\qquad
\overline{\psi}_{p_f}(y)=\psi_{p_f}^{\dagger}(y)\gamma^0\end{aligned}$$ where $\psi _{p_i}(y)$ and $\psi _{p_f}(y)$ are defined respectively in Eq. (\[3\]) and Eq. (\[4\]). Using the Fourier decomposition of the Green’s function $$G(x-y)=-\frac{1}{(2\pi)^4}\int
\frac{exp{[-i\kappa(x-y)]}}{\kappa^2+i\varepsilon}d^4\kappa\label{11}$$ the scattering matrix element becomes $$\begin{aligned}
S_{fi}&=&\int \overline{\psi}_{p_B}(x)\gamma_{\mu}\psi_t(x)\int
\int
\frac{exp{[-i\kappa(x-y)]}}{\kappa^2+i\varepsilon}\frac{d^4\kappa}{(2\pi)^4}\nonumber\\
&\times&
\overline{\psi}_{p_f}(y)\gamma^{\mu}\psi_{p_i}(y)d^4yd^4x\label{12}\end{aligned}$$ The $y$-integration can be performed easily yielding $$\int
d^4y\exp[i(\kappa+p_f-p_i)y]=(2\pi)^4\delta[\kappa-(p_i-p_f)]\label{13}$$ Now the $\kappa$-integration is done, and the $S_{fi}$ reads with the usual normalization $$\begin{aligned}
S_{fi}&=&
\frac{\overline{u}(p_f,s_f)\gamma^{\mu}u(p_i,s_i)}{\sqrt{4E_fE_iV^2}(p_i-p_f)^2}
\frac{\overline{u}(p_B,s_B)\gamma_{\mu}}{\sqrt{2E_BV}}(2\pi)^{3/2}\nonumber\\
&&\times\Phi_t(\mathbf{q})
(2\pi)\delta(E_i-E_f-E_B+|\varepsilon_b|)\label{14}\end{aligned}$$ We note $$\mathcal{F}_{\mu}=\overline{u}(p_B,s_B)\gamma_{\mu}$$ Using the standard procedures of QED [@16]\], one obtains for the spin unpolarized triple differential cross section (TDCS) $$\frac{d\overline{\sigma}}{dE_B d\Omega_f
d\Omega_B}=\frac{|\mathbf{p}_f||\mathbf{p}_B|}{|\mathbf{p}_i|
c^6}\frac{1}{2(p_i-p_f)^4}\frac{1}{2}\sum_{s_is_f}\left|\mathcal{M}_{fi}\right|^2\label{15}$$ with $$\begin{aligned}
\frac{1}{2}\sum_{s_is_f}\sum_{s_{B}}\left|\mathcal{M}_{fi}\right|^2&=&\frac{1}{2}\sum_{s_is_f}\left|\mathcal{F}_{\mu}\overline{u}(p_f,s_f)\gamma^{\mu}u(p_i,s_i)\right|^2\nonumber\\
&=&2c^2\Big\{(\mathcal{F}.p_f)(\mathcal{F}^*.p_i)+(\mathcal{F}.p_i)(\mathcal{F}^*.pf)\nonumber\\
&&-\mathcal{F}\mathcal{F}^*((p_i.p_f)-c^2)\Big\}.\label{16}
\end{aligned}$$ Using the spinor normalization condition $u^+(p,s)u(p,s')=2E\delta_{ss'}$, we find for the TDCS $$\begin{aligned}
&&\frac{d\overline{\sigma}}{dE_B d\Omega_f
d\Omega_B}=\frac{|\mathbf{p}_f||\mathbf{p}_B|}{c^4|\mathbf{p}_i|
}\frac{8}{(p_i-p_f)^4}\nonumber\\
&&\times\left[E_f(p_B.p_i)+E_i(p_B.p_f)+E_B(2(p_i.p_f)-c^2)\right]\nonumber\\
&&\times|\Phi_{1,1/2,1/2}(\mathbf{q=\Delta-p_B})|^2.\label{17}
\end{aligned}$$ Note that $$\begin{aligned}
\left.\begin{array}{c}
(p_i-p_f)=(E_i-E_f)/c-(\mathbf{p_i}-\mathbf{p_f}) \\
(p_B.p_i)=E_BE_i/c^2-\mathbf{p_B}\mathbf{p_i}\\
(p_B.p_f)=E_BE_f/c^2-\mathbf{p_B}\mathbf{p_f}\\
(p_i.p_f)=E_iE_f/c^2-\mathbf{p_i}\mathbf{p_f}
\end{array}\right.\label{18}
\end{aligned}$$
Coplanar asymmetric geometry
============================
The SRCBA theoretical model in the instantaneous approximation
--------------------------------------------------------------
In this section, the same computation is done, the main difference lying in the description of the ejected electron where now, a Sommerfeld-Maue wave function accurate to the order $Z/c$ in the relativistic corrections, is used. We have $\phi
_{f}(t,\mathbf{x})=\exp (-iE_{B}t)\psi _{p_{B}}^{(-)}(\mathbf{x})$ and $\psi _{p_{B}}^{(-)}(\mathbf{x})$ is given by : $$\begin{aligned}
&&\psi _{p_{B}}^{(-)}(\mathbf{x}) =\exp (\pi \eta _{B}/2)\Gamma
(1+i\eta _{B}) \exp
(i\mathbf{p}_{B}.\mathbf{x})\nonumber \\
&\times&\big\{\mathsf{1}_{4}-\frac{ic}{2E_{B}}\mathbf{\alpha} .\mathbf{\nabla }_{(2)}\big\}\, _{1}F_{1}(-i\eta
_{B},1,-i(p_{B}x+\mathbf{p}_{B}.\mathbf{x}))\nonumber \\
&\times&\frac{u(p_{B},s_{B})}{\sqrt{2E_{B}V}} \label{}\end{aligned}$$ normalized to the volume $V$. The Sommerfeld parameter is given by $$\eta _{B}=\frac{E_{B}}{c^{2}p_{B}} \label{}$$ The transition matrix element for the direct channel is given by $$\begin{aligned}
S_{fi} &=&-i\int d\mathbf{x}\frac{\overline{u}(p_{f},s_{f})}{\sqrt{2E_{f}V}}\gamma
_{(1)}^{0}\frac{\overline{u}(p_{B},s_{B})}{\sqrt{2E_{B}V}}\gamma
_{(2)}^{0}\Big \{\,_{1}F_{1}(i\eta _{B},\nonumber\\
&&1,i(p_{B}x+\mathbf{p}_{B}.\mathbf{x}) \mathsf{1}_{4}
-\frac{i}{2cp_{B}}(\mathbf{\alpha
.p}_{B}+p_{B}\mathbf{\alpha .}\widehat{ \mathbf{x}})\nonumber\\
&\times&_{1}F_{1}(i\eta
_{B}+1,2,i(p_{B}x+\mathbf{p}_{B}.\mathbf{x}))\Big\} \varphi
^{(\pm )}(\mathbf{x})
\exp (-i\mathbf{p}_{B}.\mathbf{x})\nonumber\\ &\times&[\exp (i\mathbf{\Delta .x})-1] \frac{8\pi^{2} }{\Delta ^{2}} \delta
(E_{f}+E_{B}-E_{i}-\varepsilon _{b})
\nonumber\\&\times&\frac{u(p_{i},s_{i})}{\sqrt{2E_{i}V}}\exp (\pi
\eta _{B}/2)\Gamma
(1-i\eta _{B}) \label{}$$ The spin unpolarized TDCS is given by $$\begin{aligned}
\frac{d\overline{\sigma} }{dE_{B}d\Omega _{B}d\Omega _{f}}&=&\frac{1}{64c^{6}\pi ^{3}}\frac{|\mathbf{p}_{f}||\mathbf{p}_{B}|}{|\mathbf{p}_{i}|}\frac{\exp
(\pi \eta _{B})}{\Delta ^{4}}\left| \Gamma (1-i\eta _{B})\right|
^{2}\nonumber\\&&
\left| \widetilde{S}_{fi}^{(1)}+\widetilde{S}_{fi}^{(2),1}+\widetilde{S}_{fi}^{(2),2}\right| ^{2} \label{27}\end{aligned}$$ for the description of the quantities $\widetilde{S}_{fi}^{(1)}$, $\widetilde{S }_{fi}^{(2),1}$, $\widetilde{S}_{fi}^{(2),2}$ and for more details see [@13].
The SRCBA theoretical model with the inclusion of the photon propagator
-----------------------------------------------------------------------
The transition matrix element $S_{fi}$ for direct channel in the non-instantaneous approximation is given by $$S_{fi}=-i\int_{-\infty }^{+\infty }dt<\phi _{f}(x)|{A\!\!\!/}(x)| \phi
_{i}(x)>\label{}$$ We replace all wave functions and ${A\!\!\!/}(x)$ in $S_{fi}$ and we get : $$\begin{aligned}
S_{fi}&=&-i\frac{4\pi2\pi\delta(E_i-E_f-E_B+|\varepsilon_b|)}{\sqrt{8V^3E_iE_fE_B}(p_i-p_f)^2}
\nonumber\\&&
\Big(\overline{u}(p_{B},s_{B})\gamma_\mu\overline{u}(p_{f},s_{f})\gamma^\mu
u(p_i,s_i)
H_1(\mathbf{q})\nonumber\\&&+\overline{u}(p_{B},s_{B})\gamma_\mu\gamma^0(\gamma^0\frac{E_B}{c}-{p\!\!\!/}_{B})
\overline{u}(p_{f},s_{f})\gamma^\mu
\nonumber\\&&u(p_i,s_i)H_2(\mathbf{q})+\overline{u}(p_{B},s_{B})\gamma_\mu\overline{u}(p_{f},s_{f})\gamma^\mu
\nonumber\\&&u(p_i,s_i)\widetilde{H}_2(\mathbf{q})\Big)\end{aligned}$$ The quantities $H_1(\mathbf{q=\Delta-p_B})$, $H_2(\mathbf{q=\Delta-p_B})$, $\widetilde{H}_2(\mathbf{q=\Delta-p_B})$ are given in [@8]. This transition matrix element contains three terms, denoted respectively by : $S_{fi}^{(1)}$, $S_{fi}^{(2,1)}$, $S_{fi}^{(2,2)}$. $$\begin{aligned}
S_{fi}^{(1)}=-i[\overline{u}(p_{B},s_{B})\gamma_\mu][\overline{u}(p_{f},s_{f})\gamma^\mu
u(p_i,s_i)]H_1(\mathbf{q)}\end{aligned}$$ $$\begin{aligned}
S_{fi}^{(2,1)}&=&-i[\overline{u}(p_{B},s_{B})\gamma_\mu\gamma^0(\gamma^0\frac{E_B}{c}-{p\!\!\!/}_{B})]
\nonumber\\ &&[\overline{u}(p_{f},s_{f})\gamma^\mu u(p_i,s_i)]
H_2(\mathbf{q)}\end{aligned}$$ $$\begin{aligned}
S_{fi}^{(2,2)}=-i[\overline{u}(p_{B},s_{B})\gamma_\mu][\overline{u}(p_{f},s_{f})\gamma^\mu
u(p_i,s_i)] \widetilde{H}_2(\mathbf{q)}\end{aligned}$$ We note $$\begin{aligned}
S_{fi}^{(1)}=-i\mathcal{A}(s_B,s_f,s_i)H_1(\mathbf{q)}\end{aligned}$$ $$\begin{aligned}
S_{fi}^{(2,1)}=-i\mathcal{B}(s_B,s_f,s_i)H_2(\mathbf{q)}\end{aligned}$$ $$\begin{aligned}
S_{fi}^{(2,2)}=-i\mathcal{C}(s_B,s_f,s_i)\widetilde{H}_2(\mathbf{q)}\end{aligned}$$ We calculate the square of $S_{fi}$. We have 9 terms :\
$\left| S _{fi}^{(1)}\right| ^{2}$,$ \left|
S_{fi}^{(2,2)}\right| ^{2}$, $\left| {S}_{fi}^{(2,1)}\right|
^{2}$, and $S _{fi}^{(1)\dagger }{S}_{fi}^{(2,1)}$, $S
_{fi}^{(2,1)\dagger }{S}_{fi}^{(1)}$, $S _{fi}^{(2,2)\dagger
}S_{fi}^{(2,1)}$, $S _{fi}^{(2,1)\dagger } S_{fi}^{(2,2)}$, $S
_{fi}^{(2,2)\dagger } {S}_{fi}^{(1)}$, as well as $S
_{fi}^{(1)\dagger } S_{fi}^{(2,2)}$ .\
The different sums over spin states are given by : $$\begin{aligned}
&&\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}\mathcal{A}^\dagger(s_B,s_f,s_i)
\mathcal{C}(s_B,s_f,s_i)\nonumber\\
&&=\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}\mathcal{C}^\dagger(s_B,s_f,s_i)
\mathcal{A}(s_B,s_f,s_i)\nonumber\\
&&=\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}|\mathcal{A}(s_B,s_f,s_i)|^2\nonumber\\
&&=\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}|\mathcal{C}(s_B,s_f,s_i)|^2 \nonumber\\
&&=8c^2\Big[E_f(p_B.p_i)+E_i(p_B.p_f)-E_Bc^2\Big]\end{aligned}$$ $$\begin{aligned}
&&\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}|\mathcal{B}(s_B,s_f,s_i)|^2\nonumber\\
&=&8(E_B^2-c^4)\Big[E_f(p_B.p_i)+E_i(p_B.p_f)-E_Bc^2\Big]\end{aligned}$$ $$\begin{aligned}
&&\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}\mathcal{A}^\dagger(s_B,s_f,s_i)
\mathcal{B}(s_B,s_f,s_i)\nonumber\\&&=
\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}\mathcal{C}^\dagger(s_B,s_f,s_i)
\mathcal{B}(s_B,s_f,s_i)\nonumber\\&&=
\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}\mathcal{B}^\dagger(s_B,s_f,s_i)
\mathcal{C}(s_B,s_f,s_i)\nonumber\\&&=
\frac{1}{2}\sum_{s_f,s_i}\sum_{s_{B}}\mathcal{B}^\dagger(s_B,s_f,s_i)
\mathcal{A}(s_B,s_f,s_i)\nonumber\\&&=2c^2\Big\{8c(p_f.p_B)(p_B.p_i)-
4c(p_f.p_i)p_B^2-\nonumber\\ &&
\frac{E_B}{c}[4E_f(p_B.p_i)+4E_i(p_B.p_f)-4E_B(p_f.p_i)]\nonumber\\&&-
\frac{4}{c}[(E_B^2-c^2p_B^2)(p_i.p_f-c^2)]\Big\}\end{aligned}$$
The spin unpolarized TDCS in the non-instantaneous approximation is given by $$\begin{aligned}
\frac{d\overline{\sigma }}{dE_{B}d\Omega _{B}d\Omega
_{f}}&=&\frac{1}{2^4c^{6}\pi ^{3}}
\frac{|\mathbf{p}_{f}||\mathbf{p}_{B}|}{|\mathbf{p}_{i}|}\frac{\exp
(\pi \eta _{B})}{(p_i-p_f) ^{4}}\left| \Gamma (1-i\eta
_{B})\right| ^{2}\nonumber\\&& \left|
S_{fi}^{(1)}+S_{fi}^{(2,1)}+S_{fi}^{(2,2)}\right| ^{2} \label{}\end{aligned}$$
Results and Discussion
======================
Before beginning the discussion of the results obtained, it is worthwhile to recall the meaning of some abbreviations that will appear throughout this section. The NRPWBA stands for the non relativistic plane wave Born approximation where non relativistic plane waves are used to describe the incident, scattered and ejected electrons. This approximation is valid only in the non relativistic regime and particularly in the case of the coplanar symmetric geometry. The NRCBA approximation stands for the non relativistic Coulomb Born approximation which is valid only in the non relativistic regime and in the case of the asymmetric coplanar geometry of Ehrhardt [@2]. The RPWBANP stands for the relativistic plane wave Born approximation without propagator, which is valid for the case of a coplanar binary geometry and uses Dirac free plane wave solutions as well as a direct interaction potential. Its results take only into account the instantaneous part of the interaction potential. The RPWBAWP stands for the relativistic plane wave Born approximation with the inclusion of the photon propagator.It is valid for the case of a coplanar binary geometry, uses Dirac free plane wave solutions and its results take into account the instantaneous part of the interaction as well as the radiation part of the interaction. The SRCBANP stands for the semi relativistic Coulomb Born approximation which is valid both for the coplanar asymmetric geometry and the binary coplanar geometry. In this approximation, the incident and scattered electrons are described by free Dirac plane waves while the ejected electron is described by a Sommerfeld-Maue wave function and the atomic hydrogen target is described by a Darwin wave function [@14]. However, its uses a direct interaction potential in the instantaneous approximation. Finally, the SRCBAWP stands for the semi relativistic Coulomb Born approximation with the full inclusion of the photon propagator. It is similar to the previous approximation except that it includes both the instantaneous and radiation parts of the interaction.\
Due to this number of approximations that may be confusing, we will focus in our discussion on the most important features of our model and restrain ourselves to consider in the non relativistic regime, the comparison of all the TDCSs in the case of the binary coplanar geometry. In the case of the coplanar asymmetric geometry, we will compare the three relevant TDCSs, the SRCBAWP, the SRCBANP and the NRCBA.
The non relativistic regime
---------------------------
### The binary coplanar geometry
![\[fig1.eps\] TDCSs scaled in $10^{-3}$ for $E_i=2700$ $eV$ and $E_B=1349.5$ $eV$, and the angles $\theta_f=45^{\circ}$, $\theta_i=\phi_i=\phi_f=0^{\circ}$. The angle $\phi_B$ is such that $\phi_B=180^{\circ}$. The curves of the six approaches overlap.](fig1){width="3" height="4.5"}
We begin our discussion by the kinematics of the process. For the binary coplanar geometry, we choose the following angular situation where $p_i$ is along the $Oz$ axis ($\theta_i=\phi_i=0^{\circ}$). For the scattered electron, we choose ($\theta_f=45^{\circ}, \phi_f=0^{\circ}$) and for the ejected electron we choose $\phi_B=180^{\circ}$) and the angle $\theta_B$ varies from $30^{\circ}$ to $60^{\circ}$. The energies of the incident and ejected electrons are respectively $E_i=2700$ $eV$ and $E_B=1349.5$ $eV$. It is expected that for such a regime and choice of geometry, relativistic effects will be small and it is indeed the case as this can be seen in Fig.1 where the six approaches give nearly the same results. Even if this is not apparent on this figure, we have plotted the results of six approaches and obtained close results. In the binary coplanar geometry, the TDCSs are well peaked around the specific angle $\theta_f$ chosen and decrease very rapidly to zero within a small angular spread (typically ten degrees). The energy of the incident electron being non relativistic (it corresponds to a relativistic parameter $\gamma_i=(1-(\beta_i)^2)^{(-1/2)}$), the description of the interaction potential by a direct non relativistic potential or the use of the photon propagator is not important in that case.
### The coplanar asymmetric geometry
In this geometry, the use of any approach involving plane waves will lead to irrelevant results since the Coulomb description of the ejected electron is needed. The angular choice is as follows : $p_i$ is along the $Oz$ axis and $\theta_i=0^{\circ},
\phi_i=0^{\circ}$. For the scattered electron, we choose ($\theta_f=3^{\circ},\phi_f=0^{\circ}$), for the ejected electron we choose $\phi_B=180^{\circ}$ and the angle $\theta_B$ varies from $-180^{\circ}$ to $180^{\circ}$.
![\[fig2.eps\] TDCS for $E_i=2700$ $eV$ and $E_B=54$ $eV$, and the angles $\theta_f=3^{\circ}$, $\theta_i=\phi_i=\phi_f=0^{\circ}$. The angle $\phi_B$ is such that $\phi_B=180^{\circ}$. The curves overlap.](fig2){width="3" height="4.5"}
The energy of the incident electron is $E_i=2700$ $eV$ and the energy of the ejected electron is $E_B=54.$ $eV$. The range of the angular variable $\theta_B$ is due to the fact that in the coplanar asymmetric geometry, one expects to see two peaks, a binary peak which is, in our case, located in the vicinity of $\theta_B=77^{\circ}$ and a recoil peak which is located in the vicinity of $\theta_B=-105^{\circ}$. As a validity check of our calculations, we have also reproduced the results of Byron and Joachain [@17] for the energies $E_i=250$ $eV$ and $E_B=5.$ $eV$. The three approaches : NRCBA, SRCBAWP and SRCBANP are shown in Fig. 2 where the spread of the three TDCSs covers the whole range that the angular variable $\theta_B$ can take and small differences begin to appear particularly at the recoil peak. However and as might be expected for such energies, the SRCBAWP and the SRCBANP remain very close because for such a regime the use of the instantaneous or the non instantaneous approximation amounts to nearly the same. Radiation effects are not important enough for such non relativistic energies.
The relativistic regime
-----------------------
### The binary coplanar geometry
For this regime, the energy of the incident electron is $E_i=511002$ $eV$ and the energy of the ejected electron is $E_B=225501$ $eV$. This value of the energy $E_i$ corresponds to a relativistic parameter $\gamma_i=2$. We choose the following angular situation where $p_i$ is along the $Oz$ axis ($\theta_i=\phi_i=0^{\circ}$).
![\[fig3.eps\] TDCS scaled in $10^{-19}$ for $E_i=510999$ $eV$ and $E_B=255499.5$ $eV$, and the angles $\theta_f=55^{\circ}$, $\theta_i=\phi_i=\phi_f=0^{\circ}$. The angle $\phi_B$ is such that $\phi_B=180^{\circ}$.](fig3){width="3" height="4.5"}
For the scattered electron, we choose ($\theta_f=55^{\circ}$, $\phi_f=0^{\circ}$) and for the ejected electron we choose $\phi_B=180^{\circ}$ and the angle $\theta_B$ varies from $30^{\circ}$ to $60^{\circ}$. Relativistic effects can no longer be neglected and also the use of the instantaneous approximation becomes itself questionable. Let us explain what is contained in Fig 3. The first important point that has to be mentioned is the following : in such a regime, the use of non relativistic approaches is not physically founded. So, we will focus mainly on the four relativistic approximations aforementioned. What appears clearly is that the four relativistic models give different results with the importance of radiation effects clearly shown. To summarize Fig 3, we can say that as regards to the SRCBAWP and SRCBANP, the instantaneous approximations is no longer valid. We have a situation where the TDCS(SRCBAWP) is higher than the TDCS(SRCBANP) and also where the TDCS(RPWBAWP) is higher than the TDCS(RPWBANP). This is a general rule and we have made many simulations to assert that it is a valid rule in the relativistic regime. However, there is no general rule as to the comparison between the TDCS(SRCBAWP) and the TDCS(RPWBAWP) as well as the comparison between the TDCS(RPWBANP) and the TDCS(SRCBANP) because another choice of ($\theta_f$, e.g $\theta_f=39^{\circ}$ will give rise to a reverse situation where $TDCS(RPWBAWP)>TDCS(SRCBAWP)$ whereas $TDCS(RPWBANP)>TDCS(SRCBAWP)$.
![\[fig3.eps\] TDCS scaled in $10^{-13}$ for the energies $E_i=511002$ $eV$ and $E_B=225501$ $eV$, and the angles $\theta_f=39^{\circ}$, $\theta_i=\phi_i=\phi_f=0^{\circ}$. The angle $\phi_B$ is such that $\phi_B=180^{\circ}$.](fig4){width="3" height="4.5"}
In Fig. 4, we only compare the two SRCBAWP and SRCBANP to assess the importance of radiation effects and as this behavior never changes, one can easily see that the use of the instantaneous approximation underestimates the value of the TDCS using the photon propagator by a factor two. The same behavior is observed with the RPWBAWP and the RPWBANP.
### The coplanar asymmetric geometry
In this geometry, the angular choice is as follows : $p_i$ is along the $Oz$ axis and $\theta_i=3^{\circ}, \phi_i=0^{\circ}$. For the scattered electron, we choose ($\theta_f=3^{\circ},
\phi_f=0^{\circ}$), for the ejected electron we choose $\phi_B=180^{\circ}$ and the angle $\theta_B$ varies from $0^{\circ}$ to $180^{\circ}$. The energy of the incident electron is $E_i=511002$ $eV$ and the energy of the ejected electron is $E_B=10220.04$ $eV$.
![\[fig3.eps\] TDCS scaled in $10^{-12}$ for $E_i=511002$ $eV$ and $E_B=10220.04$ $eV$, and the angles $\theta_f=3^{\circ}$, $\theta_i=\phi_i=\phi_f=0^{\circ}$.The angle $\phi_B$ is such that $\phi_B=180^{\circ}$.](fig5){width="3" height="4.5"}
In the relativistic regime, there is no occurrence of a recoil peak and the spread of the angular variable $\theta_B$ of the ejected electron is reduced. As expected, radiation effects are indeed important and in Fig. 5, we see that the binary peak is located in the vicinity of $\theta_B=75^{\circ}$ whereas the TDCS(SRCBANP) underestimates the TDCS(SRCBAWP) by a factor 2. The comparison with the two other approaches using relativistic plane waves is not relevant for this geometry.
Conclusion
==========
In this work, we have studied the two theoretical models (instantaneous and non-instantaneous approximations) for the relativistic ionization of atomic hydrogen by electron impact using non relativistic or relativistic plane wave functions to describe all electrons in the binary coplanar geometry and using the Sommerfed-Maue wave function to describe the ejected electron in the case of the coplanar asymmetric geometry. In all the approaches using relativistic plane waves, the use of the exact relativistic wave function for atomic hydrogen in its ground state is tractable, whereas in the asymmetric case, one can only use the semi relativistic Darwin wave function [@14] to describe the atomic target. The general conclusions that can be drawn from this work are summarized in what follows. In the non relativistic regime, the results obtained in the binary coplanar geometry give the same results for an appropriate choice of the angular parameters whereas in the coplanar asymmetric geometry, on can only compare the three approaches NRCBA, SRCBAWP and SRCBANP. This comparison has been made and it was shown that though relatively small, the presence of a recoil peak together with a visible binary peak even for an incident electron energy of $2700$ $eV$ reproduces the qualitative features of the Ehrhardt geometry [@2].\
In the relativistic regime radiation effects are important and the general rule that can be firmly asserted is the following : the instantaneous approach is always smaller than the non instantaneous approach because radiation effects can no longer be ignored. The use of the photon propagator as well as the QED formalism becomes necessary.
[90]{} I.E. McCarthy and E. Weigold, *Electron-Atom Collisions*, Cambridge University Press, 1995. H. Ehrhardt, M. Schulz, T. Tekaat, and K. Willmann, Phys. Rev. Lett. **22**, 89 (1969). G.R. Dangerfield and B.M. Spicer, J. Phys. B, **8**, 1744, (1975). D.H.H. Hoffmann, C. Brendel, H. Genz, W. Löw, S; Muller and A. Richter, Z. Phys., **A 293**, (1979). R. Anholt, Phys. Rev. **A 19**, 1004 (1979). J.H. Scofield, Phys. Rev. **A 18**, 963, (1978). B.L. Moiseiwitsch and S.G. Stockmann, J. Phys. B, **13**, 2975, (1980). B.L. Moiseiwitsch, Adv. At. Mol. Phys., **16**, 281, (1980). I. Fuss, J. Mitroy, and B. M. Spicer, J. Phys. B **15**, 3321 (1982). (1999). W. Nakel and C.T. Whelan, Phys. Rep., 315, 409, (1999). **315**, 409, (1999). H.R.J. Walters, H. Ast, C.T. Whelan, R.M. Dreizler, H. Graf, C.D. Schröter, J. Bonfert, W. Nakel, Z. Phys., **D 23**, 353, (1992). Y. Attaourti and S. Taj, Phys. Rev. A **69**, 063411 (2004). Y. Attaourti, S. Taj and B. Manaut, Phys. Rev. A **71**, 062705 (2005). J. Eichler and W.E. Meyerhof, *Relativistic Atomic Collisions*, Academic Press, 1995. F. Gross, *Relativistic Quantum Mechanics and Field Theory*, Wiley Science Paperback Series, 1999. W. Greiner, *Quantum Electrodynamics* , Second Corrected Edition, Springer, 1996. F.W. Byron Jr and C.J. Joachain, Phys. Rep. **179**, 211, (1989).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A generative probabilistic model for relational data consists of a family of probability distributions for relational structures over domains of different sizes. In most existing statistical relational learning (SRL) frameworks, these models are not projective in the sense that the marginal of the distribution for size-$n$ structures on induced substructures of size $k < n$ is equal to the given distribution for size-$k$ structures. Projectivity is very beneficial in that it directly enables lifted inference and statistically consistent learning from sub-sampled relational structures. In earlier work some simple fragments of SRL languages have been identified that represent projective models. However, no complete characterization of, and representation framework for projective models has been given. In this paper we fill this gap: exploiting representation theorems for infinite exchangeable arrays we introduce a class of directed graphical latent variable models that precisely correspond to the class of projective relational models. As a by-product we also obtain a characterization for when a given distribution over size-$k$ structures is the statistical frequency distribution of size-$k$ substructures in much larger size-$n$ structures. These results shed new light onto the old open problem of how to apply Halpern et al.’s “random worlds approach” for probabilistic inference to general relational signatures.'
author:
- |
Manfred Jaeger$^1$[^1] Oliver Schulte$^2$ $^1$Computer Science Department, Aalborg University, Aalborg, Denmark\
$^2$School of Computing Science, Simon Fraser University, Burnaby, Canada [email protected], [email protected]
bibliography:
- 'mplan.bib'
title: 'A Complete Characterization of Projectivity for Statistical Relational Models [^2]'
---
Introduction
============
Many types of generative models have been proposed for relational data in several fields, including machine learning and statistics. For i.i.d. data, a parametrized model defines a distribution over samples of a fixed size $n$, for every $n$. The analogue for generative relational models is a distribution ${Q^{(n)}}$ over complex multi-relational graphs (“worlds” in logical terminology) of a fixed size $n$, for every $n$. Research in statistical theory and discrete mathematics on the one hand, and AI and machine learning on the other hand has focussed on somewhat different aspects of relational models: the former is mostly concerned with internal model properties such as exchangeability, projectivity and behavior in the limit, whereas the latter is focussed on learning and inference tasks for one size $n$ at a time.
It is well known that in many popular statistical relational learning (SRL) frameworks the dependence of ${Q^{(n)}}$ on $n$ exhibits sometimes counter-intuitive and hard to control behavior. Most types of SRL models are not projective in the sense that the distribution ${Q^{(n)}}$ for $n$ nodes is the marginal distribution derived from the $Q^{n+1}$ distribution [@Shalizi2013; @Jaeger2018]. For exponential random graph and Markov logic network (MLN) models it has also been observed that the ${Q^{(n)}}$ tend to become degenerate as $n$ increases in the sense that the probability becomes concentrated on a few “extreme” structures [@rinaldo2009; @chatterjee2013estimating; @poole2014population]. Some authors have proposed to better control the behavior of MLNs by adjusting the model parameters as a function of $n$ [@jain2010adaptive]; however, no strong theoretical guarantees have yet been derived for such approaches.
In this paper we focus on projectivity as a very powerful condition to control the behavior of ${Q^{(n)}}$. In projective models, inferences about a fixed set of individuals are not sensitive to population size. This implies that inference trivially becomes *domain-lifted* [@broeck2011completeness], convergence of query probabilities becomes trivial, and certain statistical guarantees for learning from sub-sampled relational structures can be obtained [@Jaeger2018]. These benefits come at a certain cost in terms of expressivity: projective models are necessarily “dense” in the sense that, e.g., the expected number of edges in a projective random graph model is quadratic in $n$. In spite of these limitations, there exist projective model types such as the stochastic block model and the infinite relational model [@xu2006learning; @kemp2006learning] that have been proven very useful in practice. It thus seems very relevant to fully exploit the capabilities of projective models by developing maximally expressive projective representation, learning and inference frameworks. In this paper we take an important step in this direction by deriving a complete characterization of projective models as a certain class of directed latent variable models.
While the characterization we obtain is completely general, we approach our problem from the perspective that knowledge about the distributions ${Q^{(n)}}$ is given in the form of statistical frequencies of substructures of a small size $k$. For example, $k$ could be the maximal number of variables in an MLN formula, in which case the substructure frequencies are a sufficient statistics for learning the MLN parameters. In a somewhat different setting, $k$ can be the number of variables used in a Halpern/Bacchus-style statistical probability formula forming a statistical knowledge base [@Halpern90; @Bacchus90]. In all cases the question arises of how to generalize this knowledge to infer probabilities for specific instances (“beliefs"), either by statistical model estimation (as in most current SRL frameworks), or by inferring plausible beliefs based on invariance or maximum entropy principles, as in the random worlds approach of Bacchus et al. , and more recently in [@kern2010novel] and [@kuzelka2018relational]. A fundamental question that then arises is whether the given substructure frequencies can actually be the marginal distribution of ${Q^{(n)}}$ for large $n$. Results about the random worlds method need to be conditioned on the assumption that the statistical knowledge is “eventually consistent” [@halpern2017reasoning Chapter 11]. Similar assumptions are made in [@kuzelka2018relational]. As a by-product of our characterization of projective models we obtain that the same characterization also describes the distributions that can be induced as marginals of arbitrary ${Q^{(n)}}$.
Related Work
============
We discuss work on generative graph models related to exchangeability and projectivity, the two key properties in our study.
#### Exchangeability.
Exchangeability requires that a generative model should assign the same probability to graphs that differ only in node labellings. This is true for the large class of template-based relational models, because typical model discovery methods do not introduce templates that reference individual nodes [@Kimmig2014]. For example, they may only construct first-order logic formulas with no constant symbols. This includes most structure learning algorithms for Markov Logic Networks (e.g., [@Schulte2012]).[^3] Similarly, the sufficient statistics of exponential random graph models (e.g., the number of triangles in a graph) are typically defined without special reference to any particular node. Niepert and Van den Broeck have exploited the weaker notion of *partial exchangeability* to obtain tractable inference for certain SRL models.
#### Projectivity.
The importance of projectivity for graph modelling has been discussed previously [@Shalizi2013; @Jaeger2018]. Chatterjee and Diaconis discuss how estimation and inference in exponential random graph models depends on the sample size. Shalizi and Rinaldo give necessary and sufficient projectivity conditions for an exponential random graph model; they show that these are satisfied only in rare conditions. Jaeger and Schulte discuss a number of common SRL models, including MLNs and Relational Bayesian Networks, and show that they are projective only under restrictive conditions. Projective models used in practice factor a graph into independent components given a set of latent variables. Popular examples include the stochastic block model and generalizations [@hoff2002latent], the infinite relational model [@Orbanz2014], and recent graph neural network models such as the graph variational auto-encoder [@Kipf2016]. Our work shows that a latent conditional independence representation is not only sufficient for projectivity, but also necessary. We prove this result for a very large class of structured data, essentially general finite multi-dimensional arrays (tensors) with no restrictions on their dimensionality. Our results heavily depend on the theory of infinite exchangeable multi-dimensional arrays [@hoover1979relations; @aldous1981representations; @kallenberg2006probabilistic; @Orbanz2014]. The question of realizability of a given frequency distribution as a relational marginal has also been raised by Kuzelka et al., who then focus on approximate realizability, rather than characterizations of exact realizability.
Background
==========
Basic Definitions
-----------------
We use the following basic notation. The set of integers $\{1,\ldots,n\}$ is denoted $[n]$. For any $d\geq 1$, we write $[n]_{\neq}^d$ for the set of $d$-tuples containing $d$ distinct elements from $[n]$. The subset of $[n]_{\neq}^d$ containing tuples in which the elements appear in their natural order is denoted $\langle n \rangle^d$ (so that $\langle n \rangle^d$ corresponds to a standardized representation for the set of all $d$-element subsets of $[n]$). Extending this notation to the infinite case, we can also write $[\Nset]_{\neq}^d$ and $\langle \Nset \rangle^d$.
#### Relations and Possible Worlds.
A relational *signature* $S$ contains relations of varying arities. We refer to the maximal arity of relations contained in $S$ as the *arity of* $S$, denoted $\emph{arity}(S)$. A *possible world* $\omega$ (for $S$) specifies 1) a finite domain $D=\{d_1,\ldots,d_n\}$, 2) for each $m$-ary relation from $S$ an $m$-dimensional binary adjacency matrix. We refer to $n$ as the *size* of $\omega$, and also call $\omega$ an $n$-world. For most purposes, we can assume that $D=[n]$, or at least $D\subset \Nset$. However, even if we make this assumption for convenience of presentation, we do not generally assume that the integer label of a randomly observed domain element can also be observed. We denote by ${\Omega^{(n)}}$ the set of all possible worlds for a given signature $S$ with domain $[n]$. The relevant signature is usually implicit from the context, and not made explicit in the notation. Finally, $\Omega:=\cup_n \Omega^{(n)}$.
#### Relational Substructures.
We also require notation to refer to different types of substructures of a possible $n$-world $\omega$. For a subset $I\subset [n]$ of size $|I|=m<n$ we denote with $\omega\downarrow I$ the $m$-world induced by $I$, i.e., the possible world with domain $I$, and the relations of $\omega$ restricted to arguments from $I$. For a tuple $\boldi\in [n]^m_{\neq}$ we denote with $\omega\downarrow \boldi$ the world over the domain $[m]$ obtained by relabeling the domain elements in the sub-world induced by the set $\boldi$ as $i_h\mapsto h$ (cf. Figure \[fig:worlddefs\], top row). A little less conventional is the following concept, that will become important for our main theorem: for $m=1,\ldots, \emph{arity}(S)$ we define $D_m(\omega)$ as the *arity-$m$ data of $\omega$*. Informally speaking, $D_m(\omega)$ collects all the information from all adjacency arrays of $\omega$ that refers to exactly $m$ distinct elements. For example (cf. Figure \[fig:worlddefs\]), $D_1(\omega)$ contains the data (adjacency arrays) of all unary relations of $S$, but also the information contained on the diagonal of a two-dimensional adjacency array for a binary (edge) relation, i.e., the information about self-loops of that relation. A possible world can then also be described by the tuple $(D_m)_{m=1,\ldots,\emph{arity}(S)}$. Furthermore, $D_m(\omega)$ can be decomposed into the factors $D_m(\omega\downarrow \boldi)$, where $\boldi$ ranges over $\langle n\rangle^m$. We denote with ${\cal T}_m$ the space of possible values of $D_m(\omega\downarrow \boldi)$ ($|\boldi|=m$). A possible world $\omega\in{\Omega^{(n)}}$ then also is given by an assignement of a value in ${\cal T}_m$ for all $\boldi\in\langle n\rangle^m$ ($m=1,\ldots, \emph{arity}(S)$).
![Top left: world $\omega$ with one unary relation (black/white) and one binary (edge) relation; top middle/right: sub-worlds induced by $I=\{1,3\}$ and $\boldi=(3,1)$; second row: unary and binary data parts; bottom: spaces ${\cal T}_1,{\cal T}_2$ for the given signature.[]{data-label="fig:worlddefs"}](./worlds.pdf)
Worldlet Frequency Distributions {#sec:worldletfreqs}
================================
Many graph analysis methods examine frequent characteristic subgraphs to provide information about a larger graph. We can think of a subgraph as a template that can be instantiated multiple times in a large graph. For example, in a social network we can count the number of friendship triangles among women. Depending on the framework, such templates go by different names (e.g., graphlets, motifs, frequency subgraphs) and are represented using different syntax (e.g., SQL queries, first-order logic, semantic relationships). We observe that subgraph templates can be represented in a general syntax-independent way as the collection of fully specified graphs ${\Omega^{(k)}}$ of a fixed size $k$, where we think of $k$ as a small number (typically in the range $k=2,\ldots,5$). When seen as a subgraph pattern, we refer to a world $\omega \in {\Omega^{(k)}}$ as a . We assume that for every worldlet, the frequency of its occurrence in a larger world is available, through learning or expert elicitation (cf. [@Bacchus90]). As a notational convention, we use $k$ and $n$ to denote domain sizes of (small) worldlets and large “real” worlds, respectively. This convention only is intended to support intuitions, and does not have any strict mathematical implications.
#### Statistical Frequency Distributions.
The intuitive idea of observing random worlds by sampling subsets of larger domains can be formalized in slightly different ways, e.g. by assuming sampling with or without replacement, or by interpreting the observation as a unique world, or only an isomorphism class [@diaconis2007graph; @kuzelka2018relational]. In many aspects alternative sampling models become essentially equivalent as $n\rightarrow\infty$ [@diaconis2007graph]. We here adopt a sampling model in which an ordered sample is drawn without replacement. Thus, a sample from a world $\omega\in{\Omega^{(n)}}$ is given by one of the $n!/(n-k)!$ tuples $\boldi\in[n]^k_{\neq}$, and the observed worldlet then is $\omega\downarrow\boldi$. Note that this sampling method does not rely on observing the original labels of elements drawn from $\omega$ to obtain the labeling of elements in the sampled worldlet, and therefore also makes sense when the elements of $\omega$ can not be assumed to have (observable) integer labels. The frequency distribution obtained through this sampling method is denoted ${P}^{(k)}(\cdot|\omega)$.
Let $S=\{e\}$ consist of a single binary relation. Let $\omega\in{\Omega^{(n)}}$ be a “star” with center 1, i.e., $e$ consists of the edges $\{1\rightarrow l: l=2,\ldots,n\}$. The probability that a random draw of 2 elements contains the node 1 then is $2/n$, with equal probability that 1 is the first or second drawn element. The three worldlets $1\bullet\!\! {\white\rightarrow} \!\! \bullet 2$, $1\bullet\!\! \rightarrow\!\! \bullet 2$ and $1\bullet\!\! \leftarrow \!\! \bullet 2$ then have probabilities $1-2/n, 1/n, 1/n$ (in this order) under ${P}^{(k)}(\cdot|\omega)$.
Every world $\omega$ defines a frequency distributions ${P}^{(k)}(\cdot|\omega)$. If first a random $\omega$ is selected, we obtain a two-step sampling procedure that was first described in a more general context by Fenstad .
#### Fenstad Sampling.
Given a possible world distribution ${Q^{(n)}}$, we define the distribution ${P^{(k)}}\circ{Q^{(n)}}$ for $k$-worlds $\omega'$ as follows: $$\label{eq:twostepsampling}
({P^{(k)}}\circ{Q^{(n)}}) (\omega') := \sum_{\omega\in{\Omega^{(n)}}} {Q^{(n)}}(\omega){P^{(k)}}(\omega'\mid\omega).$$ We denote with $\Delta^{(k)}_{n}$ the set of distributions on ${\Omega^{(k)}}$ that have a representation of the form (\[eq:twostepsampling\]) for some ${Q^{(n)}}$. If $k<l<n$, then ${P^{(k)}}\circ(P^{(l)}\circ{Q^{(n)}})={P^{(k)}}\circ{Q^{(n)}}$, and thus $\Delta^{(k)}_{n}\subseteq \Delta^{(k)}_{l}$.
![Illustration of $\Delta^{(k)}_{n}$ for $k=3$ and $n=3,4,5,6$. Cf. examples \[ex:Deltakn\] and \[ex:running1\][]{data-label="fig:scatterfig"}](./samplingdistributions.png "fig:") (-185,212)[$n=3$]{} (-65,212)[$n=4$]{} (-185,93)[$n=5$]{} (-65,93)[$n=6$]{}
\[ex:Deltakn\] In this example and some of the following, we take $S$ to contain a single undirected edge relation $e$. In order to comply with our general definitions, which are based on directed relations, we consider an undirected edge $i\bullet - \bullet j$ to be a shorthand for the conjunction $i\bullet \rightarrow \bullet j$ and $i\bullet \leftarrow \bullet j$, and we assume that all worlds with uni-directional edges ($i\bullet \rightarrow \bullet j$ but not $i\bullet \leftarrow \bullet j$) or self-loops ( $i\bullet \rightarrow \bullet i$) have probability zero. Disregarding these probability zero worlds, $\Omega^{(3)}$ then contains 8 possible worlds belonging to 4 different isomorphism classes. The top row of Table \[tab:wdistributions\] depicts these isomorphism classes, together with the count of worlds in each class.
Figure \[fig:scatterfig\] illustrates for $n=3,4,5,6$ the worldlet frequency distributions ${P^{(k)}}(\cdot|\omega)$ defined by the worlds $\omega\in{\Omega^{(n)}}$. Each (blue) dot is the distribution defined by one world after projecting its 8-dimensional probability vector into 2-dimensional space. Some jitter is applied to exhibit the multiplicities of $n$-worlds defining the same distribution on worldlets of size 3. The sets $\Delta^{(k)}_{n}$ are the convex hulls of these points. The distribution marked by the (red) + in Table \[tab:wdistributions\] and Figure \[fig:scatterfig\] belongs to $\Delta^{(k)}_{n}$ for $n=3,4$, but not for $n=5,6$.
Relational Models and Distribution Families {#sec:properties}
===========================================
As our goal is to examine properties of relational models that are independent of a particular model syntax, we use a family of distributions as a semantic view of a parametrized model. The two key properties of families in our study are exchangeability and projectivity.
Distribution Families: Exchangeability, and Projectivity
--------------------------------------------------------
A *family of distributions* $\{{Q^{(n)}}: n \in\Nset\}$ specifies, for each finite domain size $n$, a distribution ${Q^{(n)}}$ on the possible world set ${\Omega^{(n)}}$.
A probability distribution ${Q^{(n)}}$ on ${\Omega^{(n)}}$ is *exchangeable*, if ${Q^{(n)}}(\omega)={Q^{(n)}}(\omega')$ whenever $\omega$ and $\omega'$ are isomorphic. A family is exchangeable, if every member of the family is exchangeable.
Intuitively a distribution family is projective if its members are mutually consistent in the sense that the world distribution over a smaller domain size is the marginal distribution over a larger one. For a precise definition, we follow our notation for relational substructures, and for each $n$-world $\omega$, write $\omega \downarrow [m]$ for the size-$m$ subworld that results from restricting $\omega$ to the first $m$ elements. A distribution ${Q^{(n)}}$ over $n$-worlds then induces a [*marginal*]{} probability for an $m$-world $\omega'$ as follows:
$${Q^{(n)}}\downarrow[m] (\omega') = \sum_{\omega \in {\Omega^{(n)}}: \omega \downarrow [m] = \omega'}\!\!\!\!\!\! {Q^{(n)}}(\omega)$$
Projectivity is the central concept for our investigation:
\[def:projective\] An exchangeable family $({Q^{(n)}})_{n\in\Nset}$ is *projective*, if for all $m<n$: ${Q^{(n)}}\downarrow [m] ={Q^{(m)}}$.
Note that in contrast to more general notions of projectivity found in the theory of stochastic processes, we here define projectivity only for exchangeable families. Exchangeability implies that the marginal distribution ${Q^{(n)}}\downarrow I$ is the same for all subsets $I$ of size $m$, and therefore we only need to consider the marginal ${Q^{(n)}}\downarrow [m]$ as a prototype.
Statistical frequency distributions ${P^{(k)}}(\cdot\mid\omega)$ always are exchangeable. As a special case, if $\omega\in{\Omega^{(n)}}$, then ${P^{(n)}}(\cdot\mid\omega)$ samples a random permutation of $\omega$, i.e., is the uniform distribution on the isomorphism class of $\omega$. It follows that distributions defined by Fenstad sampling (\[eq:twostepsampling\]) also are exchangeable, for any ${Q^{(n)}}$.
We approach the question of how to characterize and represent projective families through the more specific question of whether a given distribution ${Q^{(k)}}$ can be embedded in a projective family. The following definition provides the necessary terminology.
\[def:extendable\] Let ${Q^{(k)}}$ be an exchangeable distribution on ${\Omega^{(k)}}$. ${Q^{(k)}}$ is called
- *$n$-extendable*, if ${Q^{(k)}}\in\Delta^{(k)}_{n}$; any ${Q^{(n)}}$ that induces ${Q^{(k)}}$ via (\[eq:twostepsampling\]) is called an *extension* of ${Q^{(k)}}$.
- *extendable*, if it is $n$-extendable for all $n>k$;
- *projective extendable* if there exists a projective family $({Q^{(n)}})_n$ of extensions of ${Q^{(k)}}$.
 ($\times 1$)  ($\times 3$)  ($\times 3$)  ($\times 1$) Name
---------------------------------- ---------------------------------- ---------------------------------- ---------------------------------- ---------------
1 0 0 0 ${ 1_{E_3} }$
0 0 0 1 ${ 1_{K_3} }$
0 1/3 0 0 +
1/4 0 1/4 0 bipart
: Some example worldlet distributions[]{data-label="tab:wdistributions"}
\[ex:running1\] The rows in Table \[tab:wdistributions\] specify several exchangeable distributions on $\Omega^{(3)}$ (in the undirected graph setting, as described in Example \[ex:Deltakn\]). The numbers in the table specify the probabilities of each world in a given isomorphism class, not the total probability of the isomorphism class. The first two are the point masses on the empty graph (denoted $E_3$) and complete graph ($K_3$), respectively. If ${ 1_{E_n} }$ denotes the point mass on the empty graph of size $n$, then $( { 1_{E_n} })_n$ is a projective family. Similarly for the family $({ 1_{K_n} })_n$, and the family of mixtures $(0.5\cdot{ 1_{E_n} }+0.5\cdot{ 1_{K_n} })_n$.
The row labeled + is the distribution marked by the (red) + in the plots of Figure \[fig:scatterfig\]. If $\omega\in\Omega^{(4)}$ is the graph that contains the two edges $1\bullet\!\!-\!\!\bullet 2$ and $3\bullet\!\!-\!\!\bullet 4$, then this distribution is equal to $P^{(3)}(\cdot|\omega)$. Thus, it is 4-extendable, which is also visible in the top right panel of Figure \[fig:scatterfig\] showing that ’+’ coincides with sampling distributions induced by 4-worlds. However, ’+’ is not $n$-extendable for any $n\geq 5$. This is visible in Figure \[fig:scatterfig\] as for $n=5,6$ ’+’ lies outside the convex hull of the worldlet frequency distributions. Proposition \[prop:modularity\] below will provide a simple tool for proving the non-extendability of ’+’.
The last row in the table describes the distribution that in the limit for $n\rightarrow \infty$ is the worldlet frequency distribution defined by complete, balanced bipartite graphs, i.e., graphs whose edge set is equal to $\{i\bullet\!\!-\!\!\bullet j: 1\leq i\leq \lfloor n/2 \rfloor; \lfloor n/2 \rfloor +1\leq j \leq n \}$. It will follow from our main theorem that this distribution is projective extendable.
Domain Sampling Distributions
-----------------------------
Extendable distributions ${Q^{(k)}}$ in the sense of Definition \[def:extendable\] are mixtures of worldlet frequency distributions. An important special case is when ${Q^{(k)}}$ is a pure worldlet frequency distribution ${P^{(k)}}(\cdot|\omega)$ defined by a single world $\omega$. In that case, however, one cannot expect that ${Q^{(k)}}$ can be represented in this form with suitable $\omega$ for all $n$, because the sets $\{{P^{(k)}}(\cdot|\omega): \omega\in{\Omega^{(n)}}\}$ for different $n$ are concentrated on different grids of rational numbers, and therefore are largely disjoint (cf. Figure \[fig:scatterfig\]). Following the approach already taken by Bacchus et al. to give semantics to statistical probability terms in the random worlds approach [@BaGroHalKol92; @halpern2017reasoning] we therefore only require that ${Q^{(k)}}$ is approximately equal to some $ {P^{(k)}}(\cdot|\omega)$, with an increasing accuracy in the approximation as the size of $\omega$ increases.
\[def:dsrealizable\] Let [${Q^{(k)}}$]{} be a probability distribution on ${\Omega^{(k)}}$. We say that [${Q^{(k)}}$]{} is a *domain sampling distribution* if the following holds: for every $\epsilon >0$ there exists $n\in\Nset$, such that for every $n'\geq n$: there exists a possible $n'$-world $\omega$, so that for all $\omega'\in{\Omega^{(k)}}$: $$\label{eq:realizable}
|P^{(k)}(\omega'\mid \omega)-{Q^{(k)}}(\omega')|<\epsilon.$$
Thus, the property of being a domain sampling distribution strengthens the property of extendability in that in the representation (\[eq:twostepsampling\]) only point masses ${Q^{(n)}}={ 1_{\omega} }$ are allowed, but weakens it in that (\[eq:realizable\]) only requires approximate equality.
For the worldlet distributions of Table \[tab:wdistributions\] we have ${ 1_{E_3} }=P^{(3)}(\cdot| E_n)$ for all $n\geq 3$, so that ${ 1_{E_3} }$ is a domain sampling distribution (with zero approximation error). Similarly for ${ 1_{K_3} }$. The mixture $0.5\cdot { 1_{E_3} }+0.5\cdot { 1_{K_3} }$ is projective extendable, but not a domain sampling distribution. The distribution ’$+$’ is not a domain sampling distribution. This is indicated by Figure \[fig:scatterfig\], because already for $n=6$ the distribution is separated by a distance $\epsilon>0$ from the set $\Delta^{(3)}_6$. Because of the nested structure of the $\Delta^{(3)}_n$ there then also cannot be better approximations for larger $n>6$. The last ’bipart’ distribution in Table \[tab:wdistributions\] again is a domain sampling distribution with a non-zero approximation error that only vanishes as $n\rightarrow\infty$.
A Representation Theorem {#sec:reptheo}
========================
We now proceed to derive our main result, which is a comprehensive characterization of families $({Q^{(n)}})_n$ and worldlet marginals ${Q^{(k)}}$ with the structural properties described in Section \[sec:properties\]. We introduce a representation for projective families that is based on the analysis and representation theorems for infinite exchangeable arrays developed by Aldous and Hoover . The definitive treatment is given by Kallenberg . We therefore call the following an AHK model.
\[def:ahkmodel\] Let $S$ be a signature with maximal $\emph{arity}(S)=a\geq 1$. An AHK model for $S$ is given by
- A family of i.i.d. random variables $\{U_{\boldi}| \boldi\in \langle \Nset \rangle^m, m=0,\ldots,a\}$, where each $U_{\boldi}$ is uniformly distributed on $[0,1]$.
- A family of random variables $\{D_{\boldi}| \boldi\in \langle \Nset \rangle^m, m=1,\ldots,a\}$. For $\boldi\in \langle \Nset \rangle^m$ the variable $D_{\boldi}$ takes values in ${\cal T}_m$.
- For each $m=1,\ldots,a$ a measurable function $$\label{eq:ffunction}
f^m: [0,1]^{2^m}\rightarrow {\cal T}_m$$ so that
- for $\boldi=(i_1,\ldots,i_m)\in\langle \Nset \rangle^m$ the value of $D_{\boldi}$ is defined as $f^m(\boldU_{\boldi})$, where $$\begin{gathered}
\label{eq:Tfromf}
\boldU_{\boldi}=(U_{\emptyset},U_{i_1},\ldots,U_{i_m},U_{( i_1,i_2)},\ldots,\\
U_{( i_{m-1},i_m)},\ldots
\ldots, U_{( i_1,\ldots,i_m) }),
\end{gathered}$$ is the vector containing all $U_{\boldi'}$-variables with $\boldi'\subseteq\boldi$ in lexicographic order.
- $f^m$ is permutation equivariant, in the sense that for any permutation $\pi$ of $[m]$ $$f^m(\pi\boldU_{\boldi})=\pi f^m(\boldU)$$ where $\pi\boldU_{\boldi}$ is the permutation of $\boldU_{\boldi}$ that in the place of $U_{\boldi'}$ contains $U_{\pi\boldi'}$ with $\pi\boldi'$ the ordered tuple of the elements $\{\pi(i): i\in\boldi'\}$.
An AHK model that does not contain the $U_{\emptyset}$ variable is called an AHK$^-$ model.
![Plate representation of AHK model with $a=3$[]{data-label="fig:ahkplates"}](./ahk-plates-portrait.pdf)
Figure \[fig:ahkplates\] gives an illustration of the structure of an AHK model in plate notation. An AHK model is fully determined by the functions $\boldf:=(f^m)_{m=1\ldots,a}$, and we therefore write $\boldf$ to refer to an AHK model. By a slight abuse of notation, we also use $\boldf$ to denote the distribution defined by the model on the possible worlds over the infinite domain $\Nset$, and write $\boldf\downarrow [n]$ for the marginal on the induced sub-world over the domain $[n]$.
The following example gives a simple illustration of how the permutation equivariance condition for the functions $f^m$ ensures exchangeability.
We encode a version of the Erdős-Rényi random graph model in which any pair of nodes is connected with probability 1/2 by an edge, and that edge is given a random direction. Thus, the target distribution on worldlets of size 2 is $P(1\bullet \leftarrow \bullet 2)=P(1\bullet \rightarrow \bullet 2)=0.25$, $P(1\bullet \hspace{3mm} \bullet 2)=0.5$. The state space ${\cal T}_1$ contains the two states “self-loop” and “no self-loop”. Since self-loops have probability zero, we simply let $f^1$ be the constant function that returns “no self-loop” regardless of the input $U$-variables. The state space ${\cal T}_2$ contains the four states $1\bullet\!\! {\white\rightarrow} \!\! \bullet 2$, $1\bullet\!\! \rightarrow\!\! \bullet 2$, $1\bullet\!\! \leftarrow \!\! \bullet 2$, and $1\bullet\!\! \leftrightarrow \!\! \bullet 2$, of which only the first three have non-zero probability. Let $$\begin{gathered}
f^2(x_0,x_1,x_2,x_3):= \\
\left\{
\begin{array}{ll}
1\bullet \rightarrow \bullet 2 & \mbox{if}\ x_1 < x_2 \ \mbox{and}\ x_3 <0.5 \\
1\bullet \leftarrow \bullet 2 & \mbox{if}\ x_2 < x_1 \ \mbox{and}\ x_3 <0.5 \\
1\bullet \hspace{3mm} \bullet 2 & \mbox{otherwise.}
\end{array}
\right.
\end{gathered}$$ For clarity we here use a notation that makes it clear that the functions $f^m$ are defined on arrays of length $2^m$, and their definition distinguishes arguments by their position in the input array, not by their semantic nature as a variable $U_{\boldi'}$. For $\pi: 1\mapsto 2, 2\mapsto 1$ we then have $\pi(1\bullet \rightarrow \bullet 2 )= 1\bullet \leftarrow \bullet 2$, and $f^2(\pi\boldU_{(1,2)})=f^2(U_{\emptyset},U_2,U_1,U_{(1,2)})= \pi f^2(\boldU_{(1,2)})$. Together with the fact that the tuples $\boldU_{(1,2)}$ and $\pi\boldU_{(1,2)}$ have identical distribution, this implies that the two values $ 1\bullet \rightarrow \bullet 2$, $ 1\bullet \leftarrow \bullet 2$ of $D_{(1,2)}$ have the same probability.
Generalizing from this example, and also noting that the plate representation of the AHK models directly implies that marginals $\boldf\downarrow [n]$ simply are given by instantiating the plate model only for $\boldi\subset [n]$, we can note the following proposition.
\[prop:ahkprojective\] Let $\boldf$ be an AHK model. The marginals $\boldf\downarrow [n]$ are exchangeable, and the family $(\boldf\downarrow [n])_n$ is projective.
For a given worldlet distribution ${Q^{(k)}}$ with $k\geq \emph{arity}(S)$ we say that ${Q^{(k)}}$ *has an AHK representation*, if there exists an $\boldf$ with $\boldf\downarrow [k]={Q^{(k)}}$.
We can now formulate our main result.
\[theo:extendable2\] Let ${Q^{(k)}}$ be an exchangeable distribution on $\Omega^{(k)}$ with $k\geq \emph{arity}(S)$. For the statements
(A)
: ${Q^{(k)}}$ is a domain sampling distribution.
(B)
: ${Q^{(k)}}$ has a AHK$^-$ representation
(C)
: ${Q^{(k)}}$ is a finite mixture of domain sampling distributions
(D)
: ${Q^{(k)}}$ is extendable
(E)
: ${Q^{(k)}}$ is projective extendable
(F)
: ${Q^{(k)}}$ has a AHK representation
the following implications hold:
$$\mbox{\bf (A)}\Leftrightarrow \mbox{\bf (B)}
\Rightarrow \mbox{\bf (C)}
\Leftrightarrow \mbox{\bf (D)}
\Leftrightarrow \mbox{\bf (E)}
\Leftrightarrow \mbox{\bf (F)}$$
The full proof of the theorem is given in the extended online version of this paper (<http://arxiv.org/abs/2004.10984>).
Discussion
==========
In this section we consider some of the trade-offs between limitations in expressivity of projective models on the one hand, and gain in algorithmic and statistical tractability on the other hand. Limitations in expressivity can be considered in terms of what distributions ${Q^{(n)}}$, for a fixed $n$, can be represented, and in terms of the limitations for the family $\{{Q^{(n)}}|n\in\Nset\}$ as a whole. Considering a single distribution ${Q^{(n)}}$, we can observe a *modularity* property as described by the following proposition, and illustrated in Figure \[fig:modular\]
![Modularity of AHK models: if $\omega$ on the left has nonzero probability, then also the set of worlds $O$ on the right.[]{data-label="fig:modular"}](./worlds-extend.pdf)
\[prop:modularity\] Let $\boldf$ be an AHK model, $\omega\in{\Omega^{(n)}}$ with $\boldf\downarrow[n](\omega)>0$. Let $O\subset\Omega^{(n+1)}$ be the set of $n+1$ worlds $\omega'$ for which $\omega'\downarrow [n]=\omega'\downarrow \{1,\ldots,n-1,n+1 \} =\omega$. Then $\boldf\downarrow[n](O)>0$. Moreover, if $\boldf$ is an AHK$^-$ model, then $\omega'\downarrow [n]=\omega$ and $\omega'\downarrow \{1,\ldots,n-1,n+1 \} =\omega$ are independent events given $\omega'\downarrow [n-1]$.
Figure \[fig:modular\] illustrates the proposition with $n=3$: if the world $\omega$ on the left has nonzero probability, then also the set of 4-worlds $O$ on the right has nonzero probability. $O$ is the set of 4-worlds for which the substructures induced by $\{1,2,3\}$ and $\{1,2,4\}$ are both isomorphic to $\omega$. The dashed arc connecting nodes 3 and 4 on the right indicates that the value of $D_{(3,4)}$ determining the relations between nodes 3 and 4 can vary for different elements of $O$.
As an application of Proposition \[prop:modularity\] we can see that the ’+’ distribution of Table \[tab:wdistributions\] does not have an AHK representation, and therefore cannot be extendable (cf. Example \[ex:running1\]): letting $n=2$ and $\omega= 1\bullet\!\!-\!\!\bullet2$, we obtain from the proposition that also 3-worlds with two edges $1\bullet\!\!-\!\!\bullet2$ and $2\bullet\!\!-\!\!\bullet3$ must have nonzero probability, which is not the case for ’+’.
We now turn to structural limitations of the whole family $\{{Q^{(n)}}|n\in\Nset\}$ implied by an AHK representation. As already mentioned in the introduction, projective families generate structures that are “dense” in the limit. More precisely, if $\omega\in{\Omega^{(k)}}$ is a worldlet with $\boldf\downarrow [k] (\omega)>0$, then the expected number of $k$-tuples in worlds of size $n$ which induce sub-worlds isomorphic to $\omega$ grows linearly in $n^k$. Specifically, if graph edges have a nonzero probability at all, then the expected number of edges grows linearly in $n^2$. It must be emphasized, though, that this only imposes limits on modeling the asymptotic behavior of evolving graphs. For any fixed domain size, an AHK model can fit any observed degree distribution:
Let $n^*\in\Nset$, and let $f(d)$ ($d=0,1,\ldots n^*$) denote an out-degree distribution for directed graphs on $[n^*]$. For arbitrary $n$ we can normalize out-degrees in graphs of size $n$ via $d\mapsto d/n$. Let $F(\delta)$ ($\delta\in[0,1]$) be the cumulative distribution function obtained from $f()$ for the normalized degrees $d\mapsto d/n^*$. We now define $$\begin{gathered}
f^2(U_i,U_j,U_{(i,j)}):= \\
\left\{
\begin{array}{ll}
i\bullet\!\! \rightarrow \!\! \bullet j & \mbox{if}\ U_i\geq F(U_{(i,j)})\ \mbox{and}\ U_j < F(U_{(i,j)}) \\
i\bullet\!\! \leftarrow \!\! \bullet j & \mbox{if}\ U_j\geq F(U_{(i,j)})\ \mbox{and}\ U_i < F(U_{(i,j)}) \\
i\bullet\!\! \leftrightarrow \!\! \bullet j & \mbox{if}\ U_i\geq F(U_{(i,j)})\ \mbox{and}\ U_j \geq F(U_{(i,j)}) \\
i\bullet\!\! {\white\leftrightarrow} \!\! \bullet j & \mbox{otherwise}
\end{array}
\right.
\end{gathered}$$ Let $\delta_i$ denote the normalized out-degree of node $i$. Then for all $u\in[0,1]$ we obtain the expected normalized out-degree: $$\label{eq:expoutdeg}
E[\delta_i|U_i=u]=F^{-1}(u).$$ $U_i$ being uniformly distributed, the right-hand side of (\[eq:expoutdeg\]) is distributed with cdf $F()$, and so the expected normalized degree distribution follows $F()$. In the special case $n=n^*$ then the expected absolute degree distribution is the original $f()$.
On the positive side, we obtain significant computational and robustness advantages from the use of projective models: inference is *lifted* in the strongest possible sense that the complexity of computing a query probability for a query involving $k$ named entities is independent of the size of the domain in which the entities are embedded. For learning, projectivity is a necessary condition for consistent estimation from substructures randomly sampled from domains of unknown size. However, further conditions beyond projectivity are required to formulate and derive precise consistency guarantees [@Jaeger2018]. Statistical consistency and robustness results can therefore not be directly given for AHK models in general without first identifying a suitable effectively representable and parameterizable class of functions from which the $f^m$ can be constructed. Identifying rich and tractable such classes, and evaluating their learning capabilities empirically and theoretically is future work.
When evaluating the trade-offs of AHK models for a particular application, it must always be born in mind that the strenghts of generative, projctive models only come to bear when one needs to deal with diverse types of queries (so that a discriminative model for a fixed prediction task would be inadequate), and when one has to deal with data from domains of different and/or uncertain sizes. We note that this is basically the opposite side of the task spectrum from where many current popular node classification and link prediction problems are situated, in which both learning and inference is conducted for a fixed task on a single given graph, e.g., [@wu2020comprehensive].
Conclusion
==========
In this paper we have laid theoretical foundations for the study and application of rich classes of projective families. Bringing together research strands in statistical graph theory and statistical relational learning we have derived an explicit characterization of projective families in the form of a directed graphical (plate) model. We have shown that closely linked to projectivity is the (approximate) realizability as a statistical frequency distributions of worldlet samples drawn from large domain. These results give us a characterization of the form of statistical knowledge to which the random worlds approach of Bacchus et al. can be applied.
Interestingly, the structure of AHK models has much in common with the “independent choice logic” family of SRL frameworks [@Sato95; @Poole97; @KimDemDeRSanRoc11] that also generate random relational structures as deterministic functions of a set of a-priori independent random variables. However, the continuous nature of the $U_{\boldi}$ variables in the AHK model, and the potential need of functions $f^m$ not readily expressible in existing SRL languages pose significant challenges for the direct application of existing SRL techniques.
On the theoretical side, many interesting questions remain regarding statistical principles of model selection, and unbiasedness and consistency of estimation: for a given worldlet distribution ${Q^{(k)}}$ there will often be multiple AHK models that precisely fit ${Q^{(k)}}$ and therefore are indistinguishable based on likelihood scores. What invariance, parsimony, or plain parameter regularization principles are then most useful for model selection?
Acknowledgments {#acknowledgments .unnumbered}
===============
Oliver Schulte’s contribution was supported by a Discovery Grant from the Natural Sciences and Engineering Research Council of Canada.
[ ]{}
Details on ${\cal T}_m$, $D_{\boldi}$
=====================================
We fully formalize the definitions of ${\cal T}_m$, $D_m$ and $D_{\boldi}$.
Let $r\in S$ be a $k$-ary relation. The interpretation of $r$ in a possible $n$-world $\omega$ is given by a $k$-dimensional 0,1-valued adjacency matrix $A_r(\omega)$, which we view as a mapping $[n]^k\rightarrow \{0,1\}$. Let $m\leq k$, and denote by $[n]^k_m$ the subset of $[n]^k$ that contains all $k$-tuples with exactly $m$ distinct elements. Let $A_r(\omega)|m$ denote the restriction of $A_r(\omega)$ to tuples from $[n]^k_m$. Then $$D_m(\omega) := \{A_r(\omega)|m\ |\ r\in S: \emph{arity}(r)\geq m\}.$$ The set $\{ D_m(\omega)\ |\ m=1,\ldots,\emph{arity}(S)\}$ contains all the information of all adjacency matrices $A_r(\omega)$, and is therefore an equivalent representation of the world $\omega$.
Now let $\boldi\in \langle n\rangle^m$. Define $$D_{\boldi}(\omega) := D_m(\omega \downarrow \boldi).$$ ${\cal T}_m$ is the space of possible values for $ D_{\boldi}(\omega)$ with $\omega\in\Omega$ and $\boldi\in \langle \Nset\rangle^m$. Recall that the projection operation $\omega \downarrow \boldi$ involves a re-naming of the elements of $\boldi$ as $1,\ldots,m$. Thus, for example, for the $\omega$ of Figure \[fig:worlddefs\] we have that $D_{(1,3)}(\omega) $ is the element of ${\cal T}_2$ that is graphically represented as $1 \circ\!\! \rightarrow\!\! \circ 2$ in Figure \[fig:worlddefs\].
Additional Examples
===================
Consider the ’bipart’ distribution of Table \[tab:wdistributions\]. This is a domain-sampling distribution and has an AHK$^-$ representation given by functions $f^1, f^2$ as follows. Here $\emph{arity}(S)=2$, so AHK models only are over variables $U_{\boldi},D_{\boldi}$ with $\boldi\leq 2$. The state space ${\cal T}_1$ contains the two states “self-loop” and “no self-loop”. Since self-loops have probability zero in ’bipart’, we simply let $f^1$ be the constant function that returns “no self-loop” regardless of the input $U$-variables. The state space ${\cal T}_2$ contains the four states $1\bullet\!\! {\white\rightarrow} \!\! \bullet 2$, $1\bullet\!\! \rightarrow\!\! \bullet 2$, $1\bullet\!\! \leftarrow \!\! \bullet 2$, and $1\bullet\!\! \leftrightarrow \!\! \bullet 2$, of which only the first and last have non-zero probability. To represent this distribution, we let the $U_i$ encode the partition of the domain into two (equal) parts, and then define for $i<j$: $$\begin{gathered}
f^2(U_i,U_j):= \\
\left\{
\begin{array}{ll}
i\bullet\!\! \leftrightarrow \!\! \bullet j & \mbox{if}\ U_i<0.5<U_j\ \mbox{or}\ U_j<0.5<U_i \\
i\bullet\!\! {\white\leftrightarrow} \!\! \bullet j & \mbox{otherwise}
\end{array}
\right.
\end{gathered}$$ Thus, $ f^2$ here neither makes use of $U_{\emptyset}$, nor of $U_{(i,j)}$. If instead of a complete bipartite graph we want to model that an (undirected) edge between nodes of different components only exists with a certain probability $p<1$, we have to add to the first case in the definition of $f^2$ the condition $U_{(i,j)}<p$. It is clear that in a similar manner one can represent arbitrary stochastic block models.
Alternative Sampling Model {#sec:altsampling}
==========================
The statistical frequence distribution ${P}^{(k)}(\cdot|\omega)$ defined in Section \[sec:worldletfreqs\] is based on drawing a random ordered sample, and labeling the elements in the sample according to the order in which they were drawn. For the use in the proof of Theorem \[theo:extendable2\] we here introduce a slightly different sampling model, in which we draw an unordered sample, and elements are relabeled according to the order of their original labels. A random $k$-world sampled from a world $\omega\in{\Omega^{(n)}}$ then is given by one of the $\binom{n}{k}$ subesets $\{i_1,i_2,\cdots,i_k \}\subseteq [n]$ where w.l.o.g. we can assume $i_1<i_2<\cdots<i_k$. The substructure induced by this subset defines a worldlet over $[k]$ via the identification $i_j\mapsto j$. We denote the resulting sampling distribution by $\hat{P}^{(k)}(\cdot|\omega)$. Note that for this sampling model we need to assume that we can observe the labels of the randomly drawn elements from $\omega$ in order to determine their order and to define the re-labeling. Observe, too, that $\hat{P}^{(k)}(\cdot|\omega)$ in general is not exchangeable: if, e.g. $\omega$ is a linear chain $1\rightarrow 2 \rightarrow 3 \cdots \rightarrow n$, then $\hat{P}^{(k)}( 1\circ\!\!\rightarrow\!\! \circ2 | \omega)=(n-1)/\binom{n}{2} \neq
\hat{P}^{(k)}( 1\circ\!\!\leftarrow\!\! \circ2 | \omega)=0$.
Using $\emph{Iso}(\omega)\subset {\Omega^{(n)}}$ to denote the isomorphism class of a world $\omega\in{\Omega^{(n)}}$, we obtain the following relationship between the sampling distributions $P^k$ and $\hat{P}^{(k)}$: $$\label{eq:PandPhat}
{P}^{(k)}(\omega'|\omega)= \frac{1}{|\emph{Iso}(\omega')|}\sum_{\omega''\in\emph{Iso}(\omega')} \hat{P}^{(k)}(\omega''|\omega).$$ If $\hat{P}(\cdot|\omega)$ is exchangeable, then $\hat{P}(\cdot|\omega) ={P}(\cdot|\omega) $.
Proofs for Section \[sec:reptheo\]
==================================
In the following proofs we have to consider operations on different sets that are defined by permutations of $[n]$. We here list in detail the pertinent definitions.
Let $n\in\Nset$. A permutation $\pi$ of $[n]$ operates on the following sets as follows:
A
: on $[\Nset]_{\neq}^n$: for $\boldk=(k_1,\ldots,k_n)\in [\Nset]_{\neq}^n$ define $\pi \boldk=(k_{\pi(1)},\ldots,k_{\pi(n)})$.
B
: on $[n]^k$ with $k\geq 1$: for $\boldi=(i_1,\ldots,i_k)\in [n]^k$ define $\pi\boldi=(\pi(i_1),\ldots,\pi(i_k))$.
C
: on ${\Omega^{(n)}}$: let $\omega\in{\Omega^{(n)}}$. $\pi\omega$ then is the world in which for a $k$-ary $r\in S$: $A_r(\pi\omega)(\boldi)=1$ iff $A_r(\pi^{-1}\boldi)(\omega)=1$ ($\boldi\in[n]^k$).
D
: on ${\cal T}^n$: let $t\in {\cal T}^n$ and $r\in S$ with $\emph{arity}(r)=k\geq n$. Define $\pi t$ by: $(A_r|n)(\boldi)=1$ in $\pi t$ iff $(A_r|n)(\pi^{-1}\boldi)=1$ in $ t$ ($\boldi\in [n]^k_n$). Intuitively, $\pi t$ is obtained from $t$ by re-labeling the nodes of $t\in {\cal T}^n$ according to $\pi$ (cf. Figure \[fig:worlddefs\]).
E
: on ${2^{[n]}}$: for $\boldi=\{i_1,\ldots,i_l\}\subseteq [n]$ define $\pi\boldi:=\{\pi i_1,\ldots,\pi i_l\}$
F
: on $[2^n]$: for $\boldi\in 2^{[n]}$ let $\emph{Idx}(\boldi)$ be the index of $\boldi$ in the lexicographic ordering of $2^{[n]}$. Then, for $1\leq j\leq 2^n$: $\pi j:=\emph{Idx}(\pi \emph{Idx}^{-1}(j) )$, where $\pi \emph{Idx}^{-1}(j) $ is defined according to ${\bf E}$.
G
: on $[0,1]^{2^n}$: let $\boldx = (x_1,x_2,\ldots,x_{2^n}) \in [0,1]^{2^n} $. Then $\pi\boldx:=(x_{\pi^{-1}1},\ldots,x_{\pi^{-1}2^n})$.
According to the above definitions, an expression $\pi\boldk$ is potentially ambiguous, and depends on whether $\boldk$ is seen as an element of $[\Nset]_{\neq}^n$, $[n]^k$, or ${2^{[n]}}$. In the following, the correct reading of such expressions, if not spelled out explicitly, will be implied by the context in which the expression appears.
Now let $\pi$ be a permutation of $[n]$, let $m<n$, and $\boldi\in\langle n\rangle^m$. Then $\boldi=(i_1,\ldots,i_h)$ induces a permutation $\pi_{\boldi}$ of $[m]$ as follows: $$\label{eq:pii}
\pi_{\boldi}: h \mapsto \emph{Idx}(\pi(i_h))\hspace{10mm} (1\leq h\leq m),$$ where $ \emph{Idx}(\pi(i_h))$ is the index of $\pi(i_h)$ in the ordering of the set $\{\pi(i_h)|1\leq h\leq m \}$. The permutation $\pi_{\boldi}$ then again induces operations [**A**]{}-[**F**]{} on $[\Nset]^m_{\neq}$, …, $[0,1]^{2^m}$ according to the previous definitions.
\[ex:inducedperm\] Let $n=4$ and $\pi: 1\mapsto 3$, $2\mapsto 1$, $3\mapsto 4$, $4\mapsto 2$. Let $m=2$ and $\boldi=(1,4)$. Then $\pi\boldi =(2,3)$, and $\pi_{\boldi}: 1\mapsto 2$, $2\mapsto 1$. Note that $\pi\boldi$ being just a set (ordered tuple) does not contain all information on how $\pi$ is defined on the components of $\boldi$. However, jointly, $\pi\boldi$ and $\pi_{\boldi}$ are sufficient to reconstruct the restriction of $\pi$ to the components of $\boldi$.
[of Proposition \[prop:ahkprojective\]]{} Let $S$ be a signature with $\emph{arity}(S)=a$. Let $n\in\Nset$. Let $\omega\in{\Omega^{(n)}}$ and $\pi$ a permutation of $[n]$. We have to show that $$\label{eq:local050}
\boldf\downarrow[n](\omega)=\boldf\downarrow[n](\pi\omega).$$ Let $\boldU_{[n]}$ denote the set of $U_{\boldi}$ variables with $\boldi\in \langle n\rangle^m$ ($m\leq a$). We can then define $$\label{eq:deffomega}
\boldf^{-1}(\omega):= \{\boldu\in [0,1]^{\boldU_{[n]}}: \boldf(\boldu)=\omega\}$$ where $\boldf(\boldu)=\omega$ is shorthand for: for all $m\leq a, \boldi=(i_1,\ldots,i_m)\in \langle n\rangle^m$: $$\label{eq:local055}
f^m(\boldu_{\boldi}) = D_m(\omega\downarrow\boldi),$$ where $\boldu_{\boldi}=(u_{\emptyset},u_{i_1},\ldots,u_{(i_1,\ldots,u_m)})$. We now need to verify that $$\label{eq:local060}
\boldf({\boldu})=\omega \Leftrightarrow
\boldf(\pi{\boldu})=\pi\omega,$$ where $\pi\boldu$ is defined as the operation of $\pi$ on $[0,1]^{2^n}$ (case [**G**]{}) by viewing $\boldu\in[0,1]^{\boldU_{[n]}}$ as an elemenent of $[0,1]^{2^n}$ via the lexicographic ordering of the $u_{\boldi}\in\boldu$. This then proves (\[eq:local050\]) because the joint distribution of the $U_{\boldi}$ variables is invariant under the permutation $\pi$. The right-hand side of (\[eq:local060\]) can be re-written in the form of (\[eq:local055\]) as: for all $m\leq a$, $\boldi\in\langle n\rangle^m$: $$\label{eq:local070}
f^m((\pi\boldu)_{\boldi})=D_m(\pi\omega\downarrow\boldi)$$ Since $\pi$ operates as a bijection on $ \langle n\rangle^m$, we can change the variable quantified over from $\boldi$ to $\pi\boldi$, and obtain the equivalent condition that for all $\boldi$ $$\label{eq:local073}
f^m((\pi\boldu)_{\pi\boldi})=D_m(\pi\omega\downarrow\pi\boldi)$$ Now with $$\begin{aligned}
\label{eq:local075}
& & (\pi\boldu)_{\pi\boldi} = \pi_{\boldi}\boldu_{\boldi}\\
\label{eq:local078}
& & D_m(\pi\omega\downarrow\pi\boldi)=\pi_{\boldi} D_m(\omega\downarrow\boldi)
\end{aligned}$$ and substituting (\[eq:local055\]) in the right-hand side of (\[eq:local078\]) we obtain that (\[eq:local060\]) is equivalent that for all $\boldi$ and all $t\in {\cal T}^m$ $$\label{eq:local080}
f^m(\pi_{\boldi}\boldu_{\boldi})=\pi_{\boldi} f^m(\boldu_{\boldi}),$$ which is exactly the permutation equivariance of $f^m$.
The projectivity of the family $(\boldf\downarrow [n])_n$ follows directly from the graphical structure of the AHK model.
Let $n$ and $\pi$ as in Example \[ex:inducedperm\]. Let $S$ contain a single binary relation symbold, and let $\omega\in\Omega^{(4)}$ be the world with edges $(1,2),(2,4),(4,1)$. Then $\pi\omega$ is the world with edges $(1,2),(2,3),(3,1)$. Let $$\boldu=(u_{\emptyset},u_1,\ldots,u_4,u_{(1,2)},\ldots,u_{(2,3,4)},u_{(1,2,3,4)})\in [0,1]^{2^4}$$ such that $\boldf(\boldu)=\omega$. In particular for $\boldi=(1,4)$: $$f^2(\boldu_{\boldi})=f^2((u_{\emptyset},u_1,u_4,u_{(1,4)})= 1\circ \leftarrow \circ 2.$$ Then $\pi\boldi$ and $\pi_{\boldi}$ are as in Example \[ex:inducedperm\], and $$(\pi\boldu)_{\pi\boldi}=(u_{\emptyset},u_4,u_1,u_{(1,4)})=\pi_{\boldi}(\boldu_{\boldi}).$$ Permutation equivariance of $f^2$ then gives $$\begin{gathered}
f^2((u_{\emptyset},u_4,u_1,u_{(1,4)})) = \\
\pi_{\boldi} f^2((u_{\emptyset},u_1,u_4,u_{(1,4)})) = 1\circ \rightarrow \circ 2.
\end{gathered}$$
[of Theorem \[theo:extendable2\]]{}
$({\bf A})\Rightarrow ({\bf E})$: The proof relies on adaptations of concepts and results introduced for the study of graph limits [@borgs2008convergent; @diaconis2007graph], starting with the definition of the function ${\tau_{\rm ind}}$ as follows: let ${\cal O}:=[0,1]^{\Omega}$ (to be thought of as the space of probability assignments to all possible worlds). Then let $$\label{eq:taufunctional}
\begin{array}{lll}
{\tau_{\rm ind}}: & \Omega & \rightarrow {\cal O} \\
& \omega & \mapsto ({P^{(n)}}(\omega'|\omega))_{\omega'\in{\Omega^{(n)}},n\geq 1}
\end{array}$$ where we define ${P^{(n)}}(\omega'|\omega)=0$ when the size of $\omega$ is $<n$. Thus, for any $\omega\in\Omega$ the vector ${\tau_{\rm ind}}(\omega)$ has only finitely many non-zero components. For $o\in{\cal O}$ we denote with $o[\omega']$ the $\omega'$ component of $o$, and with $o[{\Omega^{(n)}}]$ the tuple of components for $\omega'\in {\Omega^{(n)}}$. Let $(\omega_i)_{i\in \Nset}$ be an enumeration of $\Omega$. The space ${\cal O}$ then can be equipped with the metric $d(o,o'):=\sum_i \frac{1}{2^i} d(o[\omega_i],o'[\omega_i])$, where $d$ on the right-hand side is the standard Euclidean metric on $[0,1]$. With this metric, ${\cal O}$ is a compact metric space.
Now assume that $Q^{(k)}$ is a domain sampling distribution. Let $\epsilon_n\rightarrow 0$ ($n\rightarrow\infty$). For each $n$, let $\omega_n$ be such that (\[eq:realizable\]) holds with $\omega=\omega_n$ and $\epsilon=\epsilon_n$, and such that $\omega_n\in \Omega^{(n')}$ with $n'\geq n$. The sequence $({\tau_{\rm ind}}(\omega_n))_n$ contains a convergent sub-sequence. Let $o^* \in{\cal O}$ be its limit. We claim that $o^*$ defines a projective family that extends $Q^{(k)}$:
- For each $m$: $o^*[{\Omega^{(m)}}]$ is a probability distribution on ${\Omega^{(m)}}$. Let ${\cal O}^{(m)}_{\emph{prob}}$ be the subset of $o\in {\cal O}$ for which $o[{\Omega^{(m)}}]$ is a probability distribution. Then ${\cal O}^{(m)}_{\emph{prob}}$ is closed, and ${\tau_{\rm ind}}(\omega_n)\in {\cal O}^{(m)}_{\emph{prob}}$ for all $n>m$. Thus, also $o^*\in {\cal O}^{(m)}_{\emph{prob}}$.
- For $m'>m$: $o^*[\Omega^{m'}]\downarrow m = o^*[{\Omega^{(m)}}]$. Let ${\cal O}^{(m,m')}_{\emph{proj}}$ be the subset of $o\in {\cal O}$ for which $o[\Omega^{m'}]\downarrow m = o[{\Omega^{(m)}}]$. Again, this is a closed set, ${\tau_{\rm ind}}(\omega_n)\in {\cal O}^{(m,m')}_{\emph{proj}}$ for all $n\geq m'$, and therefore also $o^*\in {\cal O}^{(m,m')}_{\emph{proj}}$.
- For each $m\geq k$: $o^*[{\Omega^{(m)}}]\downarrow k = Q^{(k)}$: this follows from the definition of the sequence $\omega_n$.
([**B**]{})$\Rightarrow$ ([**A**]{}): Let $\boldf$ be an AHK$^-$ model that represents ${Q^{(k)}}$. Let $n\geq k$ and consider $\boldf\downarrow [n]$. Let $\boldD$ be the vector of all $D_{\boldi}$-variables with $\boldi\subset [n]$, i.e., $\boldD$ represents a random $\omega\in{\Omega^{(n)}}$ drawn according to $\boldf\downarrow [n]$. For any $\boldk\in\langle n\rangle^k$ let $\boldD_{\boldk}=(D_{\boldi})_{\boldi\in\langle n\rangle^m:\boldi\subseteq\boldk;m=1,\ldots,a}$ be the vector of variables that represent the worldlet induced by $\boldk$. Now consider a fixed $\omega'\in{\Omega^{(k)}}$ and the random variables ${{\mathbbm 1}[\boldD_{\boldk}=\omega'] }$ where ${{\mathbbm 1}[] }$ is the indicator function. Since $\boldf$ represents ${Q^{(k)}}$, we have that $$\label{eq:local90}
E[{{\mathbbm 1}[\boldD_{\boldk}=\omega'] }]={Q^{(k)}}(\omega').$$ Taking the average over all $\boldk\in\langle n\rangle^k$, we obtain the random variable $$\label{eq:local100}
\frac{1}{\binom{n}{k}}\sum_{\boldk\in\langle n\rangle^k} {{\mathbbm 1}[\boldD_{\boldk}=\omega'] }=
\hat{P}^{(k)}(\omega'|\boldD),$$ with $\hat{P}$ as introduced in Appendix \[sec:altsampling\]. From (\[eq:local90\]) it follows that still $$\label{eq:local105}
E[\hat{P}^{(k)}(\omega'|\boldD)]={Q^{(k)}}(\omega').$$
The variables $\boldD_{\boldk}$ and $\boldD_{\boldk'}$ are independent according to $\boldf\downarrow [n]$ whenever $\boldk\cap\boldk'=\emptyset$(note that this is not the case in the AHK model, where the $U_{\emptyset}$ variable can induce a dependency). The family of random variables ${{\mathbbm 1}[\boldD_{\boldk}=\omega'] }$ ($\boldk\in \langle n\rangle^k$) has the same (weak) dependence structure as found in U statistics [@hoeffding1963probability]. The probability bound given by Hoeffding and extended to more general scenarios by Janson (cf. Corollary 2.2 and Example 4.1 in [@janson2004large]) provide the following bound for the distribution of the random variable defined by (\[eq:local100\]): for $t>0$ $$\label{eq:local110}
P_{\boldD}( \hat{P}^{(k)}(\omega'|\boldD) \geq {Q^{(k)}}(\omega')+t ) \leq e^{-2 \lfloor \frac{n}{k} \rfloor t^2},$$ where $P_{\boldD}$ denotes the distribution defined by $\boldf\downarrow[n]$ on $\boldD$. The same bound holds for $P_{\boldD}( \hat{P}^{(k)}(\omega'|\boldD) \leq {Q^{(k)}}(\omega')-t )$.
Now let $\epsilon>0$ as in Definition \[def:dsrealizable\] be given. Then, for any $n\in\Nset$ $$\label{eq:local120}
2\cdot|{\Omega^{(k)}}|\cdot e^{-2 \lfloor \frac{n}{k} \rfloor \epsilon^2}$$ is an upper bound for the probability that for some $\omega'\in{\Omega^{(k)}}$ the frequency $\hat{P}^{(k)}(\omega'|\boldD)$ is outside the bounds ${Q^{(k)}}(\omega')\pm \epsilon$. For all sufficiently large $n$, this bound is $<1$, which, in particular means that for all sufficiently large $n$ there exist $\omega_n\in{\Omega^{(n)}}$ with $$\label{eq:local130}
\hat{P}^{(k)}(\omega'|\omega_n)={Q^{(k)}}(\omega')\pm \epsilon$$ for all $\omega'\in{\Omega^{(k)}}$. Since ${Q^{(k)}}$ is exchangeable, it follows from (\[eq:PandPhat\]) that (\[eq:local130\]) also holds when $\hat{P}^{(k)}$ is replaced by $ {P}^{(k)}$.
([**C**]{})$\Rightarrow$ ([**E**]{}): the set of projective extendable ${Q^{(k)}}$ is convex. The implication then follows from $({\bf A})\Rightarrow ({\bf E})$.
([**D**]{})$\Rightarrow$ ([**C**]{}): The set of extendable distributions is the closed convex set ${\cal P}^{(k)}:=\cap_n \Delta^{(k)}_{n}$. Since ${\cal P}^{(k)}$ has finite dimension, each point of ${\cal P}^{(k)}$ is a finite convex combination of its extreme points. Let $p$ be an extreme point of ${\cal P}^{(k)}$. We show that $p$ is a domain-sampling distribution. Each $\Delta^{(k)}_{n}$ is a polytope whose vertices are statistical frequency distributions ${P^{(k)}}(\cdot|\omega)$ for some $\omega\in{\Omega^{(n)}}$. With Lemma \[lem:convex\] it follows that $p$ is the limit of a convergent sequence of distributions ${P^{(k)}}(\cdot|\omega_n)$, and thus a domain sampling distribution.
([**E**]{})$\Rightarrow$ ([**F**]{}): Let $\Omega^{(\Nset)}$ denote the set of infinite worlds over the node set $\Nset$. Let $\{{Q^{(n)}}| n\geq 1\}$ be a projective family extending ${Q^{(k)}}$. By the standard Kolmogorov existence theorem and the exchangeability of the ${Q^{(n)}}$, there exists a distribution $Q^{(\Nset)}$ on $\Omega^{(\Nset)}$ such that $Q^{(\Nset)}\downarrow \boldk ={Q^{(n)}}$ for any $\boldk\in [\Nset]_{\neq}^n$ ($n\geq 1$). Let $W_{\Nset}$ be a $Q^{(\Nset)}$-distributed random variable, and for $\boldk\in [\Nset]_{\neq}^m$ with $m\in [a]$ let $$\label{eq:DkDef}
\tilde{D}_{\boldk}:=D_m(W_{\Nset}\downarrow\boldk).$$ The variables $( \tilde{D}_{\boldk})_{\boldk}$ then form an $a$-dimensional exchangeable array in the sense of [@kallenberg2006probabilistic]. Even though we will later identify the variables $\tilde{D}_{\boldk}$ with the $D_{\boldk}$-variables of our target AHK representation, we distinguish them for now via the $\tilde{D}$ notation. According to the representation theorm [@kallenberg2006probabilistic Lemma 7.25] the joint distribution of the $(\tilde{D}_{\boldk})_{\boldk}$ has a representation given by
- a family of i.i.d. random variables $\{U_{\boldi}| \boldi\in \langle \Nset \rangle^m, m=0,\ldots,a\}$ as described in Definition \[def:ahkmodel\];
- functions $${f}^m:[0,1]^{2^m}\rightarrow\Omega^{(m)}\ (m=1,\ldots,a)$$
such that for $\boldk$ of length $m$: $$\label{eq:Dkvars}
\tilde{D}_{\boldk}={f}^m(\boldU_{\boldk})$$ where the arguments $\boldU_{\boldk}$ consist of all $U_{\boldk'}$ with $\boldk'\subseteq \boldk$ enumerated according to the lexicographic ordering of the indices of $\boldk'$ in $\boldk$. In detail: if $\boldk=(k_1,\ldots,k_m)$, $\boldk'=\{k_{i'_1},\ldots,k_{i'_{d'}}\}$, $\boldk''=\{k_{i''_1},\ldots,k_{i''_{d''}}\}$ with $i'_1<\cdots<i'_{d'}$ and $i''_1<\cdots<i''_{d''}$, then $U_{\boldk'}$ precedes $U_{\boldk''}$ in $\boldU_{\boldk}$ if $(i'_1,\ldots,i'_{d'})$ precedes $(i''_1,\ldots,i''_{d''})$ lexicographically. Note that when $\boldk$ is ordered as in Definition \[def:ahkmodel\], the lexicographic ordering of subsets $\boldk',\boldk''\subseteq\boldk$ according to their indices in $\boldk$ corresponds to the lexicographic ordering of the tuples $\boldk',\boldk''$ themselves.
To transform this representation into the form stated in Definition \[def:ahkmodel\] we exploit the fact that the family $(\tilde{D}_{\boldk})_{\boldk}$ not only is exchangeable, but also exhibits certain deterministic permutation invariance relationships.
Now let $\pi$ be a permutation of $[m]$. Let $\boldk\in[\Nset]_{\neq}^m$. For the variables $\tilde{D}_{\boldk}$ of (\[eq:Dkvars\]) we then have $$\label{eq:DpikDk}
\tilde{D}_{\pi\boldk}= D_m(W_{\Nset}\downarrow \pi\boldk) = \pi(\tilde{D}_{\boldk}).$$ Thus, it is sufficient to consider the $\tilde{D}_{\boldk}$ for ordered $\boldk$, and we now let for $\boldk\in\langle\Nset\rangle^m\subset [\Nset]^m_{\neq}$: $$\label{eq:defproofDk}
D_{\boldk}:=\tilde{D}_{\boldk}.$$ According to (\[eq:Dkvars\]) $$\begin{array}{l}
\tilde{D}_{\pi\boldk}= f^m( \boldU_{\pi\boldk} ) \\
\pi(\tilde{D}_{\boldk}) = \pi f^m( \boldU_{\boldk} )
\end{array}$$ Together with $$\boldU_{\pi\boldk}=\pi\boldU_{\boldk},$$ then (\[eq:DpikDk\]) gives the permutation equivariance of $f^m$: $$f^m(\pi\boldU_{\boldk})=\pi(f^m(\boldU_{\boldk})).$$
([**A**]{})$\Rightarrow$ ([**B**]{}): Let ${Q^{(k)}}$ be a domain sampling distribution. From ([**A**]{})$\Rightarrow$ ([**E**]{} )$\Rightarrow$ ([**F**]{}) we know that ${Q^{(k)}}$ has an AHK representation $\boldf$. For the purpose of this proof we have to be careful to properly distinguish between the functions $\boldf$, and the distribution they define. We write $P_{\boldf}$ for the distribution defined by $\boldf$ (undoing the slight abuse of notation introduced after the statement of Definition \[def:ahkmodel\]). When, in the following, we refer to the marginal distribution of the continuous variable $U_{\emptyset}$, then expressions of the form $P_{\boldf}(U_{\emptyset}=x|A)$ stand for the value at $x\in[0,1]$ of the density function representing the conditional distribution of $U_{\emptyset}=x$ given the event $A$. Since ${Q^{(k)}}$ is a domain sampling distribution we have that for tuples $\boldk,\boldk'\in\langle \Nset \rangle^m$ with $\boldk\cap\boldk'=\emptyset$ the variables $D_{\boldk},D_{\boldk'}$ are independent: $$\label{eq:local400}
P_{\boldf}(D_{\boldk'}|D_{\boldk})=P_{\boldf}(D_{\boldk'}).$$ We first show that the $D_{\boldk}$ are independent of $U_{\emptyset}$. Assume otherwise. Then there exists a $t\in{\cal T}^m$ such that $P_{\boldf}(D_{\boldk}=t|U_{\emptyset}=x)$ is not almost surely constant for $x\in [0,1]$. We then obtain a contradiction to (\[eq:local400\]) by first expanding: $$\begin{gathered}
P_{\boldf}(D_{\boldk'}=t|D_{\boldk}=t)=\\
\int_{[0,1]} P_{\boldf}(U_{\emptyset}=x|D_{\boldk}=t) P_{\boldf}(D_{\boldk'}=t|U_{\emptyset}=x,D_{\boldk}=t)dx.
\end{gathered}$$ Using the conditional independenc of $D_{\boldk},D_{\boldk'}$ given $U_{\emptyset}$, this simplifies to: $$\begin{gathered}
\int_{[0,1]} P_{\boldf}(U_{\emptyset}=x|D_{\boldk}=t) P_{\boldf}(D_{\boldk'}=t|U_{\emptyset}=x)dx.
\end{gathered}$$ With Bayes’s rule and $P_{\boldf}(U_{\emptyset}=x)=1$ this becomes: $$\begin{gathered}
\int_{[0,1]} \frac{P_{\boldf}(D_{\boldk}=t|U_{\emptyset}=x)}{P_{\boldf}(D_{\boldk}=t)} P_{\boldf}(D_{\boldk'}=t|U_{\emptyset}=x)dx, \\
\end{gathered}$$ and using $P_{\boldf}(D_{\boldk}=t|U_{\emptyset})=P_{\boldf}(D_{\boldk'}=t|U_{\emptyset})$ and $P_{\boldf}(D_{\boldk}=t)=P_{\boldf}(D_{\boldk'}=t)$ we finally obtain: $$\label{eq:local430}
P_{\boldf}(D_{\boldk'}=t|D_{\boldk}=t)=\int_{[0,1]} \frac{P_{\boldf}(D_{\boldk'}=t|U_{\emptyset}=x)^2}{P_{\boldf}(D_{\boldk'}=t)}dx$$ Expanding also the right-hand side of (\[eq:local400\]) we can write: $$\label{eq:local435}
P_{\boldf}(D_{\boldk'}=t)=\int_{[0,1]} P_{\boldf}(D_{\boldk'}=t|U_{\emptyset}=x)dx.$$ We then obtain $$\label{eq:local440}
P_{\boldf}(D_{\boldk'}=t|D_{\boldk}=t)>P_{\boldf}(D_{\boldk'}=t)$$ by multiplying both (\[eq:local430\]) and (\[eq:local435\]) with $ P_{\boldf}(D_{\boldk'}=t)$, and an application of Jensen’s inequality to the random variable $P_{\boldf}(D_{\boldk'}=t|U_{\emptyset})$ and the convex function $x\mapsto x^2$. The inequality is strict, because of the strict convexity of the square function, and the assumption that $P_{\boldf}(D_{\boldk'}=t|U_{\emptyset})$ is not almost surely constant.
The above argument can be carried out in exactly the same manner (only with a heavier load of notation) to also show that every finite family $D_{\boldk_1},\ldots,D_{\boldk_n}$ ($\boldk_i\in \langle \Nset\rangle^{m_i}, m_i\leq a, n\geq 1$) is independent of $U_{\emptyset}$. It then follows that in the AHK model representing ${Q^{(k)}}$ the node $U_{\emptyset}$ is redundant and can be eliminated. It remains to show that the conditional distributions of the $D_{\boldi}$ on the remaining $U_{\boldi'}$ variables $(\emptyset\neq\boldi'\subseteq\boldi)$ can still be represented by deterministic measurable, permutation-equivariant functions $f^m$.
Let $A(D_{\boldk_1},\ldots,D_{\boldk_n})\subseteq [0,1]$ be such that $P(A(D_{\boldk_1},\ldots,D_{\boldk_n}))=1$ (under the uniform distribution on $[0,1]$), and $$\label{eq:local450}
P_{\boldf}(D_{\boldk_1},\ldots,D_{\boldk_n})= P_{\boldf}(D_{\boldk_1},\ldots,D_{\boldk_n}|U_{\emptyset}=x)$$ for all $x\in A(D_{\boldk_1},\ldots,D_{\boldk_n}) $. Since there are only countably many finite families of $D_{\boldk_i}$, also the intersection of the sets $A(D_{\boldk_1},\ldots,D_{\boldk_n})$ for all such families has probability 1, and, in particular, there exists an $x^*\in[0,1]$ such that (\[eq:local450\]) holds with $x=x^*$ for all families. For every $m\leq a$ now define $$\tilde{f}^m(x_1,\ldots,x_{2^m-1}):= {f}^m(x^*,x_1,\ldots,x_{2^m-1}).$$ The $\tilde{f}^m$ (in conjunction with uniform distributions on all $U_{\boldi}$ ($\boldi\neq\emptyset$)) then represent the original distribution $P_{\boldf}$ (marginalized on all variables other than $ U_{\emptyset}$), and inherit the permutation equivariance of the $f^m$. Note that the operation of permutations $\pi$ on $[0,1]^{2^m}$ leaves the first component invariant, and therefore can also be directly seen as permutations on the domains $[0,1]^{2^m-1}$ of the $\tilde{f}^m$.
([**F**]{})$\Rightarrow$ ([**E**]{}): this follows from Proposition \[prop:ahkprojective\].
\[lem:convex\] For $n\geq 1$ let ${\cal P}_n$ be a polytope in $\Rset^k$, such that ${\cal P}_{n+1}\subseteq {\cal P}_n$ for all $n$. Let ${\cal P}:=\cap_n {\cal P}_n$, and $p$ an extreme point of ${\cal P}$. Then there exists for each $n$ a vertex $p_n$ of ${\cal P}_n$ such that $\lim_{n\rightarrow\infty} p_n =p$.
It follows from Straszewicz’s theorem that it is sufficient to consider the case that $p$ is an exposed point of ${\cal P}$, i.e., there exists a hyperplane $H$, such that ${\cal P}\cap H =\{p\}$. Let $H^+$ be the closed half-space defined by $H$ for which also ${\cal P}\cap H^+ =\{p\}$. For each $n$ then ${\cal Q}_n:={\cal P}_n \cap H^+$ is a polytope, and $\{p\}=\cap_n {\cal Q}_n$. Thus, for any sequence of vertices $q_n\in {\cal Q}_n$ we have $\lim_n q_n = p$. To conclude the proof we have to show that each ${\cal Q}_n$ has at least one vertex that is also a vertex of the original ${\cal P}_n$. For this we consider two cases: first, assume that ${\cal P}_n\cap H^+\setminus H \neq \emptyset$. Then every vertex of ${\cal P}_n$ that lies within $H^+\setminus H $ also is a vertex of ${\cal Q}_n$, and there exists at least one such vertex. Second, assume that ${\cal P}_n\cap H^+\subset H$. Then $H$ is a supporting hyperplane of ${\cal P}_n$, and all vertices of ${\cal P}_n\cap H$ are also vertices of ${\cal P}_n$.
Proposition \[prop:modularity\]: Correction and Proof
=====================================================
Proposition \[prop:modularity\] as stated in the paper contained a typo and an incorrect statement at the end. A corrected version of the theorem is as follows:
\[prop:modularity\_update\] Let $\boldf$ be an AHK model, $\omega\in{\Omega^{(n)}}$ with $p:=\boldf\downarrow[n](\omega)>0$. Let $O\subset\Omega^{(n+1)}$ be the set of $n+1$ worlds $\omega'$ for which $\omega'\downarrow [n]=\omega'\downarrow (1,\ldots,n-1,n+1 ) =\omega$. Then $\boldf\downarrow[n](O)\geq p^2$.
The corrected typo is a replacement of $\downarrow \{\ldots\}$ by $\downarrow (\ldots)$. The conditional independence claimed in the original version of the proposition does, in fact, not even hold in AHK$^-$ models. Proposition \[prop:modularity\_update\] sharpens the main statement of the original proposition slightly by providing the quantitative bound $\geq p^2$. These corrections do not affect Figure \[fig:modular\], or the application of the proposition to the ’+’ distribution of Table \[tab:wdistributions\].
Let $I=\cup_{m=0}^{\emph{arity}(S)} \langle n+1 \rangle^m$, so that the distribution over worlds $\omega'\in\Omega^{(n+1)}$ is modeled by the random variables $\{U_{\boldi},D_{\boldi}| \boldi\in I\}$. We partition $I$ into 4 subsets as follows: $I_1$ contains the $\boldi\subseteq [n-1]$ (including $U_{\emptyset}$). $I_2 $ contains the $\boldi$ which contain $n$, but not $n+1$; $I_3$ contains the $\boldi$ which contain $n+1$, but not $n$, and $I_4$ contain the $\boldi$ which contain both $n$ and $n+1$. For $k=1,\ldots, 4$ we write $\boldU_k,\boldD_k$ for the sets of variables $U_{\boldi},D_{\boldi}$ with $\boldi\in I_k$, and $\boldu_k$ to denote a set of values for $\boldU_k$. We can then decompose $\omega'\in\Omega^{(n+1)}$ as $$\label{eq:omegadecomp}
\omega'=(\omega^-,\rho_2,\rho_3,\rho_4)$$ where $\omega^-=\omega'\downarrow [n-1]$ is determined by the variables $\boldD_1$, and $\rho_k$ is the sub-structure defined by the variables $\boldD_k$ ($k=2,3,4$). Unlike the first component $\omega^-$, these sub-structures are not possible worlds in their own right. Now, let $\omega$ be as given in the proposition. Limiting the decomposition (\[eq:omegadecomp\]) to ${\Omega^{(n)}}$, we can write $\omega=(\omega^-,\rho_2)$. From now on let $\omega^-,\rho_2$ be the fixed structures defined by $\omega$, and let $\rho_3$ be the sub-structure that is isomorphic to $\rho_2$ via the re-naming $n+1\mapsto n$. Then $\omega'\in O$ iff the first three components in its decomposition are $(\omega^-,\rho_2,\rho_3)$.
The variables $\boldD_2$ that define $\rho_2$, for example, are a function of $\boldU_1$ and $\boldU_2$. We write $\boldf(\boldu_1,\boldu_2)=\rho_2$ when the values $\boldu_1,\boldu_2$ induce the substructure $\rho_2$. Similarly for other substructures. Using again ${{\mathbbm 1}[] }$ to denote the indicator function, we can now write $$\begin{gathered}
\label{eq:local600}
p=\int_{[0,1]^{\boldU_1}}\int_{[0,1]^{\boldU_2}}{{\mathbbm 1}[\boldf(\boldu_1)=\omega^-] }\cdot \\
{{\mathbbm 1}[\boldf(\boldu_1,\boldu_2)=\rho_2] }d\boldu_2 d\boldu_1.
\end{gathered}$$ and $$\begin{gathered}
\label{eq:local610}
P(\boldf^{-1}(O)) =
\int_{[0,1]^{\boldU_1}}\int_{[0,1]^{\boldU_2}}\int_{[0,1]^{\boldU_3}}{{\mathbbm 1}[\boldf(\boldu_1)=\omega^-] }\cdot \\
{{\mathbbm 1}[\boldf(\boldu_1,\boldu_2)=\rho_2] } \cdot
{{\mathbbm 1}[\boldf(\boldu_1,\boldu_3)=\rho_3] }d\boldu_3 d\boldu_2\boldu_1.
\end{gathered}$$ Since the sub-structures $\rho_2,\rho_3$ are ismorphic and given by the same functional dependency on $\boldu_1,\boldu_2$, respectively $\boldu_1,\boldu_3$, we can define: $$\begin{gathered}
\label{eq:local620}
F(\boldu_1):=
\int_{[0,1]^{\boldU_2}}
{{\mathbbm 1}[\boldf(\boldu_1,\boldu_2)=\rho_2] }d\boldu_2 = \\
\int_{[0,1]^{\boldU_3}}
{{\mathbbm 1}[\boldf(\boldu_1,\boldu_3)=\rho_3] }d\boldu_3.
\end{gathered}$$ Then (\[eq:local600\]) becomes $$\label{eq:local630}
\int_{[0,1]^{\boldU_1}} {{\mathbbm 1}[\boldf(\boldu_1)=\omega^-] } F(\boldu_1)d\boldu_1,$$ and (\[eq:local610\]) becomes $$\label{eq:local640}
\int_{[0,1]^{\boldU_1}} {{\mathbbm 1}[\boldf(\boldu_1)=\omega^-] } F^2(\boldu_1)d\boldu_1.$$ Applying Jensen’s inequality to the random variable ${{\mathbbm 1}[\boldf(\boldu_1)=\omega^-] } F(\boldu_1)$ then proves the proposition (observing that the square of the indicator function is the indicator function itself).
[^1]: Contact Author
[^2]: This Paper consists of an unaltered version of the paper published at IJCAI 2020 to which an appendix has been added. The bibliography is extended with a few additional references used in the appendix.
[^3]: An exception is the Boostr system [@Khot2013], which constructs first-order MLN formulas with constants.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose an exactly solvable model to reveal the physics of the interplay between interaction and disorder in bosonic systems. Considering interacting bosons in a double-well potential, in which disorder is mimicked by taking the energy level mismatch between the two wells to be randomly distributed, we find “two negatives make a positive" effect. While disorder or interaction by itself suppresses the phase coherence between the two wells, both together enhance the phase coherence. This model captures several striking features of the disordered Bose-Hubbard model found in recent numerical simulations. Results at finite temperatures may help explain why a recent experiment did not find any evidence for the enhancement of phase coherence in a disordered bosonic system.'
author:
- 'Qi Zhou and S. Das Sarma'
title: 'Bosons in a double-well potential: Understanding the interplay between disorder and interaction in a simple model '
---
Physics of disorder in both bosonic and fermionic systems has been attracting great interest from physicists for decades. While the effect of disorder on non-interacting particles can be well described by Anderson localization[@Anderson], understanding the interplay between interaction and disorder still remains challenging[@Review]. Recent developments in studying cold atoms provide new opportunities to reveal the nature of interacting quantum particles in the presence of disorder[@Reviewatoms]. In atomic systems, the strength of both interaction and disorder can be well controlled experimentally, unlike in solid state systems. There has been significant recent progress in studying disorder effects in both non-interacting and interacting atomic systems[@Aspect; @Ingusico1; @Ingusico2; @Demarco1; @Demarco2; @Demarco3; @Randy].
From the theoretical side, to understand even the spinless bosons in the presence of both interaction and disorder is a non-trivial problem. Scaling analysis, renormalization group theory and sophisticated numerical simulations are often used[@JS; @Scaling; @RG; @RG2; @NT; @RS; @QMC; @DMRG]. It is desirable in this context to have some exactly solvable models, which capture all the ingredients, such as interaction, disorder and finite temperature effects. In this letter, we propose and solve such a model, which makes it possible to reveal the underlying physics transparently. Our minimal model captures the interplay between interaction and disorder in bosonic systems, showing conclusively that weak disorder in the presence of interaction generically enhances the phase coherence of the system.
The model we propose describes interacting bosons in a disordered double-well potential. The Hamiltonian can be written as $$H=-t(b^\dagger_{L}b_R+c.c)+\frac{U}{2}\sum_{\sigma}n_{\sigma}(n_\sigma-1)+\frac{\epsilon}{2}(n_R-n_L),\label{h}$$ where $b^\dagger_L$($b_L$) and $b^\dagger_R$($b_R$) are the creation(annihilation) operators in the left and right well respectively, $\sigma=L,R$, $n_L=b^\dagger_Lb_L$($n_R=b^\dagger_Rb_R$) is the number operator in the left (right) well, $t$ is the tunneling amplitude between the two wells, $U>0$ is the onsite interaction, and $\epsilon$ is the energy level mismatch between the two wells. A special case of this model ($\epsilon\equiv0$) has been studied both theoretically and in cold atom experiments[@doublewell1; @doublewell2; @Ho]. In our case, $\epsilon$ is randomly distributed according to a certain probability function $P(\epsilon) $, thus simulating disorder. In experiments, this can be realized by randomly tilting the double-well potential. The thermodynamic quantities will be the ensemble averaged values. In this paper, we will focus on the case $P(\epsilon)=1/(2\Delta), \epsilon\in[-\Delta, \Delta] $, where $\Delta$ characterizes the disorder strength. Other distribution functions for $\epsilon$ do not change the qualitative conclusion presented in this paper. Eq.(\[h\]) can be viewed as a two site version of the extensively studied disordered Bose-Hubbard(BH) model with random onsite energies.
Despite the simplicity of this model, many interesting phenomena are found by solving this model with $N$ particles. An illuminating example is the case $\epsilon\equiv0 $. It has been shown that the ground state has a crossover from the coherent state to the Fock state when $U/t$ increases[@Ho], which mimics the well-known phase transition in the thermodynamic limit from a superfluid condensate to a Mott insulator state in the full BH model. The physics for either the crossover or the phase transition between the phases is the same in the minimal model and the BH model, namely, interaction suppresses number fluctuations and spatial phase coherence. In the case where $\epsilon$ is randomly distributed, we will see that the minimal model qualitatively captures the physics of interacting bosons in a disordered potential for a number of fundamental questions, even though it can not answer some quantitative questions, such as where there is a direct transition from Mott insulator to superfluid.
Before we discuss the details, we briefly summarize the main qualitative questions that we are going to answer in this paper. [**Q1**]{}: Whether disorder can enhance the phase coherence between the two wells, defined by ${C}=\langle b^\dagger_Lb_R\rangle/N=\langle b^\dagger_Rb_L\rangle/N$, in some parameter regimes, in contrast to the intuition that disorder must always destroy the phase coherence? [**Q2**]{}: Whether interaction can also enhance the phase coherence in the presence of disorder, though without disorder it is known that repulsive interaction usually suppresses phase coherence? [**Q3**]{}: In the clean system, $\epsilon\equiv0$, the compressibility of the system vanishes when the phase coherence is destroyed by interaction. Is this still true in the disordered case? [**Q4**]{}: How does the phase coherence between the two wells depend on the temperature? We will see that the answers to these four questions provide insight into some striking features of the quantum phase diagram of the disordered BH model. Our answers to these questions on the minimal model also shed light on a recent disagreement between the prediction of numerical simulations [@QMC; @DMRG]and experimental observations[@Demarco3].
[**The answer to Q1**]{}: To solve the disordered problem, we start from the case of $N$ particles in the double-well with a fixed $\epsilon$. We write the Schrödinger equation $H|\Psi\rangle=E|\Psi\rangle$ in the Fock space. Define $|l\rangle=|N_L,N_R\rangle=|\frac{N}{2}-l,\frac{N}{2}+l\rangle$, where $l=0,\pm 1, \pm 2,...\pm N/2$. For simplicity, we assume $N$ is a large even number. Expanding $|\Psi\rangle=\sum_l\psi_l|l\rangle$, we obtain $$(E-E_l)|l\rangle=-tM_{l,l+1}|l+1\rangle-tM_{l-1,l}|l-1\rangle,$$ where $E_l=U l^2+\epsilon l+\frac{U}{2}(N^2/2-N)$, $M_{l,l+1}=M_{l+1,l}=\sqrt{N/2(N/2+1)-l(l+1)}$. The eigenenergies and eigenfunctions can be easily computed by exact diagonalization. At zero temperature, the phase coherence between the two wells can be characterized as $C_{\epsilon}=\frac{1}{N}\langle b^\dagger_Lb_R\rangle_\epsilon=\frac{1}{N}\sum_lM_{l,l+1}\psi_l^0\psi_{l+1}^0$, where $\psi_l^0$ is the ground state wave function in the Fock space, $\langle {O} \rangle$ is the expectation value of the operator ${O}$ in the ground state, and the subscript implies a fixed $\epsilon$. The results for the disordered case can then be obtained by averaging those results for fixed $\epsilon$ according to the distribution function $P(\epsilon)$, $$\overline{\langle O\rangle}=\int_{-\Delta}^\Delta d\epsilon P(\epsilon){\langle O\rangle}_\epsilon.\label{ave}$$
The results for $C_\epsilon$ for both noninteracting case and finite $U$ are shown in Fig.(1A). The numerical results with the total particle number $N=100$ have been chosen to illustrate the physics, which does not depend on the exact value of the total particle number. When $U=0$, $C_\epsilon$ quickly decreases with increasing $\epsilon$, since noninteracting particles can only tunnel from one well to the other if the energy mismatch $\epsilon$ is smaller than the tunneling $t$. The behavior of $C_\epsilon$ as a function of $\epsilon$ depends explicitly on $U$, as shown in Fig.1. The dependence of $C_\epsilon$ on $\epsilon$ becomes more extended with increasing $U$, though $C_\epsilon(\epsilon=0)$ decreases with increasing $U$. More importantly, the curves become non-monotonic when the discretization of $E_l$ induced by interaction becomes significant for $U\gtrsim tN$. Local maxima of $C_\epsilon$ emerge at certain values of $\epsilon$. For those values $\epsilon^*=U( 1-2l^*)$, where $l^*$ is an integer number, $E_{l^*}=E_{l^*-1}$ is satisfied. In other words, the state $|N/2-l^*,N/2+l^*\rangle$ has the same energy with the one $|N/2-l^*+1,N/2+l^*-1\rangle$. There is no extra interaction energy cost for one particle tunneling from one well to the other. The tunneling will be enhanced at $\epsilon^*$, forming gently varied bumps on the curves for $C_\epsilon$. This non-monotonic behavior of $C_\epsilon$ becomes even more dramatic when $U$ increases. For large values of $U$, only in narrow regions near $\epsilon$, tunneling is strongly enhanced. As a result, resonance features emerge with sharp peaks located at $\epsilon^*$ sitting on the slowly decaying envelope of $C_\epsilon$. In the large $U$ limit, $E_l\gg E_{l^*}=E_{l^*-1}$ if $l\neq l^*$ or $l^*-1$, a two level approximation can be made, and it is straightforward to show that $C_{\epsilon^*}\sim M_{l^*-1,l^*}$. $C_{\epsilon^*}$ decreases slowly as $\epsilon^*$ increases, as shown in Fig.(1A).
![A: $C_\epsilon$ as a function of $\epsilon/tN$ at different interaction. B: $\overline{C}$ as a function of $\Delta/tN$. Dashed(purple), dash-double-dotted(green), solid(brown), dotted(blue), dash-dotted(red) represent $U/tN=0$, $0.3$, $1$, $2.5$, $9.95$ respectively. Insets are results for $C_\epsilon$ and $\overline{C}$ for $U/tN=0$ with the same labels as the main figures. []{data-label="fig: fig1"}](fig1){width="3.4in"}
Now we turn to the quantity $\overline{C}=\int_{-\Delta}^\Delta d\epsilon{C}_\epsilon$, which is related to the area between $\epsilon\in[-\Delta, \Delta]$ below the curve of $C_\epsilon$. When $U$ is small, averaging $C_{\epsilon}$ according to Eq.(\[ave\]) smooths out the small bumps on the curves. The non-monotonic behavior of $\overline{C}$ is largely suppressed, as shown by the solid(brown) curve in Fig.(1B). On the other hand, we have seen that the resonance feature of the curves becomes significant for large $U$. When $\Delta$ is not large enough, the contribution from regions under the peaks to the total area is dominant. As a result, peaks can still be clearly resolved for small values of $\Delta$, as shown by the dotted(blue) and dash-dotted(red) curves in Fig.(1B). For large $\Delta$, the peaks become wiggles on the top of the slowly decaying curves of $\overline{C}$. It is also quite clear that $\overline{C}$ first increases when the disorder strength $\Delta$ grows from zero, if $\Delta<\Delta^*$, where $\Delta^*=U$ is the location of the first peak. This fact shows the significant difference of the disorder effect between the non-interacting and interacting systems. In the presence of interaction, the disorder effect is non-monotonic.
![A contour plot of $\overline{C}$ as a function of $\Delta/tN$ and $U/tN$ at zero temperature with the values of contours marked down.[]{data-label="fig: fig2"}](fig2){width="1.8"}
It is helpful to look at the contours of $\overline{C}$ as a function of $\Delta/t$ and $U/t$ in order to obtain a complete picture. Fig.(2) shows that for large $U$, the contours first bend to the right hand side when $\Delta$ increases from zero. We note that the contours of $\overline{C}$ have wiggles at large $U$. These wiggles arise from the structures in $\overline{C}$ as discussed in the last paragraph. For a fixed value of $U$, away from the positions of the peaks, $\overline{C}$ changes slowly, corresponding to the parts of the contour which are nearly parallel to the $\Delta$ axis. When approaching the peaks, $\overline{C}$ quickly increases. As a result, the contours bend towards the $U$ axis, forming a wiggly shape. The topology of Fig.(2) is very similar to the nontrivial structure of the phase boundary obtained by recent numerical simulations for the disordered BH model at an integer filling[@QMC; @DMRG]. In the latter case, the phase boundary can be viewed as the contour for the order parameters $\langle b_i\rangle=0$ in the thermodynamic limit.
We emphasize that the similarity of the topology between Fig.(2) and the phase diagram of the disordered BH model is not accidental. The lattice model can be viewed as the thermodynamic limit of the two-site problem. Moreover, the physics of the interplay between disorder and interaction is the same in both cases. Interaction discretizes the energies of the Fock states in each site, and suppresses the tunneling between different sites. However, disorder introduces relative energy shifts of the Fock states at different sites. In a randomly distributed disorder potential of large enough strength, there are always possibilities for neighboring sites to have nearly degenerate Fock states. Effectively, the tunneling of the particles as well as the spatial phase coherence will then be enhanced. If the disorder strength increases further, the weight of those configurations favoring tunneling in all the configurations of the random potential decreases. The phase coherence is thus eventually suppressed by very strong disorder, as seen also in the numerical simulations of the full disordered BH model[@NT; @RS; @QMC; @DMRG]. It is worthwhile to point out that a quantitative understanding of the establishment of a long range order in the disordered BH mode requires taking into account long-range correlations beyond neighboring sites[@fullBH]. Nevertheless, our two-site model qualitatively reveals the underlying physics for the enhancement of phase coherence through weak disorder.
[**The answer to Q2**]{}: Having answered [**Q1**]{}, the answer to [**Q2**]{} becomes clear. Fig. (3A) shows the dependence of $\overline{C}$ on $U/t$ at fixed $\Delta/t$. For $\Delta=0$, interaction suppresses phase coherence monotonically. For $\Delta\neq0$, interaction first enhances phase coherence before suppressing it. These results are reminiscent of a similar behavior of the superfluid density as a function of $U/t$ in the full disordered BH model[@NT; @RS; @QMC]. In both cases, interaction screens the disordered potential, since some particles occupy the sites with lower on-site energies and thus smooth out the effective potential for the remaining particles. The spatial phase coherence is then enhanced. However, if interaction becomes very strong, interaction itself eventually destroys phase coherence or superfluid.
We have so far seen an interesting “two negatives make a positive" effect on the spatial phase coherence of a system owing to the interplay between interaction and disorder. With only disorder or interaction, the spatial phase coherence is suppressed either by the single particle localization or the emergence of Mott state. When both interaction and disorder are present, our exactly solvable model clearly shows that the spatial phase coherence is enhanced in the parameter regime where they are comparable in strength.
{width="3.4"}
[**The answer to Q3**]{}: To extract the compressibility $\kappa_\epsilon$, we first calculate the chemical potential, $\mu_\epsilon=(E^0_\epsilon(N+\delta N)-E^0_\epsilon(N))/\delta N$, where $E^0_\epsilon(N)$ is the ground state energy of $N$ particles at fixed $\epsilon$. From the dependence of $\mu_\epsilon$ on the particle number $N$, $\kappa_\epsilon=\partial N/\partial \mu_\epsilon$ can then be obtained. The ensemble averaged compressibility $\overline{\kappa}$ can be calculated in the same manner as $\overline{C}$, i.e., $\overline{\kappa}=\int_{-\Delta}^\Delta d\epsilon \kappa_\epsilon$. The dependence of $\overline{\kappa}$ on $\Delta$ is shown in Fig.(3B). For comparison, the result of $\overline{C}$ for the same $U$ is also shown.
An interesting feature of the compressibility in the minimal model is that it remains constant for large disorder strength, even though the phase coherence $C$ decreases for very large disorder. In the large $\epsilon$ limit, almost all the particles fall into one of the two wells, $E^0_\epsilon\rightarrow UN(N-1)/2-\epsilon N$. It is easy to see $\kappa_\epsilon\rightarrow U^{-1}$. The ensemble averaged value $\overline{\kappa}$ for large $\Delta$ is mainly determined by the contribution from $\kappa_\epsilon$ in the large $\epsilon$ region. As a result, $\overline{\kappa}$ also approaches a constant value $U^{-1}$ when $\Delta$ is very large. We can view the state in this regime as an analog of the bose glass phase in the thermodynamic limit, which has vanishing order parameter $\langle b_i\rangle=0$(or supefluid density $\rho_s=0$) but a finite compressibility $\kappa$. To fully understand the bose glass phase in the disordered BH model, one needs to consider correlations beyond nearest neighboring sites, which also contribute to the finite compressibility of bose glass phase, despite the absence of a long range order. Nevertheless, our two-site model provides a simple example to demonstrate why the compressibility of a disordered system can remain finite after the phase coherence has been destroyed at large disorder strength.
![Left: $\overline{C_\beta}$ as a function of $\Delta/tN$ at different temperatures for a fixed interaction $U/tN=2$. From top to bottom, $T/tN=0.4$, $1$, $2$, $3$. Right: a contour plot of $\overline{C}$ as a function of $U/tN$ and $\Delta/tN$ at the temperature $T/tN=5$. []{data-label="fig: fig4"}](fig4){width="3.4in"}
[**The answer to Q4**]{}: We have seen that increasing disorder at large interaction leads to an increase of phase coherence at zero temperature. The topology of the phase diagram at zero temperature obtained from recent numerical simulations of the full disordered BH model also indicates a transition from the insulating phase to the superfluid phase with increasing disorder strength[@QMC; @DMRG]. However, a recent experiment did not find any evidence for the increase of the condensate fraction with increasing disorder strength[@Demarco3]. We will show that the topology of the contours of $\overline{C}$ changes at finite temperatures, namely, the contours may bend to the left hand side when the disorder strength increases from zero. This fact may help understand why the experiment at finite temperatures did not observe disorder enhanced phase coherence.
The value of the phase coherence at finite temperatures can be calculated by $C_{\beta,\epsilon}=(\sum_ne^{-\beta E_n}M_{l.l+1}\psi^n_{l}\psi^n_{l+1})/Z$, where the subscript $\beta$ denotes the thermal average, $\phi^n_l$ is the $n$th eigenfunction with eigenenergy $E_n$, $Z=\sum_n e^{-\beta E_n}$, and $\overline{C}_\beta=\int _{-\Delta}^\Delta d\epsilon C_{\beta,\epsilon}$. The results for $\overline{C}_\beta$ as a function of $\Delta$ for a fixed $U$ are shown in Fig.(4A) at different temperatures. At low temperatures, the wiggles on the curve of $\overline{C}_\beta$ retain. When temperature increases, the wiggles are gradually suppressed. At high enough temperatures, the wiggles completely vanish, leading to a monotonic decrease of $\overline{C}_\beta$. Consequently, the topology of the contours of $\overline{C}_\beta$ changes completely, as shown in Fig(4B). Since the minimal model appears to describe the full disordered model rather well qualitatively, it is reasonable to assume that the phase diagram for the lattice model at finite temperatures will change in a similar manner as Fig.(4B). It has also pointed out that the weak superfluidity in the so called finger region, corresponding to the regions near the wiggles in Fig.(2) in our case, can be easily suppressed by finite temperature effects[@QMC; @Hofstetter].
As a conclusion, we have proposed and solved a simple double-well model which incorporates many key ingredients of the disordered BH model. It strikingly captures a number of novel features of the quantum phase diagram of the full BH model. The “two negatives make a positive" effect is expected to be a general feature when the strengths of disorder and interaction are comparable. We believe that our minimal model could be a simple theoretical paradigm for understanding the details of the full disordered interacting quantum phase diagram in many situations. We expect our work will stimulate more theoretical studies to go beyond the two-site model and take into account the long range efforts in a lattice for a complete understanding of the disordered BH model.
The authors acknowledge Tin-Lun Ho for helpful discussions. This work is supported by JQI-NSF-PFC and ARO-DARPA-OLE.
[99]{} P. W. Anderson, Phys. Rev. 109, 1492 (1958) N. Trivedi, in Condensed Matter Theories Vol. 12 (Nova Science publishers, 1997), pp. 141-157. A. Aspect and M. Inguscio, Physics Today , 62, 30 (2009); L. Sanchez-Palencia and M. Lewenstein, Nature Physics 6, 87 (2010) J. Billy, *et al.*, Nature 453, 891Ð894 (2008) T. Roati, *et al.*, Nature 453, 895Ð898 (2008) B. Deissler et al., Nature Physics 6, 354 - 358 (2010) D. McKay, *et al.*, Nature 453, 76 (2008) M. White, *et al.*, Phys. Rev. Lett. 102, 055301 (2009) M. Pasienski et al., Nature Physics 6, 677 - 680 (2010) Y. P. Chen, *et al.*, Phys. Rev. A 77, 033632 (2008);D. Clement et al., Phys. Rev. A 77 033631(2008) J. Biddle and S. Das Sarma, Phys. Rev. Lett. 104, 070601 (2010) M.P.A. Fisher, *et al.*, Phys. Rev. B 40, 546 (1989) E. Altman, *et al.*, Phys. Rev. Lett. 93, 150402 (2004) E. Altman, *et al.*, Phys. Rev. Lett. 100, 170402 (2008) W. Krauth, *et al.* Phys. Rev. Lett. 67, 2307 (1991) R.T. Scalettar, *et al.* Phys. Rev. Lett. 66, 3144 (1991) V. Gurarie, *et al.*, Phys. Rev. B 80, 214519 (2009) G. Roux, *et al.*, Phys. Rev. A 78, 023628 (2008) In the full disordered BH model, with increasing disorder strength, there are either successive phase transitions in which superfluid phase is present, or only one phase transition from Mott insulator to bose glass[@QMC]. Whereas phase coherence is enhanced by disorder in both cases, a long range order is established in the first case but not in the second one. Y. Shin, *et al.*, Phys. Rev. Lett. 92, 150401 (2004) B. V. Hall, *et al.*, Phys. Rev. Lett. 98, 030402 (2007) T.-L. Ho and C. V. Ciobanu, J. Low Temp. Phys. 135, 257 (2004); H. Pu, *et al.*, Phys. Rev. Lett. 89, 090401 (2002) U. Bissbort, R. Thomale, and W. Hofstetter, Phys. Rev. A 81, 063643 (2010)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Many companies rely on Cloud infrastructures for their computation, communication and data storage requirements. While Cloud services provide some benefits, e.g., replacing high upfront costs for an IT infrastructure with a pay-as-you-go model, they also introduce serious concerns that are notoriously difficult to address. In essence, Cloud customers are storing data and running computations on infrastructures that they can not control directly. Therefore, when problems arise – violations of Service Level Agreements, data corruption, data leakage, security breaches – both customers and Cloud providers face the challenge of agreeing on which party is to be held responsible. In this paper, we review the challenges and requirements for enforcing accountability in Cloud infrastructures, and argue that smart contracts and blockchain[^1] technologies might provide a key contribution towards accountable Clouds.'
author:
- 'Gabriele D’Angelo'
- Stefano Ferretti
- Moreno Marzolla
bibliography:
- 'cloud-accountability.bib'
title: |
A Blockchain-based Flight Data Recorder\
for Cloud Accountability
---
<ccs2012> <concept> <concept\_id>10002978.10002991</concept\_id> <concept\_desc>Security and privacy Security services</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002978.10003006.10003013</concept\_id> <concept\_desc>Security and privacy Distributed systems security</concept\_desc> <concept\_significance>300</concept\_significance> </concept> <concept> <concept\_id>10010520.10010521.10010537.10003100</concept\_id> <concept\_desc>Computer systems organization Cloud computing</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
sec-introduction.tex sec-background.tex sec-proposal.tex sec-case.tex sec-conclusions.tex
[^1]: **WARNING**: this paper may contain traces of silicon snake oil and palm oil.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'A graph is an efficient open domination graph if there exists a subset of vertices whose open neighborhoods partition its vertex set. We characterize those graphs $G$ for which the Cartesian product $G\,\Box\, H$ is an efficient open domination graph when $H$ is a complete graph of order at least 3 or a complete bipartite graph. The characterization is based on the existence of a certain type of weak partition of $V(G)$. For the class of trees when $H$ is complete of order at least 3, the characterization is constructive. In addition, a special type of efficient open domination graph is characterized among Cartesian products $G \,\Box\, H$ when $H$ is a $5$-cycle or a $4$-cycle.'
author:
- Tadeja Kraner Šumenjak
- 'Iztok Peterin[^1]'
- 'Douglas F. Rall[^2]'
- 'Aleksandra Tepeh[^3]'
bibliography:
- 'EOD.bib'
nocite: '[@*]'
title: 'Partitioning the vertex set of $G$ to make $G\,\Box\, H$ an efficient open domination graph'
---
Introduction {#section:intro}
============
The domination number of a graph $G$ is a classical invariant in graph theory. It is the minimum cardinality of a set $S$ of vertices for which the union of the closed neighborhoods centered in vertices of $S$ is the entire vertex set of $G$. Hence, each vertex of $G$ is either in $S$ or is adjacent to a vertex in $S$. In other words, we can say that vertices of $S$ control each vertex outside of $S$. A classical question in such a situation is: who controls the vertices of $S$? One possible solution to this dilemma is total domination. A set $D\subseteq V(G)$ is a total dominating set of $G$ if every vertex of $G$ is adjacent to a vertex of $D$. (Hence, vertices of $D$ are also controlled by $D$.)
A natural question for a graph $G$ is whether we can find a total dominating set $D$ such that the union of the open neighborhoods of the vertices in $D$ is $V(G)$ but so that these open neighborhoods also form a partition of $V(G)$. The concept has been presented under the names total perfect codes [@CoHaHe], efficient open domination [@GaSc] and exact transversals [@CoHeKe]. In the present work we follow the terminology of efficient open domination, and we say that a graph $G$ is an efficient open domination graph if $G$ has a total dominating set $D$ such that the open neighborhoods of the vertices of $D$ form a partition of $V(G)$. A similar concept for ordinary domination was first investigated by [@biggs] and [@Kra]. They call a graph 1-perfect if it contains a perfect code, that is, a set of vertices whose closed neighborhoods partition the vertex set.
The problem of establishing whether a graph $G$ is an efficient open domination graph is an $NP$-complete problem; see [@GaScSl; @McRae]. [@GaScSl] gave a recursive characterization of the class of efficient open domination trees. [@GaSc] presented various properties of efficient open domination graphs. The efficient open domination graphs that are also Cayley graphs were studied by [@Tham] and efficient open domination grid graphs by [@CoHeKe; @Dej; @KlGo]. Moreover, [@AbHamTay] characterized those direct product graphs that are efficient open domination graphs.
Several graph products have been investigated in the last few decades and a rich theory involving the structure and recognition of classes of these graphs has emerged [@ImKl]. The most studied graph products are the Cartesian, strong, direct, and lexicographic. These four are also called the *standard products*. One approach to graph products is to deduce properties of a product with respect to (the same) properties of its factors. See a short collection of these types involving total domination and perfect codes in [@DoGrSp; @Grav; @HeRa; @Ho; @JeKlSp; @KlSpZe; @KuPeYe; @KuPeYe1; @Meki; @Rall]. The domination related questions on the Cartesian product seems to be the most problematic among the standard products. We just mention Vizing’s conjecture, which says that the domination number of a Cartesian product is at least the product of the domination numbers of the two factors. Settling this conjecture is one of the most challenging problems in the area of domination (see the recent survey on Vizing’s conjecture [@BrDoKl]). Efficient open domination is no exception, which could be the reason it has not been studied intensively yet in the Cartesian product setting. Other than the results on grid graphs mentioned above, a step forward in this direction was made only recently by [@KuPeYe1] where some special types of Cartesian products were considered. In the same paper complete descriptions of efficient open domination graphs among lexicographic and strong products of graph were given.
The aim of this paper is to show how the problem of finding efficient open domination graphs among Cartesian products can be approached by partitioning the vertex set of one factor. In the next section we set the context by supplying needed definitions and previous results in this area. In Section \[sec:completefactor\] we prove that for $r \ge 3$, the graph $G\,\Box\,K_r$ has an efficient open dominating set if and only if $V(G)$ has a weak partition that satisfies certain properties. This provides a way to construct graphs with efficient open dominating sets in this family of Cartesian products. In addition we give a structural characterization of the trees $T$ such that $T\,\Box\,K_r$ has an efficient open dominating set. Section \[sec:diam2\] addresses this weak partition approach to graphs of diameter 2.
Definitions and previous results {#sec:def}
================================
Throughout the article we consider only finite, simple graphs. For most common graph theory notation and definitions we follow the book by [@ImKl]. In particular, our definitions and notation for open ($N(v)$) and closed ($N[v]$) neighborhoods of a vertex $v$, for distance ($d_{G}(u,v)$) between a pair of vertices and for the diameter ($\mathrm{diam}(G)$) of a graph are the same as theirs. The distance $d_{G}(e,v)$ between an edge $e$ and a vertex $v$ in $G$ is the shortest distance between $v$ and the two end vertices of $e$, while the distance $d_{G}(e_{1},e_{2})$ between edges $e_{1}$ and $e_{2}$ is the shortest distance between the end vertices of $e_{1}$ and the end vertices of $e_{2}$. In general, for nonempty subsets $P$ and $Q$ of $V(G)$, the distance $d_{G}(P,Q)$ between them is the shortest distance between a vertex from $P$ and a vertex from $Q$. A *weak partition* of a set $X$ is a collection of pairwise disjoint subsets of $X$ whose union is $X$. We emphasize that, in contrast to a partition, members of a weak partition are allowed to be empty. The subgraph induced by a subset $S$ of $V(G)$ is denoted by $\left\langle S\right\rangle$. A *matching* in $G$ is any (possibly empty) set of independent edges. If $r$ is a positive integer, then the vertex set of each of the graphs $P_r$, $K_r$, and $C_r$ (if $r>2$) will be the interval $[r]$ defined by $[r]=\{1,\ldots,r\}$.
Since this present work concerns total domination on Cartesian products, we include several of the important definitions here for the sake of completeness. We say that a vertex $x$ of $G$ *dominates* a vertex $y$ (equivalently, $y$ *is dominated by* $x$) if $y \in N(x)$. A subset $D$ of $V(G)$ is a *total dominating set* of $G$ if each vertex in $G$ is dominated by at least one vertex in $D$. The *total domination number* of a graph $G$ is the minimum cardinality of a total dominating set of $G$ and is denoted by $\gamma_{t}(G)$. The *Cartesian product*, $G\,\Box\, H$, of graphs $G$ and $H$ is a graph with $V(G\,\Box\, H)=V(G)\times V(H)$. Two vertices $(g,h)$ and $(g',h')$ are adjacent in $G\,\Box\, H$ whenever ($gg'\in E(G)$ and $h=h'$) or ($g=g'$ and $hh'\in E(H)$). For a fixed $h\in V(H)$ we call $G^{h}=\{(g,h)\in V(G\,\Box\, H):g\in V(G)\}$ a $G$-*layer* in $G\,\Box\, H$. Similarly, an $H$-*layer* $^{g}\!H$ for a fixed $g\in V(G)$ is defined as $^{g}\!H=\{(g,h)\in V(G\,\Box\, H):h\in V(H)\}$. Notice that the subgraph of $G\,\Box\, H$ induced by a $G$-layer or an $H$-layer is isomorphic to $G$ or $H$, respectively. The map $p_{G}:V(G\,\Box\, H)\rightarrow V(G)$ defined by $p_{G}((g,h))=g$ is called a *projection map onto* $G$. Similarly, we define $p_{H}$ as the *projection map onto* $H$. Projections are defined as maps between vertices, but frequently it is more useful to see them as maps between graphs.
A graph $G$ is an *efficient open domination graph* (shortly an *EOD-graph*) if there exists a set $D$, called an *efficient open dominating set* (shortly an *EOD-set*), for which $\bigcup _{v\in D}N(v)=V(G)$ and $N(u)\cap N(v)=\emptyset$ for every pair $u$ and $v$ of distinct vertices of $D$. Note that two different vertices of an EOD-set are either adjacent or at distance at least three. It is easy to see that the path $P_{n}$ is an EOD-graph if and only if $n\not\equiv 1 \pmod{4}$, while the cycle $C_{n}$ is an EOD-graph if and only if $n\equiv 0 \pmod{4}$. Let $G$ and $H$ be graphs such that $G\,\Box\,H$ is an EOD-graph with an EOD-set $D$. Note that the projection of an edge in $\left\langle D\right\rangle$ onto $G$ is either a vertex or an edge. When the projection of every edge in $\left\langle D\right\rangle$ onto $G$ is an edge, we say that $D$ is a *parallel EOD-set* with respect to $G$. A Cartesian product that contains a parallel EOD-set with respect to one of its factors is called a *parallel EOD-graph*.
Among the class of nontrivial Cartesian products several infinite families of EOD-graphs have been found. In [@CoHeKe; @KlGo] the authors investigated EOD-graphs among the grid graphs (that is, Cartesian products of paths). Results from both papers are merged in the following characterization.
*[@CoHeKe; @KlGo]* Let $t\geq r\geq 3$. The grid graph $P_r\,\Box\, P_t$ is an EOD-graph if and only if $r$ is an even number and $t\equiv x \pmod{r+1}$ for some $x\in \{1,r-2,r\}$.
Some partial results on EOD-graphs in the family of torus graphs (Cartesian products of cycles) were presented by [@Dej], by characterizing only those with a parallel EOD-set (there referred to as a parallel total perfect code).
*[@Dej]*\[torus-parallel\] The Cartesian product $C_{r}\,\Box\, C_{t}$ has a parallel EOD-set if and only if $r$ and $t$ are multiples of four.
[@KuPeYe1] recently continued with the study of EOD-graphs among tori and cylinders (Cartesian product of a path and a cycle).
*[@KuPeYe1]* Let $t\ge 4$. The torus $C_4\,\Box\, C_t$ is an EOD-graph if and only if $t\equiv 0 \pmod{4}$.
In addition, they proved that $C_r\,\Box\, C_t$ is not an EOD-graph if $r\in \{3,5,6,7\}$ and $t\geq r$. Based on the above observations they posed the following conjecture.
*[@KuPeYe1]* Let $r$ and $t$ be integers such that $r \ge 3$ and $t \ge 3$. The torus $C_{r}\,\Box\, C_{t}$ is an EOD-graph if and only if $r \equiv 0 \pmod{4}$ and $t \equiv 0 \pmod{4}$.
The same authors characterized the graphs $G$ for which $G\,\Box\, K_{2}$ is an EOD-graph. In order to do this they introduced the so-called zig-zag graphs, [@KuPeYe1]. Let $G$ be a graph on at least three vertices and $E'=\{e_{1},\ldots ,e_{k}\}$ a subset of $E(G)$, where $e_{i}=u_{i}v_{i}$ for every $i\in[k]$, with the following properties:
- $N(u_{i})\cap N(v_{i})=\emptyset $;
- $d_{G}(e_{i},e_{j})\geq 2$ for $1 \le i<j\le k$;
- for every $x\in V(G)-\{u_{i},v_{i}:i\in [k]\}$ there exist unique $j$ and $\ell $, $j\neq \ell $, such that $d_{G}(x,e_{j})=d_{G}(x,e_{\ell })=1$;
- for every sequence $e_{i_{1}},\ldots ,e_{i_{j}}$ of distinct edges with $j>2$ and with\
$d_{G}(e_{i_{_{\ell}}},e_{i_{_{\ell +1 \pmod{j}}}})=2$ for $\ell \in \{1,\ldots ,j\}$, $j$ must be an even number.
We call $E'$ a *zig-zag set* of $G$ and, if there exists a zig-zag set in $G$, we call $G$ a *zig-zag graph*.
\[zigzagEOD\] *[@KuPeYe1]* If $G$ is a zig-zag graph, then $G\,\Box\, K_{2}$ is an EOD-graph.
Not all EOD-graphs among $G\,\Box\, K_{2}$ are given by the above theorem. Kuziak et al. observed that for a description of all EOD-graphs among Cartesian products of graphs with $K_2$, a certain combination of zig-zag graphs and 1-perfect graphs is needed (see [@KuPeYe1] for details).
One can observe that for $r>2$, every EOD-set in $G\,\Box\, K_{r}$ is a parallel EOD-set with respect to $G$. Namely, if an edge induced by two vertices of a vertex subset $A$ of $G\,\Box\, K_{r}$ projects to a single vertex $g\in V(G)$, then the layer $^{g}\!K_r$ contains a vertex that is dominated more than once by $A$. This observation led to the idea of how to approach the problem of finding EOD-graphs among $G\,\Box\, K_{r}$ for $r>2$. This is presented in the next section.
$G\,\Box\, K_{r}$ for $r>2$ {#sec:completefactor}
===========================
In order to obtain a characterization of EOD-graphs among $G\,\Box\, K_{r}$, $r>2$, we introduce a new concept, based on a weak partition of the vertex set of $G$. As we will see in later sections, a modification of this concept can be used for the construction of EOD-graphs that are Cartesian products $G \,\Box\, H$ where $H$ belongs to several other special classes of graphs.
Let $r$ be an integer larger than 1. We call a graph $G$ a *$K_{r}$-amenable graph* if there exists a weak partition $\{V_{0},V_{1},\ldots
,V_{r}\}$ of $V(G)$, such that
- if $x\in V_{0}$, then $|N(x)\cap V_{i}|=1$ for every $i\in [r]$,
- $\left\langle V_{i}\right\rangle$ is a matching in $G$ for every $i\in [r]$,
- $\left\langle V_{1}\cup \cdots \cup V_{r}\right\rangle$ is a matching in $G$.
For the sake of clarity in the above definition we emphasize that the induced subgraphs $\left\langle V_{i}\right\rangle$ and $\left\langle V_{1}\cup \cdots \cup V_{r}\right\rangle$ do not contain any edges other than those in their perfect matchings.
We first prove that $K_{2}$-amenable graphs do not differ from zig-zag graphs.
A graph $G$ is a $K_{2}$-amenable graph if and only if $G$ is a zig-zag graph.
Let $G$ be a $K_{2}$-amenable graph with a weak partition $\{V_{0},V_{1},V_{2}\}$ of $V(G)$ that satisfies conditions (A), (B) and (C). We will show that $E'=\left\langle V_{1}\cup V_{2}\right\rangle $ is a zig-zag set of $G$ by demonstrating that conditions $(i)-(iv)$ hold. Since $\left\langle V_{1}\cup V_{2}\right\rangle$, $\left\langle V_{1}\right\rangle $ and $\left\langle
V_{2}\right\rangle $ are matchings, $E'$ is a set of edges $\{e_{1},\ldots ,e_{k}\}$. By the same argument we derive that $d_{G}(e_{i},e_{j})\geq 2$ for $i \neq j$, and thus $(ii)$ holds. Let $e_{i}=u_{i}v_{i}$ for every $i\in [k]$. If $x\in N(u_{i})\cap N(v_{i})$ for some $i\in [k]$, then $x\in V_{0}$ by the matching argument again. But this contradicts condition (A) since $|N(x)\cap V_{i}|\geq 2$ in this case. Hence $(i)$ also holds. If $x\in V(G)-\{u_{i},v_{i}:i\in [k]\}$, then $x\in V_{0}$. By (A) we have that $|N(x)\cap V_{i}|=1$ for every $i\in \{1,2\}$, which implies the existence of exactly two different edges $e_{j}$ and $e_{\ell }$ of $E'$ with $d_{G}(x,e_{j})=d_{G}(x,e_{\ell })=1$. This proves $(iii)$. To prove $(iv)$, let $e_{i_{1}},e_{i_{2}},\ldots ,e_{i_{j}}$, $j>2$, be a sequence of distinct edges with $d_{G}(e_{i_{_{\ell }}},e_{i_{_{\ell
+1 \pmod{j}}}})=2$ for $\ell \in [j]$. In addition, let $x_{\ell }$ be a common neighbor of $e_{i_{_{\ell }}}$ and $e_{i_{_{\ell +1\pmod{j}}}}$. As before, $x_{\ell }\in V_{0}$ for every $\ell \in [j]$. Without loss of generality, suppose the end-vertices of the edge $e_{i_{1}}$ belong to $V_1$. By condition (A) for the vertex $x_1$, the end-vertices of $e_{i_{2}}$ belong to $V_2$. The same argument for the vertex $x_2$ implies that the end-vertices of $e_{i_{3}}$ belong to $V_1$. Continuing this way, we get a zig-zag pattern for the end-vertices of $e_{i_{1}},e_{i_{2}},\ldots ,e_{i_{j}}$. If $j$ is an odd number, then the end-vertices of $e_{i_{1}}$ and $e_{i_j}$ are both in $V_{1}$, which gives a contradiction with condition (A) for the vertex $x_j$. Thus $j$ is an even number and $(iv)$ holds as well.
Now let $G$ be a zig-zag graph with a zig-zag set $E^{\prime
}=\{e_{1},\ldots ,e_{k}\}$ where $e_{i}=u_{i}v_{i}$. We set $V_{0}=V(G)-\{u_{i},v_{i}:i\in [k]\}$. Observe that $E'$ can be partitioned as $E'=E_1\cup \cdots \cup E_t$ such that for each $i\in [t]$, the following holds. The set $E_i$ is a maximal set of edges such that between any two distinct edges $e_{j}$ and $e_n$ from $E_i$ there exists a sequence $e_{j}=e_{j_0},e_{j_{1}},\ldots ,e_{j_{\ell}}=e_n$, $\ell \geq 1$, of distinct edges where the distance between two consecutive edges in this sequence is $2$. Such a sequence is called a [*2-step sequence of length*]{} $\ell$.
Observe that there exists a partition of $E'=E_1\cup \cdots \cup E_t$, such that $E_i$, for every $i\in [t]$, consists of a maximal set of edges such that between any two distinct edges $e_{j}$ and $e_k$ from $E_i$ there exists a sequence $e_{j}=e_{j_0},e_{j_{1}},\ldots ,e_{j_{\ell}}=e_k$, $\ell \geq 1$, of distinct edges such that the distance between two consecutive edges in this sequence is $2$ (we call such sequence a [*2-step sequence of length*]{} $\ell$). Now, in $E_i$ fix an arbitrary edge $e$. For an arbitrary edge $f$ in $E_i$ there exists a 2-step sequence between $e$ and $f$. Property $(iv)$ implies that the lengths of all different 2-step sequences between $e$ and $f$ are of the same parity. Thus, edges of $E_i$ can be partitioned into two sets $E_i^{1}$ and $E_i^{2}$. The set $E_i^{1}$ consists of $e$ and all edges $f$ for which the length of a 2-step sequence between $e$ and $f$ is even, and $E_i^{2}=E_i-E_i^{1}$. For every $i\in [t]$ let $V_i^{1}$ denote the set of end-vertices of edges in $E_i^{1}$, and $V_i^{2}$ the set of end-vertices of edges in $E_i^{2}$. Finally, let $V_{1}=V_1^{1}\cup \cdots \cup V_t^{1}$ and $V_{2}=V_1^{2}\cup \cdots \cup V_t^{2}$.
We will show that $\{V_{0},V_{1},V_{2}\}$ is a weak partition of $V(G)$ satisfying conditions (A), (B) and (C). Properties (B) and (C) clearly follow, since $d_{G}(e_{i},e_{j})\geq 2$ for every pair $e_{i},e_{j}\in E'$. To prove (A) let $x\in
V_{0}$. By $(iii)$ there exist exactly two different edges $e_{p},e_{r}\in E'$ such that $d_{G}(x,e_{p})=1=d_{G}(x,e_{r})$. Note that $e_{p}$ and $e_{r}$ belong to the same $E_i$ in the partition of $E'$. Recall that we have fixed the edge $e\in E_i$. If a 2-step sequence between $e$ and $e_{p}$ and a 2-step sequence between $e$ and $e_{r}$ have the same parity, then we obtain a contradiction with $(iv)$. Hence, end-vertices of one edge, say $e_{p}$, belong to $V_{1}$, and end-vertices of $e_{r}$ belong to $V_{2}$. Since, in addition, $N(u_{i})\cap N(v_{i})=\emptyset $, by $(i)$ for every $i$ we have $|N(x)\cap V_{1}|=1=|N(x)\cap V_{2}|$ and condition (A) holds.
\[complete\] Let $r$ be a positive integer such that $r>2$ and let $G$ be a graph. The Cartesian product $G\,\Box\, K_{r}$ is an EOD-graph if and only if $G$ is a $K_{r}$-amenable graph.
Let $G$ be a $K_{r}$-amenable graph with corresponding weak partition $\{V_{0},\ldots,V_{r}\}$ of $V(G)$. We define a subset $D$ of $V(G\,\Box\, K_{r})$ by $D=\{(g,i): i \in [r] \mbox{ and } g \in V_i\}$. It follows that $D$ contains at most one vertex from each $K_r$-layer. To prove that $G\,\Box\, K_{r}$ is an EOD-graph we will show that every vertex of $G\,\Box\, K_{r}$ is dominated by exactly one vertex of $D$. Let $i \in [r]$ and let $g \in V(G)$. First, assume that $g\in V_{0}$. By (A), the vertex $g$ has a unique neighbor $x_{i}$ in $V_{i}$. Consequently, $(g,i)$ is adjacent to $(x_{i},i)$ and $(x_{i},i)\in D$. Moreover, by the uniqueness of $x_{i}$, no other vertex of $D$ dominates $(g,i)$. Now assume that $g\in V_{i}$. Since $\left\langle V_{i}\right\rangle $ is a perfect matching, $g$ has a unique neighbor $g'$ in $V_i$. It follows that $(g',i)\in D$ and that $(g',i)$ is the only neighbor of $(g,i)$ in $D$. Finally, assume that $g \in V_j$ for some $j \in [r]$ such that $j\neq i$. By the definition of $D$ this implies that $\{(g,j)\} = D \cap\,^{g}K_{r}$. In addition, since (B) and (C) hold, $(g,i)$ has no neighbor in $G^i \cap D$. The result is that $(g,i)$ is dominated by exactly one vertex, namely $(g,j)$, of $D$. Consequently, $D$ is an EOD-set of $G \,\Box\, K_r$ and $G\,\Box\, K_{r}$ is an EOD-graph.
To prove the converse, suppose that $G\,\Box\, K_{r}$ is an EOD-graph with an EOD-set $D$. For $i\in [r]$ let $V_i=\{v\in V(G): (v,i)\in D\}$, and let $V_{0}=V(G)-(V_{1}\cup \cdots \cup V_{r})$. As we observed in Section \[sec:def\], $D$ is necessarily parallel with respect to $G$. This means that every $^{v}K_{r}$ contains at most one vertex of $D$, and we thus infer that $\{V_{0}, V_1,\ldots ,V_{r}\}$ is a weak partition of $V(G)$. We prove that conditions (A), (B), and (C) of the definition of $K_r$-amenable hold. If condition (A) is not satisfied, then there exist $x\in V_{0}$ and $i\in [r]$, such that $|N(x)\cap V_{i}|=0$ or $|N(x)\cap V_{i}|>1$. In the first case $(x,i)$ is not dominated by any vertex of $D$, and in the second case $(x,i)$ is dominated by more than one vertex of $D$. Both cases are in contradiction with the assumption that $D$ is an EOD-set of $G\,\Box\, K_{r}$. Hence, the weak partition $\{V_{0},V_1,\ldots ,V_{r}\}$ satisfies property (A). Let $i \in [r]$ and let $g \in V_i$. Since $|D \cap\, ^gK_r|\le 1$ and $(g,i)$ has exactly one neighbor in $D$, it follows that $|N(g) \cap (V_1 \cup \cdots \cup V_r)|= 1 = |N(g) \cap V_i|$. Hence, both (B) and (C) hold. Therefore, $G$ is a $K_{r}$-amenable graph.
Let $r$ be an integer larger than 1. In the rest of this section we present a recursive description of the family of all $K_{r}$-amenable trees. The following construction generalizes the construction of zig-zag trees (that is, $K_{2}$-amenable trees) from [@KuPeYe1]. We will denote by $K_{1,r}^{+}$ the tree of order $2r+1$ obtained from the star $K_{1,r}$ by subdividing each edge exactly once. It is clear that $K_{1,r}^{+}$ is a $K_{r}$-amenable tree, and the corresponding partition of $V(K_{1,r}^{+})$ is unique up to a permutation of $[r]$. We now define an infinite family $\mathcal{T}_{r}$ of trees. Each member of $\mathcal{T}_{r}$ will have a weak partition $\{V_0, V_1, \ldots, V_r\}$ of its vertex set associated with it.
Suppose that $T'$ is a tree of order $n$ such that $\{V'_{0}, V'_{1},\ldots,V'_{r}\}$ is a weak partition of $V(T')$ and that $T''$ is a tree of order $m$ such that $\{V''_{0}, V''_{1}, \ldots ,V''_{r}\}$ is a weak partition of $V(T'')$.
We say that a tree $T$ of order $n+m-2$ is obtained from $T'$ and $T''$ by a [**Type-a**]{} construction if $T$ is isomorphic to the tree formed by choosing any $i \in [r]$, any edge $u'_{i}v'_{i}$ in $\left\langle V'_{i}\right\rangle$, any edge $u''_{i}v''_{i}$ in $\left\langle V''_{i}\right\rangle$ and then identifying the vertices $u'_{i}$ with $u''_{i}$ (now called $u_{i}$) and $v'_{i}$ with $v''_{i}$ (now called $v_{i}$) to obtain the edge $u_{i}v_{i}$ in $T$. The associated weak partition $\{V_0, V_1, \ldots,V_r\}$ of $V(T)$ is defined by $V_j=V'_{j} \cup V''_{j}$ if $j \neq i$, and $V_i=(V'_{i} \cup V''_{i} \cup \{u_i,v_i\})-\{u'_{i},v'_{i}, u''_{i}, v''_{i}\}$.
A tree $S$ of order $n+m$ is obtained from $T'$ and $T''$ by a [**Type-b**]{} construction if $S$ is isomorphic to the tree formed from the union of $T'$ and $T''$ by adding an edge $xy$ for some $x\in V'_{0}$ and some $y\in V''_{0}$. The associated weak partition $\{V_0, V_1, \ldots,V_r\}$ of $V(S)$ is given by $V_i= V'_{i} \cup V''_{i}$ for $0 \le i \le r$.
The family $\mathcal{T}_{r}$ is defined recursively as follows. A tree $T$ belongs to $\mathcal{T}_{r}$ if and only if $T=K_{1,r}^{+}$ with its partition as indicated above or $T$ can be obtained from smaller trees in $\mathcal{T}_{r}$ by a finite sequence of Type-a or Type-b constructions.
\[zigzagtree\] Let $r$ be an integer such that $r\ge 2$. The path of order 2 is $K_{r}$-amenable. If $T$ is a tree of order more than 2, then $T$ is a $K_r$-amenable graph if and only if $T\in \mathcal{T}_{r}$.
Let $r$ be an integer such that $r\ge 2$. For the path of order 2, let $V_1=V(P_2)$, $V_0=\emptyset=V_i$ for $2 \le i \le r$. This weak partition $\{V_0,V_1,\ldots,V_r\}$ satisfies the definition showing that $P_2$ is a $K_r$-amenable graph. For the remainder of this proof we assume that all trees under consideration have order at least 3. As noted above, the tree $K_{1,r}^{+}$ is a $K_{r}$-amenable tree. One can conclude directly from the definitions that if $T'$ and $T''$ are both $K_{r}$-amenable trees, then a tree obtained from $T'$ and $T''$ by a Type-a or a Type-b construction is also a $K_{r}$-amenable graph. Thus, it follows by induction (on the number of Type-a and Type-b constructions) that every member of $\mathcal{T}_{r}$ is a $K_{r}$-amenable graph.
Conversely, let $T$ be a $K_{r}$-amenable tree of order at least 3 with a corresponding weak partition $\{V_{0},V_1,\ldots ,V_{r}\}$ and let $k=|V_0|$. Since $T$ has order at least 3, it follows from the definition that $k\ge 1$. We use induction on $k$ to show that $T\in \mathcal{T}_{r}$. Let $k=1$ and $V_{0}=\{v\}$. By property (A) $\deg(v)=r$; let $N(v)=\{u_{1},\ldots ,u_{r}\}$ where $u_{i}\in V_{i}$. By (B) every $u_{i}$ has a unique neighbor $w_{i}$ in $V_{i}$ and by (C) $u_{i}$ and $w_{i}$ have no neighbors in $V_{j}$ for $j\neq i$. Moreover, $u_{i}$ and $w_{i}$ have no additional neighbors in $V_{0}$ since $k=1$. Thus, $T$ is isomorphic to $K_{1,r}^{+}$ and hence $T\in \mathcal{T}_{r}$.
Now suppose that $k>1$. Note that every vertex in $V_0$ has degree at least $r$. If there exists $v\in V_{0}$ with $\deg (v)>r$, then there exists $w \in V_{0}\cap N(v)$. Let $T'$ be the component of $T-vw$ that contains $v$ and let $T''$ be the component that contains $w$. For $0 \le i \le r$, let $V_{i}'=V_{i}\cap V(T')$ and let $V_{i}''=V_{i}\cap V(T'')$. The resulting weak partitions of $V(T')$ and $V(T'')$ clearly satisfy properties (A), (B) and (C), and furthermore $|V_{0}'|<k$ and $|V_{0}''|<k$. By the induction hypothesis both $T'$ and $T''$ belong to $\mathcal{T}_{r}$. Since $T$ is obtained from $T'$ and $T''$ by a Type-b construction, it follows that $T\in \mathcal{T}_{r}$.
Now, suppose that all vertices of $V_{0}$ are of degree $r$ (and hence $\left\langle V_{0}\right\rangle$ contains no edges). Choose $u$ and $v$ from $V_{0}$ with the property that $d_{T}(u,v)$ is minimum among all different pairs of vertices from $V_{0}$. Clearly, $2\leq d_{T}(u,v)\leq 3$. Let $w$ be the neighbor of $u$ on the shortest $u,v$-path in $T$. Without loss of generality we may assume that $w\in V_{1}$. By (B), $w$ has a unique neighbor, say $w'$, in $V_{1}$. The forest $T-uw$ has two connected components. The component that contains $u$ is denoted by $T_{u}$ and the one that contains $v$ is denoted by $T''$. Let $T'$ be the tree obtained from $T_u$ by adding vertices $t$ and $t'$ and adding edges $ut$ and $tt'$. Let $V_i''=V_i\cap V(T'')$ for $0 \le i \le r$, let $V_1'=(V_1\cap V(T_u))\cup \{t,t'\}$, and let $V_i'=V_i\cap V(T_u)$ for $i=0$ and $2 \le i \le r$. Properties (A), (B) and (C) clearly hold for the above defined weak partitions of $V(T')$ and $V(T'')$. Thus, $T'$ and $T''$ are $K_{r}$-amenable trees. By the induction hypothesis, they are also in $\mathcal{T}_{r}$. Note that $T$ is isomorphic to the tree obtained from $T'$ and $T''$ by a Type-a construction that identifies $t$ with $w$, and $t'$ with $w'$. Consequently, $T \in \mathcal{T}_{r}$.
This theorem together with Theorem \[complete\] combine to give us the following characterization of those trees $T$ such that $T\,\Box\,K_r$ is an EOD-graph for $r\ge3$.
Let $r$ be a positive integer larger than 2 and let $T$ be a tree. The Cartesian product $T\,\Box\,K_r$ is an EOD-graph if and only if $T=P_2$ or $T\in \mathcal{T}_{r}$.
$G\,\Box\, H$ with diam$(H)=2$ {#sec:diam2}
==============================
In this section we consider Cartesian products of graphs where (at least) one factor has diameter $2$. Motivation for the study of such graphs arises from the previous section. An EOD-set of $G\,\Box\, H$ that is parallel with respect to $G$ when $\mathrm{diam}(H)=2$ shares an important property with such a set in $G \,\Box\, K_r$ for $r \ge 3$. This is given in the following lemma.
\[lem:atmosttwo\] Let $H$ be a graph of diameter 2 and let $G$ be a graph such that $G\,\Box\, H$ has an EOD-set $D$. For every vertex $g$ in $G$, $|D \cap\,^{g}H| \le 2$. If in addition $D$ is parallel with respect to $G$, then $|D \cap\,^{g}H| \le 1$ for every $g\in V(G)$. If $|D \cap\,^{g}H| = 2$, then the two distinct vertices in $D \cap\,^{g}H$ are adjacent.
Assume that $D$ is an EOD-set of $G \,\Box\,
H$ and suppose that $(g,u)$ and $(g,v)$ are distinct vertices in $D$. The graph $H$ has diameter 2, and this implies that $uv\in E(H)$ or $u$ and $v$ have a common neighbor $w$ in $H$. Since every vertex in $^{g}H$ is dominated exactly once by $D$, we infer that $(g,u)$ and $(g,v)$ are adjacent, and $|D \cap\,^{g}H| \le 2$. It follows immediately that if $D$ is parallel with respect to $G$, then no $H$-layer can contain two members of $D$. $\,\Box\,$
As we will see, finding an appropriate weak partition of vertices in $G$ will be useful in the characterization of (parallel) EOD-graphs among Cartesian products $G\,\Box\, H$ where $\mathrm{diam}(H)=2$. First we show that the Cartesian product of a graph of diameter $2$ and a tree on at least three vertices does not admit a parallel EOD-set with respect to the tree.
\[diameter\] Let $H$ be a graph with $\mathrm{diam}(H)=2$ and let $T$ be a tree. If $T$ is different than $K_{2}$, then $T\,\Box\, H$ does not contain a parallel EOD-set with respect to $T$.
Let $H$ be a graph with $\mathrm{diam}(H)=2$. Suppose, in order to obtain a contradiction, that there exists a tree $T$ different than $K_{2}$, such that $T\,\Box\, H$ admits a parallel EOD-set $D$ with respect to $T$.
First, we claim that $T^{h}\cap D= \emptyset$ for every non-universal vertex $h$ in $H$. If this does not hold, then there exist vertices $(u_{0},h),(v_{0},h)\in D$ which are adjacent in $T\,\Box\, H$. Since $h$ is not universal in $H$, there is $h' \in V(H)$ such that $d_H(h,h')=2$. Observe that $(u_{0},h')$ and $(v_{0},h')$ are not dominated by $(u_{0},h)$ and $(v_{0},h)$. Moreover, they are not dominated by any vertex in $^{u_0}\!H$ and $^{v_0}\!H$ (since $\mathrm{diam}(H)=2$ we have that $|^{x}H\cap D|\leq 1$ for every $x\in V(T)$ by Lemma \[lem:atmosttwo\]). Therefore, there exists a neighbor $u_{1}$ of $u_{0}$ and a neighbor $v_{1}$ of $v_{0}$, such that $(u_{0},h')$ is dominated by $(u_{1},h')\in D$ and $(v_{0},h')$ is dominated by $(v_{1},h')\in D$. Moreover, since $(u_{1},h'),(v_{1},h')\in D$, there exist $(u_{2},h'),(v_{2},h')\in D$, where $u_{2}u_{1},v_{2}v_{1}\in E(T)$. To dominate vertices $(u_{2},h)$ and $(v_{2},h)$, there must exist $(u_{3},h),(v_{3},h)\in D$ where $u_{3}u_{2},v_{3}v_{2}\in E(T)$. Continuing in this way we obtain a two-way infinite walk $\ldots
u_{2}u_{1}u_{0}v_{0}v_{1}v_{2}\ldots$ in $T$. Since $T$ is a tree, all vertices of this walk are pairwise different. But this is in contradiction with $T$ being finite, and the claim is proved.
We infer that $H$ has to contain universal vertices and that the projection of every edge in $\left\langle D \right\rangle$ onto $H$ is a universal vertex. Now, let $h$ be a universal vertex of $H$, such that $T^{h}\cap D\neq \emptyset $ and let $(u,h),(v,h)\in D$ be adjacent vertices. Together they dominate all vertices of $^{u}H$ and $^{v}H$. There also exists a non-universal vertex $h'$ in $H$ because $\mathrm{diam}(H)=2$. Since $T$ is different than $K_{2}$, at least one of $u$ and $v$, say $u$, has a neighbor $w$ in $T$. Note that $(w,h)$ is dominated by $(u,h)$, and $(w,h')$ is not dominated by $(u,h)$ nor $(v,h)$. Since $
T^{h'}\cap D=\emptyset $, there exists another universal vertex $h_{1}\in V(H)$, such that $(w,h')$ is dominated by $(w,h_{1})\in D$. This yields a final contradiction, since $(w,h)$ is dominated by both $(u,h)$ and $(w,h_{1})$ from $D$, which is not possible in an EOD-set $D$.
$G\,\Box\, K_{m,n}$\[sec:completebipartite\]
--------------------------------------------
In this subsection we give a necessary and sufficient condition on a graph $G$ such that $G\,\Box\, K_{m,n}$ is an EOD-graph for $1 \le m\le n$. The condition will be the existence of a weak partition of $V(G)$ that satisfies very specific requirements. While it may not be easy to determine whether a given graph $G$ has such a weak partition, the requirements of the weak partition will make it straightforward to construct graphs $G$ such that $G \,\Box\, K_{m,n}$ is an EOD-graph.
Since $K_{m,n}$ has diameter 2 and we are not requiring the EOD-set of $G \,\Box\, K_{m,n}$ to be parallel with respect to $G$, we will refer often to Lemma \[lem:atmosttwo\]. For ease of explanation we assume throughout this subsection that $1 \le m \le n$ and that $K_{m,n}$ has partite sets $A$ and $B$ given by $A=\{1,\ldots,m\}$ and $B=\{m+1,\ldots,m+n\}$. With this notation we let $\mathcal{C}_{m,n}$ be a weak partition of $V(G)$ containing $mn+m+n+1$ parts indexed as follows:
- $V_0,V_1,\ldots, V_m, V_{m+1},\ldots, V_{m+n}$; and
- $V_{[i,m+j]}$ for $1 \le i \le m$ and $1 \le j \le n$.
We will say that $\mathcal{C}_{m,n}$ is *$K_{m,n}$-amenable* if it is a weak partition satisfying the following conditions.
- For $1 \le i \le m+n$, the induced subgraph $\left\langle V_{i}\right\rangle$ is a matching.
- For $1 \le i \le m$ and $m+1 \le j \le m+n$, $\left\langle V_{i}\cup V_j\right\rangle$ is a matching.
- For $1 \le i<j\le m$ or $m+1\le i < j \le m+n$, each $x$ in $V_i$ has exactly one neighbor in $V_j$ and each $y$ in $V_j$ has exactly one neighbor in $V_i$.
- If $x \in V_{[i,m+j]}$ for some $1 \le i \le m$ and some $1 \le j \le n$, then $N(x) \subseteq V_0$.
- If $x \in V_0$, then $|N(x) \cap \left( \cup_{1 \le j\le n}V_{[i,m+j]} \cup V_i\right)|=1$ for $1 \le i \le m$, and\
$|N(x) \cap \left( \cup_{1 \le i\le m}V_{[i,m+j]} \cup V_{m+j}\right)|=1$ for $1 \le j \le n$.
A graph $G$ will be called *$K_{m,n}$-amenable* if $V(G)$ has a weak partition that is $K_{m,n}$-amenable. With this definition we are now able to give a constructive characterization of those graphs $G$ such that $G \,\Box\, K_{m,n}$ is an EOD-graph.
\[thm:completebiparitite\] Let $m$ and $n$ be positive integers such that $m \le n$ and let $G$ be a graph. The Cartesian product $G \,\Box\, K_{m,n}$ is an EOD-graph if and only if $G$ is $K_{m,n}$-amenable.
Assume that $G$ is $K_{m,n}$-amenable and that $\mathcal{C}_{m,n}$ is a weak partition of $V(G)$ indexed as above and satisfying the conditions (I)-(V) in the definition above. We define a subset $D$ of $V(G \,\Box\, K_{m,n})$ by specifying its intersection with each $K_{m,n}$-layer. If $r$ is an integer such that $1 \le r \le m+n$ and $g \in V_r$, then $D \cap\,^{g}K_{m,n}=\{(g,r)\}$. If $r$ and $s$ are integers with $1 \le r \le m$ and $1 \le s \le n$ such that $g \in V_{[r,m+s]}$, then $D \cap\,^{g}K_{m,n}=\{(g,r),(g,m+s)\}$. Finally, if $g \in V_0$, then $D \cap\,^{g}K_{m,n}=\emptyset$. Since $\mathcal{C}_{m,n}$ is a weak partition, the set $D$ is well-defined. We now show that $D$ is an EOD-set of $G \,\Box\, K_{m,n}$ by showing that each vertex of $G \,\Box\, K_{m,n}$ has exactly one neighbor in $D$.
Let $(x,t)$ be an arbitrary vertex in $G \,\Box\, K_{m,n}$. Assume $x \in V_0$. Suppose first that $1 \le t \le m$. By (V) there exists $y \in V(G)$ such that $\{y\}= N(x) \cap \left( \cup_{1 \le j\le n}V_{[t,m+j]} \cup V_t\right)$. This implies that $(y,t) \in D$ and that $(y,t)$ dominates $(x,t)$. Furthermore, it follows from (V) and $D \cap\,^{x}K_{m,n}=\emptyset$ that $(y,t)$ is the only neighbor of $(x,t)$ that belongs to $D$. The case $m+1 \le t\le m+n$ is similar. Assume next that $x \in V_{[r,m+s]}$ for some $r$ and $s$ such that $1 \le r \le m$ and $1 \le s \le n$. By the definition of $D$ we get that both $(x,r)$ and $(x,m+s)$ belong to $D$. Exactly one of these is adjacent to $(x,t)$. Combining this with property (IV) it follows that $(x,t)$ has exactly one neighbor in $D$. Finally, assume that $x \in V_r$ for some $r$ with $1 \le r \le m$. (The case $m+1 \le r \le m+n$ is similar.) This means that $(x,r) \in D$ and $|D \cap\,^{x}K_{m,n}|=1$. There are three subcases to consider, namely (i) $m+1 \le t \le m+n$, (ii) $t=r$, and (iii) $t \neq r$ but $1 \le t \le m$. If $m+1 \le t \le m+n$, then $(x,r)$ dominates $(x,t)$ (from within the layer $^{x}K_{m,n}$). From (I), (II) and (IV) we see that $(x,t)$ is not adjacent to any vertex in $D \cap G^t$. Thus, in subcase (i) $(x,t)$ has a unique neighbor in $D$. Assume that $t=r$. By (I) there is a unique $y \in V_r\cap N(x)$. By definition $(y,r)\in D$ and thus $(x,r)$ is dominated by $D$. Properties (I) and (IV) together imply that $(x,t)$ has no other neighbor in $D$. Finally, assume subcase (iii) holds. By (III) there exists a unique $z \in V_t \cap N(x)$. Now $(z,t)\in D$ and $(x,t)$ is dominated by $(z,t)$. Consequently, by (IV) it follows that $(z,t)$ is the only vertex in $D$ that dominates $(x,t)$. We have shown that $D$ is an EOD-set of $G \,\Box\, K_{m,n}$.
Conversely, suppose that $S$ is an EOD-set of $G\,\Box\, K_{m,n}$. Since $K_{m,n}$ has diameter 2, we apply Lemma \[lem:atmosttwo\] and conclude that $|S \cap\,^{g}K_{m,n}| \le 2$ for every vertex $g$ in $G$. We produce a weak partition $\mathcal{C}$ of $V(G)$ as follows. The sets in $\mathcal{C}$ are those in the following specifications. Note that some of these subsets might be empty.
- $V_0=\{ x \in V(G): S \cap\,^xK_{m,n}=\emptyset\}$,
- $V_i=\{ x \in V(G): S \cap\,^xK_{m,n}=\{(x,i)\}\}$ for $1 \le i \le m+n$,
- $V_{[i,m+j]}=\{ x \in V(G): S \cap\,^xK_{m,n}=\{(x,i),(x,m+j)\}\}$ for $1 \le i \le m$ and $1 \le j \le n$.
The verification that $\mathcal{C}$ is $K_{m,n}$-amenable (that is, it satisfies properties (I)-(V)) follows directly from the assumption that $S$ is an EOD-set of $G \,\Box\, K_{m,n}$ and is left to the reader.
(6,0) coordinate (a); (4,-2) coordinate (b); (-2,-2) coordinate (c); (0,0) coordinate (d); (2,0) coordinate (e); (4,2) coordinate (f); (-2,2) coordinate (g); (8,0) coordinate (u); (8,-2) coordinate (v); (10,0) coordinate (w); (10,-2) coordinate (x); (6,-2) coordinate (y); (a) – (b); (a) – (u); (a) – (y); (b) – (c); (b) – (e); (b) – (f); (c) – (d); (c) – (g); (a) – (f); (d) – (e); (d) – (g); (e) – (f); (f) – (g); (u) – (v); (u) – (w); (v) – (x); (x) – (w); (a) – (f); (a) \[fill=black\] circle (); (b) \[fill=black\] circle (); (c) \[fill=black\] circle (); (d) \[fill=black\] circle (); (e) \[fill=black\] circle (); (f) \[fill=black\] circle (); (g) \[fill=black\] circle (); (y) \[fill=black\] circle (); (u) \[fill=black\] circle (); (v) \[fill=black\] circle (); (w) \[fill=black\] circle (); (x) \[fill=black\] circle (); (6,.7) node [$0$]{}; (6,-2.7) node [$[2,5]$]{}; (-.7,0) node [$5$]{}; (2.7,0) node [$5$]{}; (-2,2.7) node [$3$]{}; (4,2.7) node [$3$]{}; (4,-2.7) node [$4$]{}; (-2,-2.7) node [$4$]{}; (8,.7) node [$1$]{}; (8,-2.7) node [$1$]{}; (10,.7) node [$2$]{}; (10,-2.7) node [$2$]{};
The graph $G$ in Figure \[fig:K23\] was constructed to have a weak partition that is $K_{2,3}$-amenable. The partite sets of $K_{2,3}$ are as in the development above, $A=\{1,2\}$ and $B=\{3,4,5\}$. For simplicity the vertices of $G$ are labeled to indicate the subset of the weak partition that contains them. For example, the vertices labeled $1$ are in $V_1$ while the vertex labeled $[2,5]$ is the only member of $V_{[2,5]}$. By Theorem \[thm:completebiparitite\] the Cartesian product $G \,\Box\, K_{2,3}$ is an EOD-graph.
$G\,\Box\, C_{r}, r\in\{4,5\}$\[5cycle\]
----------------------------------------
In this subsection we first define a type of weak partition of $V(G)$ that will enable us to characterize those Cartesian products $G\,\Box\, C_{5}$ that are parallel EOD-graphs with respect to $G$. To describe these weak partitions we need to modify Condition (C) as it was stated in Section \[sec:completefactor\] and add an additional condition. The operations on the subscripts in these new conditions are made modulo $5$.
- $\left\langle V_{i}\cup V_{i+1}\right\rangle $ is a matching in $G$ for every $i\in [5]$,
- if $x\in V_{i}$, then $|N(x)\cap V_{i+2}|=1$ and $|N(x)\cap V_{i-2}|=1$ for every $i\in [5]$.
Notice that the condition (C$'$) is weaker than (C). We say that $G$ is *$C_{5}$-parallel amenable* if there exists a weak partition $\{V_{0},V_{1},V_{2},V_{3},V_{4},V_{5}\}$ of $V(G)$ that satisfies conditions (A), (B), (C$'$) and (D).
\[5cycle1\] For any graph $G$, the Cartesian product $G\,\Box\, C_{5}$ is a parallel EOD-graph with respect to $G$ if and only if $G$ is a $C_{5}$-parallel amenable graph.
Assume first that $G$ is a $C_{5}$-parallel amenable graph and let $\{V_{0},V_{1},V_{2},V_{3},V_{4},V_{5}\}$ be a weak partition of $V(G)$ that satisfies conditions (A), (B), (C$'$) and (D). We define a subset $D$ of $V(G\,\Box\, C_{5})$ by $D=\{(g,i) : g\in V_i\,\,\mbox{for}\,\, i\in[5]\}$. Notice that $|D\cap\,^{g}C_{5}|=1$ for every $g\in V(G)-V_0$. We will show that every vertex of $G\,\Box\, C_{5}$ is dominated by exactly one vertex of $D$. Let $(g,j)$ be an arbitrary vertex of $G\,\Box\, C_{5}$.
Assume first that $g\in V_{i}$ for some $i\in [5]$. If $j \in \{i-1,i+1\}$, then $(g,j)$ is dominated by $(g,i)$. Moreover, $(g,j)$ is dominated only by $(g,i)$ in $D$, since $(g,i)$ is the only vertex in $D\cap\,^{g}C_{5}$ and (B) and (C$'$) hold. If $j=i$, then $(g,j)$ is dominated by $(g',i)$, where $gg'$ is an edge in $\left\langle V_{i}\right\rangle $ (notice that $g'$ exists by (B)). Note that $(g,j)$ is dominated only by $(g',i)$ from $D$ by (B) and the fact that $|D\cap\,^{g}C_{5}|=1$. It remains to consider $j=i+2$ and $j=i-2$. Assume $j=i+2$; the case $j=i-2$ is similar. By condition (D), $g$ has a unique neighbor $x_{i+2}\in V_{i+2}$. By definition $(x_{i+2},i+2)\in D$, and thus $(x_{i+2},i+2)$ dominates $(g,j)$. As before, since $|D\cap\,^{g}C_{5}|=1$ and since condition (D) holds, it follows that $(x_{i+2},i+2)$ is the only vertex of $D$ that dominates $(g,j)$. Hence, if $g\in V_i$ for some $i\in[5]$, then $(g,j)$ is dominated exactly once by $D$. Finally, assume that $g\in V_{0}$. By condition (A), $g$ has a unique neighbor $x_{j}\in V_{j}$. This implies that $(x_{j},j)\in D$ and that $(x_j,j)$ is the only vertex in $D$ that dominates $(g,j)$. Consequently, $D$ is a parallel EOD-set of $G\,\Box\, C_{5}$.
Conversely, let $G\,\Box\, C_{5}$ be a parallel EOD-graph and let $D$ be a parallel EOD-set with respect to $G$. Let $V_0,V_1,V_2,V_3,V_4,V_5$ be subsets of $V(G)$ defined as follows. If $|D\cap\,^gC_5|=0$, then $g\in V_{0}$. For $i\in [5]$, $g\in\,V_{i}$ if and only if $\{(g,i)\}=D\cap\,^{g}C_{5}$. By Lemma \[lem:atmosttwo\], $|D\cap\,^{g}C_{5}|\leq 1$ for every $g\in V(G)$, and thus $\{V_{0},V_1,V_2,V_3,V_4,V_{5}\}$ is a weak partition of $V(G)$. Note that only $V_0$ can be empty.
We will show that this weak partition satisfies conditions (A), (B), (C$'$) and (D). If (A) does not hold, then there exists a vertex $(g,i)$ where $|D\cap\,^{g}C_{5}|=0$ and either $|N(g)\cap V_{i}|=0$ or $|N(g)\cap V_{i}|>1$ for some $i \in [5]$. In the first case $(g,i)$ is not dominated by $D$ and in the second case $(g,i)$ is dominated by at least two vertices, both contradicting the fact that $D$ is a parallel EOD-set. Thus, (A) holds. If (B) is not satisfied, then there exists $g\in V_{i}$, for some $i\in [5]$, such that either $\deg _{\left\langle V_{i}\right\rangle }(g)=0$ or $\deg
_{\left\langle V_{i}\right\rangle }(g)>1$, which yields exactly the same contradiction as for (A). Hence, (B) is true as well. If (C$'$) does not hold, then there exist $g\in V_{i}$ and $g'\in V_{i+1}$ for some $i\in [5]$, such that $gg'\in E(G)$. We infer that $(g,i+1)$ is dominated twice, that is by $(g,i)$ and by $(g',i+1)$, which is not possible. Finally, if (D) does not hold, then for some $i\in [5]$, there exists $x\in V_{i}$ such that $|N(x)\cap V_{i+2}|\neq 1$ or $|N(x)\cap V_{i-2}|\neq 1$. Again we get that some vertex is not dominated by $D$ (if $|N(x)\cap V_{i+2}|=0=|N(x)\cap V_{i-2}|$) or that some vertex is dominated more than once by $D$ (if $|N(x)\cap V_{i+2}|>1$ or $|N(x)\cap V_{i-2}|>1$), which is not possible. This shows that (D) is also true, which completes the proof.
While the complete characterization of EOD Cartesian products where one factor is $C_4\cong K_{2,2}$ was given in Subsection \[sec:completebipartite\], here we describe all $G$ such that $G\,\Box\, C_4$ is a parallel EOD-graph with respect to $G$. For $C_{4}$ notice that computations on the subscripts are done modulo $4$ in the set $[4]$, and in this case $i+2=i-2$. Thus, we can restate condition (D) as
- if $x\in V_{i}$, then $|N(x)\cap V_{i+2}|=1$ for every $i\in [4]$.
We say that $G$ is $C_{4}$-parallel amenable if there exists a weak partition $\{V_{0},V_{1},V_{2},V_{3},V_{4}\}$ of $V(G)$ that fulfills conditions (A), (B), (C$'$) and (D$'$). The proof of the following theorem follows the same lines as the proof of Theorem \[5cycle1\] if we take into consideration computation modulo $4$ instead of modulo $5$.
\[c4\] For any graph $G$, the Cartesian product $G\,\Box\, C_{4}$ is a parallel EOD-graph with respect to $G$ if and only if $G$ is a $C_{4}$-parallel amenable graph.
For $r\in \{4,5\}$ there exist many graphs $G$ which are not $C_r$-parallel amenable, but for which $G\,\Box\, C_{4}$ is an EOD-graph (clearly $G\,\Box\, C_{4}$ is not a parallel EOD-graph with respect to $G$ in this case). One of the smallest examples is $P_3$, which is not $C_4$-parallel amenable, but $P_3\,\Box\, C_{4}$ is an EOD-graph, even a parallel EOD-graph with respect to $C_4$.
Conclusion
==========
As already mentioned, this method of defining weak partitions is most easily implemented when one of the graphs has small diameter. Despite this fact, there is no reason why one should not use it on graphs with larger diameter. We illustrate this idea on a special case from the class of cycles.
Our goal is to define a weak partition of a graph $G$ that consists of $V_0$ and a family of sets $V_A$ where $A$ is a subset of $[k]$ with certain properties. We derive these properties from the second graph in the product, which is $C_k$ now. Again we have two possibilities for an edge from $\left\langle D\right\rangle$, where $D$ is an EOD-set of $G\,\Box\, C_k$: either it projects to $C_k$ as an edge or as a vertex. If it projects to an edge in $C_k$, then $A$ must contain two consecutive elements $i$ and $i+1$. If an edge projects to a single vertex $j\in V(C_k)$, then $j\in A$ but neither $j+1$ nor $j-1$ is in $A$. Moreover, two non-consecutive elements of $A$ must differ by at least 3 modulo $k$, so that no vertex in the product is dominated more than once.
(6,0) coordinate (a); (4,-2) coordinate (b); (-2,-2) coordinate (c); (0,0) coordinate (d); (2,0) coordinate (e); (4,2) coordinate (f); (-2,2) coordinate (g); (-4,0) coordinate (h); (a) – (b); (g) – (h); (c) – (h); (b) – (c);(b) – (f);(a) – (f); (f) – (g); (d) – (e); (d) – (g); (a) – (f); (a) \[fill=black\] circle (); (b) \[fill=black\] circle (); (c) \[fill=black\] circle (); (d) \[fill=black\] circle (); (e) \[fill=black\] circle (); (f) \[fill=black\] circle (); (g) \[fill=black\] circle (); (h) \[fill=black\] circle (); (7,0) node [$[1,2]$]{}; (-1,0) node [$[1,2]$]{}; (3,0) node [$[4,5]$]{}; (-2,2.7) node [$0$]{}; (4,2.7) node [$[4,5]$]{}; (4,-2.7) node [$0$]{}; (-2,-2.7) node [$[3,6]$]{}; (-5,0) node [$[3,6]$]{};
In particular, for $C_6$ we obtain the following weak partition: $$V_0,V_1,V_2,V_3,V_4,V_5,V_6,V_{[1,2]},V_{[2,3]},V_{[3,4]},V_{[4,5]},
V_{[5,6]},V_{[6,1]},V_{[1,4]},V_{[2,5]},V_{[3,6]}\,.$$ Clearly the size of the weak partition increases with $k$. Together with this weak partition, several conditions are needed as well. For instance we need a condition similar to (A) and (V) to care about all vertices from $V_0$. As in the case of $K_{2,3}$-amenable graphs, it seems to be hard to decide whether a graph $G$ is a “$C_k$-amenable” graph. However, it is not difficult to construct (small) examples of such graphs. An example of a “$C_6$-amenable” graph is given in Figure \[fig:C6\]. The labeling follows the prescription given for Figure \[fig:K23\].
[^1]: Partially supported by the Ministry of Science of Slovenia under the grant P1-0297.
[^2]: This author was supported by the grant “Internationalisation–a pillar of development of University of Maribor.” and by a grant from the Simons Foundation (\#209654 to Douglas F. Rall).
[^3]: This author was partially supported by Slovenian research agency ARRS, program no. P1-00383, project no. L1-4292, and Creative Core-FISNM-3330-13-500033.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Optimal $B$-robust estimate is constructed for multidimensional parameter in drift coefficient of diffusion type process with small noise. Optimal mean-variance robust (optimal $V$-robust) trading strategy is find to hedge in mean-variance sense the contingent claim in incomplete financial market with arbitrary information structure and misspecified volatility of asset price, which is modelled by multidimensional continuous semimartingale. Obtained results are applied to stochastic volatility model, where the model of latent volatility process contains unknown multidimensional parameter in drift coefficient and small parameter in diffusion term.'
author:
- 'N. Lazrieva, T. Toronjadze'
title: 'Optimal Robust Mean-Variance Hedging in Incomplete Financial Markets'
---
Georgian–American University, Business School, 3 Alleyway II,\
Chavchavadze Ave. 17a, Tbilisi, Georgia, E-mail: [email protected]\
A. Razmadze Mathematical Institute, 1, M. Aleksidze St., Tbilisi, Georgia
Introduction, Motivation and Results
====================================
The hedging and pricing of contingent claims in incomplete financial markets, and dynamic portfolio selection problems are important issues in modern theory of finance. These problems are associated due to the so-called mean-variance approach.
For determining a “good” hedging strategy in incomplete market with arbitrary information structure $F=(\cF)_{0\leq t\leq T}$, one riskless asset and $d$, $d\geq 1$, risky assets, whose price process is a semimartingale $X$, the mean-variance approach suggests to use the quadratic criterion to measure the hedging error, i.e. to solve the mean-variance hedging problem introduced by Föllmer and Sondermann [@10]: $$\label{4-1.1}
\text{minimize} \;\;\; E\Bigg( H-x-\int_0^T \tht_t d X_t\Bigg)^2
\;\;\;\text{over all} \;\;\; \tht\in \Tht,$$ where contingent claim $H$ is a $\cF_T$-measurable square-integrable random variable (r.v.), $x$ is an initial investment, $\Tht$ is a class of admissible trading strategies, $T$ is an investment horizon.
The mean-variance formulation by Markowitz [@27], provides a foundation for a single period portfolio selection (see, also Merton [@28]). In recent paper of Li and Ng [@23] the concept of Markowitz’s mean-variance formulation for finding the optimal portfolio policy and determining the efficient frontier in analytical form has been extended to multiperiod portfolio selection.
As it pointed out in Li and Ng [@23] the results on multiperiod mean-variance formulation with one riskless asset can be derived using the results of the mean-variance hedging formulation.
Therefore, the mean-variance hedging is s powerful approach for both above mentioned major problems.
The problem (\[4-1.1\]) was intensively investigated in last decade (see, e.g., Dufiie and Richardson [@9], Schwezer [@37], [@38], [@39], Delbaen et al. [@8], Monat and Striker [@29], Rheinländer and Schweizer [@34], (RSch hereafter), Pham et al. [@32], Gourieroux et al. [@11] (GLP hereafter), Laurent and Pham [@19]).
A stochastic volatility model, proposed by Hull and White [@13] and Scott [@40], where the stock price volatility is an random process, is a popular model of incomplete market, where the mean-variance hedging approach can be used (see, e.g., Laurent and Pham [@19], Biagini et al. [@13], Mania and Tevzadze [@25], Pham et al. [@32]).
Consider the stochastic volatility model described by the following system of SDE $$\label{4-1.2}
\begin{gathered}
d X_t=X_t \, dR_t, \quad X_0>0, \\
d R_t=\mu_t(R_t,Y_t)\,dt +\sg_. dw_t^R, \quad R_0=0, \\
\sg_t^2=f(Y_t), \\
dY_t=a(t,Y_t;\al)\,dt+\ve\,d w_t^\sg, \quad Y_0=0,
\end{gathered}$$ where $w=(w^R, w^\sg)$ is a standard two-dimensional Wiener process, defined on complete probability space $(\Om,\cF, P)$, $F^w=(\cF_t^w)_{0\leq t\leq T}$ is the $P$-augmentation of the natural filtration $\cF_t^w=\sg(w_s, 0\leq s\leq t)$, $0\leq t\leq T$, generated by $w$, $f(\cdot)$ is a continuous one-to-one positive locally bounded function (e.g., $f(x)=e^x$), $\al=(\al_1,\dots,\al_m)$, $m\geq 1$, is a vector of unknown parameters, and $\ve$, $0<\ve\ll 1$, is a small number. Assume that the system (\[4-1.2\]) has an unique strong solution.
This model is analogous to the model proposed by Renault and Touzi [@33] (RT hereafter). The principal difference is the presence of small parameter $\ve$ in our model, which due to the assumption that the volatility of randomly fluctuated volatility process is small (see, also Sircar and Papanicolau [@41]). Thus assumption enables us to use the prices of trading options with short, nearest to the current time value maturities for volatility process filtration and parameter estimation purposes (see below). In contrast, RT [@33] needs to assume that there exist trading derivatives with any (up to the infinity) maturities.
Important feature of the stochastic volatility models is that volatility process $Y$ is unobservable (latent) process. To obtain explicit form of optimal trading strategy full knowledge of the model of the process $Y$ is necessary and hence one needs to estimate the unknown parameter $\al=(\al_1,\dots,\al_m)$, $m\geq 1$.
A variety of estimation procedures are used, which involve either direct statistical analysis of the historical data or the use of implied volatilities extracted from prices of existing traded derivatives.
For example, one can use the following method based on historical data.
Fix the time variable $t$. From observations $X_{t_0^{(n)}},\dots,X_{t_n^{(n)}}$, $0=t_0^{(n)}<\cdots<t_n^{(n)}=t$, $\max\limits_j [t_{j+1}^{(n)}-t_j^{(n)}] \to 0$, as $n\to 0$, calculate the realization of yield process $R_t=\int\limits_0^t \frac{dX_s}{X_s}$, and then calculate the sum $$S_n(t) =\sum_{j=0}^{n-1} |R_{t_{j+1}^{(n)}} -R_{t_j^{(n)}}|^2.$$
It is well-known (see, e.g., Lipster and Shiryaev [@24]) that $$S_n(t) \str{P}{\to} \int_0^t \sg_s^2\,ds \;\;\;\text{as} \;\;\; n\to \infty.$$
Since $\sg_t^2(\om)=f(Y_t)$ is a continuous process we get $$\sg_t^2(\om) =\lim_{\Dl \downarrow 0} \frac{F(t+\Dl,\om)-F(t,\om)}{\Dl}\,,$$ where $F(t,\om)=\int\limits_0^t \sg_s^2(\om)ds$.
Hence, the realization $(y_t)_{0\leq t\leq T}$ of the process $Y$ can be found by the formula $y_t=f^{-1}(\sg_t^2)$, $0\leq t\leq T$.
More sofisticated methods using the same idea can be found, e.g., in Chesney et al. [@5], Pastorello [@31].
We can use the reconstructed sample path $(y_t)$, $0\leq t\leq T$. to estimate the unknown parameter $\al$ in the drift coefficient of diffusion process $Y$.
The second, market price adjusted procedure of reconstruction the sample path of volatility process $Y$ and parameter estimate was suggested by RT [@33], where they used implied volatility data.
We present a quick review of this method, adapted to our model (\[4-1.2\]).
Suppose that the volatility risk premium $\lb^\sg\equiv 0$, meaning that the risk from the volatility process is non-compensated (or can be diversified away). Then the price $C_t(\sg)$ of European call option can be calculated by the Hull and White formula (see, e.g., RT [@33]), and Black-Scholes (BS) implied volatility $\sg^i(\sg)$ can be found as an unique solution of the equation $$C_t(\sg)=C_t^{BS}(\sg^i(\sg)),$$ where $C^{BS}(\sg)$ denotes the standard BS formula written as a function of the volatility parameter $\sg$.
Here (for further estimational purposes) only at-the-money options are used.
Under some technical assumptions (see Proposition 5.1 of RT [@33], and Bujeux and Rochet [@24] for general diffusion of volatility process) $$\label{4-1.3}
\frac{\pa\sg_t^i(\sg,\al)}{\pa \sg_t}>0$$ (remember that the drift coefficient of process $Y$ depends on unknown parameter $\al$).
Fix current value of time parameter $t$, $0\leq t\leq T$, and let $0<T_1<T_2<\cdots<T_{k-1}<t<T_k$ be the maturity times of some traded at-the-money options.
Let $\sg_{t_j^\ve}^{i^*}$ be the observations of an implied volatility at the time moments $0=t_0^\ve<t_1^\ve<\cdots<t_{[\frac{t}{\ve}]}=t$, $\max\limits_j [t_{j+1}^\ve-t_j^\ve]\to 0$, as $\ve\to 0$.
Then, using (\[4-1.3\]), and solving the equation $$\sg_{t_j^\ve}^i(\sg_{t_j^\ve},\al)=\sg_{t_j^\ve}^{i^*},$$ one can obtained the realization $\{\wt\sg_{t_j^\ve}\}$ of the volatility $(\sg_t)$, and thus, using the formula $y_{t_j^\ve}=f^{-1}(\wt\sg_{t_j^\ve}^2)$, the realization $\{y_{t_j^\ve}\}$ of volatility process $(Y_t)$, which can be viewed as the realization of nonlinear AR(1) process: $$Y_{t_{j+1}^\ve} -Y_{t_j^\ve} =a(t_j^\ve,Y_{t_j^\ve};\al)
(t_{j+1}^\ve-t_j^\ve) +\ve(w_{t_{j+1}^\ve}^\sg-w_{t_j^\ve}^\sg).$$
Using the data $\{y_{t_j^\ve}\}$ one can construct the MLE $\wh\al_t^\ve$ of parameter $\al$, see, e.g., Chitashvili et al. [@26], [@27], Lazrieva and Toronjadze [@20].
Remember the scheme of construction of MLE. Rewrite the previous AR(1) process, using obvious simple notation, in form $$Y_{j+1}-Y_j =a(t_j,Y_j;\al)\Dl +\ve \Dl w_j^\sg.$$
Then $$\begin{aligned}
\frac{\pa}{\pa y}\, P\{Y_{j+1}\leq y \mid Y_j\} & =
\frac{1}{\sqrt{2\pi\Dl\ve}}\,\exp
\left( -\frac{(y-Y_j-a(t_j,Y_j;\al)\Dl)^2}{2\ve^2\Dl} \right) \\
& =: \vf_{j+1}(y,Y_j;\al),\end{aligned}$$ and the $\log$-derivative of the likelihood process $\ell_t=(\ell_t^{(1)},\dots,\ell_t^{(m)})$ is given by the relation $$\ell_t^{(i)}=\sum_j \ell_{j+1}^{(i)}, \quad i=\ol{1,m},$$ where $$\begin{gathered}
\ell_{j+1}^{(i)}(y;\al)=\frac{\pa}{\pa \al_i}\,
\ln\vf_{j+1}(y,Y_j;\al) \\
=\frac{1}{\ve^2\Dl}\,(y-Y_j-a(t_j,Y_j;\al)\Dl)
\dot{a}^{(i)}(t_j,Y_j;\al)\Dl.\end{gathered}$$ Hence MLE is a solution (under some conditions) of the system of equations $$\frac{1}{\ve^2\Dl} \sum_j (y_{j+1}-y_j-a(t_j,y_j;\al)\Dl)
\dot{a}^{(i)}(t_j,y_j;\al)\Dl =0, \quad i=\ol{1,m},$$ where the reconstructed data $\{y_j\}=\{y_{t_j^\ve}\}$ are substituted).
Following RT [@33] let us introduce the functionals $$\begin{gathered}
HW_\ve^{-1} :\wh\al_t^\ve(p) \to \left( y_{t_j^\ve}^{(p+1)},\;\;
0\leq j\leq \left[\frac{t}{\ve}\right] \right), \\
MLE_\ve :\left( y_{t_j^\ve}^{(p+1)},\;\;
0\leq j\leq \left[\frac{t}{\ve}\right] \right)\to \wh\al_t^\ve(p+1) \end{gathered}$$ and $$\phi_\ve=MLE_\ve \circ HW_\ve^{-1}.$$
Starting with some constant initial value (or preliminary estimate obtained, e.g., from historical data) one can compute a sequence of estimates $$\wh\al_t^\ve(p+1) =\phi_\ve(\wt\al_t^\ve(p)), \quad p\geq 1.$$ If the operator $\phi_\ve$ is a strong contraction in the neighborhood of the true value of the parameter $\al^0$, for a small enough $\ve$, then one can define the estimate $\wh\al_t^\ve$ as the limits of the sequence $\{\wh \al_t^\ve(p)\}_{p\geq 1}$. It was proved in RT [@33] that $\wh \al_t^\ve$ is a strong consistent estimate of the parameter $\al$.
Return to our consideration.
Interpolating on some way the corresponding (to the estimate $\wh \al_t^\ve$) realization $\{y_{t_j^\ve}\}$ we get the reconstructed continuous sample path $(y_s)_{0\leq s\leq t}$ of the latent process $Y$, which can be used for further analysis.
Unfortunately, both described statistical procedures are highly sensitive w.r.t errors in all steps of parameter identification process.
Hence, this is a natural place for introducing the robust procedure of parameter estimates.
Suppose that the sample path $(y_s)_{0\leq s\leq t}$ comes from the observation of process $(\wt Y_s)_{0\leq s\leq t}$ with distribution $\wt P_\al^\ve$ from the shrinking contamination neighborhood of the distribution $P_\al^\ve$ of the basic process $Y=(Y_s)_{0\leq s\leq t}$. That is $$\label{4-1.4}
\frac{d \wt P_\al^\ve}{d P_\al^\ve} \,\Big|\, \cF_t^w=\cE_t(\ve N^\ve),$$ where $N^\ve=(N_s^\ve)_{0\leq s\leq t}$ is a $P_\al^\ve$-square integrable martingale, $\cE_t(M)$ is the Dolean exponential of martingale $M$.
In the diffusion-type framework (\[4-1.4\]) represents the Huber gross error model (as it explain in Remark 2.2). The model of type (\[4-1.4\]) of contamination of measures for statistical models with filtration was suggested by Lazrieva and Toronjadze [@21], [@22].
In Section 2 we study the problem of construction of robust estimates for contamination model (\[4-1.4\]).
In subsection 2.1 we give a description of the basic model and definition of consistent uniformly linear asymptotically normal (CULAN) estimates, connected with the basic model (Definition 2.1).
In subsection 2.2 we introduce a notion of shrinking contamination neighborhood, described in terms of contamination of nominal distribution, which naturally leads to the class of alternative measures (see (\[4-2.18\]) and (\[4-2.19\])).
In subsection 2.3 we study the asymptotic behaviour of CULAN estimates under alternative measures (Proposition 2.2), which is the basis for the formulation of the optimization problem.
In subsection 2.4 the optimization problem is solved which leads to construction of optimal $B$-robust estimate (Theorem 2.1).
Based on the limit theorem (subsection 2.1), one can construct the asymptotic confidence region of level $\gm$ for unknown parameter $\al$ $$\lim_{\ve\to \infty} P_\al^\ve \left(\ve^{-2}(\al-\al_t^{*,\ve})' V^{-1}
(\psi^*;\al_t^{*,\ve})(\al-\al_t^{*,\ve}) \leq \chi_\gm^2\right)=1-\gm,$$ where $\chi_\gm^2$ is a quantile of order $1-\gm$ of $\chi^2$-distribution with $m$ degree of freedom, and $V(\psi^*;\al)$ is given by (\[4-2.17\]).
This region shrinks to the estimate $\al_t^{*,0}$, as $\ve\to 0$.
Now if the coefficient $a(t,y;\al)$ in (\[4-1.2\]) is such that the solution $Y_t^\ve(\al)$ of SDE (\[4-1.2\]) is continuous w.r.t parameter $\al$ (see, e.g., Krylov [@16]), then the confidence region of parameter $\al$ is mapped to the confidence interval for $Y_t^\ve(\al)$, which shrinks to $Y_t^*=Y_t^0(\al_t^{*,0})$, Further, by the function $f$, the latter interval is mapped to the confidence interval for $\sg_t$, which shrinks to $\sg_t^*=f^{1/2}(Y_t^0(\al_t^{*,0}))$. Denote $\sg_t^0$ the center of this interval. Then the interval can be written in the form $$\sg_t=\sg_t^0 +\dl(\ve)h_t,$$ where $\dl(\ve)\to 0$, as $\ve\to 0$, and $h\in \cH$ (see (\[4-3.18\])).
Thus, we arrive at the asset price model (\[4-1.2\]) with misspecified volatility, and it is natural to consider the problem of construction of the robust trading strategy to hedge a contingent claim $H$.
We investigate this problem in the mean-variance setting in Section 3. We consider the general situation, when the asset price is modelled by $d$-dimensional continuous semimartingale and the information structure is given by some general filtration.
In subsection 3.1 we give a description of the financial market model.
In subsection 3.2 we collect the facts concerning the variance-optimal equivalent local martingale measure, which plays a key role in the mean-variance hedging approach.
In last subsection 3.3 we construct “optimal robust hedging strategy” (Theorem 3.1) by approximating the optimization problem (\[4-3.24\]) by the problem (\[4-3.26\]). As it is mentioned in Remark 3.2, such approach and term are common in robust statistic theory. In contact to optimal $B$-robustness (see Section 2), here we develop the approach, known in robust statistics as optimal $V$-robustness, see Hampel et al. [@12].
Note that our approach allows incorporating current information on the underlying model, and hence is adaptive. Namely, passing from time value $t$ to $t+\tau$, $\tau>0$, when more information about market prices are available, the asymptotic variance-covariance of the constructed estimate $\al_t^{*,\ve}$ becomes smaller, and hence the estimation procedure becomes more precise.
In the paper of Runggaldier and Zaccaria [@36] the adaptive approach to risk management under general uncertainty (restricted information) was developed. As it is mentioned in this paper there exist a series of investigations dealt with various type of adaptive approaches (see list of references in [@36]). But in all these papers (except Runggaldier and Zaccaria [@36]) the uncertainty is only in the stock appreciation rate in contrast to our consideration, where the model misspecification is due to the volatility parameter.
The consideration of misspecified asset price models was initiated by Avellaneda et al. [@1], Avellaneda and Paras [@2].
Various authors in different settings attacked the robustness problem. The method used in Section 3 was suggested by Toronjadze [@42] for asset price process modelled by the one-dimensional process. As it will be shown in Remark 3.2 below, in simplest case when asset price process is a martingale w.r.t initial measure $P$, and it is possible to find the solution of “exact” optimization problem (\[4-3.24\]), this solution coincides with the solution of an approximating optimization problem (\[4-3.26\]). In more general situation (when asset price process is not more the $P$-martingale) investigation of the problem (\[4-3.24\]) by, e.g., control theory methods seems sufficiently difficult. Anyway, we do not know the solution of the problem (\[4-3.24\]).
Return to the stochastic volatility model (\[4-1.2\]) and describe successive steps of our approach:
1\) For each current time value $t$, $0<t<T$, reconstruct the sample path $(y_s)_{0\leq s\leq t}$, using the historical data or the tradable derivatives prices;
2\) Using the approach developed in Section 2, calculate the value $\al_t^{*,\ve}$ of the robust estimate of parameter $\al$ (i.e. construct the deterministic function $t\to \al_t^{*,\ve}\in R^m$) and then find the confidence region for parameter $\al$;
3\) Based on the volatility process model find the confidence interval for $Y_t(\al)$;
4\) Denoting $a^*(t,y)=a(t,y;\al_t^{*,\ve})$, where $a(t,y;\al)$ is a drift coefficient of volatility process, consider the stochastic volatility model with misspecified asset price model and fully specified volatility process model $$\begin{gathered}
dX_t=X_t\, dR_t, \quad X_0>0, \\
dR_t=(\sg_t^0+\dl(\ve)h_t) dM_t^0, \quad R_0=0, \\
dY_t=a^*(t,Y_t)\,dt+\ve\,d w_t^\sg, \quad Y_0=0, \quad 0\leq t\leq T,\end{gathered}$$ where $$dM_t^0=k_t\,dt +d w_t^R,$$ $h\in \cH$ and $\sg_t^0$ is the center of the confidence interval of volatility.
Using Theorem 3.1 construct the optimal robust hedging strategy by the formula (\[4-3.38\]), $$\tht_t^*=\frac{1}{\sg_t^0}\, \left[ \psi_t^{1,H} +\zt_t(V_t^*-
(\psi_t^H)'U_t\right],$$ where all objects are defined in Theorem 3.1.
It should be mentioned that if one constructs a hedging strategy $\wt\tht_t^*$ by the above-given formula with $\sg_t^{*,\ve}=f^{1/2}(Y_t^\ve(\al_t^{*,\ve}))$ instead of $\sg_t^0$, then the strategies $\wt\tht_t^*$ and $\tht_t^*$ would be different, since $\sg_t^{*,\ve}\neq \sg_t^0$, in general. Hence the value $\Dl_t=|\sg_t^{*,\ve}-\sg_t^0|$ defines the correction term between the robust, $\tht_t^*$ and non-robust, $\wt\tht^*$ strategies.
In nontrivial case, when $k_t=k(Y_t)$ the variance-optimal martingale measure $\wt P$ is given by (\[4-3.17\]), $\zt_t=-k_t\cE_t(-k\cdot M^0)$ (see subsection 3.2), and the process $(X_t,Y_t)_{0\leq t\leq T}$ is the Markov process. If $H=h(X_T,Y_T)$ $(h(x,y)$ is some function), then $\wt V_t^H=E^{\wt P}(H|\cF_t^w)=E^{\wt P}(h(X_T,Y_T)|\cF_t^w)=v(t,X_T,Y_T)$ and if, e.g., $v(t,x,y)\in C^{1,2,2}$, then $v$ is an unique solution of the following partial differential equation $$\frac{\pa v}{\pa t} +a^*\,\frac{\pa v}{\pa y}+\frac{1}{2}
\left(\ve^2\,\frac{\pa^2v}{\pa y^2}+x^2v^2\,\frac{\pa^2 v}{\pa x^2}\right)=0,$$ with the boundary condition $v(T,x,t)=h(x,y)$. More general situation with nonsmooth $v$ is considered in Laurent and Pham [@19], Mania and Tevzadze [@25].
Further, one can find the Galtchouck–Kunita–Watanabe decomposition of r.v. $H$ (see, e.g., Pham et al. [@32]) putting $$\xi_t^H=\frac{\pa v(t,X_t,Y_t)}{\pa x}\,, \quad
L_T^H=\ve \int_0^T \frac{\pa v}{\pa y}\,(t,X_t,Y_t)\,d w_t^\sg,$$ and calculate $\psi_t^H$, $L_T$ and $V_t^*$ using (4.13) and (4.14) of RSch [@34].
Thus one get the explicit solution of the mean-variance hedging problem.
Finally, here is the short summary of approach:
a\) Incorporate the robust procedure in statistical analysis of volatility process. That is construct and use in the model optimal $B$-robust estimate of unknown parameter in drift coefficient of volatility process.
Parameter estimation naturally leads to the asset price model misspecification.
b\) Incorporate the second robust procedure in financial analysis of contingent claim hedging. That is construct and use for hedging purposes optimal $V$-robust trading strategy.
In our opinion this “double robust” strategy should be more attractive to protect the hedger against the possible errors.
The general asymptotic theory of estimation can be found in Ibragimov and Khas’miskii [@14]; the theory of robust statistics is developed in Hampel et al. [@12] and in Rieder [@35]; the theory of the trend parameter estimates for diffusion process with small noise is developed in Kutoyants [@18]; the book of Musiela and Rutkowsky [@30] is devoted to the mathematical theory of finance and finally, the general theory of martingales can be found in Jacod and Shiryaev [@15].
Optimal $B$-Robust Estimates
============================
Basic model. CULAN estimates
----------------------------
The basic model of observations is described by the SDE $$\label{4-2.1}
d Y_s=a(s,Y;\al)\,ds +\ve\,dw_s, \quad Y_0=0, \quad 0\leq s\leq t,$$ where $t$ is a fixed number, $w=(w_s)_{0\leq s\leq t}$ is a standard Wiener process defined on the filtered probability space $(\Om,\cF,F=(\cF_s)_{0\leq s\leq t},P)$ satisfying the usual conditions, $\al=(\al_1,\dots,\al_m)$, $m\geq 1$, is an unknown parameter to be estimated, $\al\in \cA\sbs R^m$, $\cA$ is an open subset of $R^m$, $\ve$, $0<\ve\ll 1$, is a small parameter (index of series). In our further considerations all limits correspond to $\ve\to 0$.
Denote $(C_t,\cB_t)$ a measurable space of continuous on $[0,t]$ functions $x=(x_s)_{0\leq s\leq t}$ with $\sg$-algebra $\cB_t=\sg(x:x_s,s\leq t)$. Put $\cB_s=\sg(x:x_u,u\leq s)$.
Assume that for each $\al\in \cA$ the drift coefficients $a(s,x;\al)$, $0\leq s\leq t$, $x\in C_t$ is a known nonanticipative (i.e. $\cB_s$-measurable for each $s$, $0\leq s\leq t$) functional satisfying the functional Lipshitz and linear growth conditions $\bL$: $$\begin{gathered}
|a(s,x^1;\al)-a(s,x^2;\al)| \leq L_1 \int_0^s |x_u^1-x_u^2|\,d k_u +
L_2|x_s^1-x_s^2|, \\
|a(s,x;\al)|\leq L_1 \int_0^s (1+|x_u|)\,d k_u+L_2(1+|x_s|),\end{gathered}$$ where $L_1$ and $L_2$ are constants, which do not depend on $\al$, $k=(k(s))_{0\leq s\leq t}$ is a non-decreasing right-continuous function, $0\leq k(s)\leq k_0$, $0:k_0<\infty$, $x^1,x^2\in C_t$.
Then, as it is well-known (see, e.g., Liptser and Shiryaev [@24]), for each $\al\in \cA$ the equation (\[4-2.1\]) has an unique strong solution $Y^\ve(\al)=(Y_s^\ve(\al))_{0\leq s\leq t}$, and in addition (see Kutoyants [@18]) $$\sup_{0\leq s\leq t} |Y_s^\ve(\al)-Y_s^0(\al)| \leq
C\ve \sup_{0\leq s\leq t} |w_s| \;\;\Pas,$$ with some constant $C=C(L_1,L_2,k_0,t)$, where $Y^0(\al)=(Y_s^0(\al))_{0\leq s\leq t}$ is the solution of the following nonperturbated differential equation $$\label{4-2.2}
d Y_s=a(s,Y;\al)\,ds, \quad Y_0=0.$$
Change the initial problem of estimation of parameter $\al$ by the equivalent one, when the observations are modelled according to the following SDE $$\label{4-2.3}
d X_s=a_\ve(s,X;\al)\,ds +dw_s, \quad X_0=0,$$ where $a_\ve(s,x;\al)=\frac{1}{\ve}\,a(s,\ve x;\al)$, $0\leq s\leq t$, $x\in C_t$, $\al\in \cA$.
It is clear that if $X^\ve(\al)=(X_s^\ve(\al))_{0\leq s\leq t}$ is the solution of SDE (\[4-2.3\]), then for each $s\in [0,t]$ $\ve X_s^\ve(\al)=Y_s^\ve(\al)$.
Denote by $P_\al^\ve$ the distribution of process $X^\ve(\al)$ on the space $(C_t,\cB_t)$, i.e. $P_\al^\ve$ is the probability measure on $(C_t,\cB_t)$ induced by the process $X^\ve(\al)$. Let $P^w$ be a Wiener measure on $(C_t,\cB_t)$. Denote $X=(X_s)_{0\leq s\leq t}$ a coordinate process on $(C_t,\cB_t)$, that is $X_s(x)=x_s$, $x\in C_t$.
The conditions $\bL$ guarantee that for each $\al\in \cA$ the measures $P_\al^\ve$ and $P^w$ are equivalent $(P_\al^\ve\sim P^w)$, and if we denote $z_s^{\al,\ve}=\frac{dP_\al^\ve}{dP^w}|\cB_s$ the density process (likelihood ratio process), then $$z_s^{\al,\ve}(X)=\cE_s(a_\ve(\al) \cdot X) :=
\exp \Bigg\{ \int_0^s a_\ve(u,X;\al)\, dX_u-
\frac{1}{2} \int_0^s a_\ve^2(u,X;\al)\,du\Bigg\}.$$
Introduce class $\Psi$ of $R^m$-valued nonanticipative functionals $\psi$, $\psi:[0,t]\times C_t\times \cA \to R^m$ such that for each $\al\in \cA$ and $\ve>0$ $$\begin{aligned}
1) \quad & E_\al^\ve \int_0^t |\psi(s,X;\al)|^2ds<\infty, \label{4-2.4} \\
2) \quad & \int_0^t |\psi(s,Y^0(\al);\al)|^2ds<\infty, \label{4-2.5} \\
3) \quad & \text{uniformly in $\al$ on each compact $K\sbs \cA$} \notag \\
& P_\al^\ve -\lim_{\ve\to 0} \int_0^t |\psi(s,\ve X;\al)-
\psi(s,Y^0(\al);\al)|^2 ds=0, \label{4-2.6}\end{aligned}$$ where $|\cdot|$ is an Euclidean norm in $R^m$, $P_\al^\ve-\lim\limits_{\ve\to 0} \zt_\ve=\zt$ denotes the convergence $P_\al^\ve\{|\zt_\ve-\zt|>\rho\}\to 0$, as $\ve\to 0$, for all $\rho$, $\rho>0$.
Assume that for each $s\in [0,t]$ and $x\in C_t$ the functional $a(s,x;a)$ is differentiable in $\al$ and gradient $\dot{a}=\big(\frac{\pa}{\pa \al_1}\,a,\dots,\frac{\pa}{\pa\al_m}\,a\big)'$ belongs to $\Psi$ $(\dot{a}\in \Psi)$, where the sign “$'$” denoted a transposition.
Then the Fisher information process $$I_s^\ve(X;\al) :=\int_0^s \dot{a}_\ve(u,X;\al)
[\dot{a}_\ve(u,X;\al)]'du, \quad 0\leq s\leq t,$$ is well-defined and, moreover, uniformly in $\al$ on each compact $$\label{4-2.7}
P_\al^\ve-\lim_{\ve\to 0} \ve^2 I_t^\ve(\al)=I^0(\al),$$ where $$I^0(\al):= \int_0^t \dot{a}(s,Y^0(\al);\al)
[\dot{a}(s,Y^0(\al);\al]'ds.$$
For each $\psi\in \Psi$, introduce the functional $\psi_\ve(s,x;\al):=\frac{1}{\ve} \,\psi(s,\ve x;\al)$ and matrices $\Gm_\ve^\psi(\al)$ and $\gm_\ve^\psi\al$: $$\begin{aligned}
\Gm_\ve^\psi(X;\al) & := \int_0^t \psi_\ve(s,X;\al) [\psi_\ve(s,X;\al)]'ds,
\label{4-2.8} \\
\gm_\ve^\psi(X;\al) & := \int_0^t \psi_\ve(s,X;\al) [\dot{a}_\ve(s,X;\al)]'ds.
\label{4-2.9} \end{aligned}$$
Then from (\[4-2.6\]) it follows that uniformly in $\al$ on each compact $$\begin{aligned}
& P_\al^\ve -\lim_{\ve\to 0} \ve^2 \Gm_\ve^\psi(\al) =\Gm_0^\psi(\al),
\label{4-2.10} \\
& P_\al^\ve -\lim_{\ve\to 0} \ve^2 \gm_\ve^\psi(\al) =\gm_0^\psi(\al),
\label{4-2.11} \end{aligned}$$ where the matrices $\Gm_0^\psi(\al)$ and $\gm_0^\psi(\al)$ are defined as follows $$\begin{aligned}
\Gm_0^\psi(\al) & := \int_0^t \psi(s,Y^0(\al);\al) [\psi(s,Y^0(\al);\al)]'ds,
\label{4-2.12} \\
\gm_0^\psi(\al) & := \int_0^t \psi(s,Y^0(\al);\al)
[\dot{a}(s,Y^0(\al);\al)]'ds.
\label{4-2.13} \end{aligned}$$
Note that,by virtue of (\[4-2.4\]), (\[4-2.5\]) and $\dot{a}\in \Psi$, matrices given by (\[4-2.8\]), (\[4-2.9\]), (\[4-2.12\]) and (\[4-2.13\]) are well-defined.
Denote $\Psi_0$ the subset of $\Psi$ such that for each $\psi\!\in \!\Psi_0$ and $\al\in \cA$, $\rank \Gm_0^\psi(\al)\!=m$ and $\rank \gm_0^\psi(\al)=m$.
Assume that $\dot{a}\in \Psi_0$.
For each $\psi\in \Psi_0$, define a $P_\al^\ve$-square integrable martingale $L^{\psi,\ve}(\al)= \linebreak (L_s^{\psi,\ve}(\al))_{0\leq s\leq t}$ as follows $$\label{4-2.14}
L_s^{\psi,\ve}(X;\al) =\int_0^s \psi_\ve(u,X;\al) (d X_u-\al_\ve(u,X;\al)\,du).$$
Now we give a definition of CULAN $M$-estimates.
\[d4-2.1\] An estimate $(\al_t^{\psi,\ve})_{\ve>0} =(\al_{1,t}^{\psi,\ve}, \dots, \al_{m,t}^{\psi,\ve})_{\ve>0}'$, $\psi\in \Psi_0$, is called consistent uniformly lineal asymptotically normal (CULAN) if it admits the following expansion $$\label{4-2.15}
\al_t^{\psi,\ve}=\al+[\gm_0^\psi(\al)]^{-1} \ve^2 L_t^{\psi,\ve}(\al) +
r_{\psi,\ve}(\al),$$ where uniformly in $\al$ on each compact $$\label{4-2.16}
P_\al^\ve-\lim_{\ve\to 0} \ve^{-1} r_{\psi,\ve}(\al)=0.$$
It is well-known (see Lazrieva, Toronjadze [@20]) that under the above conditions uniformly in $\al$ on each compact $$\cL\{\ve^{-1}(\al_t^{\psi,\ve}-\al) \mid P_\al^\ve\} \str{w}{\to}
N(0,V(\psi;\al)),$$ with $$\label{4-2.17}
V(\psi;\al):=[\gm_0^\psi(\al)]^{-1} \Gm_0^\psi(\al)
([\gm_0^\psi(\al)]^{-1})',$$ where $\cL(\zt|P)$ denotes the distribution of random vector $\zt$ calculated under measure $P$, symbol “$\str{w}{\to}$” denotes the weak convergence of measures, $N(0,V(\psi;\al))$ is a distribution of Gaussian vector with zero mean and covariance matrix $V(\psi;\al)$.
\[r4-2.1\] In context of diffusion type processes the $M$-estimate $(\al_t^{\psi,\ve})_{\ve>0}$ is defined as a solution of the following stochastic equation $$L_t^{\psi,\ve}(X;\al)=0,$$ where $L_t^{\psi,\ve}(X;\al)$ is defined by (\[4-2.14\]), $\psi\in \Psi_0$.
The asymptotic theory of $M$-estimates for general statistical models with filtration is developed in Chitashvili et al. [@7]. Namely, the problem of existence and global behaviour of solutions is studied. In particular, the conditions of regularity and ergodicity type are established, under which $M$-estimates have a CULAN property.
For our model, in case when $\cA=R^m$, the sufficient conditions for CULAN property take the form:
\(1) for all $s$, $0\leq s\leq t$, and $x\in C_t$ the functionals $\psi(s,x;\al)$ and $\dot{a}(s,x;\al)$ are twice continuously differentiable in $\al$ with bounded derivatives satisfying the functional Lipshitz conditions with constants, which do not depend on $\al$.
\(2) the equation (w.r.t $y$) $$\Dl(\al,y):= \int_0^t \psi(s,Y^0(\al);y)
(a(s,Y^0(\al);\al)-a(s,Y^0(\al);y))\,ds=0,$$ has an unique solution $y=\al$.
The MLE is a special case of $M$-estimates when $\psi=\dot{a}$.
\[r4-2.2\] According to (\[4-2.7\]) the asymptotic covariance matrix of MLE $(\wh\al_t^\ve)_{\ve>0}$ is $[I_0(\al)]^{-1}$. By the usual technique one can show that for each $\al\in \cA$ and $\psi\in \Psi_0$, $I_0^{-1}(\al)\leq V(\psi,\al)$ (see (\[4-2.17\])), where for two symmetric matrices $B$ and $C$ the relation $B\leq C$ means that the mattix $C-B$ is nonnegative definite.
Thus, the MLE has a minimal covariance matrix among all $M$-estimates.
Shrinking contamination neighborhoods
-------------------------------------
In this subsection we give a notion of a contamination of the basic model (\[4-2.3\]), described in terms of shrinking neighborhoods of basic measures $\{P_\al^\ve$, $\al\in \cA$, $\ve>0\}$, which is an analog of the Huber gross error model (see, e.g., Hampel et.al. [@12] and also, Remark \[r4-2.3\] below).
Let $\cH$ be a family of bounded nonanticipative functionals $h:[0,t]\times C_t\times \cA\to R^1$ such that for all $s\in [0,t]$ and $\al\in \cA$ the functional $h(s,x;\al)$ is continuous at the point $x_0=Y^0(\al)$.
Let for each $h\in \cH$, $\al\in\cA$ and $\ve>0$, $P_\al^{\ve,h}$ be a measure on $(C_t,\cB_t)$ such that $$\begin{aligned}
1) \qquad & P_\al^{\ve,h}\sim P_\al^\ve, \notag \\
2) \qquad & \frac{dP_\al^{\ve,h}}{d P_\al^\ve} =\cE_t(\ve N_\al^{\ve,h}),
\label{4-2.18}\end{aligned}$$ where $$\label{4-2.19}
3) \qquad N_{\al,s}^{\ve,h}:= \int_0^s h_s(u,X;\al)
(dX_u-a_\ve(u,X;\al)\,du),$$ with $h_\ve(s,x;\al):=\frac{1}{\ve}\,h(s,\ve x;\al)$, $0\leq s\leq t$, $x\in C_t$.
Denote $\bP_\al^{\ve,\cH}$ a class of measures $P_\al^{\ve,h}$, $h\in \cH$, that is $$\bP_\al^{\ve,\cH} =\{P_\al^{\ve,h};\;\;h\in \cH\}.$$ We call $(\bP_\al^{\ve,\cH})_{\ve>0}$ a shrinking contamination neighborhoods of the basic measures $(P_\al^\ve)_{\ve>0}$, and the element $(P_\al^{\ve,h})_{\ve>0}$ of these neighborhoods is called alternative measure (or simply alternative).
Obviously for each $h\in \cH$ and $\al\in \cA$, the process $N_\al^{\ve,h}=(N_{\al,s}^{\ve,h})_{0\leq s\leq t}$ defined by (\[4-2.19\]) is a $P_\al^\ve$-square integrable martingale. Since under measure $P_\al^\ve$ the process $\ol{w}=(\ol{w}_s)_{0\leq s\leq t}$ defined as $$\ol{w}_s:= X_s-\int_0^s a_\ve(u,X;\al)\,du, \quad 0\leq s\leq t,$$ is a Wiener process, then by virtue of the Girsanov Theorem the process $\wt w:=\ol{w}+\la \ol{w},\ve N_\al^{\ve,h}\ra$ is a Wiener process under changed measure $P_\al^{\ve,h}$. But by the definition $$\wt w_s=X_s-\int_0^s (a_\ve(u,X;\al)+\ve h_\ve(u,X;\al))\,du,$$ and hence, one can conclude that $P_\al^{\ve,h}$ is a weak solution of SDE $$d X_s=(a_\ve(s,X;\al)+\ve h_\ve(s,X;\al))\,ds +d w_s, \quad X_0=0.$$
This SDE can be viewed as a “small” perturbation of the basic model (\[4-2.3\]).
\[r4-2.3\] 1) In the case of i.i.d. observations $X_1,X_2,\dots,X_n$, $n\geq 1$, the Huber gross error model in shrinking setting is defined as follows $$f^{n,h}(x;\al):=(1-\ve_n)f(x;\al)+\ve_n h(x;\al),$$ where $f(x;\al)$ is a basic (core) density of distribution of r.v. $X_i$ (w.r.t some dominating measure $\mu$), $h(x;\al)$ is a contaminating density, $f^{n,h}(x;\al)$ is a contaminated density, $\ve_n=O(n^{-1/2})$. If we denote by $P_\al^n$ and $P_\al^{n,h}$ the measures on $(R^n,\cB(R^n))$, generated by $f(x;\al)$ and $f^{n,h}(x;\al)$, respectively, then $$\frac{dP_\al^{n,h}}{d P_\al^n} =\prod_{i=1}^n
\frac{f^{n,h}(X_i;\al)}{f(X_i;\al)} =
\prod_{i=1}^n (1+\ve_n H(X_i;\al))=\cE_n(\ve_n\cdot N_\al^{n,h}),$$ where $H=\frac{h-f}{f}$, $N_\al^{n,h}=(N_{\al,m}^{n,h})_{1\leq m\leq n}$, $N_{\al,m}^{n,h}=\sum\limits_{i=1}^m H(X_i;\al)$, $N_\al^{n,h}$ is a $P_\al^n$-martingale, $\cE_n(\ve_n N_\al^{n,h})=\prod\limits_{i=1}^n (1+\ve_n\Dl N_{\al,i}^{n,h})$ is the Dolean exponential in discrete time case.
Thus $$\label{4-2.20}
\frac{dP_\al^{n,h}}{d P_\al^n}=\cE(\ve_n\cdot N_\al^{n,n}),$$ and the relation (\[4-2.18\]) is a direct analog of (\[4-2.20\]).
2\) The concept of shrinking contamination neighborhoods, expressed in the form of (\[4-2.18\]) was proposed in Lazrieva and Toronjadze [@21] for more general situation, concerning with the contamination areas for semimartingale statistical models with filtration.
Note here that the power of the small parameter $\ve$ is crucial. One cannot consider the perturbation of measure with different power of $\ve$ if he/she wish to get nontrivial result.
In the remainder of this subsection we study the asymptotic properties of CULAN estimates under alternatives.
For this aim we first consider the problem of contiguity of measures $(P_\al^{\ve,h})_{\ve>0}$ to $(P_\al^\ve)_{\ve>0}$.
Let $(\ve_n)_{n\geq 1}$, $\ve_n \downarrow 0$, and $(\al_n)_{n\geq 1}$, $\al_n\in K$, $K\sbs \cA$ is a compact, be arbitrary sequences.
\[p4-2.1\] For each $h\in \cH$ the sequence of measures $(P_{\al_n}^{\ve_n,h})$ is contiguous to sequence of measures $(P_{\al_n}^{\ve_n})$, i.e. $$(P_{\al_n}^{\ve_n,h}) \vartriangleleft (P_{\al_n}^{\ve_n}).$$
From the predictable criteria of contiguity (see, e.g., Jacod and Shiryaev [@15]), follows that we have to verify the relation $$\label{4-2.21}
\lim_{N\to \infty} \limsup_{n\to \infty}
P_{\al_n}^{\ve_n,h} \left\{ h_t^n \left(\frac{1}{2}\right)>N\right\}=0,$$ where $h^n(\frac{1}{2})=(h_s^n(\frac{1}{2}))_{0\leq s\leq t}$ is the Hellinger process of order $\frac{1}{2}$.
By the definition of Hellinger process (see, e.g., Jacod and Shiryaev [@15]) we have $$h_t^n\left(\frac{1}{2}\right) =h_t^n \left(\frac{1}{2}\,,
P_{\al_n}^{\ve_n,h}, P_{\al_n}^{\ve_n}\right) =
\frac{1}{8} \int_0^t \left[ h(s,\ve_n X;\al_n)\right]^2ds,$$ and since $h\in \cH$, and hence is bounded, $h_t^n(\frac{1}{2})$ is bounded too, which provides (\[4-2.21\]).
\[p4-2.2\] For each estimate $(\al_t^{\ve,\psi})_{\ve>0}$ with $\psi\in \Psi_0$ and each alternative $(P_\al^{\ve,h})_{\ve>0} \in (\bP_\al^{\ve,\cH})_{\ve>0}$ the following relation holds true $$\cL\left\{ \ve^{-1} (\al_t^{\psi,\ve}-\al) \mid P_\al^{\ve,h}\right\}
\str{w}{\to} N\left( [\gm_0^\psi(\al)]^{-1} b(\psi,h;\al), V(\psi,\al)\right),$$ where $$b(\psi,h;\al):=\int_0^t \psi(s,Y^0(\al);\al) h(s,Y^0(\al);\al)\,ds.$$
Proposition \[p4-2.1\] together with (\[4-2.16\]) provides that uniformly in $\al$ on each compact $$P_\al^{\ve,h} -\lim_{\ve\to 0} \ve^{-1} r_{\psi,\ve}(\al)=0,$$ and therefore we have to establish the limit distribution of random vector $[\gm_0^\psi(\al)]^{-1}\ve L_t^{\psi,\ve}$ under the measures $(P_\al^{\ve,h})_{\ve>0}$.
By virtue of the Girsanov Theorem the process $L^{\psi,\ve}(\al)=(L_s^{\psi,\ve}(\al))_{0\leq s\leq t}$ is a semimartingale with canonical decomposition $$\label{4-2.22}
L_s^{\psi,\ve}(\al) =\wt L_s^{\psi,\ve}(\al)+b_{\ve,s}(\psi,h;\al),
\quad 0\leq s\leq t,$$ where $\wt L^{\psi,\ve}(\al)=(\wt L_s^{\psi,\ve}(\al))_{0\leq s\leq t}$ is a $P_\al^{\ve,h}$-square integrable martingales defined as follows $$\wt L_s^{\psi,\ve}(X;\al):=\int_0^s \psi_\ve(u,X;\al)\,
\left(d X_u-(a_\ve(u,X;\al)+\ve h_\ve(u,X;\al)\right)\,du,$$ and $$b_{\ve,s}(\psi,h;\al):= \ve \int_0^s \psi_\ve (u,X;\al) h_\ve(u,X;\al)\,du.$$
But $\la \wt L^{\psi,\ve}(\al)\ra_t=\Gm_\ve^\psi(\al)$, where $\Gm_\ve^\psi(\al)$ is defined by (\[4-2.8\]). On the other hand, from Proposition \[p4-2.1\] and (\[4-2.10\]) it follows that $$P_\al^{\ve,h}-\lim_{\ve\to 0} \la \ve \wt L^{\psi,\ve}(\al)\ra_t =
P_\al^{\ve,h}-\lim_{\ve\to 0} \ve^2 \Gm_\ve^\psi(\al) =
P_\al^\ve -\lim_{\ve\to 0} \ve^2 \Gm_\ve^\psi(\al)=\Gm_0^\psi(\al)$$ uniformly in $\al$ on each compact, and hence $$\label{4-2.23}
\cL \left\{ [\gm_0^\psi(\al)]^{-1} \ve \wt L_t^{\psi,\ve}(\al) \mid
P_\al^{\ve,h}\right\} \str{w}{\to} N(0,V(\psi;\al)).$$
Finally, relation (\[4-2.23\]) together with (\[4-2.22\]) and relation $$P_\tht^{\ve,h} -\lim_{\ve\to 0} \ve b_{\ve,t}(\psi,h;\al) =
\int_0^t \psi(s,Y^0(\al);\al) h(s,Y^0(\al);\al)\,ds =
b(\psi,h;\al),$$ provides the desirable results.
Optimization criteria. Construction of optimal $B$-robust estimates
-------------------------------------------------------------------
In this subsection we state and solve an optimization problem, which results in construction of optimal $B$-robust estimate.
Initially, it should be stressed that the bias vector $\wt b(\psi,h;\al):=[\gm_0^\psi(\al)]^{-1} \times \linebreak b(\psi,h;\al)$ can be viewed as the influence functional of the estimate $(\al_t^{\psi,\ve})_{\ve>0}$ w.r.t. alternative $(P_\al^{\psi,h})_{\ve>0}$.
Indeed, the expansion (\[4-2.15\]) together with (\[4-2.22\]) and (\[4-2.23\]) allows to conclude that $$\cL\left\{ \ve^{-1} (\al_t^{\psi,\ve} -\al-\ve^2[\gm_0^\psi(\al)]^{-1}
b_\ve(\psi,h;\al)) \mid P_\al^{\ve,h} \right\} \str{w}{\to}
N(0,V(\psi,\al)),$$ and, hence, the expression $$\al+\ve^2[\gm_0^\psi(\al)]^{-1} b_\ve(\psi,h;\al)-\al=
\ve^2 [\gm_0^\psi(\al)]^{-1} b_\ve(\psi,h;\al),$$ plays the role of bias on the “fixed step $\ve$” and it seems natural to interpret the limit $$P_\al^{\ve,h} -\lim_{\ve\to 0} \frac{\al+\ve^2[\gm_0^\psi(\al)]^{-1}
b_\ve(\psi,h;\al)-\al}{\ve} =[\gm_0^\psi(\al)]^{-1} b(\psi,h;\al),$$ as the influence functional.
For each estimate $(\al_t^{\psi,\ve})_{\ve>0}$, $\psi\in \Psi_0$, define the risk functional w.r.t. alternative $(P_\al^{\ve,h})_{\ve>0}$, $h\in \cH$, as follows: $$D(\psi,h;\al) =\lim_{a\to \infty} \lim_{\ve\to 0}
E_\al^{\ve,h}\left( (\ve^{-2} |\al_t^{\psi\ve}-\al|^2) \wedge a\right),$$ where $x\wedge \al=\min(x,a)$, $a>0$, $E_\al^{\ve,h}$ is an expectation w.r.t. measure $P_\al^{\ve,h}$.
Using Proposition \[p4-2.2\] it is not hard to verify that $$D(\psi,h;\al)=|\wt b(\psi,h;\al)|^2+\tr V(\psi,\al),$$ where $\tr A$ denotes the trace of matrix $A$.
By Proposition \[p4-2.2\] $$\ve^{-1} (\al_t^{\psi,\ve}-\al)\stackrel{d}{\to}
N\big(\wt b(\psi,h;\al),V(\psi;\al)\big),$$ where $\stackrel{d}{\to}$ denotes the convergence by distribution (by distribution $P_\al^{\ve,h}$ in our case), $N(\wt b,V)$ is a Gaussian random vector with mean $\wt b$ and covariation matrix $V$.
But if $\xi=(\xi_1,\dots,\xi_m)'$ is a Gaussian vector with parameters $(\mu,\sg^2)$, then $$E|\xi|^2=\sum_{i=1}^m E\xi_i^2=\sum_{i=1}^m (E\xi_i)^2+
\sum_{i=1}^m D\xi_i=|\mu|^2+\tr \sg^2,$$ as it was required.
Connect with each $\psi\in \Psi_0$ the function $\wt\psi$ as follows $$\wt\psi(s,x;\al)=[\gm_0^\psi(\al)]^{-1} \psi(s,x;\al), \quad
0\leq s\leq t, \quad x\in C_t, \quad \al\in \cA.$$ Then $\wt\psi\in \Psi_0$ and $$\gm_0^{\wt\psi}(\al)=Id,$$ where $Id$ is an unit matrix, $$V(\psi;\al)=V(\wt\psi;\al)=\Gm_0^{\wt\psi}(\al), \quad
\wt b(\psi,h;\al)=\wt b(\wt\psi,h;\al)=b(\wt\psi,h;\al).$$ Therefore $$\label{4-2.24}
D(\psi,h;\al) =D(\wt\psi,h;\al) =|b(\wt\psi,h;\al)|^2 +
\tr \Gm_0^{\wt\psi}(\al).$$
Denote $\cH_r$, a set of functions $h\in \cH$ such that for each $\al\in \cA$ $$\int_0^t |h(s,Y^0(\al);\al)|\,ds\leq r,$$ where $r$, $r>0$, is a constant.
Since, for each $r>0$, $$\sup_{h\in \cH_r} |b(\wt\psi,h;\al)|\leq const(r)
\sup_{0\leq s\leq t} |\wt\psi(s,Y^0(\al);\al)|,$$ where constant depends on $r$, we call the function $\wt \psi$ an influence function of estimate $(\al_t^{\psi,\ve})_{\ve>0}$ and a quantity $$\gm_\psi^*(\al) =\sup_{0\leq s\leq t} |\wt\psi(s,Y^0(\al);\al)|$$ is named as the (unstandardized) gross error sensitivity at point $\al$ of estimate $(\al_t^{\psi,\ve})_{\ve>0}$.
Define $$\begin{gathered}
\Psi_{0,c} =\bigg\{\psi\in \Psi_0\;:\;
\int_0^t \psi(s,Y^0(\al);\al) [\dot{a} (s,Y^0(\al);\al)]'ds=
Id, \label{4-2.25} \\
\gm_\psi^*(\al)\leq c\bigg\}, \label{4-2.26}\end{gathered}$$ where $c\in [0,\infty)$ is a generic constant.
Take into account the expression (\[4-2.24\]) for the risk functional we come to the following optimization problem, known in robust estimation theory as Hampel’s optimization problem: minimize the trace of the asymptotic covariance matrix of estimate $(\al_t^{\psi,\ve})_{\ve>0}$ over the class $\Psi_{0,c}$, that is $$\label{4-2.27}
\text{minimize} \;\;\; \int_0^t \psi(s,Y^0(\al);\al)
[\psi(s,Y^0(\al);\al)]'ds$$ under the side conditions (\[4-2.25\]) and (\[4-2.26\]).
Define the Huber function $h_c(z)$, $z\in R^m$, $c>0$, as follows $$h_c(z):= z\min\left(1,\frac{c}{|z|}\right).$$
For arbitrary nondegenerate matrix $A$ denote $\psi_c^A=h_c(A\dot{a})$.
\[t4-2.1\] Assume that for given constant $c$ there exists a nondegenerate $m\times m$-matrix $A_c^*(\al)$, which solves the equation $($w.r.t. matrix $A)$ $$\label{4-2.28}
\int_0^t \psi_c^A(s,Y^0(\al);\al) [\dot{a}(s,Y^0(\al);\al)]'ds=Id.$$
Then the function $\psi_c^{A_c^*(\al)}=h_c(A_c^*(\al)\dot{a})$ solves the optimization problem $(\ref{4-2.27})$.
We follow Hampel et al. [@12].
Let $A$ be an arbitrary $m\times m$-matrix.
Since for each $\psi\in \Psi_{0,c}$, $\int \psi(\dot{a})'=Id$, $\int \dot{a}[\dot{a}]' =I^0(\al)$ (see (\[4-2.7\])), then $$\int (\psi-A\dot{a} )(\psi-A\dot{a})' =
\int \psi\psi'-A-A'+AI^0(\al)A'$$ (here and below we use simple evident notation for integrals).
Therefore since the trace is an additive functional instead of minimizing of $\tr \int \psi \psi'$ we can minimize $$\tr \int(\psi-A\dot{a})(\psi-A\dot{a})'=\int |\psi-A\dot{a}|^2.$$
Note that for each $z$ $$\arg \min_{|y|\leq c} |z-y|^2=h_c(z).$$ Indeed, it is evident that minimizing $y$ has the form $y=\bt z$, where $\bt$, $0\leq
\bt\leq 1$, is constant. Then $$\min_{|y|\leq c} |z-y|^2=\min_{\bt\leq \frac{c}{|z|}}
(1-\bt)^2|z|^2.$$
Thus we have to find $$\arg \min_{\bt\leq \frac{c}{|z|}} (1-\bt)^2=
\min\Big(1,\frac{c}{|z|}\Big).$$
But last relation is trivially satisfied. Hence the minimizing $y^*\!=\!z\min(1,\frac{c}{|z|})$ and $$\arg \min_{|\psi|\leq c} |\psi-A\dot{a}|^2=h_c(A\dot{a}).$$
From the other side, $$|h_c(z)|^2=|z|^2I_{\{|z|\leq c\}}+
\frac{|z|^2}{|z|^2}\,c^2\,I_{(|z|\geq c)}\leq c^2.$$ Hence $$|h_c(z)|\leq c \quad \text{for all} \;\;\;z$$ and therefore $h_c(A\dot{a})$ satisfies the condition (\[4-2.26\]) for each $A$.
Now it is evident that a function $h_c(A\dot{a})$ minimizes the expression under integral sign, and hence the integral itself over all functions $\psi\in \Psi_0$ satisfying (\[4-2.26\]).
At the same time the condition (\[4-2.25\]), generally speaking, can be violated. But, since a matrix $A$ is arbitrary, we can choose $A=A_c^*(\al)$ from (\[4-2.28\]) which, of course, guarantees the validity of (\[4-2.25\]) for $\psi_c^*=\psi_c^{A_c^*(\al)}$.
As we have seen the resulting optimal influence functions $\psi_c^*$ is defined along the process $Y^0(\al)=(Y_s^0(\al))_{0\leq s\leq t}$, which is a solution of equation (\[4-2.2\]).
But for constructing optimal estimate we need a function $\psi_c^*(s,x;\al)$ defined on whole space $[0,t]\times C_t\times \cA$.
For this purpose define $\psi_c^*(s,x;\al)$ as follows; $$\label{4-2.29}
\psi_c^*(s,x;\al) =\psi_c^{A_c^*(\al)} (s,x;\al) =
h_\ve(A_c^*(\al) \dot{a}(s,x;\al)),$$ and as usual $\psi_{c,\ve}^*(s,x;\al)=\frac{1}{\ve} \psi_c^*(s,\ve x;\al)$, $0\leq s\leq t$, $x\in C_t$, $\al\in \cA$.
\[d4-2.2\] We say that $\psi_c^*(s,x;\al)$, $0\leq s\leq t$, $x\in C_t$, $\al\in \cA$, is an influence function of optimal $B$-robust estimate $(\al_t^{*,\ve})_{\ve>0}=(\al_t^{\psi_c^{*},\ve})_{\ve>0}$ over the class of CULAN estimates $(\al_t^{\psi,\ve})_{\ve>0}$, $\psi\in \Psi_{0,c}$, if the matrix $A^*(\al)$ is differentiable in $\al$.
From (\[4-2.9\]), (\[4-2.11\]), (\[4-2.28\]) and (\[4-2.29\]) it directly follows that $$\gm_0^{\psi_c^*}(\al) =P_\al^\ve -\lim_{\ve\\to 0} \ve^2 \gm_\ve^{\psi_c^*}(\al) =
\int_0^t \psi_c^*(s,Y^0(\al);\al) (\dot{a}(s,Y^0(\al);\al))'ds=Id.$$ Besides, for each alternative $(P_\al^{\ve,h})_{\ve>0}$, $h\in \cH$, according to the Proposition \[p4-2.2\] we have $$\cL \left\{\ve^{-1} (\al_t^{*,\ve}-\al) \mid P_\al^{\ve,h}\right\} \str{w}{\to}
N(b(\psi_c^*,h;\al), V(\psi_c^*;\al)) \;\;\;\text{as} \;\;\; \ve\to 0,$$ where $$b(\psi_c^*,h;\al)=\int_0^t \psi_c^*(s,Y^0(\al);\al) h(s,Y^0(\al);\al)\,ds,$$ and $V(\psi_c^*;\al)=\Gm_0^{\psi_c^*}(\al)$.
Hence, the risk functional for estimate $(\al_t^{*,\ve})_{\ve>0}$ is $$D(\psi_c^*,h;\al)=|b(\psi_c^*,h;\al)|^2 +\tr \Gm_0^{\psi_c^*},
\quad h\in \cH,$$ and the (unstandardized) gross error sensitivity of $(\al_t^{*,\ve})_{\ve>0}$ is $$\gm_{\psi_c^*}(\al) =\sup_{0\leq s\leq t} |\psi_c^*(s,Y^0(\al);\al)|\leq c.$$
From above reasons, we may conclude that $(\al_t^{*,\ve})_{\ve>0}$ is the optimal $B$-robust estimate over the class of estimates $(\al_t^{\psi,\ve})_{\ve>0}$, $\psi\in \Psi_{0,c}$ in the following sense: the trace of asymptotic covariance matrix of $(\al_t^{*,\ve})_{\ve>0}$ is minimal among all estimates $(\al_t^{\psi,\ve})_{\ve>0}$ with bounded by constant gross error sensitivity, that is $$\Gm_0^{\psi_c^*}(\al) \leq \Gm_0^\psi(\al) \;\;\;\text{for all} \;\;\;
\psi\in \Psi_{0,c}\,. \;\;\qed$$
Note that for each estimate $(\al_t^{\psi,\ve})_{\ve>0}$ and alternatives $(P_\al^{\ve,h})_{\ve>0}$, $h\in \cH$, the influence functional is bounded by $const(r) \cdot c$. Indeed, we have for $\psi\in \Psi_{0,c}$, $$\sup_{h\in \cH_r} |b(\psi,h;\al)|\leq const(r) \cdot c:= C(r,c),$$ and since from (\[4-2.24\]) $$\inf_{\psi\in \Psi_{0,c}} \sup_{h\in \cH_r} D(\psi,h;\al) \leq
C^2(r,c)+\tr \Gm_0^{\psi_c^*}(\al),$$ we can choose “optimal level” of truncation, minimizing the expression $$C^2(r,c)+\tr \Gm_0^{\psi_c^*}(\al)$$ over all constants $c$, for which the equation (\[4-2.28\]) has a solution $A_c^*(\al)$. This can be done using the numerical methods.
For the problem of existence and uniqueness of solution of equation (\[4-2.28\]) we address to Rieder [@35].
In the case of one-dimensional parameter $\al$ (i.e. $m=1$) the optimal level $c^*$ of truncation is given as an unique solution of the following equation (see Lazrieva and Toronjadze [@21], [@22]) $$r^2c^2 =\int_0^t [\dot{a} (s,Y^0(\al);\al)]_{-c}^c \dot{a}(s,Y^0(\al);\al)\,ds-
\int_0^t ([\dot{a}(s,Y^0(\al);\al)]_{-c}^c)^2 \,ds,$$ where $[x]_a^b=(x\wedge b)\vee a$ and the resulting function $$\psi^*(s,x;\al) =[\dot{a}(s,x;\al)]_{-c^*}^{c^*} , \quad
0\leq s\leq t, \quad x\in C_t,$$ is $(\Psi_0,\cH_r)$ optimal in the following minimax sense: $$\sup_{h\in \cH_r} D(\psi^*,h;\al) =\inf_{\psi\in \Psi}
\sup_{h\in \cH_r} D(\psi,h;\al).$$
Optimal Mean-Variance Robust Hedging
====================================
A financial market model
------------------------
Let $(\Om,\cF,F=(\cF_t)_{0\leq t\leq T},P)$ be a filtered probability space with filtration $F$ satisfying the usual conditions, where $T\in (0,\infty]$ is a fixed time horizon. Assume that $\cF_0$ is a trivial and $\cF_T=\cF$.
There exist $d+1$, $d\geq 1$ primitive assets: one bound, whose price process is assumed to be 1 at all times and $d$ risky assets (stocks), whose $R^d$-valued price process $X=(X_t)_{0\leq t\leq T}$ is a continuous semimartingale given by the relation: $$\label{4-3.1}
dX_t =\diag (X_t) \,dR_t, \quad X_0>0,$$ where $\diag(X)$ denotes the diagonal $d\times d$-matrix with diagonal elements $X^1,\dots, X^d$, and the yield process $R=(R_t)_{0\leq t\leq T}$ is a $R^d$-valued continuous semimartingale satisfying the stricture condition (SC). That is (see Schweizer [@38]) $$\label{4-3.2}
d R_t=d\la \wt M\ra_t\lb_t+d\wt M_t, \quad R_0=0,$$ where $\wt M=(\wt M_t)_{0\leq t\leq T}$ is a $R^d$-valued continuous martingale, $\wt M\in \cM_{0,\loc}^2(P)$, $\lb=(\lb_t)_{0\leq t\leq T}$ is a $F$-predictable $R^d$-valued process, and the mean-variance tradeoff (MVT) process $\wt \cK=(\wt\cK_t)_{0\leq t\leq T}$ of process $R$ $$\label{4-3.3}
\wt\cK_t:=\int_0^t \lb_s' d\la \wt M\ra_s\lb_s =\la \lb'\cdot \wt M\ra_t<\infty
\;\;\Pas, \;\; t\in [0,T].$$
\[r4-3.1\] Remember that all vectors are assumed to be column vectors.
Suppose that the martingale $\wt M$ has the form $$\label{4-3.4}
\wt M=\sg\cdot M,$$ where $M=(M_t)_{0\leq t\leq T}$ is a $R^d$-valued continuous martingale, $M\in \cM_{0,\loc}^2(P)$ with $d\la M^i,M^j\ra_t=I_{i_j}^{d\times d} d C_t$, $I^{d\times d}$ is the identity matrix, $C=(C_t)_{0\leq t\leq T}$ is a continuous increasing bounded process with $C_0=0$.
Further, let $\sg=(\sg_t)_{0\leq t\leq T}$ is a $d\times d$-matrix valued, $F$-predictable process with $\rank(\sg_t)=d$ for any $t$, $P$-a.s., the process $(\sg_t^{-1})_{0\leq t\leq T}$ is locally bounded, and $$\label{4-3.5}
\int_0^T \sg_t\,d\la M\ra_t \sg_t'<\infty \;\; \Pas$$
Assume now that the following condition be satisfied:
There exist fixed $R^d$-valued, $F$-predictable process $k=(k_t)_{0\leq t\leq T}$ such that $$\label{4-3.6}
\lb=\lb(\sg)=(\sg')^{-1}k.$$
In the case from (\[4-3.2\]) we get $$\begin{gathered}
dR_t=d\la \wt M\ra_t \lb_t +d\wt M_t =\sg_td\la M\ra_t \sg_t'
(\sg_t')^{-1} k_t+\sg_t d M_t \notag \\
=\sg_t(d\la M\ra_tk_t+dM_t), \label{4-3.7}\end{gathered}$$ and $$\begin{gathered}
\wt\cK_t =\int_0^t \lb_s' d\la \wt M\ra_s\lb_s =
\int_0^t k_t'((\sg_t')^{-1})' \sg_t d\la \wt M\ra_t
\sg_t'(\sg_t')^{-1} k_t \\
=\int_0^t k_t' d\la M\ra_t k_t=\la k\cdot M\ra_t:=\cK_t.\end{gathered}$$
From (\[4-3.3\]) we have $$\label{4-3.8}
\cK_t<\infty \;\; \Pas \;\;\text{for all} \;\; t\in [0,T].$$
Thus, of we introduce the process $M^0=(M_t^0)_{0\leq t\leq T}$ by the relarion $$\label{4-3.9}
d M_t^0=d\la M\ra_t k_t +d M_t, \quad M_0^0=0,$$ then the MVT process $\cK=(\cK_t)_{0\leq t\leq T}$ of $R^d$-valued semimartingale $M^0$ is finite, and hence $M^0$ satisfies SC.
Finally, the scheme (\[4-3.1\]), (\[4-3.2\]), (\[4-3.4\]), (\[4-3.6\]) and (\[4-3.9\]) can be rewritten in the following form $$\label{4-3.10}
\begin{aligned}
dX_t & =\diag (X_t) \,dR_t, \quad X_0>0, \\
dR_t & =\sg_t\,d M_t^0, \quad R_0=0, \\
d M_t^0 & =d\la M\ra_t k_t+d M_t, \quad M_0=0,
\end{aligned}$$ where $\sg$ and $k$ satisfy (\[4-3.5\]) and (\[4-3.8\]), respectively.
This is our financial market model.
Characterization of variance-optimal ELMM (equivalent local martingale measure)
-------------------------------------------------------------------------------
A key role in mean-variance hedging plays variance-optimal ELMM (see, e.g., RSch [@34], GLP [@11]). Here we collect some facts characterizing this measure.
We start with remark that the sets ELMMs for processes $X$, $R$ and $M^0$ form (\[4-3.10\]) coincide. Hence we can and will consider the simplest process $M^0$.
Introduce the notation $$\cM_2^e:= \left\{ Q\sim P\,:\, \frac{dQ}{dP}\in L^2(P), \;\; M^0 \;\;
\text{is a $Q$-local martingale}\right\},$$ and suppose that $$(c.1) \hskip+3cm \cM_2^e \neq \vnth. \hskip+5cm$$ The solution $\wt P$ of the optimization problem $$E\cE_T^2(\cM^Q)\to \inf_{Q\in \cM_2^e}$$ is called variance-optimal ELMM.
Here $$\frac{dQ}{dP}\Big|_{\cF_T} =\cE_T(M^Q),$$ and $(\cE_t(M^Q))_{0\leq t\leq T}$ is the Dolean exponential of martingale $M^Q$.
It is well-known (see, e.g., Schweizer [@38], [@39]) that under condition (c.1) variance-optimal ELMM $\wt P$ exist.
Denote $$\wt z_T:=\frac{d\wt P}{dP} \Big|_{\cF_T} \,,$$ and introduce RCLL process $\wt z=(\wt z_t)_{0\leq t\leq T}$ by the relation $$\wt z_t=E^{\wt P} (\wt z_T/\cF_T), \quad 0\leq t\leq T.$$
Then, by Schweizer [@38], [@39] $$\label{4-3.11}
\wt z_T=\wt z_0+\int_0^T \zt_t'\,d M_t^0,$$ where $\zt=(\zt_t)_{0\leq t\leq T}$ is the $R^d$-valued $F$-predictable process with $$\int_0^T \zt_t'\,d\la M\ra_t \zt_t<\infty,$$ and the process $\big(\int\limits_0^t \zt_s' dM_s^0\big)_{0\leq t\leq T}$ is a $\wt P$-martingale.
Relation (\[4-3.11\]) easily implies that the process $\wt z$ is actually continuous.
Suppose, in addition to (c.1), that the following condition is satisfied:
(c.$^*$) all $P$-local martingales are continuous.
This technical assumption is satisfied in stochastic volatility models, where $F=F^w$ is the natural filtration generated by the Wiener process.
It shown in Mania and Tevzadze [@35], Mania et al. [@26] that under conditions (c.1) and (c$^*$) density $\wt z_T$ of variance optimal ELMM is uniquely characterized by the relation $$\label{4-3.12}
\wt z_T =\frac{\cE_T((\vf-k)'\cdot M^0)}{E\cE_T((\vf-k)'\cdot M^0)}\,,$$ where $\vf$ together with the pair $(L,c)$ is the unique solution of the following equation $$\label{4-3.13}
\frac{\cE_T((\vf-2k)'\cdot M)}{\cE_T(L)} =c\cE_T^2(-k'\cdot M),$$ where $L\in M_{0,\loc}^2(P)$, $\la L,M\ra=0$, $c$ is a constant.
Moreover, the process $\zt=(\zt_t)_{0\leq t\leq T}$ from (\[4-3.11\]) has the form $$\label{4-3.14}
\zt_t=(\vf_t-k_t)\cE_t((\vf-k)'\cdot M^0).$$
Here $\vf=(\vf_t)_{0\leq t\leq T}$ is a $R^d$-valued, $F$-predictable process with $$\int_0^T \vf_t'\,d\la M\ra_t\vf_t<\infty.$$
Let $\tau$ be $F$-stopping time.
Denote $\la k'\cdot M\ra_{T\tau}=\la k'\cdot M\ra_T-\la k'\cdot M\ra_\tau$.
\[p4-3.1\]
$1.$ Equation $(\ref{4-3.13})$ is equivalent to equation $$\label{4-3.15}
\frac{\cE_T(\vf'\cdot M^*)}{\cE_T(L)} =ce^{\la k'\cdot M\ra_T},$$ where the $R^d$-valued process $M^*=(M_t^*)_{0\leq t\leq T}$ is given by the relation $$d M_t^*=2d\la M\ra_t k_t +dM_t, \quad M_0^*=0.$$
$2.$ [a)]{} If there exists the martingale $m=(m_t)_{0\leq t\leq T}$, $m\in \cM_{0,\loc}^2(P)$ such that $$\label{4-3.16}
e^{-\la k'\cdot M\ra_T} =c+m_T, \quad \la m,M\ra=0,$$ then $\vf\equiv 0$ and $L_T=\int\limits_0^T \frac{1}{c+m}\,dm_s$ solve the equation $(\ref{4-3.15})$.
In this case $$\label{4-3.17}
\wt z_T=\frac{\cE_T(-k'\cdot M^0)}{E\cE_T(-k'\cdot M^0)}\,,$$ process $\zt=(\zt_t)_{0\leq t\leq T}$ from $(\ref{4-3.11})$ is equal to $$\zt_t=-k_t\cE_t(-k'\cdot M^0),$$ and $$E\left[ \left( \frac{\wt z_T}{\wt z_\tau}\right)^2 \Big/ \cF_\tau\right] =
\frac{1}{E(e^{-\la k'\cdot M\ra_{T\tau}} /\cF_\tau)}\,.$$
[b)]{} If there exist $R^d$-valued $F$-predictable process $\ell=(\ell_t)_{0\leq t\leq T}$, $\int\limits_0^T \ell_t'd \la M\ra \ell_t <\infty$ and $$e^{\la k'\cdot M\ra_T}=c+\int_0^T \ell_t'\,d M_t^*,$$ then $L\equiv 0$ and $\vf_t=\frac{\ell_t}{c+\int_0^t \ell_s'dM_s^*}$ solve the equation $(\ref{4-3.15})$.
In this case $$\wt z_T=\cE_T(-k'\cdot M) \;\; (:=\wh z_T, \;\;
\text{the density of minimal martingale measure} \;\; \wh P),$$ and $$E\left(\left( \frac{\wt z_T}{\wt z_\tau}\right)^2 \Big/ \cF_\tau\right) =
E^{P^*} (e^{\la k'\cdot M\ra_{T\tau}} \big/ \cF_\tau),$$ where $dP^*=\cE_T(-2k'\cdot M) dP$.
1\. By the Yor formula $$\begin{gathered}
\cE_T\left(\vf-2k)'\cdot M\right)=
\cE_T(\psi'\cdot M-2k'\cdot M) \\
=\cE_T\Bigg( \vf'\cdot \bigg( M+2\int_0^\cdot d\la M\ra_t k_t \bigg) -
2\int_0^\cdot \psi_t' d\la M\ra_t k_t-2k'\cdot M\Bigg) \\
=\cE_T(\vf'\cdot M^*)\cE_T(-2k'\cdot M),\end{gathered}$$ and $$\cE_T^2(-k'\cdot M)=\cE_T(-2k'\cdot M) e^{\la k'\cdot M\ra_T}.$$ Assertion follows.
2\. a) Note at first that $\la L,M\ra=0$. Further, by the formula we can write $$\ln(c+m_t)-\ln c=\int_0^t \frac{1}{c+m_s}\,d m_s -
\frac{1}{2} \int_0^t \frac{1}{(c+m_s)^2}\, d\la m\ra_s.$$
Hence $$e^{\ln(c+m_T)-\ln c} =\cE_T(L),$$ and thus $$\cE_T(L) =\frac{c+m_T}{c} =\frac{e^{-\la k'\cdot M\ra_T}}{c}\,.$$
Finally, by the Bayes rule and the Girasnov Theorem $$\begin{gathered}
E\left( \left( \frac{\wt z_T}{\wt z_\tau}\right)^2 \Big/\cF_\tau\right)=
\frac{E(\cE_T(-2k'\cdot M)e^{-\la k'\cdot M\ra_T}/\cF_\tau)}
{E^2(\cE_T(-k'\cdot M)e^{-\la k'\cdot M\ra_T}/\cF_\tau)}\\
=\frac{E^*(c+m_T/\cF_\tau)\cE_T^2(-k'\cdot M)}
{(\wh E(c+m_\tau/\cF_\tau))^2\cE_T^2(-2k'\cdot M)} =
\frac{c+m_\tau}{(c+m_\tau)^2}\cdot e^{\la k'\cdot M\ra_\tau} \\
=\frac{1}{E(e^\la k'\cdot M\ra_{T\tau}/\cF_\tau)}\,.\end{gathered}$$
The proof of case 2 b) is quite analogous.
Proposition is proved.
Misspecified asset price model and robust hedging
-------------------------------------------------
Denote by $\Ball_L(0,r)$, $r\in [0,\infty)$ the closed $r$-radius ball in the space $L=L_\infty(dt\times dP)$, with the center at the origin, and let $$\begin{gathered}
\cH:= \big\{ h=\{h_{ij}\}, \; i,j=\wh{1,d}:h \;\text{is $F$-predictable} \;\;
d\times d\text{-matrix} \notag \\
\text{valued process,} \;\; \rank(h)=d, \;\; h_{ij}\in \Ball_L(0,r), \;\;
r\in [0,\infty)\big\}.
\label{4-3.18}\end{gathered}$$
Class $\cH$ is called the class of alternatives.
Fix the value of small parameter $\dl>0$, as well as $d\times d$-matrix valued, $F$-predictable process $\sg^0=(\sg_t^0)_{0\leq t\leq T}= (\{\sg_{ij,t}^0\}, \,1\leq i,j\leq d)_t$ such that $|\sg_{ij,t}^0|\leq const$, $\fa i,j,t$, the matrix $(\sg^0)^2=\sg^0(\sg^0)'$ is uniformly elliptic, i.e. for each vector $v_t=(v_t',\dots,v_t^d)$ with probability 1 $$\label{1.4.4}
\sum_{i,j=1}^d (\sg^0)_{ij,t}^2 v_t^i v_t^j \geq c\sum_{i=1}^d |v_t^i|^2,
\quad c>0, \quad 0\leq t\leq T,$$ and denote $$\label{4-3.20}
A_\dl=\{\sg: \sg=\sg^0+\dl h, \;\; h\in \cH\}.$$
\[1.4.1\] Every $\sg$ from the class $A_\dl$ for sufficiently small $\dl$ is $F$-predictable $d\times d$-valued process with bounded elements and the matrix $\sg^2=\sg\sg'$ is uniformly elliptic.
The process $\sg$ is $F$-predictable as linear combination of $F$-predictable processes. Further, $$|\sg_{ij,t}|=|\sg_{ij,t}^0+\dl h_{ij,t}|\leq const+\dl r, \quad
0<\dl\ll 1.$$
From (\[1.4.4\]) and (\[4-3.20\]) for each vector $\nu_t=(\nu_t^1,\dots,\nu_t^d)$ we have $$\begin{aligned}
\sum_{i,j=1}^d(\sg^2)_{ij,t}\nu_t^i\nu_t^j & =
\sum_{i,j=1}^d(\sg^0+\dl h)(\sg^0+\dl h)_{ij,t}'\nu_t^i\nu_t^j\notag\\
& =\sum_{i,j=1}^d (\sg^0(\sg^0)')_{ij,t}\nu_t^i\nu_t^j+
\dl\sum_{i,j=1}^d (\sg^0h')_{ij,t}\nu_t^i\nu_t^j \notag \\
& \quad +\dl\sum_{i,j=1}^d (h(\sg^0)')_{ij,t}\nu_t^i\nu_t^j+
\dl^2\sum_{i,j=1}^d(hh')_{ij,t}\nu_t^i\nu_t^j. \label{1.4.6}\end{aligned}$$
Note now that the elements of matrices $\sg^0$ and $h$ are bounded. Hence choosing $\dl$ sufficiently small we get $$\max\left(\dl|(\sg^0h')_{ij,t}|, \dl|(h(\sg^0))_{ij,t}|,
\dl^2|(hh')_{ij,t}|\right)\leq \frac{\ve}{3}\,.$$
Therefore from (\[1.4.4\]) and (\[1.4.6\]) we get $$\begin{gathered}
\sum_{i,j=1}^d \sg_{ij,t}^2 \nu_t^i\nu_t^j\geq
(c-const \cdot \ve)\sum_{i,j=1}^d|\nu_t^i|^2 \;\;\;\text{for each $\ve>0$}. \end{gathered}$$
Proposition is proved.
Consider the set of processes $\{R^\sg (\text{or}\; X^\sg)$, $\sg\in A_\dl\}$, which represents the misspecified of asset price model.
Define the class of admissible trading strategies $\Tht=\Tht(\sg^0)$.
\[1.4.2\] For each $R^d$-valued $F$-predictable process $\tht=(\tht_t)_{0\leq t\leq T}$ and for each $\sg\in A_\dl$, $\dl>0$, $$\begin{aligned}
aE\int_0^T|\tht_t|^2dC_t & \leq E\int_0^T \tht_t'\sg_td\la M\ra_t\sg_t'\tht_t \\
& =E\int_0^T \tht_t'\sg_t\sg_t'\tht_tdC_t\leq
AE\int_0^T|\tht_t|^2dC_t,\end{aligned}$$ where the constants $a$, $A$ are such that $0<a\leq A<\infty$, and the parameter $\dl>0$ is sufficiently small.
Remember that $d\la M\ra_t=d\la M^i,M^j\ra_t=I_{ij}^{d\times d} dC_t$. Hence $$E\int_0^T \tht_t'\sg_td\la M\ra_t\sg_t'\tht_t=
E\int_0^T \tht_t'\sg_t\sg_t'\tht_t\,dC_t.$$
Further, since $\sg=\sg^0+\dl h$ and elements of matrices $\sg^0$ and $h$ are bounded, then the same is true for the elements of matrix $\sg$ with $0\leq \dl\leq const$. Thus using the inequality $ab\leq 2(a^2+b^2)$ we get $$E\int_0^T \tht_t'\sg_t\sg_t'\tht_t\,dC_t\leq AE\int_0^T|\tht_t|^2dC_t.$$ On the other hand, by Proposition \[1.4.1\] the matrix $\sg^2=\sg\sg'$ is uniformly elliptic for sufficiently small $\dl$, which yields the first inequality.
\[d4-3.1\] The class $\Tht=\Tht(\sg^0)$ is a class of $R^d$-valued $F$-predictable processes $\tht=(\tht_t)_{0\leq t\leq T}$ such that $$\label{4-3.21}
E\int_0^T |\tht_t|^2 d C_t<\infty.$$
Let $\tht\in \Tht$ be the dollar amount (rather than the number of shares) invested in the stock $X^\sg$, $\sg\in A_\dl$. Then for each $\sg\in A_\dl$ the trading gains induced by the self-financing portfolio strategy associated to $\tht$ has the form $$\label{4-3.22}
G_t(\sg,\tht)=\int_0^t \tht_s'\,d R_s^\sg, \quad 0\leq t\leq T,$$ where $R^d=(R_t^d)_{0\leq t\leq T}$ is the yield process given by (\[4-3.10\]).
Introduce the condition:
(c.2) There exists ELMM $\ol{Q}$ such that the density process $z=z^{\ol{Q}}$ satisfies the reverse Hölder inequality $R_2(P)$, see definition in RSch [@34].
It is well-known that under the conditions (c.1) and (c.2) the density process $\wt z=(\wt z_t)_{\leq t\leq T}$ of the variance-optimal ELMM satisfies $R_2(P)$ as well, see Dolean et al. [@8].
Now under the conditions (c.1) and (c.2) the r.v. $G_T(\sg,\tht)\in L^2(P)$, $\fa \sg\in A_\dl$, and the space $G_T(\sg,\Tht)$ is closed in $L^2(P)$, $\fa \sg\in A_\dl$ (see, e.g., Theorem 2 of RSch [@34]).
A contingent claim is an $\cF_T$-measurable square-integrable r.v. $H$, which models the payoff from a financial product at the maturity date $T$.
The problem we are interested in is to find the robust hedging strategy for a contingent claim $H$ in the above described incomplete financial market model with misspecified asset price process $X^\sg$, $\sg\in A_\dl$, using mean-variance approach.
For each $\sg\in A_\dl$, the total loss of a hedger, who starts with the initial capital $x$, uses the strategy $\tht$, believes that the stock price process follows $X^\sg$, and has to pay a random amount $H$ at the date $T$, is $H$-$x$-$G_T(\sg,\tht)$.
Denote $$\label{4-3.23}
\cJ(\sg,\tht):=E(H-x-G_T(\sg,\tht))^2.$$
One setting of the robust mean-variance hedging problem consist in solving the optimization problem $$\label{4-3.24}
\text{minimize} \;\; \sup_{\sg\in A_\dl} \cJ(\sg,\tht) \;\;
\text{over all strategies} \;\; \tht\in \Tht.$$
We “slightly” change this problem using the approach developed in Toronjadze [@42] which based on the following approximation $$\begin{aligned}
\sup_{\sg\in A_\dl} \cJ(\sg,\tht) & =\exp \big\{ \sup_{h\in \cH}
\ln\cJ(\sg^0+\dl h,\tht)\big\} \\
& \simeq \exp \bigg\{ \sup_{h\in \cH} \bigg[ \ln \cJ(\sg^0,\tht) +
\dl\,\frac{D\cJ(\sg^0,h,\tht)}{\cJ(\sg^0,\tht)}\bigg]\bigg\} \\
& =\cJ(\sg^0,\tht) \exp \bigg\{ \dl\sup_{h\in \cH}
\frac{D\cJ(\sg^0,h,\tht)}{\cJ(\sg^0,\tht)}\bigg\},\end{aligned}$$ where $$D\cJ(\sg^0,h,\tht):= \frac{d}{d\dl}\,\cJ(\sg^0+\dl h,\tht) |_{\dl=0} =
\lim_{\dl\to 0} \frac{\cJ(\sg^0+\dl h,\tht)-\cJ(\sg^0,\tht)}{\dl}\,,$$ is the Gateaux differential of the functional $\cJ$ at the point $\sg^0$ in the direction $h$.
Approximate (in leading order $\dl$) the optimization problem (\[4-3.24\]) by the problem $$\begin{gathered}
\text{minimize} \;\; \cJ(\sg^0,\tht) \exp \bigg\{ \dl\sup_{h\in \cH}
\frac{D\cJ(\sg^0,h,\tht)}{\cJ(\sg^0,\tht)}\bigg\} \notag \\
\text{over all strategies} \;\; \tht\in \Tht. \label{4-3.25}\end{gathered}$$
Note that each solution $\tht^*$ of the problem (\[4-3.25\]) minimizes $\cJ(\sg^0,\tht)$ under the constraint $$\sup_{h\in \cH} \frac{D\cJ(\sg^0,h,\tht)}{\cJ(\sg^0,\tht)} \leq c:=
\sup_{h\in \cH} \frac{D\cJ(\sg^0,h,\tht^*)}{\cJ(\sg^0,\tht^*)}\,.$$
This characterization of an optimal strategy $\tht^*$ of the problem (\[4-3.25\]) leads to the
\[d4-3.2\] The trading strategy $\tht^*\in \Tht$ is called optimal mean-variance robust trading strategy against the class of alternatives $\cH$ if it is a solution of the optimization problem $$\begin{gathered}
\text{minimize} \;\; \cJ(\sg^0,\tht)\;\;\text{over all strategies} \;\;
\tht\in \Tht, \;\; \text{subject to constraint} \notag \\
\sup_{h\in \cH} \frac{D\cJ(\sg^0,h,\tht)}{\cJ(\sg^0,\tht)}\leq c, \label{4-3.26}\end{gathered}$$ where $c$ is some generic constant.
\[r4-3.3\] In contrast to “mean-variance robust” trading strategy which associates with optimization problem (\[4-3.24\]) and control theory, we find the “optimal mean-variance robust” strategy in the sense of Definition \[d4-3.2\]. Such approach and term are common in robust statistics theory (see, e.g., Hampel et al. [@12], Rieder [@35]).
Does the suggested approach provide “good” approximation? Consider the case.
[**Diffusion model with zero drift.**]{} Let a standard Wiener process $w=(w_t)_{0\leq t\leq T}$ be given on the complete probability space $(\Om,\cF,P)$. Denote by $F^w=(\cF_t^w$, $0\leq t\leq T)$ the $P$-augmentation of the natural filtration $\cF_t^w=\sg(w_s$, $0\leq s\leq t)$, $0\leq t\leq T$, generated by $w$.
Let the stock price process be modeled by the equation $$ dX_t^\sg=X_t^\sg\cdot \sg_t\,dw_t, \;\; X_0^\sg>0, \;\; 0\leq t\leq T,$$ where $\sg\in A_\dl$ with $$\int\limits_0^T (\sg_t^0)^2\,dt<\infty$$ and $h\in \Ball_{L_\infty(dt\times dP)}(0,r)$, $0<r<\infty$. All considered processes are real-valued.
Denote by $R^\sg$ the yield process, i.e., $$ dR_t^\sg=\sg_t\,dw_t, \;\; R_0^\sg=0, \;\; 0\leq t\leq T.$$
The wealth at maturity $T$, with the initial endowment $x$, is equal to $$V_T^{x,\tht}(\sg)=x+\int_0^T \tht_t\,dR_t^\sg.$$ Let, further, the contingent claim $H$ be $\cF_T^w$-measurable $P$-square-integrable r.v.
Consider the optimization problem (\[4-3.24\]). It is easy to see that if $\sg\in A_\dl$; then $$\sg_t^0-\dl r\leq \sg_t\leq \sg_t^0+\dl r, \;\; 0\leq t\leq T, \;\;
\Pas,$$
By the martingale representation theorem $$ H=EH+\int_0^T \vf_t^H\,dw_t, \;\; \Pas,$$ where $\vf^H$ is the $F^w$-predictable process with $$\label{ch12}
E\int_0^T (\vf_t^H)^2\,dt<\infty.$$ Hence $$E\big(H-V_T^{x,\tht}(\sg)\big)^2=
(EH-x)^2+E\int_0^T (\vf_t^H-\sg_t\tht_t)^2\,dt.$$ From this it directly follows that the process $$\begin{aligned}
\sg_t^{*}(\tht) & =(\sg_t^0-\dl r)
I_{\{\frac{\vf_t^H}{\tht_t}\geq \sg_t^0\}}
I_{\{\tht_t\neq 0\}} \notag \\
& +(\sg_t^0+\dl r)I_{\{\frac{\vf_t^H}{\tht_t}<\sg_t^0\}}
I_{\{\tht_t\neq 0\}}, \;\; 0\leq t\leq T,\label{ch13}\end{aligned}$$ is a solution of the optimization problem $$\text{maximize} \;\; E\big(H-V_T^{x,\tht}(\sg)\big)^2 \;\;
\text{over all} \;\; \sg\in A_\dl, \;\;
\text{with a given} \;\; \tht\in \Tht.$$ It remains to minimize (w.r.t. $\tht$) the expression $$E\int_0^T \big(\vf_t^H-\sg_t^{*}(\tht)\tht_t\big)^2\,dt.$$ From (\[ch13\]) it easily follows that the equation (w.r.t. $\tht$) $$\vf_t^H-\sg_t^{*}(\tht)\tht_t=0,$$ has no solution, but $$\label{ch14}
\tht_t^{*}=\frac{\vf_t^H}{\sg_t^0}\,I_{\{\sg_t^0\neq 0\}}, \;\;
0\leq t\leq T,$$ solves problem. We assume that $0/0:=0$.
Consider now the optimization problem (\[4-3.26\]).
For each fixed $h$ $$\begin{aligned}
J(\sg,\tht) & =E\Big(H-x-\int_0^T \tht_t\,dR_t^\sg\Big)^2 \\
& =E\bigg(H-x-\int_0^T \tht_t\sg_t^0\,dw_t-
\dl\int_0^T \tht_th_t\,dw_t\bigg)^2 \\
& =J(\sg^0,\tht)-2\dl E\bigg[\Big(EH-x+
\int_0^T \big(\vf_t^H-\tht_t\sg_t^0\big)\,dw_t\Big)
\int_0^T \tht_th_t\,dw_t\bigg] \notag \\
& \quad +\dl^2E\int_0^T \tht_t^2h_t^2\,dt,\end{aligned}$$ and hence $$\label{ch15}
DJ(\sg^0,h;\tht)=
2E\int_0^T \big(\tht_t\sg_t^0-\vf_t^H\big)\tht_th_t\,dt,$$ as follows from (\[ch12\]), the definition of the class $\cH$ and the estimation $$\begin{gathered}
\Big(E\int_0^T \big(\tht_t\sg_t^0-\vf_t^H\big)\tht_th_t\,dt\Big)^2\leq
E\int_0^T \big(\tht_t\sg_t^0-\tht_t^H\big)^2\,dt\;
E\int_0^T \tht_t^2h_t^2\,dt \notag \\
\leq {\operatorname{const}}\cdot r^2\bigg(E\int_0^T \tht_t^2(\sg_t^0)^2\,dt+
E\int_0^T (\vf_t^H)^2\,dt\bigg)E\int_0^T \tht_t^2\,dt<\infty.\label{ch16}\end{gathered}$$ Since, further, $DJ(\sg^0,h;\tht)=0$ for $h\equiv 0$, using (\[ch16\]) we get $$0\leq \sup_{h\in \cH} DJ(\sg^0,h;\tht)<\infty.$$ Hence we can take $0\leq c<\infty$ in problem (6). Now if we substitute $\tht^{*}$ from (\[ch14\]) into (\[ch15\]), we get $DJ(\sg^0,h;\tht^{*})=0$ for each $h$, and thus $$\frac{\sup\limits_{h\in \cH}
DJ(\sg^0,h;\tht^{*})}{J(\sg^0,\tht^{*})}=0.$$ If we recall that $\tht^{*}=\arg\min\limits_{\tht\in \Tht_{A_\dl}} J(\sg^0,\tht)$, we get that $\tht^{*}$ defined by (\[ch14\]) is a solution of this optimization problem as well.
Thus we prove that
*[(a)]{} the mean-variance robust trading strategy $\tht^{*}=(\tht_t^{*})_{0\leq t\leq T}$ for the optimization problem $(\ref{4-3.24})$ is given by the formula $$\tht_t^{*}=\frac{\vf_t^H}{\sg_t^0}\,I_{\{\sg_t^0\neq 0\}};$$*
[(b)]{} at the same time this strategy is an optimal mean-variance robust trading strategy for the optimization problem $(\ref{4-3.26})$.
Hence in this case the suggested approach leads to the perfect solution of initial problem $(\ref{4-3.24})$.
To solve the problem (\[4-3.26\]) in general case we need to calculate $D\cJ(\sg^0,h,\tht)$. Suppose that $k=(k_t)_{0\leq t\leq T}=(k_{i,t},\, 1\leq i\leq d)_{0\leq t\leq T}$ from (\[4-3.10\]) is such that $|k_{i,t}|\leq const$ $\fa i,t$.
Following RSch [@34] and GLP [@11] introduce the probability measure $\wt Q\sim P$ on $\cF_T$ by the relation $$\label{4-3.27}
d\wt Q=\frac{\wt z_T}{\wt z_0} \,d\wt P \;\;\;\Big(\text{and hence} \;\;
d\wt Q=\frac{\wt z_T^2}{\wt z_0} \, dP\Big).$$ Using Proposition 5.1 of GLP [@11] we can write $$\begin{gathered}
\cJ(\sg,\tht)= E\,\frac{\wt z_T^2}{\wt z_0^2}\,
\frac{\wt z_0^2}{\wt z_T^2}\, \Bigg( H-x-\int_0^T \tht_t'\,d R_t^\sg\Bigg)^2
\notag \\
= \wt z_0^{-1} E^{\wt Q} \, \frac{\wt z_0^2}{\wt z_T^2}\,
\Bigg( H-x-\int_0^T \tht_t'\sg_t \,d M_t^0\Bigg)^2 \notag \\
=\wt z_0^{-1} E^{\wt Q} \Bigg( \frac{H\wt z_0}{\wt z_T} -x-
\int_0^T \psi_t^0(\sg)\, d\,\frac{\wt z_0^2}{\wt z_t^2} -
\int_0^T \left( \psi_t^1(\sg)\right)' \,
d\,\frac{M_t^0}{\wt z_t}\,\wt z_0\Bigg)^2 \notag \\
:= \ol{\cJ}(\sg,\psi^0,\psi^1) \;\;\;(\text{or} \;\; \ol{\cJ}(\sg,\psi) \;\;
\text{with} \;\; \psi=(\psi^0,\psi^1)'\,), \label{4-3.28}\end{gathered}$$ where $$\label{4-3.29}
\begin{gathered}
\psi_t^1=\psi_t^1(\sg)=\sg_t'\tht_t, \\
\psi_t^0=\psi_t^0(\sg)=\int_0^t \tht_s'\sg_s d M_s^0-
\tht_t'\sg_t M_t^0, \;\;\; 0\leq t\leq T.
\end{gathered}$$ Thus $$\psi_t^1(\sg) =\psi_t^1(\sg^0)+\dl\psi_t^1(h), \quad
\psi_t^0(\sg) =\psi_t^0(\sg^0)+\dl\psi_t^0(h).$$
Let (following RSch [@34]) $$\label{4-3.30}
\frac{H}{\wt z_T}\,\wt z_0 =E\left( \frac{H}{\wt z_T}\, \wt z_0\right) +
\int_0^T (\psi_t^H)'dU_t+L_T,$$ be the Galtchouk–Kunita–Watanabe decomposition of r.v. $\frac{H}{\wt z_T}\,\wt z_0$ w.r.t. $R^{(d+1)}$-valued $\wt Q$-local martingale $U=\big( \frac{\wt z_0}{\wt z}, \frac{M^0}{\wt z}\,\wt z_0\big)'$, where $\psi^H=(\psi^{0,H},\psi^{1,H})'\in L^2(U,\wt Q)$, the space of $F$-predictable processes $\psi$ such that $\int \psi' dU\in \cM^2(\wt Q)$ of martingale, and $L\in \cM_{0,\loc}^2(\wt Q)$, $L$ is $\wt Q$-strongly orthogonal to $U$.
Remember that $$\label{4-3.31}
\psi=(\psi^0,\psi^1)'.$$ Then, using (\[4-3.28\]), (\[4-3.29\]) and (\[4-3.30\]) we can write for each $h$ $$\begin{gathered}
\cJ(\sg^0+\dl h,\psi) =\cJ(\sg^0,\psi)+\dl\cdot 2\wt z_0^{-1} \notag \\
\times E^{\wt Q} \Bigg\{ \bigg[
\bigg( x-E^{\wt Q} \,\frac{H}{\wt z_T}\,\wt z_0\bigg) -L_T +
\int_0^T (\psi_t(\sg^0)-\psi_t^H)'dU_t\bigg]
\int_0^T (\ol{\psi}_t(h))' dU_t\Bigg\} \notag \\
+\dl^2 \wt z_0^{-1} E^{\wt Q} \Bigg[ \int_0^T (\ol{\psi}_t(h))' dU_t\Bigg]^2\notag\\
=\cJ(\sg^0,\psi)+\dl\cdot 2\wt z_0^{-1} E^{\wt Q}
\Bigg[ \int_0^T (\psi_t(\sg^0)-\psi_t^H)' dU_t
\int_0^T (\ol{\psi}_t(h))' dU_t\Bigg] \notag \\
+\dl^2 \wt z_0^{-1} E^{\wt Q} \Bigg[ \int_0^T (\ol{\psi}_t(h))' dU_t\Bigg]^2.
\label{4-3.32}\end{gathered}$$ Using Proposition 8 of RSch [@34] we have for each $h$ $$\frac{\wt z_0}{\wt z_T}\,G_r(h,\Tht)=\Bigg\{ \int_0^T ({\psi}(h))' dU_t:
{\psi}(h) \in L^2(U,\wt Q)\Bigg\},$$ and hence by (\[4-3.22\]) $$\begin{gathered}
E^{\wt Q} \Bigg(\int_0^T ({\psi}_t(h))' dU_t\Bigg)^2 \notag \\
=E^{\wt Q}\,\frac{\wt z_0^2}{\wt z_T^2}\, G_T^2(h,\tht)= \wt z_0 EG_T^2(h,\tht) =
\wt z_0 E\Bigg( \int_0^T \tht_t'\, dR_t^h\Bigg)^2 \notag \\
=\wt z_0 E\Bigg( \int_0^T \tht_t' h_t d M_t^0\Bigg)^2 =
\wt z_0 E \Bigg( \int_0^T \tht_t' h_t d \la M\ra_tk_t +
\int_0^T \tht_t' h_t d M_t\Bigg)^2 \notag \\
\leq const \Bigg[ E \Bigg( \int_0^T |\tht_t' h_t d \la M\ra_t k_t| \Bigg)^2 +
E\Bigg( \int_0^T \tht_t' h_t d M_t\Bigg)^2 \Bigg] \notag \\
\leq const \;r^2 E\int_0^T |\tht_t|^2 d C_t<\infty. \label{4-3.33}\end{gathered}$$
Further, $$\begin{gathered}
\Bigg( E^{\wt Q} \Bigg[ \int_0^T (\psi_t(\sg^0)-\psi_t^H)' dU_t
\int_0^T (\psi_t(h))' dU_t \Bigg]\Bigg)^2 \notag \\
\leq E^{\wt Q} \Bigg( \int_0^T (\psi_t(\sg^0)-\psi_t^H)' dU_t \Bigg)^2
E^{\wt Q} \Bigg( \int_0^T (\psi_t(h))' dU_t \Bigg)^2<\infty.\label{4-3.34}\end{gathered}$$ From these estimates we conclude that: $$\label{4-3.35}
1) \;\;\; D\ol{\cJ} (\sg^0,h,\psi)=
2\wt z_0^{-1} E^{\wt Q} \int_0^T (\psi_t(\sg^0)-\psi_t^H)'
d\la U\ra_t \psi_t(h)<\infty,\;\;$$ thanks to (\[4-3.33\]).
2\) $D\ol{\cJ}(\sg^0,h,\psi)|_{h\equiv 0} =0$, since $\psi(0)=0$ by (\[4-3.31\]) and (\[4-3.29\]).
Thus $$\label{4-3.36}
\sup_{h\in \cH} D \ol{\cJ}(\sg^0,h,\psi)\geq 0.$$
3\) From (\[4-3.34\]) and (\[4-3.33\]) we get $$\begin{gathered}
(D\ol{\cJ}(\sg^0,h,\psi))^2 \leq const \;\wt z_0^{-2} r^2 \\
\times E^{\wt Q} \int_0^T (\psi_t(\sg^0)-\psi_t^H)' d\la U\ra_t
(\psi_t(\sg^0)-\psi_t^H)
E\int_0^T |\tht_t|^2 d C_t<\infty.\end{gathered}$$ Thus $|D\ol{\cJ}(\sg^0,h,\psi)|$ is estimated by the expression which does not depend on $h$, and is equal to zero if we substitute $\psi_t(\sg^0)\equiv \psi_t^H$, $0\leq t\leq T$.
Hence, by (\[4-3.36\]) $$\label{4-3.37}
0\leq \sup_{h\in \cH} D\ol{\cJ}(\sg^0,h,\psi)|_{\psi\equiv \psi^H} \leq
\sup_{h\in \cH} |D\ol{\cJ}(\sg^0,h,\psi)|\big|_{\psi\equiv \psi^H}=0$$
Further, from (\[4-3.36\]) follows that we can take $c\in [0,\infty)$ in (\[4-3.26\]).
Now substituting $\psi\equiv \psi^H$ into $\ol{\cJ}(\sg^0,\psi)$ and $D\ol{\cJ}(\sg^0,h,\psi)$ we get $$\ol{\cJ}(\sg^0,\psi^H) =\min_\psi \ol{\cJ}(\sg^0,\psi) =\wt z_0^{-1} (E^{\wt P} H-x)^2+
\wt z_0^{-1} E^{\wt Q} L_T^2$$ (see Lemma 5.1 of GLP [@11]) and $$\sup_{h\in \cH} \frac{D \ol{\cJ}(\sg^0,h,\psi^H)}{\ol{\cJ}(\sg^0,\psi^H)}=0.$$
Hence the constraint of problem (\[4-3.26\]) is satisfied.
\[r4-3.4\] If $x=E^{\wt P}H$ and $L_T\equiv 0$, then we get $$\frac{D \ol{\cJ}(\sg^0,h,\psi^H)}{\ol{\cJ}(\sg^0,\psi^H)}=\frac{0}{0}$$ which is assumed to be zero, since if we consider the shifted risk functional $\wt\cJ=\ol{\cJ}+1$, the optimization problem and the optimal trading strategy will not change, but $D\wt\cJ(\sg^0,h,\psi^H)=D\ol{\cJ}(\sg^0,h,\psi^H)=0$ and $\wt\cJ(\sg^0,\psi^H)=1$.
Finally, using Proposition 8 of RSch [@34] we arrive at the following
\[t4-3.1\] In Model $(\ref{4-3.10})$ under conditions [(c.1)]{} and [(c.2)]{} the optimal mean-variance robust trading strategy $($in the sense of Definition $\ref{d4-3.1})$ is given by the formula $$\label{4-3.38}
\tht_t^* =((\sg_t^0)')^{-1} [\psi_t^{1,H} +\zt_t(V_t^*-(\psi_t^H)'U_t)],
\quad 0\leq t\leq T,$$ where $$\begin{gathered}
\psi_t^H=(\psi_t^{0,H},\psi_t^{1,H})', \quad
U_t=\left( \frac{\wt z_0}{\wt z_t}\,,\frac{M_t^0}{\wt z_t}\,\wt z_0 \right)', \\
V_t^*=\frac{\wt z_0}{\wt z_t} \Bigg( x+\int_0^t (\psi_t^H)' dU_t\Bigg),\end{gathered}$$ $\psi_t^H$ and $\zt_t$ are given by the relations $(\ref{4-3.30})$ and $(\ref{4-3.11})$, respectively, $\wt z_t$ is defined in $(\ref{4-3.11})$.
[99]{} M. Avellaneda, A. Levy and A. Paras, Pricing and hedging derivative securities in markets with uncertain volatilities. *Appl. Math. Finance* **2** (1995), 73–88. M. Avellaneda and A. Paras, Managing the volatility risk of portfolios of derivative securities: The Lagrangian uncertain volatility model. *Appl. Math. Finance* **3** (1996), 21–52.
F. Biagini, P. Guasoni and M. Pratelli, Mean-variance hedging for stochastic volatility models. *INFORMS Applied Probability Conference* (*Ulm,* 1999). *Math. Finance* **10** (2000), No. 2, 109–123.
I. Bajeux-Besnainou and J. Rochet, Dynamic spanning: Are options an appropriate instrument? *Math. Finance* **6** (1996), No. 1, 1–16.
M. Chesney, R. J. Elliott, D. Madan and H. Yang, Diffusion coefficient estimation and asset pricing when risk premia and sensitivities are time varying. *Math. Finance* **3** (1993), No. 2, 85–99.
R. Chitashvili, N. Lazrieva and T. Toronjadze, Asymptotic theory of $M$-estimators in general statistical models. Part I. On asymptotic behaviour of estimators under model perturbance. *Centrum voor Woskunde en Informatica Amsterdam,* Report BS-R9019, 1990.
R. Chitashvili, N. Lazrieva and T. Toronjadze, Asymptotic theory of $M$-estimators in general statistical models. Part II. On asymptotic behaviour of estimators in the presence of nuisance parameter. *Centrum voor Woskunde en Informatica Amsterdam,* Report BS-R9020, 1990.
F. Delbaen, P. Monat, W. Schachermayer, M. Schweizer and C. Stricker, Weighted norm inequalities and hedging in incomplete markets. *Finance Stoch.* **1** (1997), No. 3, 181–227.
D. Duffie and H. R. Richardson, Mean-variance hedging in continuous time. *Ann. Appl. Probab.* **1** (1991), No. 1, 1–15.
H. Föllmer and D. Sondermann, Hedging of nonredundant contingent claims. *Contributions to mathematical economics,* 205–223, *North-Holland, Amsterdam,* 1986.
C. Gourieroux, J. P. Laurent and H. Pham, Mean-variance hedging and numéraire. *Math. Finance* **8** (1998), No. 3, 179–200.
F. R. Hampel, E. M. Ronchetti, P. J. Rousseeuw and W. A. Stahel, Robust statistics. The approach based on influence functions. *Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical Statistics. John Wiley* [&]{} *Sons, Inc., New York,* 1986.
J. Hull and A. White, The pricing of options on assets with stochastic voltilities. *J. Finance* **3** (1987), 281–300. I. A. Ibragimov, R. Z. Has’minskiĭ, Statistical estimation. Asymptotic theory. (Translated from the Russian) *Applications of Mathematics,* 16. *Springer-Verlag, New York-Berlin,* 1981.
J. Jacod and A. N. Shiryaev, Limit theorems for stochastic processes. *Grundlehren der Mathematischen Wissenschaften* \[*Fundamental Principles of Mathematical Sciences*\], 288. *Springer-Verlag, Berlin,* 1987.
N. V. Krylov, Controlled diffusion processes. (Translated from the Russian) *Applications of Mathematics,* 14. *Springer-Verlag, New York-Berlin,* 1980.
Yu. Kutoyants, Identification of dynamical systems with small noise. *Mathematics and its Applications,* 300. *Kluwer Academic Publishers Group, Dordrecht,* 1994.
J. P. Laurent and H. Pham, Dynamic programming and mean-variance hedging. *Finance Stoch.* **3** (1999), No. 1, 83–110.
N. L. Lazrieva and T. A. Torondzhadze, Asymptotic properties of an $M$-estimate in a general statistical experiment scheme. (Russian) *Statistics and control of random processes* (Russian) (*Preila,* 1987), 105–112, *Nauka, Moscow,* 1989.
N. Lazrieva and T. Toronjadze, Robust estimators in statistical models with filtration. Shrinking neighbourhoods. *Seminarberichte, Fachbereich Mathematik, Fernuniversität, Hagen, Germany* **48** (1994), 50-68.
N. Lazrieva and T. Toronjadze, Robust estimators in statistical models associated with semimartingales. *Proc. A. Razmadze Math. Inst.* **118** (1998), 73–100.
D. Li and W.-L. Ng, Optimal dynamic portfolio selection: multiperiod mean-variance formulation. *Math. Finance* **10** (2000), No. 3, 387–406.
R. S. Liptser and A. N. Shiryayev, Statistics of random processes. I. General theory. (Translated from the Russian) *Applications of Mathematics,* Vol. 5. *Springer-Verlag, New York-Heidelberg,* 1977.
M. Mania and R. Tevzadze, A semimartingale Bellman equation and the variance-optimal martingale measure. *Georgian Math. J.* **7** (2000), No. 4, 765–792.
M. Mania, M. Santacroce and R. Tevzadze, A semimartingale backward equation related to the $p$-optimal martingale measure and the lower price of a contingent claim. *Stochastic processes and related topics* (*Siegmundsburg,* 2000), 189–212, *Stochastics Monogr.,* 12, *Taylor* [&]{} *Francis, London,* 2002.
H. M. Markowitz, Mean-variance analysis in portfolio choice and capital markets. *Basil Blackwell, Oxford,* 1987.
R. C. Merton, An analytic derivation of the efficient portfolio frontier. *J. Financ. Qualitat. Anal.* **7** (1972), 1851–1872.
P. Monat and C. Stricker, Föllmer-Schweizer decomposition and mean-variance hedging for general claims. *Ann. Probab.* **23** (1995), No. 2, 605–628.
M. Musiela and M. Rutkowski, Martingale methods in financial modelling. *Applications of Mathematics (New York),* 36. *Springer-Verlag, Berlin,* 1997.
S. Pastorello, Diffusion coefficient estimation and asset pricing when risk premia and sensitivities are time varying: A comment. *Math. Finance* **6** (1996), No. 1, 111–117.
H. Pham, T. Rheinländer and M. Schweizer, Mean-variance hedging for continuous processes: new proofs and examples. *Finance Stoch.* **2** (1998), No. 2, 173–198.
E. Renault and N. Touzi, Option hedging and implied volatilities in a stochastic volatility model. *Math. Finance* **6** (1996), No. 3, 279–302.
T. Rheinländer and M. Schweizer, On $L^2$-projections on a space of stochastic integrals. *Ann. Probab.* **25** (1997), No. 4, 1810–1831.
H. Rieder, Robust asymptotic statistics. *Springer Series in Statistics. Springer-Verlag, New York,* 1994.
W. J. Runggaldier and A. Zaccaria, A stochastic control approach to risk management under restricted information. *INFORMS Applied Probability Conference* (*Ulm,* 1999). *Math. Finance* **10** (2000), No. 2, 277–288.
M. Schweizer, Mean-variance hedging for general claims. *Ann. Appl. Probab.* **2** (1992), No. 1, 171–179.
M. Schweizer, Approximating random variables by stochastic integrals. *Ann. Probab.* **22** (1994), No. 3, 1536–1575.
M. Schweizer, Approximation pricing and the variance-optimal martingale measure. *Ann. Probab.* **24** (1996), No. 1, 206–236.
L. Scott, Option pricing when the variance changes randomly: theory, estimation and application. *J. Financ. Qualitat. Anal.* **22** (1987), 419–438
K. R. Sircar and G. G. Papanicolaou, Stochastic volatility, smile and asymptotics. *Appl. Math. Finance* **6** (1999), No. 2, 107–145.
T. Toronjadze, Optimal mean-variance robust hedging under asset price model misspecification. *Georgian Math. J.* **8** (2001), No. 1, 189–199.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Ultra-high-energy cosmic rays (UHECRs) can be accelerated by tidal disruption events of stars by black holes. We suggest a novel mechanism for UHECR acceleration wherein white dwarfs (WDs) are tidally compressed by intermediate-mass black holes (IMBHs), leading to their ignition and subsequent explosion as a supernova. Cosmic rays accelerated by the supernova may receive an energy boost when crossing the accretion-powered jet. The rate of encounters between WDs and IMBHs can be relatively high, as the number of IMBHs may be substantially augmented once account is taken of their likely presence in dwarf galaxies. Here we show that this kind of tidal disruption event naturally provides an intermediate composition for the observed UHECRs, and suggest that dwarf galaxies and globular clusters are suitable sites for particle acceleration to ultra-high energies.'
author:
- Rafael $^1$
- 'Joseph Silk$^{1,2,3,4}$'
bibliography:
- 'references.bib'
title: 'Ultra-High-Energy Cosmic Rays from Tidally-Ignited White Dwarfs'
---
Introduction {#sec:intro}
============
An open question in astrophysics is related to the origin and acceleration mechanisms of the ultra-high-energy cosmic rays (UHECRs). Their extreme energy ($E \gtrsim 10^{18}\;\text{eV}$) poses a problem for simple models of particle acceleration. Measurements by the Pierre Auger Observatory favor a mixed composition at energies $E \gtrsim 10^{18.6} \; \text{eV}$, which becomes increasingly heavier with energy [@auger2014a; @auger2014b]. The recent combined spectrum-composition fit by the Auger Collaboration also suggests a mixed composition at the sources [@auger2017a], although the results are model-dependent.
At energies below $\sim \, 0.1 \; \text{EeV}$ ($1 \; \text{EeV} \equiv 10^{18} \; \text{eV}$), the cosmic-ray spectrum can be described by the usual supernova paradigm: supernovae accelerate CR protons up to energies $\sim 10^{15} \; \text{eV}$; this leads to the so-called *knee* of the cosmic-ray spectrum. Heavier nuclei of atomic number $Z$ can be accelerated up $E_{\text{max}} \sim Z \times 10^{15} \; \text{eV}$; for iron ($Z=26$) this would imply a feature in the spectrum coinciding with the *second knee*. CRs of energies above the so-called *ankle* ($E \sim 4 \times 10^{18} \; \text{eV}$) are believed to be mostly of extragalactic origin. Betwen the second knee and the ankle there is a clear need for a local (possibly galactic) component [@hillas2005a; @thoudam2016a], whose acceleration mechanism remains to be found. Extensions of the usual acceleration mechanism by supernovae have been suggested to explain this component (see e.g. Refs. [@voelk1988a; @bell2001a]).
A difficulty common to many UHECR acceleration models concerns the efficiency at which a source accelerates heavy nuclei. Gamma-ray bursts (GRBs), for example, are inefficient, as opposed to unipolar inductors operating in pulsars and magnetars. Outflows of active galactic nuclei are also seemingly suitable accelerator sites, although the abundance of nuclei such as iron is low.
Tidal disruption events (TDEs) by supermassive black holes (SMBHs) have been suggested as the underlying mechanism for particle acceleration to ultra-high energies [@farrar2009a], and invoked to explain anisotropy results by the Pierre Auger Observatory and Telescope Array (TA) [@pfeffer2017a].
In this paper we focus on a different kind of TDE in which a white dwarf (WD) is disrupted and ignited by an intermediate-mass black hole (IMBH).
IMBHs are hypothetical objects whose existence have not been unambiguously confirmed. They provide a natural link between stellar-mass and supermassive black holes. They are believed to be found in dwarf galaxies, with an occupation fraction for active IMBHs of order 1% in both x-ray [@miller2015a; @baldassare17] and optical surveys [@moran2014a], as well as in globular clusters [@kiziltan2017a; @perera2017a; @oka2017a]. A substantial number of AGNs in dwarf galaxies may be missed because of soft x-ray absorption [@chen2017a]. Many of the dwarf “problems”, ranging from frequency, absence of cusps, too-big-to-fail and baryon fraction, can be accounted for by feedback from IMBH in the early gas-rich phase of dwarf galaxy evolution [@silk17a]. We infer that quiescent IMBHs are likely to occur in at least 10% of observed dwarfs as well as in massive globular clusters, and moreover are plausibly expected to be produced in many of the failed dwarfs predicted by galaxy formation theory in the context of the $\Lambda$CDM model, if only as relic building blocks of SMBHs [@rashkov2014a].
The typical IMBH mass lies in the range between those of stellar mass and supermassive black holes, i.e., $\sim 10^2 - 10^5 \rm M_\odot $. WDs are the end product of main-sequence stars with masses up to $\sim 8 \ \rm M_\odot .$ Due to the abundance of WDs, in dense environments they are prone to collide and/or merge with other objects. In some cases, WDs can be disrupted by IMBHs and explode as supernovae. The observation of the gamma-ray burst GRB060218 and its associated supernova, SN2006aj [@campana2006a], has been interpreted by @shcherbakov2013a as such an encounter. A similar interpretation for the object Swift J1644+57 has also been suggested by @krolik2011a.
The model {#sec:model}
=========
If a star of mass $M_\star$ is close to a black hole of mass $M_{bh}$, tidal forces may deform the star, stretching it along the orbital direction and squeezing it in the perpendicular direction. The net effect of this deformation is an increase in pressure which may be enough to trigger nuclear burn, causing the explosion of the star as a low-luminosity supernova [@carter1982a]. As suggested by @luminet1989a [@luminet1989b], this can happen for white dwarfs with masses between $0.07 M_\odot$ and $0.50 M_\odot$ that are not able to normally burn helium but which can be ignited by tidal disruptions. Similarly, for stars of $M_\star \gtrsim 0.5 M_\odot$, TDEs can trigger the burning of carbon and oxygen, of which white dwarfs in this mass range are largely composed.
Stars can be tidally disrupted by BHs if their distance is $$r_{t} \lesssim 0.1 M_{bh, 3}^{1/3} M_\star^{-1/3} R_{\star, -2} \; R_\odot$$ where $R_{\star, -2}$ is the stellar radius in units of $10^{-2}R_{\odot}$, $M_\star$ is the mass of the star in solar masses, and $M_{bh, 3}$ is the BH mass in units of $1000 M_\odot$. Let $r_p$ be the pericenter distance between the star and the black hole. The encounter strength ($\beta$) is usually defined as the ratio between the tidal radius $r_t$ and the pericenter distance $r_p$. The Schwarzschild radius of the black hole is $R_g = 2GM_{bh}/c^2$, thus constraining the maximum mass of a black hole for which the star can be tidally disrupted, to [@rosswog2009a] $$M_{bh,max} \approx 1.2 \times 10^{5} R_{\star,-2}^{3/2} M_{\star}^{-1/2} \; M_\odot.$$ Note that SMBHs with masses $M_{bh} \gtrsim 10^6 M_\odot$ cannot tidally disrupt white dwarfs.
Once ignited, the star will unavoidably become a supernova. The relative abundances of elements synthesized by white dwarfs in tidally-induced supernova events will be used to fix the initial composition of UHECRs in our model. Naturally, these abundances depend upon the initial composition of the star. Low-mass WDs ($0.07M_\odot \lesssim M_\star \lesssim 0.5M_\odot$) are mainly composed of helium, whereas those with masses $M_\star \gtrsim 0.5M_\odot$ have as their primary composition intermediate-mass nuclei such as carbon and oxygen. We now introduce two simple models, with the abundances ($a_i$) of each species obtained from estimates by @rosswog2009a\
[*model I*]{}: WDs with masses $M_\star < 0.5M_\odot$ are composed initially of helium, with abundances $a_\text{He} = 0.861$, $a_\text{C} = 0.12$, $a_\text{Si} = 0.11$, and $a_\text{Fe} = 0.016$;\
[*model II*]{}: WDs with masses $M_\star > 0.5M_\odot$ are composed initially of equal amounts of carbon and oxygen, with abundances $a_\text{He} = 0.568$, $a_\text{C} = 0.044$, $a_\text{O} = 0.096$, $a_\text{Ne} = 0.002$, $a_\text{Mg} = 0.002$, $a_\text{Si} = 0.114$, and $a_\text{Fe} = 0.173$.\
Model I is for a helium WD of $M_\star = 0.2M_\odot$ disrupted by a BH of $M_{bh} = 10^3M_\odot$, whereas model II is for a carbon/oxygen WD of $M_{\star} = 1.2M_\odot$ encountering a black hole of $M_{bh} = 500M_\odot$. The mass of the black hole and white dwarf, as well as the proximity of the encounter, are expected to affect the relative abundances of each element, so that models I and II are intended to be representative cases rather than an exact model.
We now want to estimate the luminosity in UHECRs for such encounters. For this, we follow @evans2015a and @shcherbakov2013a.
Debris of the explosion of the star will be accreted onto the BH at a rate $$\label{eq:accretionRate}
\dot{M}_{bh} \equiv \frac{dM_{bh}}{dt} = \frac{1}{3} \frac{M_\star}{t_{fb}} \left( \frac{t}{t_{fb}} \right)^{-\frac{5}{3}},$$ where $t_{fb}$ is the characteristic time for bound material to fall into the black hole, given by $$\label{eq:fbTime}
t_{fb} \approx 100 \alpha^{-3} R_{\star, -2}^{3/2} M_{bh,3}^{1/2} M_\star^{-1} \; \text{s},$$ where $\alpha \equiv \beta \sqrt{\xi} / \kappa$, with $\kappa \sim 0.5$ being a model-dependent parameter, and $\xi \equiv \Delta R_\star / R_\star$ designating the deformation factor for a spread $\Delta R_\star$. In WD-ignition scenarios, $\beta$ is typically $\sim 10$. Simulations find $\xi \sim 4$ [@evans2015a]. Therefore, our benchmark value is $\alpha \sim 40$.
Combining equations \[eq:accretionRate\] and \[eq:fbTime\], we can write $\dot{M}_{max} \equiv M_\star/(3t_{fb})$, which reads $$\dot{M}_{max} \approx 10^4 \alpha^{3} R_{\star,-2}^{-3/2} M_\star^2 M_{bh,3}^{-1/2} \; M_\odot \, \text{yr}^{-1}.
\label{eq:Mdotmax}$$ Note that the accretion rate decays with $t^{-5/3}$. This is believed to be a universal property of tidal disruption events [@rees1988a; @evans1989a; @phinney1989a; @laguna1993a; @rosswog2009a].
As the material from the supernova explosion falls into the black hole, part of the magnetic field of the former star is advected toward the event horizon, increasing the total magnetic energy in the vicinity of the black hole, yielding $$B_{bh} \sim 4 \times 10^{11} \dot{M}_{bh,4}^{1/2} M_{bh,3}^{-1} \; \text{G},
\label{eq:Bbh}$$ where $\dot{M}_{bh,4} \equiv \dot{M}_{bh} \times 10^{4} \; M_\odot \, \text{yr}^{-1}$. The magnetic field $B_{bh}$ is composed of a turbulent random component arising from magnetorotational instability [@balbus1991a; @hawley1991a; @balbus1998a], and the regular (poloidal) component generated via dynamo amplification of seed fields [@brandenburg2005a]. These fields can reach values of $B_{bh} \sim 10^{12}\;\text{G}$ [@brandenburg2005a], which can be maintained during a short period of time.
In Eq. \[eq:fbTime\], there is an $\alpha^3$ dependence. In studies by @stone2013a and @guillochon2013a, $\alpha$ has been found to be independent of the encounter strength $\beta$. These differences, however, do not invalidate our results. They would cause a change of one order of magnitude in the estimates of Eq. \[eq:Bbh\] for $\beta \sim 10$, and a factor of a few for $\beta \lesssim 5$.
We now assume that this is a Kerr black hole and that energy can be extracted from the BH via a mechanism such as @blandford1982a or @blandford1977a. For details, the reader can refer to e.g. [@blandford1977a; @rees1982a; @mckinney2004a]. The jet is launched due to the presence of a poloidal magnetic field, as shown in general relativistic magnetohydrodynamical simulations performed by @tchekhovskoy2011a and @mckinney2012a.
The main difference between this and other tidal disruption models found in the literature is the ignition of the WD. About 10% of the material of the supernova explosion is accreted onto the BH, powering the jet. The remaining material contains cosmic rays accelerated by the supernova. When these cosmic rays cross the jet, they can receive an energy boost of $\sim \gamma^2$, as proposed by @caprioli2015a for the case of AGN jets accelerating cosmic rays produced within their host galaxies; here $\gamma \approx \Psi^{-1}$ is the Lorentz factor of the jet, and $\Psi$ its opening angle. Because the accelerated cosmic rays are not part of the accretion flow, in this scenario, acceleration would not occur near the source, but rather in optically thin regions away from the base of the jet. Consequently, photodisintegration is not expected to be relevant, implying that the primary composition of the UHECRs remains virtually unaltered. The jet itself may also accelerate cosmic rays to ultra-high energies, but in this case interactions with the envelope could heavily degrade the energy of the escaping particles. Nevertheless, this is important to determine the neutrino and gamma-ray signatures of the relevant processes.
Validating the model
====================
It is well-known that supernovae can accelerate cosmic rays up to energies of at least $\sim Z \times 10^{15} \; \text{eV}$ via diffusive shock acceleration (DSA). Features in the cosmic-ray spectrum, namely the second knee and the ankle, provide a very strong case for the existence of a local (possibly galactic) component of CRs that cannot be explained with simple linear DSA theory [@hillas2005a; @thoudam2016a], thus requiring an extension of the DSA or possibly a completely new acceleration mechanism. Although the exact details of such processes are not relevant for our phenomenological description, there are interesting possibilities that naturally extend the spectrum of cosmic rays accelerated by supernovae beyond the second knee that should be mentioned. One was suggested by Bell & Lucek [@bell2001a] (see also Refs. [@lucek2000a; @bell2004a]) and consists of the non-linear amplification of magnetic fields due to the backreaction of the streaming CRs on the medium. Other mechanisms predict acceleration in winds of companion objects [@voelk1988a; @crocker2015a]. The mechanisms commonly found in the literature, including the aforementioned ones, predict $E_{max} \sim Z \times 10^{17} \; \text{eV}$. Following the supernova explosion, there are two groups of CRs that are accelerated. The first one is composed by galactic CRs present in the interstellar medium (hereafter referred to as *ISM composition*), and the second relates to the abundances of nuclei synthesized in the explosion (designated as *SN composition*). While for $E \lesssim Z \times 10^{15} \; \text{eV}$, the acceleration of ISM-composition CRs via DSA (or a similar mechanism) is virtually guaranteed, the same is not true for the case of SN composition as the synthesized nuclei would not be accelerated by forward shocks, but rather by reverse shocks, whose efficiency is about 10% the efficiency of the former. The viability of particle acceleration in reverse shocks has been demonstrated in many studies [@ellison2005a; @zirakashvili2010a; @telezhinsky2012a; @lee2014a]. Because DSA of CRs with ISM composition is not enough to explain the composition of galactic cosmic rays, an additional mechanism such as reverse shocks is required, whereby non-ISM nuclei can be produced and subsequently accelerated [@parizot2014a]. As a matter of fact, particle acceleration in reverse shocks has been proposed in Refs. [@ramaty1996a; @lingenfelter1998a] to explain a long-standing problem related to the abundance of elements such as lithium, beryllium, and boron in galactic cosmic rays. While this is rather speculative, it is not unreasonable to expect a contribution from nuclei synthesized by a supernova. From the theoretical viewpoint, particle acceleration could also take place in black-hole winds produced via a mechanism similar to a generalization of advection-dominated accretion flows [@blandford1999a; @tombesi2015a].
Now the main ingredients of our model are available: cosmic-ray nuclei are synthesized in the SN and subsequently accelerated to energies of at least $E_{max} \sim \text{a few} \times 10^{17} \; \text{eV}$; an accretion-powered jet is launched by the BH following the tidal disruption event. We now invoke a one-shot mechanism inspired by Ref. [@caprioli2015a], whereby CRs crossing the jet receive an energy boost proportional to $\Gamma^2$, with $\Gamma \sim 10$ being the Lorentz factor of the jet. A slightly higher $E_{max}$ than predicted by linear DSA theory and/or a jet Lorentz factor of $\Gamma \approx 15$ are sufficient to explain the existence of CRs at energies $\sim 10^{20} \; \text{eV}$.
Not all IMBH-WD disruptions will lead to supernovae, but up to $\sim 10\%$ of them will. The Pierre Auger Observatory has derived lower bounds on the density of sources based on the absence of clustering in the arrival directions of UHECRs; the least stringent bound is $n_0 \gtrsim 6 \times 10^{-6} \; \text{Mpc}^{-3}$ and the most stringent $n_0 \gtrsim 7 \times 10^{-4} \; \text{Mpc}^{-3}$, at 95% confidence level, depending on how sources are distributed and on the strength of extragalactic magnetic fields [@auger2013a]. Under the simple assumption that the rate of IMBH-WD encounters is constant over time, we obtain a true density $n \sim 10^{-1} \; \text{Mpc}^{-3}$ averaged over a Hubble time, using estimates from Ref. [@baumgardt2004a]. Due to deflections in intervening galactic and extragalactic magnetic fields, the apparent rate of these events would be $\dot{n}_{s} \sim n / \delta t$ [@waxman1996a; @takami2012a], where $\delta t$ is the typical spread in arrival times, given by [@waxman1996a] $$ \delta t \simeq 10^{7} B_\text{nG}^2 \ell_\text{Mpc} E_{20}^{-2} \;\, \text{yr},
\label{eq:tdelay}$$ and where $B_\text{nG} = B / \text{nG}$ is the magnetic field in nG, $\ell_\text{Mpc}$ is the coherence length in Mpc, and $E_{20}$ is the cosmic-ray energy in units of $10^{20} \; \text{eV}$. This equation holds for sources closer than $D \lesssim 75 \; \text{Mpc}$, which corresponds to the interaction horizon at ultra-high energies. Within a sphere of radius 75 Mpc, most of the volume is likely occupied by voids (for a detailed discussion see e.g. Ref. [@alvesbatista2017a]), and we conservatively estimate $\dot{n}_s \sim 10^{-8} \; \text{Mpc}^{-3}\,\text{yr}^{-1}$ using the upper limit for magnetic fields in voids, $B \sim 1 \; \text{nG}$, assuming a coherence length $\ell \simeq 1 \; \text{Mpc}$ [@planck2016a].
For a typical IMBH of 1000$M_\odot$, and a WD of $1 M_{\odot}$ and $10^{-2} R_\odot$, the maximum accretion rate is given by Eq. \[eq:Mdotmax\]. The energy injection rate ($\dot{E}$) is $$\dot{E} = \dot{n}_s E_{CR} = \frac{n}{\delta t} \epsilon E_{tot} ,$$ with $\epsilon$ being the efficiency of conversion of the total energy of the progenitor star ($E_{CR} \equiv \epsilon E_{tot}$) into cosmic rays. The total energy, $E_{tot}$, can be as high as $\sim 10^{51} \; \text{erg}$. Estimates by @katz2009a suggest $\dot{E} \sim 10^{44} \; \text{erg} \, \text{Mpc}^{-3} \, \text{yr}^{-1}$, so that in order to satisfy energy requirements, one would need $\epsilon / \delta t \gtrsim 10^{-6} \; \text{yr}^{-1}$. We estimate that $\epsilon \ll 10\%$. Recalling that $\delta t \propto B^2 \ell$ (from Eq. \[eq:tdelay\]), a simple decrease in the magnetic field by less than one order of magnitude would already render our model viable.
While we could have derived the efficiency factor from considerations about the mass function of black holes, rates of IMBH-WD TDEs, etc, the uncertainties related to these quantities and the likely dependence of $\epsilon$ on other parameters justify our approach.
Using the abundances of the elements that compose the WD, in addition to the elements synthesized in the supernova, we can fix the UHECR composition. As a proxy for the composition, we use the depth of the maximum of the air shower ($\langle X_{max} \rangle$). Therefore, the only free parameters of the model are the spectral index of the injected cosmic rays ($\alpha$), the maximum rigidity to which they are accelerated ($\mathcal{R}_{max}$), and the source evolution, assumed here to follow the star formation rate. This is a major source of uncertainty in our model, as the source evolution is a combination of the evolution of WDs, which follow the star formation rate, and of IMBHs, which is not well known [@volonteri2011a].
In order to illustrate the predictions made by the model, we compute the expected spectrum and composition by simulating the propagation of UHECRs from the sources to Earth. Assuming that the sources are randomly distributed in the comoving volume, we simulate the propagation of UHECRs using the CRPropa 3 code [@alvesbatista2016a]. We consider all relevant interactions and energy loss processes, namely: photopion production, pair production, photodisintegration, nuclear decay, and adiabatic losses due to the expansion of the universe. The main target for interactions are the cosmic microwave background (CMB) and the extragalactic background light (EBL). In particular, we use the EBL model by @gilmore2012a. The simulations use as input the fraction of each nuclear species following. The spectrum has the form $$\dfrac{dN_{inj}}{dE} = \sum\limits_{i=1}^{26} J_0 a_i E_{inj}^{-\alpha} f(E_{inj}, Z_i),
\label{eq:dNdE}$$ where $a_i$ is the relative abundance of each nucleus of atomic number $Z$, $J_0$ is the overall normalization factor, and the function $f(E_{inj})$ is 1 for $E_{inj}<Z \mathcal{R}_{max}$ and $\exp(1 - E_{inj}/(Z \mathcal{R}_{max}))$ otherwise.
Note that forward and reverse shocks may have different maximal rigidities ($\mathcal{R}_{max}$), spectral indices ($\gamma$), and injected elemental abundances ($a_i$). Moreover, these parameters may be different for the two source populations (helium and carbon-oxygen WDs). Therefore, the actual spectrum is a combination of Eq. \[eq:dNdE\] for forward and reverse shocks, as well as for both source populations.
It is worth stressing that the combined spectrum-composition fit [@auger2017a] of the data collected by the Pierre Auger Observatory suggests hard spectral indices, which is consistent with a contribution of reverse shocks to the total flux, as forward shocks tend to have softer spectral indices compared to reverse shocks [@ptuskin2013a].
The results are shown in Fig. \[fig:speccomp\] for both models. In order to obtain $\langle X_{max} \rangle$, we have used the parametrization given by @auger2013b, assuming the hadronic interaction model EPOS-LHC [@pierog2015a] for the development of cosmic-ray showers in the atmosphere.
{width=".495\textwidth"} {width=".495\textwidth"}
Note that Fig. \[fig:speccomp\] is merely for illustration purposes. In reality, $\mathcal{R}_{max}$ and $\alpha$ depend on the acceleration model, being the maximal attainable energy limited by the magnetic field capable of confining cosmic rays within the source environment. The spectrum and composition should be simultaneously constrained as done by the @auger2017a. Nevertheless, due to the many uncertainties involved in such a combined fit, whose evaluation is beyond the scope of this work, we simply present a couple of scenarios to qualitatively discuss the model. Other combinations of $\alpha$, $\mathcal{R}_{max}$, and source evolution may lead to better agreement with the measurements. Our model was devised to hold for $E \gtrsim 10^{18.6} \; \text{eV}$, since this is the energy at which the galactic and extragalactic components of the cosmic-ray spectrum overlap, according to some models [@deligny2014a].
These predictions are subject to uncertainties related both to the propagation of UHECRs from their sources to Earth and to the relative abundance of each element synthesized in the supernova. The latter has been discussed by @rosswog2009a and is inherent to the modeling of the explosion. The former depends on parameters such as the exact distribution of sources, photonuclear cross sections, and the EBL model adopted, as shown in Ref. [@alvesbatista2015a]. Moreover, magnetic fields may also play a role in the spectral shape, although to which extent it is uncertain.
Gravitational waves produced by these kinds of events are unlikely to be detectable by Advanced LIGO, since they would be at the experimental sensitivity limit, amplitude $\sim 10^{-22}$ and frequency $\sim 10 \; \text{Hz}$) [@ligo2016a].
A detailed calculation of the electromagnetic signatures of IMBH-WD tidal disruption events can be found in @macleod2014a, and suggests signatures in x-rays.
Discussion and Conclusions
==========================
We have presented an original model for UHECR acceleration in IMBH-WD TDEs with ignition of the white dwarf, that naturally explains the intermediate-mass composition of UHECRs observed by Auger. We have described the acceleration of cosmic rays to ultra-high energies by assuming a one-shot reacceleration mechanism in jets generated via TDEs of WDs by IMBHs. Cosmic rays are first accelerated by the supernova and subsequently they are reaccelerated by the accretion-powered jet. A number of mechanisms for cosmic-ray acceleration by supernovae and reacceleration by jets could be invoked. Nevertheless, one should keep in mind that in order for intermediate/heavy nuclei to escape, the source environment has to be optically thin.
Promising environments to search for IMBH-WD systems are globular clusters. Globular clusters are abundant in the universe and even in the Milky Way [@rashkov2014a]. About 10% of all stars disrupted in globular clusters are WDs. Moreover, N-body simulations by @baumgardt2004a suggest that one in every ten of these clusters contain IMBHs. Recent observations confirm that nearby massive globular clusters may contain IMBH [@kiziltan2017a; @perera2017a; @oka2017a].
Other interesting candidates are dwarf galaxies, as they may host IMBHs that are possibly required to seed the formation of supermassive black holes [@volonteri2003a; @moran2014a]. The large number of relic dwarfs expected in CDM-motivated galaxy formation models provides a potentially rich source of IMBHs that have been proposed to play an important role in determining the Milky Way galaxy’s baryon fraction [@peirani2012a]. Moreover the inefficient formation of the central SMBH by merging IMBHs should leave a significant population of surviving IMBHs [@islam03a]. The Milky Way may consequently harbor IMBHs in regions where WDs are fairly abundant. Therefore, one can expect a few IMBH-WD encounters within our galaxy contributing to the observed UHECR spectrum.
A possible tracer of IMBH-WD tidal disruption events are ultra-luminous x-ray sources (ULXs), whose luminosities ($\gtrsim 10^{39} \; \text{erg}\,\text{s}^{-1}$) exceed those of any stellar process known. @shcherbakov2013a has interpreted the x-ray flare from GRB060218 as such an event. In particular, the observation of an associated supernova (SN2006aj) strengthens this interpretation.
Detailed numerical simulations of IMBH-WD encounters could be used to improve the estimates of the composition of the exploding star, thus fixing the composition of UHECRs in the neighborhood of sources. Electromagnetic, neutrino, and gravitational-wave counterparts of such events could provide a way to test acceleration models in IMBH-WD systems. Nevertheless, because tidal ignitions of WDs by IMBHs are transient events, we do not expect to see UHECRs in coincidence with photons, neutrinos, and gravitational waves, since time delays would be incurred by intervening magnetic fields. In general because of the short time-scales and higher luminosities of white dwarf TDEs [@krolik2012a], they are expected to be prominent in flux-limited transient surveys [@baumgardt2006a].
RAB acknowledges support from the John Templeton Foundation. The work of JS has been supported in part by ERC Project No. 267117 (DARK) hosted by the Pierre and Marie Curie University-Paris VI, Sorbonne Universities and CEA-Saclay.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Junquan Tao[^1] on behalf of the CMS collaboration'
title: Search for rare and exotic Higgs Boson decay modes
---
Introduction {#intro}
============
The discovery of the Higgs boson with a mass of 125.09 $\pm$ 0.24 ${\mathrm{{GeV}}}$ [@Aad:2012tfa; @Chatrchyan:2012xdj; @Chatrchyan:2013lba] at the Large Hadron Collider (LHC) has generated great interest in exploring its properties. Many measurements from LHC have confirmed that the Higgs boson has properties, including spin, CP, and coupling strengths, that are compatible with those expected for the Higgs boson of the SM [@Englert:1964et; @Higgs:1964pj; @Guralnik:1964eu]. However the scalar sector is not well known experimentally yet, and current measurements could still accommodate for large contributions of new physics in this sector. Constraints on new physics from the LHC are still relatively loose [@Khachatryan:2016vau], which leaves room for beyond the standard model (BSM) physics. Many Higgs rare decays of Higgs boson in SM have yet to be observed, which may be sensitive to new physics if additional Higgs couplings exist. Observing exotic decays of the Higgs boson would be a striking direct evidence for the existence of physics beyond the SM. The latest results in the search for rare and exotic Higgs boson decays in proton-proton collision events collected with the CMS detector at the LHC will be summarized in this paper. The searches are performed for an extensive set of decay modes of Higgs boson including $\mathrm{H}\rightarrow{\rm X (X \rightarrow2\ell)\gamma}$ ($X= {\rm Z}, \gamma^* $ and $\ell={\rm e},\mu$), $\mathrm{H}\rightarrow{ \mu\mu / {\rm e}{\rm e}}$, invisible decays, lepton flavour violating decays and Higgs decay to light scalars or pseudo-scalars, with the full LHC run-1 data collected during 2011 and 2012 at $\sqrt{s}=7-8~{\mathrm{{TeV}}}$ and with the run-2 data at $\sqrt{s}=13~{\mathrm{{TeV}}}$ collected during 2015 and 2016. The searches are generally performed in a model independent approach, and exclusion limits in terms of production cross section times the corresponding decay branching ratio are presented. Many of the results are interpreted according to beyond-SM (BSM) Higgs scenarios, which include the Two Higgs Doublet Model (2HDM) and the Singlet Model.
$\mathrm{H}\rightarrow{\rm X \gamma}$ {#HXg}
=====================================
Within the SM, the partial width for the ${\ensuremath{{H}}}\to{\ensuremath{\mathrm{Z}}\xspace}\gamma$ decay channel ($\Gamma_{{\ensuremath{\mathrm{Z}}\xspace}\gamma}$) is rather small, resulting in a branching fraction between 0.11% and 0.25% in the 120 $-$ 160 ${\mathrm{{GeV}}}$ [@Cahn:1978nz; @Bergstrom:1985hp] mass range. A measurement of $\Gamma_{{\ensuremath{\mathrm{Z}}\xspace}\gamma}$ provides important information on the underlying dynamics of the Higgs sector because it is induced by loops of heavy charged particles, just as for the ${\ensuremath{{H}}}\to\gamma\gamma$ decay channel. This paper summarizes the most recent search for a Higgs boson in the ${\ensuremath{{H}}}\rightarrow{\ensuremath{\mathrm{Z}}\xspace}\gamma$ final state at the LHC in the 120 $-$ 160 ${\mathrm{{GeV}}}$ mass range, with the ${\ensuremath{\mathrm{Z}}\xspace}$ boson decaying into an electron or a muon pair [@Chatrchyan:2013vaa]. This is a clean final-state topology with an effective mass peak resolution of about 1-3%. Events were collected at center-of-mass energies of 7 ${\mathrm{{TeV}}}$ and 8 ${\mathrm{{TeV}}}$, corresponding to integrated luminosities of 5.0 ${\mathrm{fb}^{-1}}$ and 19.6 ${\mathrm{fb}^{-1}}$, respectively. The selected events are required to have opposite-sign electron or muon pairs. The mass spectrum for all channels combined is shown in the left plot of Fig. \[fig:HZg\]. No excess above standard model predictions has been found in the 120–160 ${\mathrm{{GeV}}}$ mass range and the first limits on the Higgs boson production cross section times the ${\ensuremath{{H}}}\to{\ensuremath{\mathrm{Z}}\xspace}\gamma$ branching fraction at the LHC have been derived, as shown in the right plot of Fig. \[fig:HZg\]. For a standard model Higgs boson mass of 125 ${\mathrm{{GeV}}}$ the expected limit at the 95$\%$ confidence level is 10 and the observed limit is 9.5.
![Left plot shows the $m_{\ell\ell\gamma}$ spectrum in the electron and the muon channels for the 7 and 8 ${\mathrm{{TeV}}}$ data combined, without weighting by the expected signal to background ratio of the individual data samples. Also shown is the expected signal due to a 125 ${\mathrm{{GeV}}}$ standard model Higgs boson, scaled by 75, and the sum of the individual fits made to the data for each channel and event class. The uncertainty band reflects the statistical uncertainty from the fits to the data. The exclusion limit on the cross section times the branching fraction of a Higgs boson decaying into a ${\ensuremath{\mathrm{Z}}\xspace}$ boson and a photon divided by the SM value is shown in the right plot. The red line represents the same cross section times the branching fraction as the SM prediction [@Chatrchyan:2013vaa].[]{data-label="fig:HZg"}](HZg_combinedData_All_2GeV.pdf "fig:"){width="48.00000%"} ![Left plot shows the $m_{\ell\ell\gamma}$ spectrum in the electron and the muon channels for the 7 and 8 ${\mathrm{{TeV}}}$ data combined, without weighting by the expected signal to background ratio of the individual data samples. Also shown is the expected signal due to a 125 ${\mathrm{{GeV}}}$ standard model Higgs boson, scaled by 75, and the sum of the individual fits made to the data for each channel and event class. The uncertainty band reflects the statistical uncertainty from the fits to the data. The exclusion limit on the cross section times the branching fraction of a Higgs boson decaying into a ${\ensuremath{\mathrm{Z}}\xspace}$ boson and a photon divided by the SM value is shown in the right plot. The red line represents the same cross section times the branching fraction as the SM prediction [@Chatrchyan:2013vaa].[]{data-label="fig:HZg"}](HZg_limitplot_AnalyticPDF_VBFTag.pdf "fig:"){width="51.00000%"}
The rare decay into the $\ell\ell\gamma$ final state of the Higgs boson is a rich source of information that can enhance our understanding of its basic properties and probe novel couplings predicted by extensions of the standard model of particle physics. The search for a Higgs boson decay ${\ensuremath{{H}}}\to\gamma^*\gamma\to\ell\ell\gamma$ is performed using proton-proton collision data recorded with the CMS detector at the LHC at a centre-of-mass energy of 8 ${\mathrm{{TeV}}}$, corresponding to an integrated luminosity of 19.7 ${\mathrm{fb}^{-1}}$ [@Khachatryan:2015lga]. No excess above the background predictions has been found in the three-body invariant mass range $120<m_{\ell\ell\gamma}<150~{\mathrm{{GeV}}}$. Limits on the Higgs boson production cross section times the ${\ensuremath{{H}}}\to\gamma^*\gamma\to\ell\ell\gamma$ branching fraction divided by the SM values have been derived, as shown in the left plot of Fig. \[fig:Hgg\]. The observed limit for $m_{\ensuremath{{H}}}=125~{\mathrm{{GeV}}}$ is about 6.7 times the SM prediction. Limits at 95% CL on $\sigma({\ensuremath{{p}}}{\ensuremath{{p}}}\to {\ensuremath{{H}}})\,\mathcal{B}({\ensuremath{{H}}}\to\mu\mu\gamma)$ for a narrow resonance are also obtained in the muon channel, as shown in the right plot of Fig. \[fig:Hgg\]. The observed limit for $m_{\ensuremath{{H}}}= 125~{\mathrm{{GeV}}}$ is 7.3 ${\mathrm{fb}}$. In addition, a search is performed for ${\ensuremath{{H}}}\to {\ensuremath{{J}\hspace{-.08em}/\hspace{-.14em}\psi}\xspace}\gamma\to\mu\mu\gamma$ decay for $m_{\ensuremath{{H}}}=125~{\mathrm{{GeV}}}$, which is sensitive to the Higgs boson coupling to charm quark and a promising way to access the couplings of the Higgs boson to the second generation quarks at the LHC [@Khachatryan:2015lga]. Events consistent with the ${\ensuremath{{J}\hspace{-.08em}/\hspace{-.14em}\psi}\xspace}$ in dimuon invariant mass are used to set a 95% CL limit on the branching fraction $\mathcal{B}({\ensuremath{{H}}}\to{\ensuremath{{J}\hspace{-.08em}/\hspace{-.14em}\psi}\xspace}\gamma) < 1.5\times10^{-3}$, that is, 540 times the SM prediction for $m_{\ensuremath{{H}}}=125~{\mathrm{{GeV}}}$.
![(Left) The 95% CL exclusion limit, as a function of the mass hypothesis, $m_{\ensuremath{{H}}}$, on $\sigma/\sigma_{SM}$, the cross section times the branching fraction of a Higgs boson decaying into a photon and a lepton pair with $m_{\ell\ell} < 20{\mathrm{{GeV}}}$, divided by the SM value. (Right) The 95% CL exclusion limit on $\sigma({\ensuremath{{p}}}{\ensuremath{{p}}}\to {\ensuremath{{H}}})\,\mathcal{B}({\ensuremath{{H}}}\to\mu\mu\gamma)$, with $m_{\mu\mu} < 20{\mathrm{{GeV}}}$, for a Higgs-like particle, as a function of the mass hypothesis, $m_{\ensuremath{{H}}}$ [@Khachatryan:2015lga].[]{data-label="fig:Hgg"}](Hgg_limits_combo.pdf "fig:"){width="49.00000%"} ![(Left) The 95% CL exclusion limit, as a function of the mass hypothesis, $m_{\ensuremath{{H}}}$, on $\sigma/\sigma_{SM}$, the cross section times the branching fraction of a Higgs boson decaying into a photon and a lepton pair with $m_{\ell\ell} < 20{\mathrm{{GeV}}}$, divided by the SM value. (Right) The 95% CL exclusion limit on $\sigma({\ensuremath{{p}}}{\ensuremath{{p}}}\to {\ensuremath{{H}}})\,\mathcal{B}({\ensuremath{{H}}}\to\mu\mu\gamma)$, with $m_{\mu\mu} < 20{\mathrm{{GeV}}}$, for a Higgs-like particle, as a function of the mass hypothesis, $m_{\ensuremath{{H}}}$ [@Khachatryan:2015lga].[]{data-label="fig:Hgg"}](Hgg_limits_xsBR_catEB_mu.pdf "fig:"){width="49.00000%"}
$\mathrm{H}\rightarrow{ \mu\mu / {\rm e}{\rm e}}$ {#Hll}
=================================================
For a Higgs boson mass, ${\ensuremath{m_{\ensuremath{{H}}}}\xspace}$, of 125 ${\mathrm{{GeV}}}$, the SM prediction for the Higgs to ${\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}$ branching fraction, ${\ensuremath{\mathcal{B}}\xspace}({\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace})$, is among the smallest accessible at the CERN LHC, $2.2\times 10^{-4}$ [@Denner:2011mq], while the SM prediction for [$\mathcal{B}$]{}([${\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}$]{}) of approximately $5\times10^{-9}$ is inaccessible at the LHC. Experimentally, however, ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace}$ and ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace}$ are the cleanest of the fermionic decays. The clean final states allow a better sensitivity, in terms of cross section, $\sigma$, times branching fraction, ${\ensuremath{\mathcal{B}}\xspace}$, than ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{\tau^{+}\tau^{-}}\xspace}}\xspace}$. In addition, a measurement of the ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace}$ decay probes the Yukawa coupling of the Higgs boson to second-generation fermions, an important input in understanding the mechanism of electroweak symmetry breaking in the SM [@Plehn:2001qg; @Han:2002gp]. Deviations from the SM expectation could also be a sign of BSM physics [@Vignaroli:2009vt; @Dery:2013rta]. The [${\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}$]{}search is performed on data corresponding to integrated luminosities of $5.0\pm0.1$ ${\mathrm{fb}^{-1}}$ at a centre-of-mass energy of 7 ${\mathrm{{TeV}}}$ and $19.7\pm0.5$ ${\mathrm{fb}^{-1}}$ at 8 ${\mathrm{{TeV}}}$, while the ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace}$ search is only performed on the 8 ${\mathrm{{TeV}}}$ data [@Khachatryan:2014aep]. Results are presented for Higgs boson masses between 120 and 150 ${\mathrm{{GeV}}}$. Events are split into categories corresponding to different production topologies and dilepton invariant mass resolutions. The signal strength is then extracted using a simultaneous fit to the dilepton invariant mass spectra in all of the categories. No significant ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace}$ signal is observed. Upper limits are set on the signal strength at the 95% CL, as shown in the left plot of Fig. \[fig:Hll\]. The combined observed limit on the signal strength, for a Higgs boson with a mass of 125 ${\mathrm{{GeV}}}$, is 7.4, while the expected limit is $6.5^{+2.8}_{-1.9}$. Assuming the SM production cross section, this corresponds to an upper limit of 0.0016 on ${\ensuremath{\mathcal{B}}\xspace}({\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace})$. For a Higgs boson mass of 125 ${\mathrm{{GeV}}}$, the best fit signal strength is $0.8^{+3.5}_{-3.4}$. In the ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace}$ channel, SM Higgs boson decays are far too rare to detect, and no signal is observed. For a Higgs boson mass of 125 ${\mathrm{{GeV}}}$, a 95% CL upper limit of 0.041 pb is set on $\sigma {\ensuremath{\mathcal{B}}\xspace}({\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace})$ at 8 ${\mathrm{{TeV}}}$, as shown in the right plot of Fig. \[fig:Hll\]. Assuming the SM production cross section, this corresponds to an upper limit on ${\ensuremath{\mathcal{B}}\xspace}({\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace})$ of 0.0019, which is approximately $3.7\times10^5$ times the SM prediction. For comparison, the ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace}$ observed 95% CL upper limit on $\sigma {\ensuremath{\mathcal{B}}\xspace}({\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace})$ as shown in the middle plot of Fig. \[fig:Hll\] is 0.033 pb (using only 8 ${\mathrm{{TeV}}}$ data), which is 7.0 times the expected SM Higgs boson cross section.
![ Mass scan for the background-only expected and observed combined exclusion limits (left). Exclusion limits on $\sigma {\ensuremath{\mathcal{B}}\xspace}$ are shown for ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace}$ (middle), and for ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace}$ (right), both for 8 ${\mathrm{{TeV}}}$. Theoretical uncertainties on the cross sections and branching fraction are omitted, and the relative contributions of GF, VBF, and VH are as predicted in the SM [@Khachatryan:2014aep]. []{data-label="fig:Hll"}](Hll_CombSplitAll_7P8TeV.pdf "fig:"){width="32.50000%"} ![ Mass scan for the background-only expected and observed combined exclusion limits (left). Exclusion limits on $\sigma {\ensuremath{\mathcal{B}}\xspace}$ are shown for ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace}$ (middle), and for ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace}$ (right), both for 8 ${\mathrm{{TeV}}}$. Theoretical uncertainties on the cross sections and branching fraction are omitted, and the relative contributions of GF, VBF, and VH are as predicted in the SM [@Khachatryan:2014aep]. []{data-label="fig:Hll"}](Hll_xsbr_CombSplitAll_8TeV.pdf "fig:"){width="32.50000%"} ![ Mass scan for the background-only expected and observed combined exclusion limits (left). Exclusion limits on $\sigma {\ensuremath{\mathcal{B}}\xspace}$ are shown for ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{{\mu^+}}}{\ensuremath{{\mu^-}}}}\xspace}}\xspace}$ (middle), and for ${\ensuremath{{\ensuremath{{H}}}\to{\ensuremath{{\ensuremath{\mathrm{e}^+}}{\ensuremath{\mathrm{e}^-}}}\xspace}}\xspace}$ (right), both for 8 ${\mathrm{{TeV}}}$. Theoretical uncertainties on the cross sections and branching fraction are omitted, and the relative contributions of GF, VBF, and VH are as predicted in the SM [@Khachatryan:2014aep]. []{data-label="fig:Hll"}](Hll_xsbr_EE_8TeV.pdf "fig:"){width="32.50000%"}
Invisible decays {#Hinv}
=================
A number of models for physics beyond the SM allow for invisible decay modes of the Higgs boson, such as decays to neutralinos in supersymmetric models [@Belanger:2001am] or graviscalars in models with extra spatial dimensions [@Giudice:2000av; @Dominici:2009pq]. More generally, invisible Higgs boson decays can be realised through interactions between the Higgs boson and dark matter (DM) [@Shrock:1982kd]. Direct searches for invisible decays of the Higgs boson increase the sensitivity to the invisible Higgs boson width beyond the indirect constraints. The typical signature at the LHC is a large missing transverse momentum recoiling against a distinctive visible system. Firstly a combination of searches for invisible decays of the Higgs boson using data collected during 2011, 2012, and 2015 are presented [@Khachatryan:2016whc]. The data collected with the CMS detector at the LHC correspond to integrated luminosities of 5.1, 19.7, and 2.3 ${\mathrm{fb}^{-1}}$ at centre-of-mass energies of 7, 8, and 13 ${\mathrm{{TeV}}}$, respectively. The combination includes searches targeting Higgs boson production in the ZH mode, in which a Z boson decays to $\ell^{+}\ell^{-}$ or ${\ensuremath{{b\overline{b}}}\xspace}$, and the qqH mode, which is the most sensitive channel. The combination also includes the first searches at CMS targeting VH production, in which the vector boson decays hadronically, and the ggH mode in which the Higgs boson is produced in association with jets. No significant deviations from the SM predictions are observed and upper limits are placed on the branching fraction for the Higgs boson decay to invisible particles. The combination of all searches yields an observed (expected) upper limit on ${\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}$ of ${\ensuremath{0.24}}$ $({\ensuremath{0.23}})$ at the 95% confidence level, assuming SM production of the Higgs boson.
Sear for Higgs boson to invisible particles in final states with an energetic jet (Monojet) or a hadronically decaying W or Z boson (Mono-V) based on the 2016 data sample of proton-proton collisions at $\sqrt{s} = 13~\mathrm{TeV}$ corresponding to an integrated luminosity of $35.9~\mathrm{fb}^{-1}$ is also presented [@CMS:2017tbk]. The observed (expected) 95% CL upper limit on the invisible branching fraction of the Higgs boson, $\sigma \times {\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}/ \sigma_{\textrm{SM}}$, is found to be 53% (40%). The limits are summarized in Fig. \[fig:Hinv\] (Middle). Additionally search for Higgs boson decaying invisibly and produced in association with the Z boson has been updated based on the 2016 data sample with an integrated luminosity of $35.9~\mathrm{fb}^{-1}$ [@Sirunyan:2017qfc]. The 95% CL median expected and observed upper limits on the production cross section times branching fraction, $\sigma_{ZH} \times C$, computed with the asymptotic $CL_{s}$ method are shown in Figure \[fig:Hinv\] (Right) for the ${\ensuremath{{\ensuremath{E_\mathrm{T}}}^{\mathrm{miss}}}}$-shape analysis. Assuming the SM production rate, the 95% observed (expected) CL upper limit on ${\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}$ is 0.45 (0.44) using the ${\ensuremath{{\ensuremath{E_\mathrm{T}}}^{\mathrm{miss}}}}$-shape analysis, and 0.40 (0.42) using the multivariate analysis. The $\mathrm{gg} \to Z(\ell\ell){\ensuremath{{H}}}$ process has been considered only for the 125 GeV mass point.
![ (Left) Observed and expected 95% CL limits on $\sigma \times {\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}/ \sigma_{\textrm{SM}}$ for individual combinations of categories targeting qqH, VH, and ggH production, and the full combination assuming a Higgs boson with a mass of 125 ${\mathrm{{GeV}}}$ [@Khachatryan:2016whc]. (Middle) Expected (dotted black line) and observed (solid black line) 95% CL upper limits on the invisible branching fraction of the 125 ${\mathrm{{GeV}}}$ SM-like Higgs boson. Limits are shown for the monojet and mono-V categories separately, and also for their combination [@CMS:2017tbk]. (Right) Expected and observed 95% CL upper limits on the production cross section times branching fraction, $\sigma_{ZH} \times {\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}$ as a function of the Higgs boson mass [@Sirunyan:2017qfc]. []{data-label="fig:Hinv"}](Hinv_combined111215.pdf "fig:"){width="41.00000%"} ![ (Left) Observed and expected 95% CL limits on $\sigma \times {\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}/ \sigma_{\textrm{SM}}$ for individual combinations of categories targeting qqH, VH, and ggH production, and the full combination assuming a Higgs boson with a mass of 125 ${\mathrm{{GeV}}}$ [@Khachatryan:2016whc]. (Middle) Expected (dotted black line) and observed (solid black line) 95% CL upper limits on the invisible branching fraction of the 125 ${\mathrm{{GeV}}}$ SM-like Higgs boson. Limits are shown for the monojet and mono-V categories separately, and also for their combination [@CMS:2017tbk]. (Right) Expected and observed 95% CL upper limits on the production cross section times branching fraction, $\sigma_{ZH} \times {\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}$ as a function of the Higgs boson mass [@Sirunyan:2017qfc]. []{data-label="fig:Hinv"}](higgsInv_brazilian.pdf "fig:"){width="28.50000%"} ![ (Left) Observed and expected 95% CL limits on $\sigma \times {\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}/ \sigma_{\textrm{SM}}$ for individual combinations of categories targeting qqH, VH, and ggH production, and the full combination assuming a Higgs boson with a mass of 125 ${\mathrm{{GeV}}}$ [@Khachatryan:2016whc]. (Middle) Expected (dotted black line) and observed (solid black line) 95% CL upper limits on the invisible branching fraction of the 125 ${\mathrm{{GeV}}}$ SM-like Higgs boson. Limits are shown for the monojet and mono-V categories separately, and also for their combination [@CMS:2017tbk]. (Right) Expected and observed 95% CL upper limits on the production cross section times branching fraction, $\sigma_{ZH} \times {\ensuremath{\mathcal{B}({\ensuremath{{H}}}\to \text{inv})}\xspace}$ as a function of the Higgs boson mass [@Sirunyan:2017qfc]. []{data-label="fig:Hinv"}](Hinv_ana_hzinv_met_nj_from110to1000_logx1_logy0.pdf "fig:"){width="28.50000%"}
Lepton flavour violating decays {#Hlfv}
===============================
In the standard model (SM), lepton flavour violating (LFV) decays of the Higgs boson are forbidden. Such decays can occur naturally in models with more than one Higgs boson doublet [@Bjorken:1977vt]. Based on $\sqrt{s}=8{\mathrm{{TeV}}}$ and 19.7 ${\mathrm{fb}^{-1}}$ proton-proton collision data, CMS published the results of the search for a LFV decay of a Higgs boson with ${\ensuremath{m_{\ensuremath{{H}}}}\xspace}=125$ ${\mathrm{{GeV}}}$ in three channels, ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}$ [@Khachatryan:2015kon], ${\ensuremath{{H}}}\to {\ensuremath{{e}}}{\ensuremath{{\tau}}}$ and ${\ensuremath{{H}}}\to {\ensuremath{{e}}}{\ensuremath{{\mu}}}$ [@Khachatryan:2016rke]. The results from the ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}$ channel [@Khachatryan:2015kon] combined the ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}_{{\ensuremath{{e}}}}$ and ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\ensuremath{{\tau}}}_\mathrm{h}}\xspace}$ decays, where ${\ensuremath{{\tau}}}_{{\ensuremath{{e}}}}$ and ${\ensuremath{{\ensuremath{{\tau}}}_\mathrm{h}}\xspace}$ are tau leptons reconstructed in the electronic and hadronic decay channels, respectively. The results show an excess of data with respect to the SM background-only hypothesis at $M_{{\ensuremath{{H}}}} =125{\mathrm{{GeV}}}$ with a significance of $2.4$ standard deviations ($\sigma$). The collinear mass $M_\text{col}$, which provides an estimator of the reconstructed Hmass using the observed decay products, is shown in Figure \[fig:Hmt\_mcol\_all\_global\_weighted\] (Left). A constraint is set on the branching fraction $\mathcal{B}({\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}})<1.51\%$ at 95% confidence level (CL), while the best fit branching fraction is $\mathcal{B}({\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}})=(0.84^{+0.39}_{-0.37})\%$. Based on the 2016 data sample of proton-proton collisions at $\sqrt{s} = 13~\mathrm{TeV}$ corresponding to an integrated luminosity of $35.9~\mathrm{fb}^{-1}$, CMS updated the search for LFV decays of the Higgs boson with $M_{{\ensuremath{{H}}}}=125$ ${\mathrm{{GeV}}}$. The updated search was performed in four decay channels, ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}_{{\ensuremath{{e}}}}$, ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\ensuremath{{\tau}}}_\mathrm{h}}\xspace}$, ${\ensuremath{{H}}}\to {\ensuremath{{e}}}{\ensuremath{{\tau}}}_{{\ensuremath{{\mu}}}}$, ${\ensuremath{{H}}}\to {\ensuremath{{e}}}{\ensuremath{{\ensuremath{{\tau}}}_\mathrm{h}}\xspace}$, where ${\ensuremath{{\tau}}}_{{\ensuremath{{e}}}}$,${\ensuremath{{\tau}}}_{{\ensuremath{{\mu}}}}$ and ${\ensuremath{{\ensuremath{{\tau}}}_\mathrm{h}}\xspace}$ correspond to the electronic, muonic and hadronic decay channels of $\tau$ leptons, respectively [@CMS:2017onh]. No evidence is found for LFV Higgs boson decays. The observed (expected) limits on the branching fraction of the Higgs boson to $\mu\tau$ and to ${\ensuremath{{e}}}\tau$ are found to be less than 0.25(0.25)% and 0.61(0.37)%, respectively.
{width="51.50000%"} {width="45.00000%"}
The ATLAS Collaboration reported searches for ${\ensuremath{{H}}}\to {\ensuremath{{e}}}{\ensuremath{{\tau}}}$ and ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}$, finding no significant excess of events over the background expectation [@Aad:2016blu; @Aad:2015gha]. The best fit branching fractions of ${\ensuremath{{H}}}\to {\ensuremath{{e}}}{\ensuremath{{\tau}}}$ and ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}$ obtained from CMS Collaboration and ATLAS Collaboration are summarized in Figure \[fig:H\_LFV\_Summary\].
{width="48.00000%"} {width="48.00000%"}
Decay to light scalars or pseudo-scalars {#Haa}
========================================
There are many well-motivated models that predict the existence of decays of the Higgs boson to non-SM particles such as lighter scalars or pseudoscalars [@Dermisek:2006wr; @Chang:2008cw; @Curtin:2013fra]. Both 2HDM and NMSSM may contain a light enough pseudoscalar state (${\ensuremath{\mathrm{a}}\xspace}$), which can yield a large ${\ensuremath{{\mathrm{h}}}\xspace}\to{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace}$ branching fraction. Four searches for decays of the 125 ${\mathrm{{GeV}}}$ Higgs boson to pairs of lighter scalars or pseudoscalars are performed at CMS with $\sqrt{s}=8{\mathrm{{TeV}}}$ and 19.7 ${\mathrm{fb}^{-1}}$ proton-proton collision data [@Khachatryan:2015wka; @Khachatryan:2017mnf], ${\ensuremath{{{\ensuremath{{\mathrm{h}}}\xspace}}\to{{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace}}\to4{\ensuremath{{\tau}}}}\xspace}$, ${\ensuremath{{{\ensuremath{{\mathrm{h}}}\xspace}}\to{{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace}}\to2{\ensuremath{{\mu}}}2{{\ensuremath{b}\xspace}}}\xspace}$, ${\ensuremath{{{\ensuremath{{\mathrm{h}}}\xspace}}\to{{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace}}\to2{\ensuremath{{\mu}}}2{\ensuremath{{\tau}}}}\xspace}$ and ${\ensuremath{{{\ensuremath{{\mathrm{h}}}\xspace}}\to{{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace}}\to4{\ensuremath{{\mu}}}}\xspace}$, where the symbol ${\ensuremath{\mathrm{a}}\xspace}$ refers to both the light scalar and light pseudoscalar for notational simplicity. The data were found to be compatible with SM predictions. Upper limits are set on the product of the cross section and branching fraction for several signal processes. Searches for non standard decays of the SM-like Higgs boson to a pair of light pseudoscalar bosons are interpreted in the context of 2HDM+S. Because $\mathcal{B}({\ensuremath{\mathrm{a}}\xspace}\to{\ensuremath{{\tau}}}^{+}{\ensuremath{{\tau}}}^{-})$ is directly proportional to $\mathcal{B}({\ensuremath{\mathrm{a}}\xspace}\to{\ensuremath{{\mu}}}^+{\ensuremath{{\mu}}}^-)$ in any type of 2HDM+S and so is $\mathcal{B}({\ensuremath{\mathrm{a}}\xspace}\to {\ensuremath{b}\xspace}\overline{{\ensuremath{b}\xspace}})$ in type-1 and -2, the results of all analyses can be expressed as exclusion limits on $\frac{\sigma({\ensuremath{{\mathrm{h}}}\xspace})}{\sigma_{\textrm{SM}}} \, \mathcal{B}({\ensuremath{{\mathrm{h}}}\xspace}\to {\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace})\, \mathcal{B}^2({\ensuremath{\mathrm{a}}\xspace}\to{\ensuremath{{\mu}}}^+{\ensuremath{{\mu}}}^-)$, as illustrated in Fig. \[fig:compa\_modelindependent\]. The exact value of $\mathcal{B}({\ensuremath{\mathrm{a}}\xspace}\to{\ensuremath{{\mu}}}^+{\ensuremath{{\mu}}}^-)$ depends on the type of 2HDM+S, on $\tan\beta$ and on the pseudoscalar boson mass. No significant excess of events is observed for any of the five analyses.
With $\sqrt{s}=13{\mathrm{{TeV}}}$ and 2.8 ${\mathrm{fb}^{-1}}$ proton-proton collision data during the 2015 data taking campaign, CMS updated the search for $h \rightarrow 2a + X \rightarrow 4\mu +X$ where $X$ denotes possible additional particles from cascade decays of a Higgs boson [@CMS:2016tgd], motivated by NMSSM and Dark SUSY models [@ArkaniHamed:2008qn; @Baumgart:2009tn; @Falkowski:2010cm]. These data are used to search for new light bosons with a mass in the range $0.25-8.5~\mathrm{GeV}/c^2$ decaying into muon pairs. No excess is observed in the data, and a model-independent upper limit on the product of the cross section, branching fraction and acceptance is derived. The results are interpreted in the context of two benchmark models, namely, the next-to-minimal supersymmetric standard model, and dark SUSY models including those predicting a non-negligible light boson lifetime, with the details in [@CMS:2016tgd].
![Expected and observed 95% CL exclusion limits on $({\sigma_{{\ensuremath{{\mathrm{h}}}\xspace}}}/{\sigma_{\textrm{SM}}}) \, \mathcal{B}({\ensuremath{{\mathrm{h}}}\xspace}\to{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace})\, \mathcal{B}^2({\ensuremath{\mathrm{a}}\xspace}\to{\ensuremath{{\mu}}}^+{\ensuremath{{\mu}}}^-)$ for various exotic ${\ensuremath{{\mathrm{h}}}\xspace}$ boson decay searches performed with data collected at 8 ${\mathrm{{TeV}}}$ with the CMS detector, assuming that the branching fractions of the pseudoscalar boson to muons, ${\ensuremath{{\tau}}}$ leptons and [$b$]{}quarks follow the assumption in [@Khachatryan:2017mnf], which implies that the limit shown for ${\ensuremath{{{\ensuremath{{\mathrm{h}}}\xspace}}\to{{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace}}\to2{\ensuremath{{\mu}}}2{{\ensuremath{b}\xspace}}}\xspace}$ is valid only in type-1 and -2 2HDM+S [@CMS:2016tgd]. []{data-label="fig:compa_modelindependent"}](Haa_Figure_007.pdf){width="50.00000%"}
Summary {#sum}
=======
Searches for rare and exotic decays of the 125 GeV Higgs boson performed with data collected with the CMS experiment have been presented. The analyzed data correspond to the full LHC run-1 dataset collected during 2011 and 2012 at $\sqrt{s}=7-8~{\mathrm{{TeV}}}$ and to the run-2 data at $\sqrt{s}=13{\mathrm{{TeV}}}$ collected during 2015 and 2016. Rare Higgs decays are extremely sensitive to new physics if additional Higgs couplings exist. Many rare decays have not been observed yet, but may become observable in the next few years. Exotic Higgs decays would bring direct evidence of such new physics. Tight limits on ${\ensuremath{{H}}}\to {\ensuremath{{e}}}{\ensuremath{{\tau}}}$ and ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}$ have been set by CMS using data collected in 2016 and the $2.4\sigma$ excess observed in run-1 with ${\ensuremath{{H}}}\to {\ensuremath{{\mu}}}{\ensuremath{{\tau}}}$ decay has been ruled out. No hint for new physics has been found when looking for Higgs invisible decays and ${\ensuremath{{\mathrm{h}}}\xspace}\to{\ensuremath{\mathrm{a}}\xspace}{\ensuremath{\mathrm{a}}\xspace}$ searches. The results are interpreted according to beyond-SM (BSM) Higgs scenarios, which include the Two Higgs Doublet Model (2HDM) and the Singlet Model. Stringent limits have been set on the existence of such processes.
I would like to thank the ICFNP2017 organizers for their hospitality and the wonderful working environment. I acknowledge the support from National Natural Science Foundation of China (No. 11505208 and No. 1161101027), China Ministry of Science and Technology (No. 2013CB838700).
G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Lett. B [**716**]{} (2012) 1 S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**716**]{} (2012) 30 S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], JHEP [**1306**]{} (2013) 081 F. Englert and R. Brout, Phys. Rev. Lett. [**13**]{} (1964) 321. P. W. Higgs, Phys. Rev. Lett. [**13**]{} (1964) 508. G. S. Guralnik, C. R. Hagen and T. W. B. Kibble, Phys. Rev. Lett. [**13**]{} (1964) 585. G. Aad [*et al.*]{} \[ATLAS and CMS Collaborations\], JHEP [**1608**]{} (2016) 045
R. N. Cahn, M. S. Chanowitz and N. Fleishon, Phys. Lett. [**82B**]{} (1979) 113. L. Bergstrom and G. Hulth, Nucl. Phys. B [**259**]{} (1985) 137 Erratum: \[Nucl. Phys. B [**276**]{} (1986) 744\]. S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**726**]{} (2013) 587 V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**753**]{} (2016) 341 A. Denner, S. Heinemeyer, I. Puljak, D. Rebuzzi and M. Spira, Eur. Phys. J. C [**71**]{} (2011) 1753
T. Plehn and D. L. Rainwater, Phys. Lett. B [**520**]{} (2001) 108 T. Han and B. McElrath, Phys. Lett. B [**528**]{} (2002) 81 N. Vignaroli, Phys. Rev. D [**80**]{} (2009) 095023
A. Dery, A. Efrati, Y. Hochberg and Y. Nir, JHEP [**1305**]{} (2013) 039
V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**744**]{} (2015) 184
G. Belanger, F. Boudjema, A. Cottrant, R. M. Godbole and A. Semenov, Phys. Lett. B [**519**]{} (2001) 93 G. F. Giudice, R. Rattazzi and J. D. Wells, Nucl. Phys. B [**595**]{} (2001) 250 D. Dominici and J. F. Gunion, Phys. Rev. D [**80**]{} (2009) 115006 R. E. Shrock and M. Suzuki, Phys. Lett. [**110B**]{} (1982) 250. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], JHEP [**1702**]{} (2017) 135 CMS Collaboration, CMS-PAS-EXO-16-048, http://cds.cern.ch/record/2264698.
A. M. Sirunyan [*et al.*]{} \[CMS Collaboration\], arXiv:1711.00431 \[hep-ex\].
J. D. Bjorken and S. Weinberg, Phys. Rev. Lett. [**38**]{} (1977) 622. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**749**]{} (2015) 337 V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**763**]{} (2016) 472 CMS Collaboration, CMS-PAS-HIG-17-001, http://cds.cern.ch/record/2264540. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Eur. Phys. J. C [**77**]{} (2017) no.2, 70 G. Aad [*et al.*]{} \[ATLAS Collaboration\], JHEP [**1511**]{} (2015) 211 R. Dermisek and J. F. Gunion, Phys. Rev. D [**75**]{} (2007) 075019 S. Chang, R. Dermisek, J. F. Gunion and N. Weiner, Ann. Rev. Nucl. Part. Sci. [**58**]{} (2008) 75 D. Curtin [*et al.*]{}, Phys. Rev. D [**90**]{} (2014) no.7, 075004 V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**752**]{} (2016) 146 V. Khachatryan [*et al.*]{} \[CMS Collaboration\], arXiv:1701.02032 \[hep-ex\]. CMS Collaboration, CMS-PAS-HIG-16-035, http://cds.cern.ch/record/2232052.
N. Arkani-Hamed, D. P. Finkbeiner, T. R. Slatyer and N. Weiner, Phys. Rev. D [**79**]{} (2009) 015014 M. Baumgart, C. Cheung, J. T. Ruderman, L. T. Wang and I. Yavin, JHEP [**0904**]{} (2009) 014 A. Falkowski, J. T. Ruderman, T. Volansky and J. Zupan, JHEP [**1005**]{} (2010) 077
[^1]:
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study a class of quantum measurements that furnish probabilistic representations of finite-dimensional quantum theory. The Gram matrices associated with these Minimal Informationally Complete quantum measurements (MICs) exhibit a rich structure. They are “positive” matrices in three different senses, and conditions expressed in terms of them have shown that the Symmetric Informationally Complete measurements (SICs) are in some ways optimal among MICs. Here, we explore MICs more widely than before, comparing and contrasting SICs with other classes of MICs, and using Gram matrices to begin the process of mapping the territory of all MICs. Moreover, the Gram matrices of MICs turn out to be key tools for relating the probabilistic representations of quantum theory furnished by MICs to quasi-probabilistic representations, like Wigner functions, which have proven relevant for quantum computation. Finally, we pose a number of conjectures, leaving them open for future work.'
author:
- 'John B. DeBrota'
- 'Christopher A. Fuchs'
- 'Blake C. Stacey'
bibliography:
- 'mic\_facts.bib'
title: |
Triply Positive Matrices and Quantum Measurements\
Motivated by QBism
---
Introduction {#sec:intro}
============
Let $\mathcal{H}_d$ be a $d$-dimensional complex Hilbert space, and let $\{E_i\}$ be a set of positive semidefinite operators on that space which sum to the identity: $$\sum_{i=1}^N E_i = I.$$ The set $\{E_i\}$ is a *positive-operator-valued measure* (POVM), which is the mathematical representation of a measurement process in quantum theory. Each element in the set — called an *effect* — stands for a possible outcome of the measurement [@Nielsen:2010 § 2.2.6]. A POVM is said to be *informationally complete* (IC) if the operators $\{E_i\}$ span the space of Hermitian operators on $\mathcal{H}_d$, and an IC POVM is said to be *minimal* if it contains exactly $d^2$ elements. For brevity, we can call a minimal informationally complete POVM a MIC.
Consider the Gram matrix of MIC elements, that is, the matrix $G$ whose entries are given by $$[G]_{ij} := \tr (E_i E_j).$$ A matrix $G$ of this type is “positive” in three senses of the word. First, each element is nonnegative. Second, $G$ is positive definite, since it is the Gram matrix of a basis on the operator space. Third, it is constructed from inner products of objects that are themselves positive semidefinite.
Of particular note among MICs are those which enjoy the symmetry property $$[G]_{ij} = [G_{\rm SIC}]_{ij} := \frac{1}{d^2} \frac{d\delta_{ij} + 1}{d+1}.$$ These are known as *symmetric* informationally complete POVMs, or SICs for short [@Zauner:1999; @Renes:2004; @Scott:2010a; @Fuchs:2017a]. In addition to their purely mathematical properties, SICs are of central interest to the technical side of QBism, a research program in the foundations of quantum mechanics [@Fuchs:2014b; @Fuchs:2013; @Healey:2016; @Fuchs:2016a]. Investigations motivated by foundational concerns led to the discovery that SICs are in many ways optimal among MICs [@Appleby:2014b; @Appleby:2015; @DeBrota:2018]. In this paper, we elaborate upon some of those results and explore the conceptual context of MICs more broadly.
In quantum physics, the *Born Rule* is a key step in the calculation of probabilities. The common way of presenting the Born Rule suggests that it fixes probabilities in terms of more fundamental quantities, namely quantum states and measurement operators. QBism, however, has promoted a change of viewpoint. From this new perspective, the Born Rule should be thought of as a *consistency condition between the probabilities assigned in diverse scenarios* — for instance, probabilities assigned to the outcomes of complementary experiments. The bare axioms of probability theory do not themselves impose relations between probabilities given different conditionals: In the abstract, nothing ties together $P(E|C_1)$ and $P(E|C_2)$. Classical intuition suggests one way to fit together probability assignments for different experiments, and quantum physics implies another. The discrepancy between these standards encapsulates how quantum theory departs from classical expectations [@Fuchs:2017b]. By studying the full lay of the land of MICs, we hope to find a probabilistic representation of the Born Rule which picks out the cleanest statement of the quantum-classical divide in all its nuance.
In the next section, we introduce the fundamentals of quantum information theory and the necessary concepts from linear algebra to prove basic results that apply to all MICs. We also deduce a condition in terms of matrix rank for when a set of vectors in $\mathbb{C}^d$ can be fashioned into a MIC. Then, in Section \[sec:constructions\], we show how to construct several classes of MICs explicitly, and we start the process of exploring the triply-positive matrices associated with them. We carry on with this task in Section \[sec:optimal\], showing that in multiple ways, the SICs are optimal among MICs. To put this in a wider context, in Section \[sec:numerics\] we investigate the triply-positive matrices of randomly-chosen MICs. The empirical eigenvalue distributions we find have intriguing features, not all of which have been explained yet. In Section \[sec:qreps\], we relate the probabilistic representations of quantum theory furnished by MICs to *quasi*-probabilistic representations, such as Wigner functions, potentially relevant to the theory of quantum computation [@Veitch:2014]. Thanks to this relation, we expect that the close study of MICs will be beneficial for understanding which resources are necessary to give quantum computation its go.
Basic Properties of MICs {#sec:basics}
========================
We begin by briefly establishing the necessary notions from quantum information theory on which this paper is grounded. In quantum physics, each physical system is associated with a complex Hilbert space. Often, in quantum information theory, the Hilbert space of interest is taken to be finite-dimensional. We will denote the dimension throughout by $d$. A *quantum state* is a positive semidefinite operator of unit trace. If an experimenter ascribes the quantum state $\rho$ to a system, then she finds her probability for the $i^{\rm th}$ outcome of the measurement modeled by the POVM $\{E_i\}$ via the Hilbert–Schmidt inner product: $$p(E_i) = \tr(\rho E_i).$$ This formula is a standard presentation of the Born Rule. The condition that the $\{E_i\}$ sum to the identity ensures that the resulting probabilities are properly normalized.
If the operators $\{E_i\}$ span the space of positive semidefinite operators, then the operator $\rho$ can be reconstructed from its inner products with them. In other words, the state $\rho$ can be calculated from the probabilities $\{p(E_i)\}$, meaning that the measurement is “informationally complete” and the state $\rho$ can, in principle, be dispensed with. Any MIC can thus be considered a “Bureau of Standards” measurement, that is, a reference measurement in terms of which all states and processes can be understood [@Fuchs:2002]. Writing a quantum state $\rho$ is often thought of as specifying the “preparation” of a system, though this terminology is overly restrictive, and the theory applies just as well to physical systems that were not processed on a laboratory workbench [@Fuchs:2011c].
The extreme points in the space of quantum states are the rank-1 projection operators: $$\rho = \ketbra{\psi}{\psi}.$$ These are idempotent operators; that is, they all satisfy $\rho^2 =
\rho$.
Given any MIC $\{E_i\}$, we can always write its elements as unit-trace positive semidefinite operators with appropriate scaling factors: $$E_i := e_i \rho_i, \hbox{ where } e_i = \tr E_i.$$ If the operators $\rho_i$ are all rank-1 projectors, we will refer to the set $\{E_i\}$ as a *rank-1 MIC.* We will call a MIC *equal-weight* when the coefficients $e_i$ are all equal. The condition that the elements sum to the identity then fixes $e_i =
1/d$. Equal-weight MICs are of physical interest, since they represent quantum measurements that have no intrinsic bias: They map the “garbage state” $(1/d)I$ to the flat probability distribution over $d^2$ outcomes.
\[dblystoch\] If a MIC $\{E_i\}$ is equal-weight, then its $G$ matrix is proportional to a *doubly stochastic* matrix.
$G$ is always symmetric, because the trace is cyclic. We can therefore sum over either index: $$\sum_j [G]_{ij}
= \sum_j \tr(E_i E_j) = \tr\left( E_i \sum_j E_j \right)
= \tr E_i
= e_i
= \frac{1}{d}.$$ The sum over any row or column of the matrix $dG$ is therefore 1.
In order to establish some basic properties that hold for all MICs, we first recall a result of linear algebra.
Let $A$ and $B$ be positive semidefinite operators. If $\text{\rm tr}AB=0$ then $AB=0$. \[psdortho\]
Write $B=\sum_i b_i \ketbra{i}{i}$. Then $0=\text{tr}AB=\sum_i b_i \bra{i}A\ket{i}$. Since $A$ and $B$ are positive semidefinite, $b_i$ and $\bra{i}A\ket{i}$ are nonnegative, so each term in the sum must equal zero. Thus, for each $i$, either $b_i=0$ or $\bra{i}A\ket{i}=0$. Thus $A$ and $B$ have orthogonal supports and $AB=0$.
Next, we recall the concept of a *dual basis.* Given a basis for a vector space, the dual of that basis is a set such that the inner products of a vector with the elements of the dual basis provide the coefficients in the expansion of that vector in terms of the original basis. In the familiar case when the original basis is orthonormal, the dual basis coincides with it: When we write a vector $\vec{v}$ as an expansion over the unit vectors $(\hat{x}, \hat{y}, \hat{z})$, the coefficient of $\hat{x}$ is simply the inner product of $\hat{x}$ with $\vec{v}$. One consequence of this definition is that if we expand the original basis in terms of itself, $$E_i = \sum_j (\tr E_i \widetilde{E}_j) E_j,$$ linear independence of the $\{E_i\}$ implies that $$\tr E_i \widetilde{E}_j = \delta_{ij}.$$
With the following convention, one can “vectorize” an operator: $$\kket{A}:=\sum_i (A\otimes I) \ket{i}\ket{i}\;.$$ The vectorized operator inner product is equal to the standard Hilbert–Schmidt inner product, so $$[G]_{ij}=\tr E_iE_j=\bbraket{E_i}{E_j}.$$ A quantity of central importance is the *frame operator* $$\label{frameop}
\mathcal{F}:=\sum_i\kketbra{E_i}{E_i}.$$ The frame operator is fundamental because it allows us to move between a basis and its dual, $$\kket{E_i}=\mathcal{F}\kket{\widetilde{E}_i}.$$ From this and Eq. (\[frameop\]), we see $$\sum_i\kketbra{\widetilde{E}_i}{E_i}=\sum_i\kketbra{E_i}{\widetilde{E}_i}=I_{d^2}.$$ Acting on one side by $\kket{E_k}$ and the other by $\bbra{\widetilde{E}_l}$ demonstrates that $$[G^{-1}]_{ij}=\tr\widetilde{E}_i\widetilde{E}_j=\bbraket{\widetilde{E}_i}{\widetilde{E}_j}\;.$$ It is also easy to check that $$\mathcal{F}^{-1}=\sum_i\kketbra{\widetilde{E}_i}{\widetilde{E}_i}\;.$$
\[isospec\] The frame operator and the Gram matrix of a MIC have the same spectrum.
To see this, form a projector out of the state $\sum_i\kket{E_i}\ket{i}$ where $\ket{i}$ is an orthonormal basis in $\mathcal{H}_{d^2}$ and perform partial traces over each subsystem. The results are $G^\text{T}$ and $\mathcal{F}$, and so, by the Schmidt theorem, the eigenvalue spectra of $\mathcal{F}$ and $G$ are equal.
We can now proceed to prove
\[noprops\] No element in a MIC can be proportional to an element of the MIC’s dual basis.
Writing the MIC as $\{E_i\}$ and the dual basis as $\{\widetilde{E}_i\}$, we can without loss of generality pick $E_1$ to be the element proportial to its dual, $\widetilde{E}_1$ (noting that $E_1$ will necessarily be orthogonal to all the other dual elements). So, assume that $E_1=\alpha \widetilde{E}_1$. A general operator $A\in\mathcal{L}(\mathcal{H}_d)$ may be written as a linear combination of dual basis elements, $$A=\sum_{k=1}^{d^2}(\text{tr}AE_k)\widetilde{E}_k,$$ so we may write the element $E_1$ as $$E_1=\sum_{k=1}^{d^2}(\text{tr}E_1E_k)\widetilde{E}_k.$$ Since $\text{tr}E_1^2=\alpha\text{tr}E_1\widetilde{E}_1=\alpha$, we have $(\text{tr}E_1^2)\widetilde{E}_1=\alpha(1/\alpha)E_1=E_1$, and it follows that $$E_1=E_1+\sum_{k\geq 2}^{d^2}(\text{tr}E_1E_k)\widetilde{E}_k\implies 0 = \sum_{k\geq 2}^{d^2}(\text{tr}E_1E_k)\widetilde{E}_k.$$ As the $\widetilde{E}_k$ are linearly independent, we must have $\text{tr}E_1E_k=0$ for all $k\neq 1$. By Lemma \[psdortho\], this implies that $E_1E_k=0$ for all $k\neq 1$. This implies that the $d^2-1$ remaining $E_k$ are operators on a $d-\text{rank}(E_1)$ dimensional subspace. But $$\text{dim}\left[\mathcal{L}\left(\mathcal{H}_{d-\text{rank}(E_1)}\right)\right]\leq(d-1)^2<d^2-1,$$ so they cannot be linearly independent.
Theorem \[noprops\] has physical meaning. In classical probability theory, we grow accustomed to orthonormal bases. For example, imagine an object that can be in any one of $N$ distinct configurations. When we write a probability distribution over these $N$ alternatives, we are encoding our expectations about which of these configurations is physically present — about the “physical condition” of the object, as Einstein would say [@Stacey:2018], or in more modern terminology, about the object’s “ontic state” [@Spekkens:2007]. We can learn everything there is to know about the object by measuring its “physical condition”, and any implementation of such an ideal measurement is represented by conditional probabilities that are 1 in a single entry and 0 elsewhere. In other words, the map from the object’s physical configuration to the reading on the measurement device is, at its most complicated, a permutation of labels. Without loss of generality, we can take the vectors that define the ideal measurement to be the vertices of the probability simplex: The measurement basis is identical with its dual, and the dual-basis elements simply label the possible “physical conditions” of the object which the measurement reads off.
In the quantum theory, by contrast, we cannot construct a MIC that has an element which is even proportional to an element in the dual. This stymies the identification of the dual-basis elements as intrinsic “physical conditions” ready for a measurement to read.
\[noprojs\] No effect of a MIC can be an unscaled projector.
Suppose $E_1=P$ is a projection operator. Then $${I}=P+\sum_{i\geq 2}^{d^2}E_i.$$ Multiplying on the left by $P$, $$P=P+\sum_{i\geq 2}^{d^2}PE_i.$$ Canceling $P$ and taking the trace implies that $$\sum_{i\geq 2}\text{tr}PE_i=0.$$ Since every element of this sum must be nonnegative, we have for all $i\geq2$, $$\text{tr}PE_i=0.$$ By Lemma \[psdortho\], the remaining $d^2-1$ POVM elements are operators on a $d-\text{rank}(P)$ dimensional subspace. But as before, $$\text{dim}\left[\mathcal{L}\left(\mathcal{H}_{d-\text{rank}(P)}\right)\right]\leq(d-1)^2<d^2-1,$$ so they cannot be linearly independent.
\[nopovm\] No elementwise rescaling of a proper subset of a MIC may form a POVM.
Since a MIC is a linearly independent set, the identity element is uniquely formed by the defining expression $${I}=\sum_{i=1}^{d^2}E_i.
\label{povmdef}$$ If a linear combination of a proper subset $\mathcal{S}$ of the MIC elements could be made to also sum to the identity, $${I}=\sum_{i\in \mathcal{S}}\alpha_i E_i,
\label{lincomb}$$ then subtracting from implies $$0=\sum_{i\in\mathcal{S}}(1-\alpha_i)E_i+\sum_{i\notin\mathcal{S}}E_i$$ which is a violation of linear independence.
No two elements in a $d=2$ MIC may be orthogonal under the Hilbert–Schmidt inner product.
An orthogonal pair of elements in dimension $2$ may be rescaled such that they sum to the identity element. Therefore, by Theorem \[nopovm\], they cannot be elements of a MIC.
As with Theorem \[noprops\], these results have physics implications. For much of the history of quantum mechanics, one type of POVM had special status: the *von Neumann measurements,* which consist of $d$ elements given by the projectors onto the vectors of an orthonormal basis of $\mathbb{C}^d$. Indeed, in older books, these are the only quantum measurements that are considered (often being defined as the eigenbases of Hermitian operators called “observables”). We can now see that, from the standpoint of informational completeness, the von Neumann measurements are rather pathological: There is no way to build a MIC by augmenting a von Neumann measurement with additional outcomes. Later, we will see how to correct for the way that the von Neumann POVMs fall short of ideal (§\[sec:constructions\]).
These results prompt a question: May any two elements of a MIC in arbitrary dimension be orthogonal? In other words, can any entry in a $G$ matrix equal zero? We answer this question in the affirmative with an explicit example of a rank-$1$ MIC in dimension $3$ with $7$ orthogonal pairs.
When multiplied by $1/3$, the following set of rank-$1$ projectors form a MIC in dimension $3$ with $7$ orthogonal pairs.
$$\begin{split}
&\left\{\begin{bmatrix}
1 & 0 & 0 \\
0&0&0\\
0&0&0
\end{bmatrix},
\begin{bmatrix}
0 & 0 & 0 \\
0&1&0\\
0&0&0
\end{bmatrix},
\begin{bmatrix}
\frac{1}{2} & 0 & \frac{1}{2} \\
0&0&0\\
\frac{1}{2}&0&\frac{1}{2}
\end{bmatrix},
\begin{bmatrix}
0 & 0 & 0 \\
0&\frac{1}{2}&\frac{1}{2}\\
0&\frac{1}{2}&\frac{1}{2}
\end{bmatrix},
\begin{bmatrix}
\frac{1}{2} & 0 & \frac{i}{2} \\
0&0&0\\
-\frac{i}{2}&0&\frac{1}{2}
\end{bmatrix},
\begin{bmatrix}
0 & 0 & 0 \\
0&\frac{1}{2}&\frac{i}{2}\\
0&-\frac{i}{2}&\frac{1}{2}
\end{bmatrix},
\begin{bmatrix}
\frac{1}{3} & \frac{i}{3} & -\frac{i}{3} \\
-\frac{i}{3}&\frac{1}{3}&-\frac{1}{3}\\
\frac{i}{3}&-\frac{1}{3}&\frac{1}{3}
\end{bmatrix},\right.\\
&\left. \begin{bmatrix}
\frac{5}{8} & -\frac{1}{8}-\frac{i}{4} & -\frac{3}{8}-\frac{i}{8} \\
-\frac{1}{8}+\frac{i}{4}&\frac{1}{8}&\frac{1}{8}-\frac{i}{8}\\
-\frac{3}{8}+\frac{i}{8}&\frac{1}{8}+\frac{i}{8}&\frac{1}{4}
\end{bmatrix},
\begin{bmatrix}
\frac{1}{24} & \frac{1}{8}-\frac{i}{12} & -\frac{1}{8}-\frac{i}{24} \\
\frac{1}{8}+\frac{i}{12}&\frac{13}{24}&-\frac{7}{24}-\frac{3i}{8}\\
-\frac{1}{8}+\frac{i}{24}&-\frac{7}{24}+\frac{3i}{8}&\frac{5}{12}
\end{bmatrix}\right\}.
\end{split}$$
These are projectors onto the following vectors in $\mathcal{H}_d$: $$\begin{split}
&\left\{(1,0,0),(0,1,0),\frac{1}{\sqrt{2}}(1,0,1),\frac{1}{2}(0,1,1),\frac{1}{\sqrt{2}}(1,0,-i),\frac{1}{\sqrt{2}}(0,1,-i),\frac{1}{\sqrt{3}}(1,-i,i),\right.\\
&\left.\frac{1}{\sqrt{40}}(5,-1+2i,-3+i),\frac{1}{\sqrt{24}}(1,3+2i,-3+i)\right\}.
\end{split}$$ The Hilbert–Schmidt Gram matrix of the MIC elements is
$$\begin{bmatrix}
\frac{1}{9}&0&\frac{1}{18}&0&\frac{1}{18}&0&\frac{1}{27}&\frac{5}{72}&\frac{1}{216}\\
0&\frac{1}{9}&0&\frac{1}{18}&0&\frac{1}{18}&\frac{1}{27}&\frac{1}{72}&\frac{13}{216}\\
\frac{1}{18}&0&\frac{1}{9}&\frac{1}{36}&\frac{1}{18}&\frac{1}{36}&\frac{1}{27}&\frac{1}{144}&\frac{5}{432}\\
0&\frac{1}{18}&\frac{1}{36}&\frac{1}{9}&\frac{1}{36}&\frac{1}{18}&0&\frac{5}{144}&\frac{1}{48}\\
\frac{1}{18}&0&\frac{1}{18}&\frac{1}{36}&\frac{1}{9}&\frac{1}{36}&0&\frac{5}{144}&\frac{1}{48}\\
0&\frac{1}{18}&\frac{1}{36}&\frac{1}{18}&\frac{1}{36}&\frac{1}{9}&\frac{1}{27}&\frac{1}{144}&\frac{5}{432}\\
\frac{1}{27}&\frac{1}{27}&\frac{1}{27}&0&0&\frac{1}{27}&\frac{1}{9}&\frac{1}{54}&\frac{1}{18}\\
\frac{5}{72}&\frac{1}{72}&\frac{1}{144}&\frac{5}{144}&\frac{5}{144}&\frac{1}{144}&\frac{1}{54}&\frac{1}{9}&\frac{1}{27}\\
\frac{1}{216}&\frac{13}{216}&\frac{5}{432}&\frac{1}{48}&\frac{1}{48}&\frac{5}{432}&\frac{1}{18}&\frac{1}{27}&\frac{1}{9}\\
\end{bmatrix}.
\label{gram}$$
The process of finding this example led us to formulate the following:
A rank-1 MIC in dimension 3 can have no more than 7 pairs of orthogonal elements.
Our next result characterizes when it is possible to build a rank-1 POVM out of a set of vectors. We will then see what additional conditions must be met in order to obtain a POVM that is a rank-1 MIC.
Consider a set of $N$ normalized vectors $\ket{\phi_i}$ in $\mathcal{H}_d$ and weights $0\leq e_i\leq 1$. Then $E_i=e_i\ketbra{\phi_i}{\phi_i}$ forms a rank-$1$ POVM iff the Gram matrix $[g]_{ij}=\sqrt{e_ie_j}\braket{\phi_i}{\phi_j}$ is a rank-$d$ projector.
Suppose $g$ is a rank-$d$ projector. This implies $N\geq d$ and $$\text{tr}g=d=\sum_{i=1}^N e_i.$$ Then the state $$\label{combined}
\sum_i\sqrt{\frac{e_i}{d}}\ket{\phi_i}\ket{i},$$ where $\ket{i}$ is any orthonormal basis in $\mathcal{H}_N$, is a normalized state in $\mathcal{H}_d\otimes\mathcal{H}_N$. Form the projector $$\sum_{i,j=1}^N\frac{\sqrt{e_ie_j}}{d}\left(\ketbra{\phi_i}{\phi_j}\otimes\ketbra{i}{j}\right).$$ If we trace over the first subsystem we obtain $$\sum_{i,j=1}^N\frac{\sqrt{e_ie_j}}{d}\braket{\phi_j}{\phi_i}\ketbra{i}{j}=\frac{1}{d}g^\text{T}.$$ If we trace over the second subsystem we obtain $$\sum_{i=1}^N\frac{e_i}{d}\ketbra{\phi_i}{\phi_i}.$$ By the Schmidt theorem, these partial traces must have the same nonzero spectrum. This implies that $$\sum_{i=1}^N e_i\ketbra{\phi_i}{\phi_i}$$ is a matrix with $d$ eigenvalues equal to $1$. As it is an operator on $\mathcal{H}_d$, it must be the identity. Thus the $E_i$ form a rank-$1$ POVM.
Conversely, suppose the $E_i$ form a rank-$1$ POVM. This implies $N\geq d$. The trace of the expression $$\sum_{i=1}^Ne_i\ketbra{\phi_i}{\phi_i}=I$$ reveals that $\sum_i e_i=d$. Following the same argument as before with the state , we see that $$\sum_{i,j=1}^N\frac{\sqrt{e_ie_j}}{d}\braket{\phi_j}{\phi_i}\ketbra{i}{j}=\frac{1}{d}g^\text{T}.$$ and $$\sum_{i=1}^N\frac{e_i}{d}\ketbra{\phi_i}{\phi_i}=\frac{1}{d}I.$$ must have the same nonzero spectrum. Thus, there are $d$ nonzero eigenvalues of $g^\text{T}$, all equal to $1$. This must also be true of $g$, so $g$ is a rank-$d$ projector.
What further distinguishes a rank-$1$ MIC? In addition to $g$ being a rank-$d$ projector, we need $N=d^2$ and $G$ full-rank, that is, we need the elements of the POVM to be linearly independent and span operator space. What is the relationship between $g$ and $G$? There may be several of interest, but certainly one worth exploring is via elementwise multiplication, sometimes called the Hadamard product [@Horn:1994]: $$\label{hadamard}
g \circ g^*=G.$$ For any two matrices $A$ and $B$, the Hadamard product satisfies the rank inequality $$\label{rank}
\text{rank}(A\circ B) \leq \text{rank}(A)\text{rank}(B).$$ If $A$ and $B$ are positive semidefinite then their Hadamard product is also positive semidefinite (this is the Schur product theorem). In this case we also have $$\text{det}(A\circ B)\geq \text{det}(A)\text{det}(B).$$ If $N\leq d^2$, $g$ and $g^*$ saturate the inequality iff they are formed from a rank-$1$ MIC. This is not the case if $N>d^2$; if the POVM contains a set which spans operator space, $G$ will be rank $d^2$. In the end, the rank inequality is another way of expressing what differentiates informationally complete POVMs from non-informationally complete POVMs, and then what further distinguishes minimal informationally complete POVMs from those with more than $d^2$ entries.
Equation is a special Hadamard product because it results in an elementwise absolute value squared operation. This raises the possibility that optimality among MICs might relate to interesting properties of the Hadamard absolute value [@Horn:1994].
Explicit Constructions of MICs {#sec:constructions}
==============================
The MICs that have attracted the most interest are the SICs, which in many ways are the optimal MICs [@Appleby:2014b; @Appleby:2015; @DeBrota:2018; @Fuchs:2003; @Scott:2006]. SICs were studied as mathematical objects (under the name “complex equiangular lines”) before their importance for quantum information was recognized [@Delsarte:1975; @Hoggar:1981; @Coxeter:1991; @Hoggar:1998]. Prior to SICs becoming a physics problem, constructions were known for dimensions $d = 2$, 3 and 8. Exact solutions for SICs are now known in the following dimensions: $$d = 2\hbox{--}21, 23, 24, 28, 30, 31, 35, 37, 39, 43, 48, 53,
120, 124, 195, 323.$$ The expressions for these solutions grow complicated quickly, but there is hope that they can be substantially simplified [@Appleby:2018]. Numerical solutions have been extracted, to high precision, in the following dimensions: $$d = 2\hbox{--}151, 168, 172, 199, 224, 228, 255, 259, 288, 292,
323, 327, 489, 528, 725, 844, 1155, 2208.$$ Both the numerical and the exact solutions have been found in irregular order and by various methods. Particular credit should be given to Andrew Scott for solo work [@Scott:2017], for collaborations with Markus Grassl [@Scott:2010a; @Grassl:2017] and for contributing code used by other researchers [@Fuchs:2017a].
Together, these results have created the community sentiment that SICs *should* exist for every finite value of $d$. To date, however, a general proof is lacking. The current frontier of SIC research extends into algebraic number theory [@Appleby:2013; @Appleby:2016; @Bengtsson:2016; @Appleby:2017b; @Kopp:2018], which among other things has led to a method for uplifting numerical solutions to exact ones [@Appleby:2017]. The topic has begun to enter the textbooks for physicists [@Bengtsson:2017] and for mathematicians [@Waldron:2018].
SICs can be considered equivalently as sets of effects, of rank-1 projectors or of vectors: $$E_i = \frac{1}{d}\Pi_i, \hbox{ where } \Pi_i = \ketbra{\pi_i}{\pi_i}.$$ It is difficult to find a meaningful visualization of structures in high-dimensional complex vector space. However, for the $d = 2$ case, an image is available. Any quantum state for a 2-dimensional system can be written as an expansion over the Pauli matrices: $$\rho = \frac{1}{2}\left(I + x\sigma_x + y\sigma_y +
z\sigma_z\right).$$ The coefficients $(x,y,z)$ are then the coordinates for $\rho$ in the *Bloch ball*. The surface of this ball, the *Bloch sphere,* lives at radius 1 and is the set of pure states. In this picture, the quantum states $\{\Pi_i\}$ comprising a SIC form a regular tetrahedron; for example, $$\Pi_{s,s'} = \frac{1}{2}\left(I + \frac{1}{\sqrt{3}}\left(s\sigma_x + s'\sigma_y + ss'\sigma_z\right)\right),$$ where $s$ and $s'$ take the values $\pm 1$.
The matrix $G_{\rm SIC}$ has the spectrum $$\lambda(G_{\rm SIC}) = \left(
\frac{1}{d}, \frac{1}{d(d+1)}, \ldots, \frac{1}{d(d+1)}
\right).$$ The flatness of this spectrum will turn out to be significant; we will investigate this point in depth in the next section.
The matrix $G_{\rm SIC}$ is real, symmetric and positive definite. Its inverse $G_{\rm SIC}^{-1}$ is also real, symmetric and positive definite, and all its off-diagonal entries are nonpositive. Thus, $G_{\rm SIC}^{-1}$ is a Stieltjes matrix [@Horn:1994]. It also belongs to the (unilluminatingly named) class of $L$-matrices, since while the off-diagonal entries are nonpositive, the diagonal entries are all positive [@Young:1971]. The question of how widely these properties hold over the set of all MICs is largely unexplored.
It is possible to construct a MIC for any arbitrary dimension $d$. One way to do so was found by Caves, Fuchs and Schack in the course of proving a quantum version of the de Finetti theorem [@Caves:2002c]. (For background on this theorem, a key result in probability theory, see [@Stacey:2015 §5.3] and [@Diaconis:2017].) We will refer to these as the *orthocross MICs.* To construct an orthocross MIC in dimension $d$, first pick an orthonormal basis $\{\ket{j}\}$. This is a set of $d$ objects, and we want a set of $d^2$, so our first step is to take all possible combinations: $$\Gamma_{jk} := \ketbra{j}{k}.$$ The orthocross MIC will be built from a set of $d^2$ rank-1 projectors $\{\Pi_\alpha\}$, the first $d$ of which are given by $$\Pi_\alpha = \Gamma_{\alpha\alpha}.$$ Then, for $\alpha = d+1, \ldots, \frac{1}{2}d(d+1)$, we take all the quantities of the form $$\frac{1}{2}\left(\ket{j} + \ket{k}\right)
\left(\bra{j} + \bra{k}\right)
= \frac{1}{2} (\Gamma_{jj} + \Gamma_{kk} + \Gamma_{jk} + \Gamma_{kj}),$$ where $j < k$. We construct the rest of the $\{\Pi_\alpha\}$ similarly, by taking all quantities of the form $$\frac{1}{2}\left(\ket{j} + i\ket{k}\right)
\left(\bra{j} - i\bra{k}\right)
= \frac{1}{2} (\Gamma_{jj} + \Gamma_{kk} - i\Gamma_{jk} + i\Gamma_{kj}),$$ where again the indices satisfy $j < k$. Thus, the set $\{\Pi_\alpha\}$ contains the projectors onto the original orthonormal basis, as well as projectors built from the “cross terms”.
The operators $\{\Pi_\alpha\}$ are all positive semidefinite, and the set is linearly independent. The final step is to transform this set into a POVM, which requires making them sum to the identity. We do this by constructing $$\Omega := \sum_{\alpha=1}^{d^2} \Pi_\alpha,$$ which is easily shown to be a positive definite operator, and thus invertible. Multiplying both sides of the above equation by $\Omega^{-1/2}$ from the left and the right, we find that $$I = \sum_{\alpha=1}^{d^2} \Omega^{-1/2} \Pi_\alpha \Omega^{-1/2}.$$ The operators that appear in this sum, $$E_\alpha := \Omega^{-1/2} \Pi_\alpha \Omega^{-1/2},$$ inherit the rank and linear independence properties of the original projectors $\{\Pi_\alpha\}$, and by construction they sum to the identity, thereby constituting a POVM.
The operator $\Omega$ has a comparatively simple matrix representation: The elements along the diagonal are all equal to $d$, the elements above the diagonal are all equal to $\frac{1}{2}(1-i)$, and the rest are $\frac{1}{2}(1+i)$, as required by $\Omega =
\Omega^\dag$. The matrix $\Omega$ is not quite a circulant matrix, thanks to that change of sign, but it can be turned into one by conjugating with a diagonal unitary matrix. Consequently, the eigenvalues of $\Omega$ can be found explicitly via discrete Fourier transformation. The result is that, for $m = 0,\ldots,d-1$, $$\lambda_m = d + \frac{1}{2}\left(\cot \frac{\pi(4m+1)}{4d} - 1 \right).$$ This mathematical result has a physical implication [@Fuchs:2002].
The probability of any outcome $E_\alpha$ of an orthocross MIC, given any quantum state $\rho$, is bounded above by $$P(E_\alpha) \leq \left[d - \frac{1}{2}\left(1 + \cot
\frac{3\pi}{4d}\right)\right]^{-1} < 1.$$
The maximum of $\tr(\rho E_\alpha)$ over all $\rho$ is bounded above by the maximum of $\tr(\Pi E_\alpha)$, where $\Pi$ ranges over the rank-1 projectors. In turn, this is bounded above by the maximum eigenvalue of $E_\alpha$. We then invoke that $$\lambda_{\rm max}(E_\alpha)
= \lambda_{\rm max}(\Omega^{-1/2}\Pi_\alpha\Omega^{-1/2})
= \lambda_{\rm max}(\Pi_\alpha \Omega^{-1} \Pi_\alpha)
\leq \lambda_{\rm max} (\Omega^{-1}).$$ The desired bound then follows.
Note that all the entries in the matrix $2\Omega$ are Gaussian integers. Consequently, all the coefficients in the characteristic polynomial of $2\Omega$ will be Gaussian integers, and so the eigenvalues of $2\Omega$ will be algebraic integers. This is an example of how, in the study of MICs, number theory becomes relevant to physically meaningful quantities — in this case, a bound on the maximum probability of a reference-measurement outcome. Number theory has also turned out to be very important for SICs, in a much more sophisticated way [@Appleby:2013; @Appleby:2016; @Bengtsson:2016; @Appleby:2017b; @Kopp:2018].
The following conjectures about orthocross MICs have been motivated by numerical investigations. We expect that they are “textbook exercises of the future”, in that with the proper hint, their proofs might become relatively straightforward.
The entries in the $G$ matrices for orthocross MICs can become arbitrarily small with increasing $d$, but no two elements of an orthocross MIC can be exactly orthogonal.
For any orthocross MIC, $G^{-1}$ is not a Stieltjes matrix.
For any orthocross MIC, the entries in $G^{-1}$ are integers or half-integers.
As mentioned above, SIC existence is an open question. It is much easier to construct a symmetric MIC when the elements are not required to be rank-1. One such class of measurements are the *Wigner MICs* [@Appleby:2007a]. These were defined by Appleby in terms of the Weyl–Heisenberg group, which is most conveniently written in terms of its generators. Let $\{\ket{j}: j = 0,\ldots,d-1\}$ be an orthonormal basis, and define $\omega = e^{2\pi i/d}$. Then the operator $$X\ket{j} = \ket{j+1},$$ where addition is interpreted modulo $d$, effects a cyclic shift of the basis vectors. The Fourier transform of the $X$ operator is $$Z \ket{j} = \omega^j \ket{j},$$ and together these operators satisfy the Weyl commutation relation $$ZX = \omega XZ.$$ The Weyl–Heisenberg displacement operators are $$D_{k,l} := (-e^{\pi i/d})^{kl} X^k Z^l,$$ and together they satisfy the conditions $$D_{k,l}^\dag = D_{-k,-l},\ D_{k,l}D_{m,n} = (-e^{\pi i/d})^{lm-kn}
D_{k+m,l+n}.$$ Each $D_{k,l}$ is unitary and a $d^{\rm th}$ root of the identity. The Weyl–Heisenberg group is the set of all operators $(-e^{\pi i/d})^m
D_{k,l}$ for arbitrary integers $m$, and it is projectively equivalent to $\mathbb{Z}_d \times \mathbb{Z}_d$.
Let the operator $B$ be constructed as $$B := \frac{1}{d+1} \sum_{k,l} D_{k,l},$$ and define $B_{k,l}$ to be its conjugate under a Weyl–Heisenberg displacement operator: $$B_{k,l} := D_{k,l} B D_{k,l}^\dag.$$ The elements of the Wigner MIC have rank $(d+1)/2$, and are defined by $$E_{k,l} := \frac{1}{d^2} \left(I + \frac{1}{\sqrt{d+1}} B_{k,l}\right).$$ A Wigner MIC exists in any odd dimension $d$.
The quantities $$W_{k,l} := (d+1) \tr(E_{k,l} \rho) - \frac{1}{d}$$ are *quasiprobabilities*: They can be negative, but the sum over all of them is unity. The quasiprobability function $\{W_{k,l}\}$ is known as the *Wigner function* of the quantum state $\rho$. The Wigner function is the prototype for an entire genre of quasiprobability representations of quantum mechanics, a topic to which we will return in Section \[sec:qreps\].
So far, we have not imposed any additional structure upon our Hilbert space. However, in practical applications, one might have additional structure in mind, such as a preferred factorization into a tensor product of smaller Hilbert spaces. For example, a register in a quantum computer might be a set of $N$ physically separate qubits, yielding a joint Hilbert space of dimension $d = 2^N$. In such a case, a natural course of action is to construct a MIC for the joint system by taking the tensor product of multiple copies of a MIC defined on the component system: $$E_{j_1,j_2,\ldots,j_N} := E_{j_1} \otimes E_{j_2} \otimes \cdots \otimes E_{j_N}.$$ Since a collection of $N$ qubits is a natural type of system to consider for quantum computation, we define the $N$-qubit *tensorhedron MIC* to be the tensor product of $N$ individual qubit SICs.
The Gram matrix of an $N$-qubit tensorhedron MIC is the tensor product of $N$ copies of the Gram matrix for the qubit SIC out of which the tensorhedron is constructed.
Consider the two-qubit tensorhedron MIC, whose elements are given by $$E_{d(j-1)+j'} := \frac{1}{4}\Pi_j \otimes \Pi_{j'},$$ with $\{\Pi_j\}$ being a qubit SIC. The Gram matrix for the tensorhedron MIC has entries $$[G]_{d(j-1)+j',d(k-1)+k'} = \frac{1}{16}
\tr[(\Pi_j\otimes\Pi_{j'}) (\Pi_k\otimes\Pi_{k'})].$$ We can group together the projectors that act on the same subspace: $$[G]_{d(j-1)+j',d(k-1)+k'} = \frac{1}{16}
\tr(\Pi_j \Pi_k \otimes \Pi_{j'}\Pi_{k'}).$$ Now, we distribute the trace over the tensor product, obtaining $$[G]_{d(j-1)+j',d(k-1)+k'} = \frac{1}{16}
\frac{2\delta_{jk}+1}{3} \frac{2\delta_{j'k'} + 1}{3}
= [G_{\rm SIC}]_{jk} [G_{\rm SIC}]_{j'k'},$$ which is just the definition of the tensor product: $$G = G_{\rm SIC} \otimes G_{\rm SIC}.$$ This extends in the same fashion to more qubits.
The spectrum of the Gram matrix for an $N$-qubit tensorhedron MIC contains only the values $$\lambda = \frac{1}{2^N}\frac{1}{3^m},\ m = 0,\ldots,N.$$
This follows readily from the linear-algebra fact that the spectrum of a tensor product is the set of products $\{\lambda_i \mu_j\}$, where $\{\lambda_i\}$ and $\{\mu_j\}$ are the spectra of the factors.
We can also deduce properties of MICs made by taking tensor products of MICs that have orthogonal elements. Let $\{E_j\}$ be a $d$-dimensional MIC with Gram matrix $G$, and suppose that exactly $N$ elements of $G$ are equal to zero. The tensor products $\{E_j \otimes
E_{j'}\}$ construct a $d^2$-dimensional MIC, the entries in whose Gram matrix have the form $[G]_{jk} [G]_{j'k'}$, as above. This product will equal zero when either factor does, meaning that the Gram matrix of the tensor-product MIC will contain $2d^4 N - N^2$ zero-valued entries. It seems plausible that in prime dimensions, where tensor-product MICs cannot exist, the possible number of zeros is more tightly bounded, but this remains unexplored territory.
SICs are Optimal among MICs {#sec:optimal}
===========================
What might it mean for a MIC to be the best among all MICs? In order to provide a quantitative answer to this question, we codify an ideal that a MIC should approach. In essence, we want to find a MIC that furnishes a probabilistic representation of quantum theory which looks as close to classical probability as is mathematically possible. The residuum that remains — the unavoidable discrepancy that even the most clever choice of MIC cannot eliminate — is a signal of what is truly *quantum* about quantum mechanics.
We set the stage with a preliminary result.
Given a MIC $\{E_i\}$ with Gramian $G$, a quantum state is pure if and only if its probabilistic representation satisfies $$\sum_{ij} p(E_i) p(E_j) [G^{-1}]_{ij} = 1.$$
Fix the Hilbert-space dimension $d$, and let $\{E_i\}$ be a MIC. Any MIC is an operator basis by virtue of linear independence. Denote by $\{\widetilde{E}_i\}$ the dual basis. We can expand any quantum state $\rho$ in terms of the dual basis, and the expansion coefficients are the inner products with the elements of the original basis: $$\rho = \sum_i (\tr E_i \rho) \widetilde{E}_i.$$ By the Born Rule, the coefficients are probabilities: $$\rho = \sum_i p(E_i) \widetilde{E}_i.$$ Now, recall that while $\tr \rho = 1$ holds for any quantum state $\tr
\rho^2 = 1$ holds if and only if that operator is a “pure” state, i.e., a rank-1 projector. These operators are the extreme points of quantum state space; all other quantum states are convex combinations of them. In terms of the MIC’s dual basis, the pure-state condition is $$\sum_{ij} p(E_i)p(E_j) \tr \widetilde{E}_i \widetilde{E}_j = 1.$$ The “metric” in this quadratic form is the Gram matrix of the dual basis. Frame theory tells us that this is the inverse of the Gram matrix of the MIC itself: $$\sum_{ij} p(E_i) p(E_j) [G^{-1}]_{ij} = 1,
\label{eq:QM-condition}$$ as desired.
We can find an analogue for this condition in classical probability theory. The extreme points in a classical theory are the vertices of the probability simplex. The “ideal of the detached observer” (as Pauli phrased it [@Fuchs:2017b]) is a measurement that reads off the system’s point in phase space, call it $\lambda_i$, without disturbance. So, if we apply the reference measurement to each of a pair of identically prepared systems, we expect that we will obtain the same outcome both times. In other words, our “collision probability” is unity: $$\sum_i p(\lambda_i)^2 = 1.
\label{eq:cprob-condition}$$ Now, we recall our goal: What is the unavoidable residuum that separates quantum from classical? In other words, how closely can we make the quantum condition on extreme points, Eq. (\[eq:QM-condition\]), resemble the classical one, Eq. (\[eq:cprob-condition\])? This depends on how close we can bring the matrix $G^{-1}$ to the identity.
It makes sense for a reference process to be intrinsically unbiased. In this case, we have another reason to do so, because we’re trying to get as close as we can to the classical theory in terms of collision probability. So, let’s say we want the collision probability for the garbage state to be what it would be for the classical ideal, i.e., $1/d^2$. We get this when the elements of the reference measurement $\{E_i\}$ are equally weighted.
So, how close can we bring $G^{-1}$ to the identity, by choosing an appropriate equal-weight MIC? We know the answer to this from a recent theorem [@DeBrota:2018]: The best choice is a SIC.
\[U-norm\] Let $G$ be the Gram matrix of an equal-weight MIC, and let $\norm{\cdot}$ be any unitarily invariant norm (i.e., any norm where $\norm{A} = \norm{UAV}$ for arbitrary unitaries $U$ and $V$). Then $$\left\|{I-\frac{1}{d}G^{-1}}\right\|\geq \left\|{I - \frac{1}{d}G_{\rm SIC}^{-1}}\right\|\;,$$ with equality if and only if the MIC is a SIC.
This is a special case of Theorem 1 in [@DeBrota:2018].
Theorem \[U-norm\] applies to all unitarily invariant norms. This includes the Frobenius norm, the trace norm, the operator norm and all the other Schatten $p$-norms, as well as the Ky Fan $k$-norms. With respect to all of these distance measures, the SICs bring $G^{-1}$ as close to the identity as possible. Theorem \[U-norm\] was originally proven for foundational reasons [@DeBrota:2018], but it turns out to answer in the affirmative a conjecture regarding a practical matter of quantum computation [@Veitia:2018 §VII.A].
So far, we have investigated the matrix $G^{-1}$. What can we say about the matrix $G$ directly? This is easiest to explore when the MIC elements are all proportional to rank-1 projection operators.
No equal-weight MIC composed of rank-1 elements can be closer to an orthonormal basis than a SIC is, when measured by Frobenius distance between the MIC’s Gram matrix and the identity.
Let $\{E_i\}$ be a MIC on $d$-dimensional Hilbert space, and assume that this MIC is equal-weight, that is, $E_i = (1/d) \rho_i$ for some rank-1 projection operators $\{\rho_i\}$. Consider the quantity $$F := \sum_{ij} \left(\delta_{ij} - \tr E_i E_j\right)^2,$$ which is the squared Frobenius distance between the MIC’s Gram matrix and the identity. If the MIC could be an orthonormal basis of the operator space, then the distance $F$ would vanish. We can split $F$ into two sums, as follows: $$F = \sum_i \left(1 - \tr E_i^2\right)^2
+ \sum_{i \neq j} \left(\tr E_i E_j\right)^2.$$ The equal-weight and rank-1 conditions let us simplify the first sum, yielding $$F = d^2 \left(1 - \frac{1}{d^2}\right)^2
+ \sum_{i \neq j} \left(\tr E_i E_j\right)^2.$$ Applying the Cauchy–Schwarz inequality to the remaining sum, we get that $$F \geq d^2 \left(1 - \frac{1}{d^2}\right)^2
+ \frac{1}{d^4 - d^2}
\left(\sum_{i \neq j} \tr E_i E_j\right)^2,$$ with equality iff all the terms in the sum are equal, i.e., $\tr E_i
E_j$ is constant for all $i \neq j$. Because the $\{E_i\}$ must sum to the identity, $$\sum_{i \neq j} \tr E_i E_j
= \sum_i \tr E_i (I - E_i)
= d - 1.$$ Consequently, $$F \geq \left(d-\frac{1}{d}\right)^2
+ \frac{d-1}{d^4 - d^2},$$ with equality iff the MIC is symmetric, i.e., $\tr E_i E_j$ is constant for all $i \neq j$. The value of this constant is fixed by dividing its sum evenly across all contributions: $$\tr E_i E_j = \frac{1}{d^2}\frac{1}{d^2-1} \sum_{i \neq j} \tr E_i E_j
= \frac{1}{d^2} \frac{1}{d+1}.$$ In terms of the projection operators $\{\rho_i\}$, $$\tr \rho_i \rho_j = \frac{d\delta_{ij} + 1}{d+1}.$$ Thus, the Gram matrix of an equal-weight, rank-1 MIC saturates the lower bound on its squared Frobenius distance to the identity iff the MIC is a SIC.
The rank-1 condition can in fact be relaxed, at the expense of a slightly more elaborate proof. However, the equal-weight condition is not optional: The Gram matrix of a MIC can be made closer to the identity by taking the weights $\{e_i\}$ to be biased.[^1]
One property of the $G$ matrices that we can deduce is a constraint on their eigenvalues, expressed in the language of *majorization* [@Horn:1994; @Marshall:2011]. Let $x$ and $y$ be two vectors of $N$ entries each, and suppose that each element in $x$ and in $y$ is nonnegative. Write $x^\downarrow$ and $y^\downarrow$ for the vectors made by sorting $x$ and $y$ in nonincreasing order. Then $x$ weakly majorizes $y$ if $$\sum_{i=1}^k x^\downarrow_i \geq \sum_{i=1}^k y^\downarrow_i$$ for all $k = 1,\ldots,N$. We write this condition as $x \succ_w y$. If the $k = N$ condition is satisfied with equality, then $x$ majorizes $y$, which we write as $x \succ y$. Majorization is a partial order on vectors of nonnegative numbers. Heuristically speaking, $x$ majorizes $y$ if $y$ is a flatter vector than $x$.
To understand the context in which our next theorem applies, we first need a preliminary result:
The Gram matrix $G$ of any rank-1 MIC has a trace equal to or larger than 1.
By definition, $$\tr G = \sum_i \tr E_i^2 = \sum_i e_i^2 \tr \rho_i^2.$$ If the MIC elements are proportional to rank-1 projectors, then $\rho_i^2 = \rho_i$, and $$\tr G = \sum_i e_i^2.$$ Applying the Cauchy–Schwarz inequality, $$\tr G \geq \frac{1}{d^2} \left(\sum_i e_i\right)^2.$$ Recalling that the $e_i$ sum to $d$ because the $E_i$ sum to the identity, we obtain that $$\tr G \geq 1,$$ as desired.
\[majorization\] Let $G$ be the Gram matrix of a MIC which satisfies $\tr G\geq 1$ and let $\lambda$ denote its spectrum. Then $\lambda\succ_w\lambda_\text{SIC}$ with equality iff the MIC is a SIC.
We may write $E_i=e_i\rho_i$ where $e_i$ is a positive number and $\rho_i$ is a quantum state. Since the frame operator has the same spectrum as the Gram matrix, we may establish a lower bound on the largest eigenvalue of the Gram matrix: $$\label{minmaxeval}
\lambda_1\geq\frac{1}{d}\bbra{I}\mathcal{F}\kket{I}=\frac{1}{d}\sum_j(\tr E_j)^2=\frac{1}{d}\sum_j e_j^2(\tr \rho_j)^2 = \frac{1}{d}\sum_je_j^2\geq\frac{1}{d^3}\left(\sum_je_j\right)^{\!2}=\frac{1}{d}\;,$$ where the second inequality has equality iff $e_j=1/d$ for all $j$. The spectrum of the Gram matrix for a SIC reveals that this lower bound is tight. Now we have $$\lambda\succ\left(\lambda_1,\frac{\tr G-\lambda_1}{d^2-1},\ldots,\frac{\tr G-\lambda_1}{d^2-1}\right)\succ_w\left(\frac{1}{d},\frac{1}{d(d+1)},\ldots,\frac{1}{d(d+1)}\right)=\lambda_\text{SIC}\;.$$ The majorization follows because the second list is the flattest distribution which sums to $\tr G=\sum_j\lambda_j$ having largest eigenvalue $\lambda_1$. The weak majorization follows if the inequality $$\label{secondmaj}
\lambda_1+n\left(\frac{\tr G-\lambda_1}{d^2-1}\right)\geq\frac{1}{d}+n\left(\frac{1}{d(d+1)}\right)$$ holds when $1/d\leq \lambda_1\leq\tr G$, $0\leq n\leq d^2-1$, $\tr G\geq 1$, and $d>1$. Equation is algebraically equivalent to $$(d\lambda_1-1)(d^2-1-n)+nd(\tr G-1)\geq 0\;,$$ which is clearly satisfied by the conditions. When both majorizations are equalities, the state $\frac{1}{\sqrt{d}}\kket{I}$ is the eigenvector of the frame operator corresponding to the maximal eigenvalue of $1/d$. As indicated by the spectrum, in the diagonalizing basis the frame operator takes the form $$\mathcal{F}=\frac{1}{d(d+1)}(I_{d^2}+\kketbra{I}{I})\;.$$ We know from prior work (see [@Appleby:2015 Corollary 1], and also Lemma \[equivlemma\] in the appendix) that this implies that $dE_j$ forms a set of SIC projectors.
Specific examples of SICs have turned out to have interesting properties from a physical standpoint, that is, for purposes of quantum information and computation. For example, SICs in dimensions $d = 2$, $d = 3$ and $d = 8$ have been identified as optimal resources for quantum-computation protocols [@Veitch:2014; @Stacey:2016c; @Howard:2016]. A SIC in dimension $d = 4$ has surprising properties with regard to quantum entanglement [@Zhu:2010a]. Similarly, SICs emerge as significant in the study of quantum cloning. Earlier work [@Fuchs:1998] identified two measures for evaluating a cloning device, the *average global fidelity* and the *average local fidelity*. These two measures imply different standards of “best”; that is, they are maximed for differing values of the inner product between pure quantum states. Using the average global fidelity, the optimum (where “two states are most quantum with respect to each other”) occurs when the inner product $\tr
\rho\sigma$ is $1/3$. On the other hand, using the average local fidelity, the optimum occurs when the inner product is $1/4$. Therefore, a qubit SIC is a maximal set of states that pairwise optimize the average global fidelity, while a qutrit SIC does the same for average local fidelity. The results of this section indicate that it is not just specific examples of SICs that have interest for quantum information theory, but the entire class of measurement.
Computational Overview of MIC Gramians {#sec:numerics}
======================================
In order to explore the realm of MICs more broadly, and to connect them with other areas of mathematical interest, it is worthwhile to generate MICs *randomly* and study the $G$ matrices that result. One way to do so is to pick a random vector in a $d$-dimensional Hilbert space and take its orbit under the Weyl–Heisenberg group. The rank-1 projectors constructed from this orbit, rescaled by $1/d$, form an equal-weight MIC. (More precisely, they will do so except for initial vectors chosen from a set of measure zero.) The eigenvalue spectra of the $G$ matrices for these Weyl–Heisenberg MICs display intriguing patterns, as shown in Figure \[fig:histograms\].
Any stochastic matrix has an eigenvector with eigenvalue 1. Since these MICs are equal-weight, $dG$ is always doubly stochastic, and so $G$ will always have an eigenvalue equal to $1/d$. Generating large numbers of random Weyl–Heisenberg MICs, we can observe patterns in the histograms of their eigenvalues. The value $1/d$ appears as a peak at the upper extreme of the spectrum. Each histogram shows an exponential decay, trailing off with increasing eigenvalue, but the features in the small-eigenvalue portion appear to be dimension-dependent.
The plateau in the eigenvalue distribution for $d = 3$, seen in Figure \[fig:histograms\](b), is related to the existence of a continuous family of unitarily inequivalent SICs in that dimension [@Tabia:2013; @Hughston:2016].
[.45]{} ![\[fig:histograms\] Empirical histograms of eigenvalues for random Weyl–Heisenberg MICs in dimensions $d = 2$ through $d
= 8$. The peak at eigenvalue $1/d$ is straightforward to explain, the behavior at low eigenvalues less so. Note in particular the broad plateau for $d = 3$ and the apparent small-eigenvalue peaks in $d \geq 5$.](d2whHistogram.png "fig:"){width="\linewidth"}
[.45]{} ![\[fig:histograms\] Empirical histograms of eigenvalues for random Weyl–Heisenberg MICs in dimensions $d = 2$ through $d
= 8$. The peak at eigenvalue $1/d$ is straightforward to explain, the behavior at low eigenvalues less so. Note in particular the broad plateau for $d = 3$ and the apparent small-eigenvalue peaks in $d \geq 5$.](d3whHistogram.png "fig:"){width="\linewidth"}
[.45]{} ![\[fig:histograms\] Empirical histograms of eigenvalues for random Weyl–Heisenberg MICs in dimensions $d = 2$ through $d
= 8$. The peak at eigenvalue $1/d$ is straightforward to explain, the behavior at low eigenvalues less so. Note in particular the broad plateau for $d = 3$ and the apparent small-eigenvalue peaks in $d \geq 5$.](d4whHistogram.png "fig:"){width="\linewidth"}
[.45]{} ![\[fig:histograms\] Empirical histograms of eigenvalues for random Weyl–Heisenberg MICs in dimensions $d = 2$ through $d
= 8$. The peak at eigenvalue $1/d$ is straightforward to explain, the behavior at low eigenvalues less so. Note in particular the broad plateau for $d = 3$ and the apparent small-eigenvalue peaks in $d \geq 5$.](d5whHistogram.png "fig:"){width="\linewidth"}
[.45]{} ![\[fig:histograms\] Empirical histograms of eigenvalues for random Weyl–Heisenberg MICs in dimensions $d = 2$ through $d
= 8$. The peak at eigenvalue $1/d$ is straightforward to explain, the behavior at low eigenvalues less so. Note in particular the broad plateau for $d = 3$ and the apparent small-eigenvalue peaks in $d \geq 5$.](d6whHistogram.png "fig:"){width="\linewidth"}
[.45]{} ![\[fig:histograms\] Empirical histograms of eigenvalues for random Weyl–Heisenberg MICs in dimensions $d = 2$ through $d
= 8$. The peak at eigenvalue $1/d$ is straightforward to explain, the behavior at low eigenvalues less so. Note in particular the broad plateau for $d = 3$ and the apparent small-eigenvalue peaks in $d \geq 5$.](d7whHistogram.png "fig:"){width="\linewidth"}
[.45]{} ![\[fig:histograms\] Empirical histograms of eigenvalues for random Weyl–Heisenberg MICs in dimensions $d = 2$ through $d
= 8$. The peak at eigenvalue $1/d$ is straightforward to explain, the behavior at low eigenvalues less so. Note in particular the broad plateau for $d = 3$ and the apparent small-eigenvalue peaks in $d \geq 5$.](d8whHistogram.png "fig:"){width="\linewidth"}
MICs, Q-reps, and SIMs {#sec:qreps}
======================
Fix an equal-weight MIC $\{E_i\}$, and consider some *other* measurement $\{B_j\}$, which may or may not be a MIC. The Born Rule tells us that $$p(B_j) = \tr(B_j \rho).$$ Writing $B_j = b_j \pi_j$ for some quantum state $\pi_j$, and expanding both $\pi_j$ and $\rho$ in terms of the MIC’s dual basis $\{\widetilde{E}_i\}$, we find that $$p(B_j) = b_j \sum_{m,n} \tr(\pi_j E_m) \tr(\rho E_n)
\tr(\widetilde{E}_m \widetilde{E}_n).$$ As before, the last trace is given by the inverse of the MIC’s Gram matrix. The other two traces are the Born Rule probabilities for the events $E_m$ and $E_n$ given the quantum states $\pi_j$ and $\rho$: $$p(B_j) = b_j \sum_{m,n} p(E_m|\pi_j) p(E_n|\rho) [G^{-1}]_{mn}.$$ The inverse Gramian $G^{-1}$ is playing the role of the metric in a bilinear form: $$p(B_j) = b_j {\bf p}_{\pi_j}^{\rm T} G^{-1} {\bf p}_\rho.$$ Thanks to the informational completeness of our reference measurement $\{E_i\}$, we can calculate the probability for any other event *in terms of probabilities,* bypassing if we wish the operator $\rho$.
For an equal-weight MIC, the matrix $dG$ is doubly stochastic (Lemma \[dblystoch\]). Its inverse is therefore doubly quasi-stochastic: The rows and columns each sum to 1, but the elements are not confined to the unit interval. Therefore, in the formula $$p(B_j) = b_j d\,{\bf p}_{\pi_j}^{\rm T} (dG)^{-1} {\bf p}_\rho,
\label{eq:gauge-freedom}$$ we can act with $(dG)^{-1}$ either to the left or to the right and preserve normalization. This necessarily introduces negativity somewhere: The linear map $(dG)^{-1}$ sends some probabilities to quasi-probabilities, vectors that are properly normalized but that lie outside the probability simplex. The form of the expression (\[eq:gauge-freedom\]) means that we have a certain “gauge freedom” about where the negativity can occur [@Stacey:2018b]. We can put it into the states by acting to the right, or we can put it into the effects by acting to the left. We can also “split the map down the middle” by writing $$(dG)^{-1} = (dG)^{-1/2} (dG)^{-1/2}$$ and acting with one factor in either direction. This has the intriguing consequence that the final probability $p(D_j)$ is *zero* exactly when the two transformed vectors are *orthogonal.*
This brings us into the territory of quasi-probability representations of quantum theory constructed on orthogonal operator bases [@Zhu:2016a; @DeBrota:2017]. A *Q-rep* is an orthogonal Hermitian operator basis $\{F_i\}$ constrained to sum to the identity: $\sum_iF_i=I$. All orthogonal bases are self-dual, meaning that each vector is proportional to its dual vector, $F_j=\alpha_j\widetilde{F}_j$. From the sum constraint, we can see that $\tr \widetilde{F}_i=1$ for all $i$. Define $f_i:=\tr F_i$. Then $\sum_i f_i=d$. If $f_i=c$ for all $i$, we say the Q-rep is equal weight. Because it is orthogonal, we have $\tr
F_iF_j=c_i\delta_{ij}$. From the sum constraint, $$\sum_i\tr F_iF_j=\tr F_j\implies\sum_ic_i\delta_{ij}=f_j\implies c_j=f_j\;.$$ Thus $\tr F_iF_j=f_i\delta_{ij}$, and we can see that the dual basis satisfies $F_j=f_j\widetilde{F}_j$. Contrast this with Theorem \[noprops\], which shows that such proportionality never holds for MICs.
There is a very natural way in which MICs relate to Q-reps. This connection is simplest and likely of most interest in the equal weight situation where $f_i=1/d$ so we will first make this special case explicit. The vectorized notation and the frame operator introduced above will turn out to be quite useful.
The definition of a Q-rep is motivated by MICs, with the positivity condition relaxed. If treated like a POVM, they can be used to define quasiprobabilities. Recall that, in the equal weight case, $dG$ is a doubly stochastic matrix (Lemma \[dblystoch\]). If a matrix $A$ is column quasistochastic then $A^{-1}$ is as well, $$\sum_k[A^{-1}]_{kj}=\sum_{ik}[A]_{ik}[A^{-1}]_{kj}=\sum_i\delta_{ij}=1\;,$$ and similarly if $A$ is row quasistochastic. Further, $A$ raised to any power remains row or column quasistochastic: When raised to any power, one of the eigenvalues of the result remains equal to $1$ and the eigenvectors are unchanged, thus, there remains a flat right or left eigenvector with eigenvalue equal to $1$ indicating that the result remains row or column quasistochastic. With these facts in hand, we can see that the following construction provides us with a Q-rep from an equal weight MIC: $$\label{qrepfrommic}
F_i := \frac{1}{\sqrt{d}}\sum_j[G^{-1/2}]_{ij}E_j\;.$$ Summing over $i$ we get the identity because $\frac{1}{\sqrt{d}}G^{-1/2}$ is doubly quasistochastic and the $E_i$ form a POVM. Orthogonality may be checked directly: $$\tr F_i F_j=\frac{1}{d}\sum_{kl}[G^{-1/2}]_{ik}[G^{-1/2}]_{jl}\tr E_kE_l=\frac{1}{d}\sum_{kl}[G^{-1/2}]_{ik}[G]_{kl}[G^{-1/2}]_{jl}=\frac{1}{d}\delta_{ij}\;,$$ where we used the fact that $G$ is symmetric. In addition to sharing the same spectrum (Lemma \[isospec\]), the Gram matrix and the frame operator share the following nice relation.
The inverse square root of the Gram matrix enjoys the following representation in terms of the frame operator: $$[G^{-1/2}]_{ij}=\bbra{\widetilde{E}_i}\mathcal{F}^{-1/2}\kket{E_j}.$$
$$\begin{split}
[G^{-1}]_{jk}&=\bbraket{\widetilde{E}_j}{\widetilde{E}_k}=\bbra{\widetilde{E}_j}\mathcal{F}^{-1/2}\mathcal{F}\mathcal{F}^{-1/2}\kket{\widetilde{E}_k}=\sum_i \bbra{\widetilde{E}_j}\mathcal{F}^{-1/2}\kketbra{E_i}{E_i}\mathcal{F}^{-1/2}\kket{\widetilde{E}_k}\\
&=\sum_i \bbra{\widetilde{E}_j}\mathcal{F}^{-1/2}\kketbra{E_i}{\widetilde{E}_i}\mathcal{F}\mathcal{F}^{-1/2}\mathcal{F}^{-1}\kket{E_k}=\sum_i \bbra{\widetilde{E}_j}\mathcal{F}^{-1/2}\kketbra{E_i}{\widetilde{E}_i}\mathcal{F}^{-1/2}\kket{E_k}\\
&\implies [G^{-1/2}]_{ij}=\bbra{\widetilde{E}_i}\mathcal{F}^{-1/2}\kket{E_j}.
\end{split}$$
Inserting this into the vectorized form of and removing the resolution of identity that appears, gives us the nice identities $$\label{eq:vectorized-qrepfrommic}
\kket{F_i}=\frac{1}{\sqrt{d}}\sum_j[G^{-1/2}]_{ij}\kket{E_j}=\frac{1}{\sqrt{d}}\mathcal{F}^{-1/2}\kket{E_i}\;.$$
In , the essential facts were that we mix together the MIC elements with a matrix which squares to $G^{-1}$ and that when scaled by $1/\sqrt{d}$ is doubly quasistochastic. It turns out that another square root of $G^{-1}$, aside from the principal square root, also satisfies these desiderata. Define the matrix $$S := -I+\frac{2}{d^2}J\;,$$ where $J$ is the Hadamard identity, that is, the matrix of all $1$s. $S$ is a real symmetric unitary, which moreover commutes with all doubly stochastic matrices, so $G^{-1/2}S$ is another square root of $G^{-1}$ preserving the properties that we need.
For any equal weight MIC, we may construct the two Q-reps $$\label{eq:principal-and-shifted}
\kket{F_i}:=\frac{1}{\sqrt{d}}\sum_j[G^{-1/2}]_{ij}\kket{E_j}\quad \text{and } \kket{F_i^\text{S}}:=\frac{1}{\sqrt{d}}\sum_j[G^{-1/2}S]_{ij}\kket{E_j}\;,$$ which we will call the *principal* Q-rep and the $\textit{shifted}$ Q-rep respectively.
The relation between $\{F_i\}$ and $\{F_i^\text{S}\}$ is an elementwise affine transformation: $$\kket{F^\text{S}_i}=-\kket{F_i}+\frac{2}{d^2}\kket{I}\;.$$
If the MIC is a SIC, $E_i=\frac{1}{d}\Pi_i$, we have $$\frac{1}{\sqrt{d}}[G_{\rm SIC}^{-1/2}]_{ij}=\sqrt{d+1}\delta_{ij}+\frac{1}{d^2}(1-\sqrt{d+1}),$$ and so we get the following two Q-reps: $$\label{sicframes}
F_j=
\frac{1}{d}\left(\sqrt{d+1}\right)\Pi_j+\frac{1}{d^2}\left(1-\sqrt{d+1}\right)
{I} \quad \text{and } F^\text{S}_j=-
\frac{1}{d}\left(\sqrt{d+1}\right)\Pi_j+\frac{1}{d^2}\left(1+\sqrt{d+1}\right)
{I}\;.$$ These are the two Q-reps identified by Zhu [@Zhu:2016a] for a different reason.[^2] Given a Q-rep, the *ceiling negativity* of a quantum state $\rho$ is the magnitude of the most negative entry in the quasi-probability vector that represents $\rho$. Maximizing the ceiling negativity over all quantum states yields the ceiling negativity of the Q-rep. The two Q-reps $\{F_j\}$ and $\{F^\text{S}_j\}$ provide, respectively, the *lower and upper bounds* on the ceiling negativity over all equal-weight Q-reps in dimension $d$. Our orthogonalization procedure, Eq. (\[eq:principal-and-shifted\]), sets Zhu’s result in a more broad conceptual context: Zhu’s Q-reps are the output of applying to a SIC a procedure that works for any MIC.
This connection can be generalized beyond the equal weight case. For notational simplicity, we introduce the matrix $\Phi:=AG^{-1}$ where $[A]_{ij}=e_i\delta_{ij}$ where $e_i:=\tr E_i$. Noting that the square root of $\Phi$ with all positive eigenvalues may be written as $\sqrt{\Phi}=A^{1/2}\left(A^{1/2}G^{-1}A^{-1/2}\right)^{1/2}A^{-1/2}$, it is relatively straightforward to show that $f_i=e_i$. The principal Q-rep satisfies $$\kket{F_i}=\sum_j[\sqrt{\Phi}]_{ij}\kket{E_j}=\mathcal{A}^{-1/2}\kket{E_i}$$ where $\mathcal{A}:=\sum_i\frac{1}{e_i}\kketbra{E_i}{E_i}$ is the frame operator for the rescaled basis $\frac{1}{\sqrt{e_i}}E_i$ and the shifted Q-rep satisfies a weight-dependent shifting $$\kket{F_i^\text{S}}=-\kket{F_i}+\frac{2}{d}e_i\kket{I}\;.$$
What about going in the other direction, from a Q-rep to a MIC? Naïvely, we could try to invert the matrix multiplication in Eq. (\[eq:vectorized-qrepfrommic\]) and obtain $\kket{E_i}$ in terms of $\kket{F_i}$. However, without an independent expression for the frame operator $\mathcal{F}$, we lack the information to carry this out. It is nevertheless possible to construct a MIC from a Q-rep in a canonical way, following a procedure suggested by Marcus Appleby.
Given a Q-rep $\{F_i\}$, let $f$ be the minimal (i.e., most strongly negative) eigenvalue of all the elements of the Q-rep. Then $$H_i := \frac{F_i-fI}{1-d^2f}
\label{eq:SIM-from-Qrep}$$ defines a MIC with equal pairwise inner products $\tr H_i H_j$ and constant $\tr H_i^2$.
The essential task to be done is to remove the negativity from the Q-rep operators, while preserving the property that they sum to the identity. First, let $f$ be the minimal eigenvalue in the Q-rep $\{F_i\}$. Now define $\{H\}_i$ as stated. By inspection we see that the $\{H_i\}$ sum to the identity. Since we are subtracting the smallest eigenvalue of all of the $\{F_i\}$, we know that every $H_i$ is positive semidefinite. Thus, the $\{H_i\}$ form a MIC. The angle between them is $$\begin{split}
\tr H_i H_j&=(\frac{1}{1-d^2f})^2\tr(F_i-fI)(F_j-fI)=\left(\frac{1}{1-d^2f}\right)^2\left(\tr F_iF_j-f\tr F_i-f\tr F_j+df^2\right)\\
&=\left(\frac{1}{1-d^2f}\right)^2\left(\frac{1}{d}\delta_{ij}-\frac{2f}{d}+df^2\right)=\frac{\delta_{ij}-2f+d^2f^2}{d-2d^3f+d^5f^2}\;,
\end{split}$$ so it takes one constant value when $i = j$ and another constant value when $i\neq j$.
Appleby and Graydon introduced the term *SIM* for a symmetric measurement of arbitrary rank; a rank-1 SIM is a SIC [@Graydon:2016; @Graydon:2016a]. Eq. (\[eq:SIM-from-Qrep\]) yields a SIM for any Q-rep.
Conclusions
===========
We have surveyed the domain of Minimal Informationally Complete quantum measurements. Central to understanding these objects are their Gram matrices, which are “triply positive” (having nonnegative entries, being positive definite and constructed from positive semidefinite operators). These matrices quantify how close the representation of quantum theory that a MIC furnishes is to classical probability theory. The minimal separation, we have seen, is brought about when the MIC is a SIC. The Gram matrices are also the key ingredient in converting MICs to Q-reps, thanks to the role they play in the orthogonalization procedure.
Many properties of MIC Gram matrices remain unknown. Numerical investigations have, in some cases, outstripped the proving of theorems, resulting in the conjectures we have enumerated. Another avenue for potential future exploration is the application of Shannon theory to MICs, i.e., studying the probabilistic representations of quantum states using entropic measures [@Buck:2000]. In the case of SICs, this has already yielded intriguing connections among information theory, group theory and geometry [@Stacey:2016c; @Stacey:2016b; @Slomczynski:2014; @Szymusiak:2014; @Szymusiak:2015]. The analogous questions for other classes of MICs remain open for investigation.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research was supported in part by the John E. Fetzer Memorial Trust, the John Templeton Foundation, and grants FQXi-RFP-1612 and FQXi-RFP-1811B of the Foundational Questions Institute and Fetzer Franklin Fund, a donor advised fund of Silicon Valley Community Foundation. The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the John Templeton Foundation. We thank Marcus Appleby, Lane Hughston, Peter Johnson, Steven van Enk and Huangjun Zhu for discussions.
Equivalent conditions for symmetric operator bases
==================================================
Here, we prove the lemma that was necessary in the final step of proving Theorem \[majorization\]. This result was Corollary 1 in [@Appleby:2015] (a follow-up to [@Appleby:2011a]), which derives it as a special case of a more general theorem about Hermitian operator bases. We now prove this special case of interest directly, following a suggestion by Zhu.
\[equivlemma\] Let $L_j$ for $j=1,\ldots,d^2$ be $d^2$ Hermitian operators on $\mathcal{H}_d$, $P_s$ denote the projector into the symmetric subspace of $\mathcal{H}_d\otimes\mathcal{H}_d$, and $\mathbf{I}$ denote the identity superoperator. The following equations are equivalent: $$\sum_{j=1}^{d^2}L_j\otimes L_j=\frac{2d}{d+1}P_s,
\label{eq1}$$ $$\sum_{j=1}^{d^2} \kketbra{L_j}{L_j}=\frac{d}{d+1}(\mathbf{I}+\kketbra{I}{I}),
\label{eq2}$$ $$\textnormal{tr}L_jL_k=\frac{1}{d+1}(d\delta_{jk}+\textnormal{tr}L_j\textnormal{tr}L_k).
\label{eq3}$$ \[equiv\]
The space $\mathcal{L}(\mathcal{H}_d)\otimes\mathcal{L}(\mathcal{H}_d)$ is isomorphic to $\mathcal{L}(\mathcal{L}(\mathcal{H}_d))$ under the map $A\otimes B\to\kketbra{A}{B^\dag}$ for $A,B \in \mathcal{L}(\mathcal{H}_d)$. The equivalence of and follows from computing the image of $ P_s=\frac{1}{2}(I+\sum\ketbra{ij}{ji})$ under this map. Explicitly, for an orthonormal basis $\ket{i}$ of $\mathcal{H}_d$, $$P_s=\frac{1}{2}\left(I\otimes I+\sum_{i,j=1}^d\ketbra{i}{j}\otimes\ketbra{j}{i}\right)\to\frac{1}{2}\left(\kketbra{I}{I}+\sum_{i,j=1}^d\kketbra{(\ketbra{i}{j})}{(\ketbra{i}{j})}\right)=\frac{1}{2}(\mathbf{I}+\kketbra{I}{I}).$$ Now applying the isomorphism to immediately gives .
To see the equivalence of and , first note the following identities obtained from partial tracing and tracing : $$\label{id1}
\sum_{j=1}^{d^2}(\text{tr}L_j)L_j=\frac{d}{d+1}\text{tr}_1\left(I\otimes I+\sum_{i,j=1}^d\ketbra{i}{j}\otimes\ketbra{j}{i}\right)=\frac{d}{d+1}\left(dI+\sum_{i,j,k=1}^d\braket{k}{i}\braket{j}{k}\ketbra{j}{i}\right)=dI$$ and $$\label{id2}
\sum_{j=1}^{d^2}(\text{tr}L_j)^2=d^2.$$ Define $L_j'=L_j-x(\text{tr}L_j)I$, where $x$ is a parameter to be specified shortly. Then we have $$\label{equivtoe2}
\begin{split} \sum_{j=1}^{d^2}\kketbra{L'_j}{L'_j}&=\sum_{j=1}^{d^2}\kketbra{L_j}{L_j}-x\sum_{j=1}^{d^2}(\text{tr}L_j)(\kket{L_j}\bbra{I}+\kket{I}\bbra{L_j})+x^2\sum_{j=1}^{d^2}(\text{tr}L_j)^2\kketbra{I}{I}\\
&=\sum_{j=1}^{d^2}\kketbra{L_j}{L_j}+(x^2d^2-2xd)\kketbra{I}{I},
\end{split}$$ where we applied and in the second step. From this it’s easy to see that if we choose $$x=\frac{1}{d}\left(1\pm\sqrt{\frac{1}{d+1}}\right),$$ then $x^2d^2-2xd=-d/(d+1)$. Making this substitution and inserting into , we can see that is equivalent to $$\sum_{j=1}^{d^2}\kketbra{L'_j}{L'_j}=\frac{d}{d+1}\mathbf{I}$$ or that $$L''_j\equiv\sqrt{\frac{d+1}{d}}L'_j$$ satisfies $$\sum_{j=1}^{d^2}\kketbra{L''_j}{L''_j}=\mathbf{I}.$$ In a finite dimensional Hilbert space, this means that $L''_j$ forms an orthonormal basis. Thus, $$\begin{split}
\delta_{jk}&=\text{tr}{L''_jL''_k}=\frac{d+1}{d}\text{tr}L'_jL'_k=\frac{d+1}{d}\text{tr}\left( (L_j-x(\text{tr}L_j)I)(L_k-x(\text{tr}L_k)I) \right)\\
&=\frac{d+1}{d}\left(\text{tr}L_jL_k+(x^2d-2x)\text{tr}L_j\text{tr}L_k\right) =\frac{d+1}{d}\left(\text{tr}L_jL_k-\frac{1}{d+1}(\text{tr}L_j)(\text{tr}L_k)\right),
\end{split}$$ which is equivalent to .
[^1]: An earlier paper by one of us (CAF) and a collaborator [@Fuchs:2013] made the claim that the equal-weight condition could be derived by minimizing the squared Frobenius distance; this is erroneous. For the purposes of that earlier paper, it is sufficient to impose by hand the requirement that the MIC be equal-weight, since it is naturally desirable that a reference measurement have no intrinsic bias. Having made this extra proviso, the conceptual conclusions of that work are unchanged. Moreover, the same conceptual conclusions can be made much more robustly by using a refined version of the argument [@DeBrota:2018].
[^2]: Zhu calls equal weight Q-reps “NQPRs” and prefers to report the dual basis elements. In his notation, $Q_j^{-}=dF_j$ and $Q_j^{+}=dF_j^\text{S}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The discovery that the number of physically consistent string vacua is on the order of $10^{500}$ has prompted several statistical studies of string phenomenology. Contained here is one such study that focuses on the Weakly Coupled Free Fermionic Heterotic String (WCFFHS) formalism. Presented are systematic extensions of the well-known NAHE (Nanopoulos, Antoniadis, Hagelin, Ellis) set of basis vectors, which have been shown to produce phenomenologically realistic models. Statistics related to the number of $U(1)$’s, the specific gauge groups and their factors, non-Abelian singlets, and spacetime supersymmetries (ST SUSYs) are discussed for the full range of models produced. These findings are compared with prior results of other large-scale investigations. Statistical coupling between the gauge groups and the number of ST SUSYs is also discussed. In particular, for order-3 extensions there is a correlation between the appearance of exceptional groups and enhanced ST SUSY. Also discussed are some three-generation GUT models found among the data sets. These models are unique because they come from basis vectors which still have a geometric interpretation – there are no “rank-cuts" in these models.'
author:
- Timothy Renner
- Jared Greenwald
- Douglas Moore
- Gerald Cleaver
title: Initial Systematic Investigations of the Landscape of Low Layer NAHE Extensions
---
Introduction
============
Physically consistent string/M-theory derived models number at least approximately $10^{500}$ [@Bousso:2000; @Ashok:2003]. The large number of vacua has prompted both computational and analytical examinations of the the “landscape" of possible string vacua, e.g. [@Dijkstra:2004; @Donagi:2004; @Valandro:2008; @Balasubramanian:2008; @Lebedev:2008; @Gmeiner:2008; @Dienes:2008; @Gabella:2008; @Donagi:2008]. The Weakly Coupled Free Fermionic Heterotic String (WCFFHS)[@Antoniadis:1986; @Antoniadis:1987; @Kawai:1986_2; @Kawai:1988] approach to string model construction has produced some of the most phenomenologically realistic string models to date [@Cleaver:1999; @Lopez:1992; @Faraggi:1989; @Faraggi:1992; @Antoniadis:1990; @Leontaris:1999; @Faraggi:1991; @Faraggi:1992_2; @Faraggi:1992_3; @Faraggi:1991_2; @Faraggi:1991_3; @Faraggi:1995; @Faraggi:1996; @Cleaver:1997; @Cleaver:1997_2; @Cleaver:1997_3; @Cleaver:1998; @Cleaver:1998_2; @Cleaver:1998_3; @Cleaver:1998_4; @Cleaver:1998_5; @Cleaver:1999_2; @Cleaver:1999_3; @Cleaver:1999_4; @Cleaver:2000; @Cleaver:2000_2; @Cleaver:2001; @Cleaver:2001_2; @Cleaver:2002; @Cleaver:2002_2; @Cleaver:2002_3; @Perkins:2003; @Perkins:2005; @Cleaver:2008; @Greenwald:2009; @Faraggi:2010a; @Gato-Rivera:2010a; @Gato-Rivera:2010b; @Gato-Rivera:2010c; @Gato-Rivera:2011a; @Faraggi:2011a; @Cleaver:2011]. Outlined in this section are the basic steps needed to construct models within this formalism. Also detailed are the methodology of systematic searches and the difficulties they pose, and an explanation of the NAHE set of basis vectors used to construct realistic string models. Finally, results of the preliminary order-2 and order-3 extension to the NAHE set are presented, with an emphasis on GUT group models.
Weakly Coupled Free Fermionic Heterotic String Model Building
-------------------------------------------------------------
The heterotic theories consist of closed strings which have independent sets of left and right moving modes. The left moving modes are ten dimensional superstrings, while the right moving modes are 26 dimensional bosonic strings. WCFFHS models are constructed by specifying the phases that fermion modes gain when parallel transported around non-contractible loops on a genus-1 world sheet of a string. The number of possible values these phases may take is referred to as the order, while the number of “basis vectors" of phases used to specify the model is referred to as the layer. The linear combinations of the basis vectors produce sectors from which the physical states of the model are built. The force and matter content is then determined from the symmetries of these states.
Modular invariance highly constrains the possible components of the basis vectors. A consistent set of basis vectors $\{\vec{\alpha}^B\}$ is such that: $$\begin{aligned}
N_{ij}\vec{\alpha}^B_i\cdot \vec{\alpha}^B_j&=&0\pmod{4},\\
N_{i}\vec{\alpha}^B_i\cdot \vec{\alpha}^B_i&=&0\pmod{8} ({\text{for even\ } N_i}),\end{aligned}$$ where $N_{ij}$ is the least common multiple of the orders of the basis vectors $\vec{\alpha}^B_{i}$ and $\vec{\alpha}^B_{j}$, and the dot product is Lorentz, with a relative minus sign between the left-moving and the right-moving contributions. Additionally,
*The number of simultaneous periodic modes for any three basis vectors must be even.*
Taking all possible linear combinations of the basis vectors builds the full “geometry" of the model, designated by sectors $\{\vec{\alpha}\}$. States from a sector $\vec{\alpha}$ are designated by their charge vectors $\vec{Q}_{\vec{\alpha}} $, defined by $$\vec{Q}_{\vec{\alpha}} = \frac{1}{2}\vec{\alpha} + \vec{F},$$ where $\vec{F}$ is all combinations of 0,$\pm$1. The massless (with regard to the Planck scale) states are such that $$\begin{aligned}
\vec{Q}_L^2=2,\\
\vec{Q}_R^2=4,\end{aligned}$$ in a real fermion basis. In a complex basis, the values are halved. A physical state in a model a state must survive all GSO projections (GSOPs) from basis vectors $\vec{\alpha}^B_j$. These require that $$\vec{\alpha}^B_j\cdot\vec{Q}_{\vec{\alpha}} = \sum_i a_ik_{ji} + s_j\pmod{2},$$ where $a_i$ are the coefficients on the basis vectors used to construct $\vec{\alpha}$, $s_j$ is 0 (1) for bosonic (fermionic) basis vectors $a^B_j$ , and $k_{ji}$ is an element of the GSO coefficient matrix.
The GSO matrix must also follow its own modular invariance constraints, given by $$\begin{aligned}
N_j k_{ij} &=&0\pmod{2},\\
k_{ij} + k_{ji} &=& \frac{1}{2} \vec{\alpha}^B_i \cdot \vec{\alpha}^B_j\pmod{2},\quad {\rm and}\\
k_{ii} + k_{i1} &=& \frac{1}{4} \vec{\alpha}^B_i \cdot \vec{\alpha}^B_i - s_i\pmod{2}.\end{aligned}$$
Challenges to Systematic Searches
---------------------------------
Systematic searches are designed to do a “complete" scan of the possible consistent sets of basis vectors and GSO coefficients that define models within a chosen parameter space. In the discussions herein, models are classified into the order (the number of available twists for right-moving modes) and the layer (the number of basic vectors). Searches with random sampling have already been performed [@Dienes:2006; @Dienes:2007_2; @Dienes:2007; @Assel:2010]. However, random sampling methods have several difficulties which are non-trivial to address [@Dienes:2007]. Systematic methods avoid these inherent difficulties because models are not “sampled" - they form a complete set. Described in this section are the motivations and challenges to systematic searches of WCFFHS models including model building speed, model data storage, and model comparison.
The Scale of Systematic WCFFHS Searches
---------------------------------------
The scope of the systematic WCFFHS model searches can be limited by several factors. The first, and most obvious, factor is that of model building speed. Because the number of models to be built can get quite large, minimization of construction speed is essential. Assuming the model building process takes around 1 second, a data set with 1,000,000 models would take about 28 hours. As the number of “layers" of basis vectors in a model increases, the number of total models to build grows geometrically; each of the 1,000,000 models now has on the order of 1,000,000 new basis vectors (likely more than that) with modular invariance. Thus, the total number of models in the data set of the next layer is an estimated $10^{12}$. At one model every second, the layer-2 data set would be completed in approximately 3000 years.
Several steps can be taken to reduce the time required for model construction, both from an analytical and implementation standpoint. From the implementation side, understanding optimization techniques when writing the model building software is essential. Minimizing file and screen i/o, as well as the amount of copying done in memory are crucial steps needed to reduce the computing time.
Another computational concern is one of data storage. While a “master atlas" of basis vectors and the models to which they map is desirable, it is not feasible. The estimated average amount of space per model is 0.725 kB. For the 1,000,000 model data set, this amounts to 725 MB. The next layer would use approximately $725\times10^3$ TB, making a pure “atlas" of WCFFHS models unreasonable to pursue. There are two solutions to this problem. One is to gather statistics as the models are generated without keeping the models themselves in memory. While this approach is less taxing on system resources, it does have disadvantages. In particular, one must know the statistics to be gathered on the models prior to runtime. Any additional statistics would require a second run. Moreover, the total model set would have to be statistically analyzed, as opposed to the distinct models only, as the models are not being kept in memory. Double-counts that result from the construction method itself cannot be avoided with this approach. The other approach is to count the distinct models only, writing those to a file that serves as a small repository. While a full discussion of uniqueness in WCFFHS models is discussed in the next section, it suffices to say there are disadvantages to this approach as well. In particular, the definition of uniqueness in these models can be somewhat nebulous at times. Additionally, each unique model that is found must be compared with the other unique models that have been constructed, making each model require more computing time to complete.
Analytic techniques can also be used to reduce the number of redundant models produced. As these techniques are related to uniqueness in WCFFHS phenomenology, discussion of such analysis will be deferred until after uniqueness has been addressed.
Uniqueness in WCFFHS Models {#sec: Uniqueness_in_WCFFHS_Models}
---------------------------
The previously described computational limitations force the analysis to be only on models that are considered distinct. However, the definition of distinctness amongst WCFFHS models is not always clear. In particular the “amount" of phenomenology done to distinguish the models from one another must be balanced with the amount of computing time required to perform such analyses, as well as how easily the analyses can be automated.
Consider two different basis vector sets producing the same gauge and non-Abelian matter content. Even though the actual charge vectors and gauge group eigenvalues for the two models may be different, the overarching group structures are identical. This could be due to a string-scale type symmetry within the construction method itself, in which case there would be some transformation that could be applied to one of the basis vector sets to reproduce the other. The $U(1)$ charges of those two models, once diagonalized, would be the same. Such a symmetry in the construction method could in principle be revealed through an analytic study. However, it could also be the case that these models have non-equivalent $U(1)$ structures, and that the GSOPs eliminated the states that were different between the models. Such models will not have identical superpotentials or D- and F-flat directions, all of which have significant impact on the phenomenological viability of a model. For the systematic searches in this study the $U(1)$ charges of the matter representations will not be considered (although they will in future studies). Thus, two models will be considered identical herein if they have the same gauge and non-Abelian matter content and the same number of non-Abelian singlets.
This approach has a caveat, however. As the models become increasingly complex, comparing two models becomes increasingly difficult to automate. Consider the following two “toy" models whose particle content is presented in Tables \[tab: Toy\_Models\].
QTY $SU(4)$ $SU(4)$ $SU(4)$ $SO(10)$ $E_8$
----- --------- --------- --------- ---------- -------
1 4 4 1 1 1
1 4 1 4 1 1
1 1 4 4 1 1
2 1 6 1 10 1
2 1 1 6 10 1
: Two models illustrating the inherent difficulty in comparing WCFFHS models.
\
QTY $SU(4)$ $SU(4)$ $SU(4)$ $SO(10)$ $E_8$
----- --------- --------- --------- ---------- -------
1 4 4 1 1 1
1 4 1 4 1 1
1 1 4 4 1 1
2 6 1 1 10 1
2 1 1 6 10 1
: Two models illustrating the inherent difficulty in comparing WCFFHS models.
\[tab: Toy\_Models\]
It is clear that these two models are equivalent if two of the $SU(4)$ groups are switched. However, a simple boolean comparison of the gauge groups and matter states by a computer program will result in these models being counted as distinct. The root of the problem lies in the fact that the three $SU(4)$ gauge groups are not identical — there are different matter representations that transform under them. The most obvious solution would be to perform brute force permutations on the identical gauge groups and resorting the matter representations. Such an approach would work, but takes a significant amount of computing time. Comparisons are the most called operation in a systematic search, however, and thus need to be as fast as possible in order for the search to be efficient.
The solution implemented in this study is to propose and use a conjecture that defines uniqueness in a slightly stronger way. The conjecture is that identical models will have matter that fits into the same “classes" of representations. These matter representation classes are formed by ignoring the singlets (after verifying equal numbers of singlets in models under cpmparison) and considering only at the dimension of the non-Abelian representations and the gauge groups under which they transform. The classes of matter representations for the two example models are presented in Table \[tab: Toy\_Model\_Classes\].
[|ccc|]{} QTY&Model 1 Classes & Model 2 Classes\
\
1&(4,4)&(4,4)\
1&(4,4)&(4,4)\
1&(4,4)&(4,4)\
\
2&(6,10)&(6,10)\
2&(6,10)&(6,10)\
\[tab: Toy\_Model\_Classes\]
Now it is clear the two models are likely equivalent. This would imply that modular invariance prevents models that have only one representation switched from those of a known modular invariant model.
QTY $SU(4)$ $SU(4)$ $SU(4)$ $SO(10)$ $E_8$
----- --------- --------- --------- ---------- -------
2 4 4 1 1 1
1 4 1 4 1 1
2 1 6 1 10 1
2 1 1 6 10 1
: Another “toy" model, declared non-existant by the conjecture about matter representation classes.
\[tab: Impossible\_Toy\_Model\]
By this conjecture, the modular invariance of Table \[tab: Toy\_Models\] models would imply the Table \[tab: Impossible\_Toy\_Model\] model is not modular invariant. More theoretical work will need to be done to prove or disprove this conjecture.
The conjecture does not fully remedy the problem of a preferred gauge group ordering. Generally, matter representations are limited by the charges they carry due to the masslessness constraints of the fermion states. Representations of larger dimension tend to carry smaller dimensional charges under the other groups in the model, if at all. Certain gauge groups produce representations of similar dimension, however, which still place an ordering on the gauge groups within a class of matter representation. In particular, the 4- and 5-dimensional representations of $SO(5)$, the 4- and 6- dimensional representations of $SU(4)$, and the 2- and 3-dimensional representations of $SU(2)^{(2)}$ can still cause double counting amongst models. To completely remove the dependence on ordering in WCFFHS models, the problem must be reduced to a counting problem. In addition to proving or disproving the conjecture, the following possibility could also be explored: Two models with identical gauge groups and ST SUSYs are identical if they have the same total number of distinct matter representation classes, the same number of distinct matter representation classes, and the same number of total fermions.
This would remove the ordering dependence from the model comparison. Until it is proven, however, there is a risk of undercounting the models. Rather than risking this undercounting with the systematic studies presented herein, models in smaller data sets were examined, and duplicates were removed by hand. This will provide a systematic uncertainty estimate for statistics from larger data sets.
The NAHE Set {#sec: The_NAHE_Set}
============
The NAHE [@Antoniadis:1989; @Faraggi:1992] set is a set of five order-2 basis vectors which have served as a common basis set for phenomenologically realistic WCFFHS models. These basis vectors are given in Table \[tab: NAHE\_Set\_BVs\].
Sec $N_R$ $\psi$ $x^{12}$ $x^{34}$ $x^{56}$ $\overline{\psi}^{~1,...,5}$ $\overline{\eta}^{~1}$ $\overline{\eta}^{~2}$ $\overline{\eta}^{~3}$ $\overline{\phi}^{~1,...,8}$
-------------------- ------- -------- ---------- ---------- ---------- ------------------------------ ------------------------ ------------------------ ------------------------ ------------------------------
$\vec{\mathds{1}}$ 2 1 1 1 1 1,...,1 1 1 1 1,...,1
$\vec{S}$ 2 1 1 1 1 0,...,0 0 0 0 0,...,0
$\vec{b}_1$ 2 1 1 0 0 1,...,1 1 0 0 0,...,0
$\vec{b}_2$ 2 1 0 1 0 1,...,1 0 1 0 0,...,0
$\vec{b}_3$ 2 1 0 0 1 1,...,1 0 0 1 0,...,0
: The basis vectors and GSO coefficients of the NAHE set arranged into sets of matching boundary conditions. $N_R$ is the order of the right mover. The elements $\psi$, $\overline{\psi}^{~i}$, $\overline{\eta}^{~i}$, and $\overline{\phi}^{~i}$ are expressed in a complex basis, while $x^i$, $y^i$, $w^i$, $\overline{y}^{~i}$, and $\overline{w}^{~i}$ are expressed in a real basis.
\
Sec $N_R$ $y^{~1,2}w^{~5,6}||\overline{y}^{~1,2}\overline{w}^{~5,6}$ $y^{~3,...,6}||\overline{y}^{~3,...,6}$ $w^{~1,...,4}||\overline{w}^{~1,...,4}$
-------------------- ------- ------------------------------------------------------------ ----------------------------------------- -----------------------------------------
$\vec{\mathds{1}}$ 2 1,...,1$~||~$1,...,1 1,...,1$~||~$1,...,1 1,...,1$~||~$1,...,1
$\vec{S}$ 2 0,...,0$~||~$0,...,0 0,...,0$~||~$0,...,0 0,...,0$~||~$0,...,0
$\vec{b}_1$ 2 0,...,0$~||~$0,...,0 1,...,1$~||~$1,...,1 0,...,0$~||~$0,...,0
$\vec{b}_2$ 2 1,...,1$~||~$1,...,1 0,...,0$~||~$0,...,0 0,...,0$~||~$0,...,0
$\vec{b}_3$ 2 0,...,0$~||~$0,...,0 0,...,0$~||~$0.,,,.0 1,...,1$~||~$1,...,1
: The basis vectors and GSO coefficients of the NAHE set arranged into sets of matching boundary conditions. $N_R$ is the order of the right mover. The elements $\psi$, $\overline{\psi}^{~i}$, $\overline{\eta}^{~i}$, and $\overline{\phi}^{~i}$ are expressed in a complex basis, while $x^i$, $y^i$, $w^i$, $\overline{y}^{~i}$, and $\overline{w}^{~i}$ are expressed in a real basis.
$k_{ij}$ = $\left(\begin{tabular}{c|ccccc}
&$$&$$&$\_1$&$\_2$&$\_3$\\
\hline
$$&1&0&1&1&1\\
$$&0&0&0&0&0\\
$\_1$&1&1&1&1&1\\
$\_2$&1&1&1&1&1\\
$\_3$&1&1&1&1&1\\
\end{tabular}\right)$
\[tab: NAHE\_Set\_BVs\]
The massless particle spectrum is given in Table \[tab: NAHE\_Set\_Particles\].
**QTY** $SU(4)$ $SU(4)$ $SU(4)$ $SO(10)$ $E_8$
--------- ---------------- ---------------- ---------------- ---------- -------
2 $\overline{4}$ $1$ $1$ $16$ $1$
2 $1$ $\overline{4}$ $1$ $16$ $1$
2 $1$ $1$ $\overline{4}$ $16$ $1$
2 $1$ $1$ $4$ $16$ $1$
1 $1$ $1$ $6$ $10$ $1$
2 $1$ $4$ $1$ $16$ $1$
1 $1$ $6$ $1$ $10$ $1$
1 $1$ $6$ $6$ $1$ $1$
2 $4$ $1$ $1$ $16$ $1$
1 $6$ $1$ $1$ $10$ $1$
1 $6$ $1$ $6$ $1$ $1$
1 $6$ $6$ $1$ $1$ $1$
: The matter content of the model produced by the NAHE set.
\[tab: NAHE\_Set\_Particles\]
The NAHE set particle content, in addition to the particles listed in Table \[tab: NAHE\_Set\_Particles\], contains an N=1 ST SUSY. The observable sector of the NAHE set is an $SO(10)\times SU(4)^3$ GUT group with three sixteen dimensional $SO(10)$ matter representations serving as the matter generations. Each generation is charged under a different $SU(4)$ gauge group, and there are two copies of each representation. There are two distinct representations with $SO(10)$ charge 16 and $SU(4)$ charge 4, since the $SU(4)$ charge has a barred and unbarred representation. This brings the total number of copies for each generation up to four. In addition, the dimension of the $SU(4)$ charge itself is counted as being a set of copies of each generation, so the total number of copies of each generation is sixteen. There are no matter representations charged under the $E_8$ group, so it is the designated hidden sector for this model. Extensions of the NAHE set do not necessarily keep the designated observable sectors in the model. A model could, rather than break down the $SO(10)$ gauge group, break the $E_8$ into an $E_6$, producing an $E_6$ observable sector. In the past, when models were constructed individually, this was not common. Most individually constructed NAHE based models broke the $SO(10)$ gauge group into a Pati-Salam ($SU(4)\otimes SU(2)\otimes SU(2)$) group, $SU(5)\otimes U(1)$, or the MSSM gauge group ($SU(3)\otimes SU(2)\otimes U(1)$). In a systematic search that is no longer guaranteed, since the basis vectors are not organized by the researcher, but rather by the computer program. The notation for the next sections will change slightly; from here on the term layer will refer to the number of basis vectors after the initial set of five NAHE vectors rather than the total number of basis vectors.
Statistics for Order-2 Layer-1 {#sec: NAHE_O2L1_Statistics}
==============================
The first set of statistics to be reported here are for extensions to the NAHE set with a single basis vector of order 2. (Thus, extensions of this class can at best reduce the number of copies of each generation from 16 to 8.) The GSO coefficients were fixed for the NAHE set to those presented in Table \[tab: NAHE\_Set\_BVs\]. The GSO coefficients for the extended basis vector were systematically generated such that all possible combinations consistent with modular invariance were built. This study was repeated for the NAHE set without the ST SUSY generating basis vector $\vec{S}$, to determine the effect of ST SUSY on the models produced.
With $\vec{S}$
--------------
There were 439 unique models produced out of 1,945,088 total models. Approximately 9.5% of the models in the data set without rank-cuts were duplicates, and 13% of the models with rank-cuts were duplicates. All duplicates were removed prior to the statistical analysis. The frequency of the individual groups appearing in the unique models is presented in Table \[tab: NAHE\_O2L1\_Gauge\_Groups\].
Gauge Group Number of Unique Models % of Unique Models
--------------- ------------------------- --------------------
$SU(2)$ 365 83.14%
$SU(2)^{(2)}$ 73 16.63%
$SU(4)$ 338 76.99%
$SU(6)$ 2 0.4556%
$SU(8)$ 2 0.4556%
$SO(5)$ 155 35.31%
$SO(8)$ 141 32.12%
$SO(10)$ 160 36.45%
$SO(12)$ 2 0.4556%
$SO(14)$ 3 0.6834%
$SO(16)$ 147 33.49%
$SO(18)$ 1 0.2278%
$SO(20)$ 2 0.4556%
$SO(22)$ 1 0.2278%
$SO(24)$ 1 0.2278%
$SO(26)$ 1 0.2278%
$E_6$ 1 0.2278%
$E_7$ 142 32.35%
$E_8$ 144 32.8%
$U(1)$ 332 75.63%
: The frequency of the individual gauge groups amongst the unique models for the NAHE + O2L1 data set. Gauge groups at Ka$\check{c}$-Moody level higher than 1 are denoted with a superscript indicating the Ka$\check{c}$-Moody level.
\[tab: NAHE\_O2L1\_Gauge\_Groups\]
The first item of note is how many models retain at least one of the original gauge groups from the NAHE set. Approximately $77\%$ of the models kept at least one $SU(4)$ gauge group, while $\approx~36\%$ of the models kept their $SO(10)$ and $\approx~33\%$ kept their $E_8$. The most common gauge group in this set is $SU(2)$, which is expected, since it is the lowest rank non-Abelian gauge group. About $17\%$ have an $SU(2)^{(2)}$ gauge group. In these models, this happens when a left moving mode is paired with a right moving mode. As mentioned earlier, left-right paired elements reduce the rank of the gauge lattice of the model and, hence, are referred to as rank cuts. The non-simply laced gauge group $SO(5)$ also appears due to rank cuts.
Also of interest is the number of models with a $U(1)$ gauge group, which for this data set is quite high. $U(1)$’s can be problematic when dealing with deeper phenomenology in a model. The more $U(1)$’s present in a model the more likely the model is to have anomalous charge. This anomalous charge must be dealt with by finding D- and F-flat directions of the potential. However, the more $U(1)$ charges present, the more flat directions a model is likely to have, enabling more flexibility when giving mass to observable sector charged exotics. Discussion of anomalous $U(1)$ charges and flat directions of these models will not be reserved for future papers.
Though most of the models have smaller individual gauge group components, some models have gauge group enhancements. There are some models with $SO(18)$, $SO(20)$, $SO(22)$, $SO(24)$, and $SO(26)$ gauge groups. Those groups have rank 9, 10, 11, 12, and 13, respectively, making them higher rank than any one of the NAHE set gauge groups. This occurs when an added basis vector bridges the gap between the mutually orthogonal sets of states of the original five basis vectors, unifying the root spaces of the individual groups into one larger group. For order-2 models, however, it is clear this is not common.
Another way of measuring the enhancements that occur is looking at the number of gauge group factors in each model. Those are plotted in Figure \[fig: NAHE\_O2L1\_Gauge\_Group\_Factors\].
\[ybar, ylabel = Number of Distinct Models, xlabel = Number of Gauge Group Factors\] coordinates[ (4,2) (5,36) (6,49) (7,74) (8,68) (9,87) (10,48) (11,36) (12,23) (13,11) (14,5) ]{};
There is a definite peak at nine gauge group factors, much higher than the five initial gauge groups present in the NAHE set. Note that there are not many models with fewer than five gauge group factors. This implies that though there are several enhanced groups of higher rank than the initial NAHE set, the other gauge groups in the model remain broken.
Relevant GUT groups and the number of unique models containing those groups is presented in Table \[tab: NAHE\_O2L1\_GUT\_Groups\].
GUT Group Number of Unique Models % of Unique Models
----------------------------------- ------------------------- --------------------
$E_6$ 1 0.2278%
$SO(10)$ 160 36.45%
$SU(5)\otimes U(1)$ 0 0%
$SU(4)\otimes SU(2)\otimes SU(2)$ 243 55.35%
$SU(3)\otimes SU(2)\otimes SU(2)$ 0 0%
$SU(3)\otimes SU(2)\otimes U(1)$ 0 0%
: The number of unique models containing GUT groups for the NAHE + O2L1 data set.
\[tab: NAHE\_O2L1\_GUT\_Groups\]
The relatively low number of $SU(n+1)$ groups (excluding $SU(4)$) explains the lack of GUT group models in this data set.
The number of ST SUSYs are plotted against the number of unique models in Figure \[fig: NAHE\_O2L1\_ST\_SUSYs\].
\[nodes near coords, ybar, ylabel = Number of Distinct Models, xlabel = Number of ST SUSYs\] coordinates[ (0,223) (1,215) (2,1) (4, 0) ]{};
Many of the basis vector extensions did not alter the ST SUSY, while about half reduced it. More interestingly, there is one model with enhanced ST SUSY. The basis vector for that model is presented in Table \[tab: NAHE\_O2\_L1\_N2\_ST\_SUSY\].
Sec O $\psi$ $x^{12}$ $x^{34}$ $x^{56}$ $\overline{\psi}^{~1,...,5}$ $\overline{\eta}^{~1}$ $\overline{\eta}^{~2}$ $\overline{\eta}^{~3}$ $\overline{\phi}^{~1,...,8}$
----------- --- -------- ---------- ---------- ---------- ------------------------------ ------------------------ ------------------------ ------------------------ ------------------------------
$\vec{v}$ 2 1 1 1 1 0,...,0 1 1 0 0,...,0
: A basis vector and $k_{ij}$ matrix row which produces an enhanced ST SUSY when added the NAHE set.
\
Sec O $y^{~1,2}w^{~5,6}||\overline{y}^{~1,2}\overline{w}^{~5,6}$ $y^{~3,...,6}||\overline{y}^{~3,...,6}$ $w^{~1,...,4}||\overline{w}^{~1,...,4}$
----------- --- ------------------------------------------------------------ ----------------------------------------- -----------------------------------------
$\vec{v}$ 2 0,0,1,1$||$1,1,1,1 0,0,1,1$||$1,1,1,1 0,...,0$||$0,...,0
: A basis vector and $k_{ij}$ matrix row which produces an enhanced ST SUSY when added the NAHE set.
\
$k_{\vec{v},j}$ = (0, 0, 0, 1, 1) \[tab: NAHE\_O2\_L1\_N2\_ST\_SUSY\]
The particle content of this model is presented in Table \[tab: NAHE\_O2\_L1\_N2\_ST\_SUSY\_Particles\].
**QTY** $SU(4)$ $SU(4)$ $SU(4)$ $SO(10)$ $E_8$
--------- --------- ---------------- ---------------- ----------------- -------
1 $1$ $\overline{4}$ $1$ $\overline{16}$ $1$
1 $1$ $\overline{4}$ $1$ $16$ $1$
1 $1$ $1$ $\overline{4}$ $\overline{16}$ $1$
1 $1$ $1$ $\overline{4}$ $16$ $1$
1 $1$ $1$ $4$ $\overline{16}$ $1$
1 $1$ $1$ $4$ $16$ $1$
1 $1$ $4$ $1$ $\overline{16}$ $1$
1 $1$ $4$ $1$ $16$ $1$
2 $1$ $6$ $6$ $1$ $1$
2 $6$ $1$ $1$ $10$ $1$
: The particle content of the $N=2$ ST SUSY NAHE based model.
\[tab: NAHE\_O2\_L1\_N2\_ST\_SUSY\_Particles\]
The gauge groups of this model are identical to those of the NAHE set, but with fewer matter representations, particularly with regard to the $SU(4)$ charges. The enhanced ST SUSY comes from a new gravitino generating sector $\vec{b}_1+\vec{b}_2+\vec{v}$, which contributes a single gravitino state to the model. The other gravitino comes from $\vec{S}$. This example highlights the importance of systematic searches; the enhanced ST SUSYs come from very specific basis vectors that combine with the NAHE set to provide unexpected phenomenology. Though there are not a statistically significant number of models with this property, these models can highlight subtleties in the WCFFHS formulation that may go unnoticed in a random search.
The number of $U(1)$ gauge groups are plotted against the number of unique models in Figure \[fig:NAHE\_O2L1\_U1\_Factors\].
\[ybar, ylabel = Number of Distinct Models, xlabel = Number of $U(1)$ Factors\] coordinates[ (0,107) (1,164) (2,106) (3,48) (4,14) ]{};
The greater number of $U(1)$ factors present in the model, the greater that model’s capacity for carrying anomalous charge. There are relatively few $U(1)$’s in the models of this class. This is likely the result of the basis vectors having only periodic phases — nonzero, non-periodic phases break $SO(2n)$ groups into $SU(n-1)\otimes U(1)$ groups in most cases. Thus, there are not many $U(1)$ groups expected or found in this data set.
Another phenomenological property which might have statistical significance is the number of non-Abelian singlets, as non-Abelian singlets often carry observable sector hypercharge. However, no such particle with this property has yet been observed. Additionally, singlet particles contribute to the mass-energy density of the model. Too many non-Abelian singlets could result in a mass-energy density higher than observed values, also producing bad phenomenology. The number of non-Abelian singlets for this data set is plotted in Figure \[fig:NAHE\_O2L1\_NA\_Singlets\].
\[ybar, ylabel = Number of Distinct Models, xlabel = Number of NA Singlets\] coordinates[ (0,162)(2,37)(4,82)(6,47)(8,61)(10,10)(12,20)(14,6)(16,13)(24,1)]{};
Without $\vec{S}$
-----------------
Also of interest is the effect of the $\vec{S}$ vector in the set of NAHE basis vectors. Not only does the $\vec{S}$ vector generate ST SUSY, it also adds a degree of freedom to the $k_{ij}$ matrix. It stands to reason that removing $\vec{S}$ will also have an effect on the massless gauge and matter content in addition to the supersymmetry. There are 282 unique models in the set of 1,940,352 consistent models, in contrast to the 439 models in the data set with $\vec{S}$. About 8.4% of the models without rank cuts had duplicates, while 9.4% of the models with rank cuts were duplicates of other models produced. All duplicates were removed from the statistics to follow. Moreover the 282 models in this set do not all belong to the $N=0$ models in Figure \[fig: NAHE\_O2L1\_ST\_SUSYs\]. In fact, the two data sets have only 8 models in common, implying that the presence or non-presence of $\vec{S}$ has an effect on more than the ST SUSY.
The gauge content of the order-2 layer-1 NAHE extensions without the $\vec{S}$ vector are presented in Table \[tab: NAHE\_NSV\_O2L1\_Gauge\_Groups\].
Gauge Group Number of Unique Models % of Unique Models
--------------- ------------------------- --------------------
$SU(2)$ 233 82.62%
$SU(2)^{(2)}$ 67 23.76%
$SU(4)$ 212 75.18%
$SU(6)$ 2 0.7092%
$SU(8)$ 2 0.7092%
$SO(5)$ 107 37.94%
$SO(8)$ 89 31.56%
$SO(10)$ 100 35.46%
$SO(12)$ 2 0.7092%
$SO(14)$ 3 1.064%
$SO(16)$ 83 29.43%
$SO(18)$ 1 0.3546%
$SO(20)$ 2 0.7092%
$SO(22)$ 1 0.3546%
$SO(24)$ 1 0.3546%
$SO(26)$ 1 0.3546%
$E_6$ 1 0.3546%
$E_7$ 95 33.69%
$E_8$ 98 34.75%
$U(1)$ 209 74.11%
: The gauge content of the NAHE + O2L1 data set without $\vec{S}$.
\[tab: NAHE\_NSV\_O2L1\_Gauge\_Groups\]
A side-by-side comparison of the gauge group content percentages is shown in Table \[tab: NAHE\_O2L1\_Gauge\_Comparison\].
Gauge Group With $\vec{S}$ Without $\vec{S}$
--------------- ---------------- -------------------
$SU(2)$ 83.14% 82.62%
$SU(2)^{(2)}$ 16.63% 23.76%
$SU(4)$ 76.99% 75.18%
$SU(6)$ 0.4556% 0.7092%
$SU(8)$ 0.4556% 0.7092%
$SO(5)$ 35.31% 37.94%
$SO(8)$ 32.12% 31.56%
$SO(10)$ 36.45% 35.46%
$SO(12)$ 0.4556% 0.7092%
$SO(14)$ 0.6834% 1.064%
$SO(16)$ 33.49% 29.43%
$SO(18)$ 0.2278% 0.3546%
$SO(20)$ 0.4556% 0.7092%
$SO(22)$ 0.2278% 0.3546%
$SO(24)$ 0.2278% 0.3546%
$SO(26)$ 0.2278% 0.3546%
$E_6$ 0.2278% 0.3546%
$E_7$ 32.35% 33.69%
$E_8$ 32.8% 34.75%
$U(1)$ 75.63% 74.11%
: A side-by-side comparison of the gauge content for NAHE + O2L1 with and without $\vec{S}$.
\[tab: NAHE\_O2L1\_Gauge\_Comparison\]
The similarities between the gauge content of these two data sets are striking. It is tempting to assume that ignoring ST SUSYs when determining uniqueness will result in the data sets being identical, since the ST SUSY generator $\vec{S}$ is the only real difference between the data sets. This is not the case. In fact, ignoring ST SUSYs when comparing the intersection of these two sets of models gives the same number of models common to both sets: 8. This implies that the matter representations are affected by whether $\vec{S}$ is in the set of basis vectors making up a model. What is likely occurring in the models with $\vec{S}$ is that the sector coming from $\vec{S}$ is contributing non-adjoint representations to the matter states of the model. In other words, gravitinos are not the only fermion states coming from $\vec{S}$. In order to begin fully mapping the heterotic landscape the full effect of leaving out the $\vec{S}$ vector should be examined, as its presence may affect the likelihood of models with three chiral generations. Other statistics from this data set will now be presented for completeness.
The number of gauge group factors, ST SUSYs, and $U(1)$ factors are plotted in Figure \[fig: NAHE\_NSV\_O2L1\_Stats\]. None of the models had any non-Abelian singlets, suggesting that the non-Abelian singlets may be coming from the $\vec{S}$ sector. The occurrances of GUT groups are charted in Table \[tab: NAHE\_NSV\_O2L1\_GUT\_Groups\].
\
GUT Group Number of Unique Models % of Unique Models
----------------------------------- ------------------------- --------------------
$E_6$ 1 0.3546%
$SO(10)$ 100 35.46%
$SU(5)\otimes U(1)$ 0 0%
$SU(4)\otimes SU(2)\otimes SU(2)$ 156 55.32%
$SU(3)\otimes SU(2)\otimes SU(2)$ 0 0%
$SU(3)\otimes SU(2)\otimes U(1)$ 0 0%
: The number of unique models containing GUT groups for the NAHE + O2L1 data set without $\vec{S}$.
\[tab: NAHE\_NSV\_O2L1\_GUT\_Groups\]
Statistics for Order-3 Layer-1 {#sec: NAHE_O3L1_Statistics}
==============================
The next set of statistics to be reported are for extensions to the NAHE set with a single basis vector of right moving order-3. The order-3 basis vectors added are fermion sectors; the left movers are order-2. Since the orders of the left and right movers are not the same, the total order of the basis vector extensions is 6. The difference between order-3 basis vectors of this type and true order-6 basis vectors is that all six possibilities for the phases will appear in an order-6 right mover, while only three phases appear in an order-3 right mover. The coefficients generating the sectors from these basis vectors still range from 0 to 5, however. This has interesting effects on the fermion spectrum, which have been noted in [@Renner:2011]. Statistics will be presented for models of this type both with and without $\vec{S}$.
With $\vec{S}$
--------------
The presence of $\vec{S}$ in the NAHE set for this search causes any order-3 basis vector with the same left mover as $\vec{S}$ to be inconsistent. This is due to the $\mathds{Z}_2^L || \mathds{Z}_3^R$ symmetries of the left and right mover. Adding $\vec{S}$ to three times the basis vector extension results in a second $\vec{0}$ sector, which means the set of basis vectors is not linearly independent. Moreover, since the right mover does not have any periodic phases, there can be no rank-cutting in these models. Non-simply laced gauge groups and higher level Ka$\check{c}$-Moody algebras are not present.
Despite having no linearly independent basis vectors with the same left movers as $\vec{S}$, this data set contains quite a bit more distinct models. There were 373,152 models in this set, but only 3,036 were unique. This relatively (compared to the order-2 extension) high number of unique models suggests that the periodic/anti-periodic phases of the order-2 models have a redundancy not present in models of this type. Based on the double counting of the order-2 models without rank-cuts, the estimated systematic uncertainty for the order-3 statistics is 10%. The gauge content of these models is presented in Table \[tab: NAHE\_O3L1\_Gauge\_Groups\].
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(2)$ 2587 85.21%
$SU(3)$ 923 30.4%
$SU(4)$ 2241 73.81%
$SU(5)$ 543 17.89%
$SU(6)$ 735 24.21%
$SU(7)$ 215 7.082%
$SU(8)$ 460 15.15%
$SU(9)$ 76 2.503%
$SU(10)$ 76 2.503%
$SU(11)$ 17 0.5599%
$SU(12)$ 41 1.35%
$SU(13)$ 3 0.09881%
$SU(14)$ 5 0.1647%
$SO(8)$ 860 28.33%
$SO(10)$ 659 21.71%
$SO(12)$ 400 13.18%
$SO(14)$ 372 12.25%
$SO(16)$ 260 8.564%
$SO(18)$ 11 0.3623%
$SO(20)$ 33 1.087%
$SO(22)$ 5 0.1647%
$SO(24)$ 15 0.4941%
$SO(26)$ 3 0.09881%
$E_6$ 193 6.357%
$E_7$ 147 4.842%
$E_8$ 80 2.635%
$U(1)$ 2955 97.33%
: The gauge group content of the NAHE + O3L1 data set.
\[tab: NAHE\_O3L1\_Gauge\_Groups\]
The most noticeable difference between Table \[tab: NAHE\_O3L1\_Gauge\_Groups\] and Table \[tab: NAHE\_O2L1\_Gauge\_Groups\] is the number of $SU(n+1)$-type gauge groups. The GSO projections of the new sector break the untwisted ($\vec{0}$) sector from $SO(44)$ to smaller $SO(2n)$-type groups. Phases that are neither periodic nor anti-periodic transform these groups from $SO(2n)$ to $SU(n)\otimes U(1)$. Since all of the phases in this set fall into that category, the $SO(2n)$-type groups appear when the states from the added sector are projected out by certain $k_{ij}$ matrix choices. The $SU(n+1)$-type groups are created when the contributions from the added sector are left in the model. The number of gauge group factors per model are plotted in Figure \[fig: NAHE\_O3L1\_Gauge\_Group\_Factors\].
\[ybar, ylabel = Number of Distinct Models, xlabel = Number of Gauge Group Factors\] coordinates[ (4,8) (5,98) (6,127) (7,209) (8,267) (9,321) (10,553) (11,473) (12,477) (13,251) (14,200) (15,32) (16,20) ]{};
As with Figure \[fig: NAHE\_O2L1\_Gauge\_Group\_Factors\], there are still very few models which have less than five gauge group factors. There are also peaks at 10, 11 and 12 gauge group factors, as opposed to just a single peak at 9 in Figure \[fig: NAHE\_O2L1\_Gauge\_Group\_Factors\].
Relevant GUT groups and number of unique models containing those groups are presented in Table \[tab: NAHE\_O3L1\_GUT\_Groups\]. The number of ST SUSYs for the order-3 data set is plotted in Figure \[fig: NAHE\_O3L1\_ST\_SUSYs\]. Notice that a statistically significant number of models have enhanced ST SUSY. This trend is expected, as every odd-ordered right mover with a massless fermion left mover will produce an additional gravitino generating sector in the model.
GUT Group Number of Unique Models % of Unique Models
----------------------------------- ------------------------- --------------------
$E_6$ 193 6.36%
$SO(10)$ 659 21.71%
$SU(5)\otimes U(1)$ 543 17.89%
$SU(4)\otimes SU(2)\otimes SU(2)$ 1648 54.28%
$SU(3)\otimes SU(2)\otimes SU(2)$ 628 20.69%
$SU(3)\otimes SU(2)\otimes U(1)$ 775 25.53%
: The number of unique models containing GUT groups for the NAHE + O3L1 data set.
\[tab: NAHE\_O3L1\_GUT\_Groups\]
\[nodes near coords, enlargelimits=0.15, ybar, ylabel = Number of Distinct Models, xlabel = Number of ST SUSYs\] coordinates[ (0,1445) (1,1305) (2,286) (4, 0) ]{};
Table \[tab: NAHE\_O3L1\_Gauge\_Groups\] also makes apparent the number of models containing $U(1)$ gauge groups. The number of $U(1)$’s per model are plotted in Figure \[fig: NAHE\_O3L1\_U1\_Factors\].
\[ybar, ylabel = Number of Distinct Models, xlabel = Number of $U(1)$ Factors\] coordinates[ (0,81) (1,304) (2,518) (3,365) (4,751) (5,586) (6,184) (7,199) (8,36) (9,12) ]{};
It is clear that most models have multiple $U(1)$ factors, and that there are more $U(1)$ factors per model for this data set than the O2L1 data set. The number of non-Abelian singlets are plotted in Figure \[fig: NAHE\_O3L1\_NA\_Singlets\].
\[ybar, ylabel = Number of Distinct Models, xlabel = Number of NA Singlets\] coordinates[ (0,1220)(1,18)(2,326)(3,42)(4,577)(5,2)(6,176)(7,12)(8,309)(9,2)(10,46)(11,1)(12,105)(13,7)(14,22)(15,2)(16,65)(17,2)(18,13)(19,1)(20,19)(21,1)(22,10)(24,20)(26,7)(28,3)(30,4)(32,8)(34,1)(36,5)(38,2)(40,3)(42,2)(48,2)(50,1)]{};
As with the order-2 extensions, the distribution of non-Abelian singlets drops off sharply after 0, indicating the NAHE-base single-layer models do not have a tendency to produce many non-Abelian singlets.
Without $\vec{S}$
-----------------
Removing $\vec{S}$ from the NAHE set for the order-2 layer-1 extensions had interesting consequences on the available matter sectors. For the order-3 layer-1 extensions the effect is expected to be more drastic, as linear independence prevented any models with the same left mover as $\vec{S}$ to exist in a model. There should be less of an impact on the ST SUSY, however, since order-3 basis vectors with massless left movers always produce their own gravitino generating sector. The lower number of possibilities for $k_{ij}$ values will also have an effect on the number of models in the set.
There are 447 unique models in this set out of 870,688 consistent models, and of those 146 models also belong to the data set with $\vec{S}$. Based on the number of duplicates in the O2L1 data set (without $\vec{S}$), the estimated systematic uncertainty for these statistics is 10%. There is significantly more overlap between the two sets, yet there are also significantly fewer unique models. The gauge group content of the NAHE+O3L1 without $\vec{S}$ is presented in Table \[tab: NAHE\_NSV\_O3L1\_Gauge\_Groups\].
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(2)$ 368 82.33%
$SU(3)$ 128 28.64%
$SU(4)$ 313 70.02%
$SU(5)$ 70 15.66%
$SU(6)$ 96 21.48%
$SU(7)$ 26 5.817%
$SU(8)$ 71 15.88%
$SU(9)$ 15 3.356%
$SU(10)$ 14 3.132%
$SU(11)$ 3 0.6711%
$SU(12)$ 8 1.79%
$SU(13)$ 1 0.2237%
$SU(14)$ 1 0.2237%
$SO(8)$ 124 27.74%
$SO(10)$ 97 21.7%
$SO(12)$ 53 11.86%
$SO(14)$ 49 10.96%
$SO(16)$ 42 9.396%
$SO(18)$ 3 0.6711%
$SO(20)$ 6 1.342%
$SO(22)$ 1 0.2237%
$SO(24)$ 2 0.4474%
$SO(26)$ 1 0.2237%
$E_6$ 37 8.277%
$E_7$ 32 7.159%
$E_8$ 16 3.579%
$U(1)$ 430 96.2%
: The gauge group content of the NAHE + O3L1 data set without $\vec{S}$.
\[tab: NAHE\_NSV\_O3L1\_Gauge\_Groups\]
A brief comparison between Tables \[tab: NAHE\_NSV\_O3L1\_Gauge\_Groups\] and \[tab: NAHE\_O3L1\_Gauge\_Groups\] makes it clear the presence of $\vec{S}$ did not significantly affect the gauge groups, as was the case with the order-2 extensions. For completeness the occurances of the GUT groups and other relevant statistics for this data set will be presented. The occurrances of the GUT groups in this data set are tabulated in Table \[tab: NAHE\_NSV\_O3L1\_GUT\_Groups\]. The number of gauge group factors, $U(1)$ factors, ST SUSYs, and non-Abelian singlets are plotted in Figure \[fig: NAHE\_NSV\_O3L1\_Statistics\].
Gauge Group Number of Unique Models % of Unique Models
----------------------------------- ------------------------- --------------------
$E_6$ 37 8.277%
$SO(10)$ 97 21.7%
$SU(5)\otimes U(1)$ 70 15.66%
$SU(4)\otimes SU(2)\otimes SU(2)$ 220 49.21%
$SU(3)\otimes SU(2)\otimes SU(2)$ 81 18.12%
$SU(3)\otimes SU(2)\otimes U(1)$ 100 22.37%
: The occurrances of the GUT groups for the NAHE + O3L1 data set without $\vec{S}$.
\[tab: NAHE\_NSV\_O3L1\_GUT\_Groups\]
\
\
Models With GUT Groups {#sec: NAHE_Models_With_GUT_Groups}
======================
The next several sections will outline statistics on models containing GUT groups from the NAHE + O2L1 and NAHE + O3L1 data sets (with $\vec{S}$). The GUT groups to be examined are $E_6$, $SO(10)$, $SU(5)\otimes U(1)$, $SU(4)\otimes SU(2) \otimes SU(2)$ (Pati-Salam), $SU(3)\otimes SU(2)\otimes SU(2)$ (Left-Right Symmetric), and $SU(3)\otimes SU(2)\otimes U(1)$ (MSSM). In addition to the spread of statistics presented in sections \[sec: NAHE\_O2L1\_Statistics\] and \[sec: NAHE\_O3L1\_Statistics\], the number of chiral fermion generations will be counted, along with any exotics which carry observable sector charge. These statistics will be gathered for all possible observable sector configurations. Thus, if there is more than one copy of any GUT group in a model (as is often the case), all possible choices for each group forming the observable sector will be examined.
The definition of a chiral matter generation will be presented with each group for clarity, and only the “net" chiral generations will be counted. That is, if there are an equal number of barred and unbarred generations, the model will have no net chiral generations. Any “net" generations which cannot be paired must remain massless until the Higgs boson gains a VEV at the TeV scale.
The statistics gathered on the chiral generations here are not enough to qualify a model or set of models as being “realistic." This study looks only to examine the basic components that sometimes lead to realistic models. Actually determining whether or not a model is realistic requires detailed analysis of the $U(1)$ charges, finding the superpotential, and finding the D- and F-flat directions. Progress is being made to automate the above steps and integrate the deeper phenomenological components of WCFFHS model building into the FF Framework. For the present analysis, however, discussion of these aspects of WCFFHS phenomenology is omitted.
$E_6$ Models
------------
Each SM generation of fermions fits into a 27 dimensional representation of $E_6$, so the number of net chiral matter generations is given by $$|N_{27} - N_{\overline{27}}|.\label{eqn: E6_Net_Generations}$$ Additionally, for large GUT groups, states which transform under the “hidden sector” can be treated as being multiple copies of an observable generation. The hidden sector groups can be broken somewhat easily by adding basis vectors, so for certain GUT groups the dimension of the hidden sector charge is treated as the number of duplicate observable generations.
Statistics will now be presented for $E_6$ models coming from single layer extensions to the NAHE set. There was only one model with an $E_6$ group in the NAHE + O2L1 data set, but there were 193 models with $E_6$ in the NAHE + O3L1 data set. The number of net chiral fermion generations with and without hidden sector duplicates is plotted in Figure \[fig: E6\_NAHE\_O3L1\_OS\_Statistics\] along with the number of observable sector charged exotics.
\
The hidden sector gauge group content of these models is presented in Table \[tab: E6\_NAHE\_O3L1\_Gauge\_Groups\].
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(2)$ 165 85.49%
$SU(3)$ 54 27.98%
$SU(4)$ 113 58.55%
$SU(5)$ 39 20.21%
$SU(6)$ 43 22.28%
$SU(7)$ 5 2.591%
$SU(8)$ 10 5.181%
$SO(8)$ 35 18.13%
$SO(10)$ 41 21.24%
$SO(12)$ 12 6.218%
$SO(16)$ 19 9.845%
$E_8$ 11 5.699%
$U(1)$ 193 100%
: The hidden sector gauge group content of models containing $E_6$ within the NAHE + O3L1 data set.
\[tab: E6\_NAHE\_O3L1\_Gauge\_Groups\]
Of note in this table is the presence of $SU(5)$ and $SU(3)$ hidden sector gauge groups in large quantities. This result stands in contrast to those found in the order-2 search performed in [@Dienes:2006], and is likely due to the extension being an order-3 basis vector. Similar conclusions were also found in [@Moore:2011]. Order-3 basis vector extensions have very different properties from order-2 and order-4 basis vectors. Figures \[fig: E6\_NAHE\_O3L1\_OS\_Statistics\] shows that the distribution of net fermion generations tends to be at zero; either no $27$’s are produced, or every $27$ is accompanied by a $\overline{27}$ for those models. For models with more than zero chiral fermion generations, the number of generations per model is even both with and without hidden sector duplicates. It is also apparent from Figure \[fig: E6\_NAHE\_O3L1\_OS\_Statistics\] that the $E_6$ charged fermions do not couple to the hidden sector in most of these models. These examinations make it clear that, though most models do not contain exotic states, there are never the correct number of chiral fermion generations to produce a realistic model.
For completeness the rest of the statistics for this data set will be presented. The number of gauge group factors per model, the number of $U(1)$ factors, the number of ST SUSYs, and the number of non-Abelian singlets are presented in Figure \[fig: E6\_NAHE\_O3L1\_Statistics\].
\
$SO(10)$ Models
---------------
In the models containing $SO(10)$, the chiral fermion generations are defined to be in the 16 dimensional representations. Thus, the number of net chiral fermion generations is given by $$|N_{16} - N_{\overline{16}}|.\label{eqn: SO10_Net_Generations}$$ The NAHE + O2L1 data set has 160 models with $SO(10)$, while the NAHE + O3L1 data set has 659 models. The number of chiral fermion generations with hidden sector charges is plotted in Figure \[fig: SO10\_NAHE\_Generations\] for the NAHE + O2L1 and NAHE + O3L1 data sets. The number of chiral fermion generations without hidden sector charges is also plotted in Figure \[fig: SO10\_NAHE\_Generations\] for the order-3 $SO(10)$ models. There were no order-2 $SO(10)$ models with more than zero chiral fermion generations. The hidden sector gauge groups for the NAHE + O2L1 $SO(10)$ models are presented in Table \[tab: SO10\_NAHE\_O2L1\_Gauge\_Groups\]. The NAHE + O3L1 $SO(10)$ model hidden sector gauge groups are presented in Table \[tab: SO10\_NAHE\_O3L1\_Gauge\_Groups\]. The number of observable exotic states for these models is plotted in Figure \[fig: SO10\_NAHE\_Exotics\] for the order-2 and order-3 models.
\
Gauge Group Number of Unique Models % of Unique Models
--------------- ------------------------- --------------------
$SU(2)$ 121 75.62%
$SU(2)^{(2)}$ 25 15.62%
$SU(4)$ 111 69.38%
$SO(5)$ 59 36.88%
$SO(8)$ 1 0.625%
$SO(14)$ 1 0.625%
$SO(16)$ 56 35%
$SO(20)$ 1 0.625%
$SO(22)$ 1 0.625%
$E_7$ 51 31.87%
$E_8$ 51 31.87%
$U(1)$ 102 63.75%
: Hidden sector gauge groups for $SO(10)$ models in the NAHE + O2L1 data set.
\[tab: SO10\_NAHE\_O2L1\_Gauge\_Groups\]
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(2)$ 570 86.49%
$SU(3)$ 126 19.12%
$SU(4)$ 456 69.2%
$SU(5)$ 74 11.23%
$SU(6)$ 132 20.03%
$SU(7)$ 20 3.035%
$SU(8)$ 64 9.712%
$SU(9)$ 5 0.7587%
$SU(10)$ 15 2.276%
$SU(12)$ 12 1.821%
$SO(8)$ 105 15.93%
$SO(12)$ 75 11.38%
$SO(14)$ 34 5.159%
$SO(16)$ 69 10.47%
$SO(20)$ 13 1.973%
$SO(22)$ 5 0.7587%
$E_6$ 41 6.222%
$E_7$ 29 4.401%
$E_8$ 20 3.035%
$U(1)$ 604 91.65%
: Hidden sector gauge groups for $SO(10)$ models in the NAHE + O3L1 data set.
\[tab: SO10\_NAHE\_O3L1\_Gauge\_Groups\]
As with the $E_6$ data sets, there are no models with three chiral generations. The number of generations in any of the models presented is either zero or even. The distribution of charged exotics is more spread out than it was for the $E_6$ models. This is an artifact of the NAHE set gauge group; the $SO(10)$ states in that set are charged under the three $SU(4)$ gauge groups. Most observable $SO(10)$ states tend to keep some of those charges under the hidden sector. The remaining statistics for these models are presented in Figure \[fig: SO10\_NAHE\_O2L1\_Statistics\] for the order-2 models.
\
The remaining statistics for the order-3 models with $SO(10)$ are presented in Figure \[fig: SO10\_NAHE\_O3L1\_Statistics\].
\
$SU(5)\otimes U(1)$ Models
--------------------------
The (flipped) $SU(5)$ GUT group’s matter generations are split into multiple representations of the $SU(5)$ group. An anti-lepton doublet and the up-type quarks are placed in a $\overline{5}$ representation, while the right-handed neutrino, the anti-quark doublet, and the down-type quarks appear in a 10 dimensional representation of $SU(5)$. Thus, a generation is formed by pairing the 10-reps with the $\overline{5}$-reps. The net number of generations for an $SU(5)\otimes U(1)$ model is given by $$|\text{min}(N_{10},N_{\overline{5}}) -\text{min}(N_{\overline{10}},N_5)|.\label{eqn: SU5_Net_Generations}$$ There were no order-2 NAHE based models containing $SU(5)\otimes U(1)$, but there were 543 order-3 models with this GUT group. The hidden sector gauge groups of those models are presented in Table \[tab: SU5\_NAHE\_O3L1\_Gauge\_Groups\]. The number of net chiral generations with and without hidden sector duplicates are plotted in Figure \[fig: SU5\_NAHE\_O3L1\_OS\_Statistics\], along with the number of exotic states with observable sector charges.
\
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(2)$ 449 82.69%
$SU(3)$ 468 86.19%
$SU(4)$ 284 52.3%
$SU(6)$ 52 9.576%
$SU(7)$ 50 9.208%
$SU(8)$ 52 9.576%
$SU(9)$ 22 4.052%
$SU(10)$ 5 0.9208%
$SO(8)$ 57 10.5%
$SO(10)$ 74 13.63%
$SO(12)$ 35 6.446%
$SO(14)$ 72 13.26%
$E_6$ 39 7.182%
$E_7$ 10 1.842%
: The hidden sector gauge groups of the $SU(5)\otimes U(1)$ models in the NAHE + O3L1 data set.
\[tab: SU5\_NAHE\_O3L1\_Gauge\_Groups\]
The most striking feature of this data set is that there are models with three chiral generations both with and without hidden sector duplicates. The significance of this finding is that these models do not have rank cuts, and thus carry a geometric interpretation. This implies that these models may be written in another construction method, particularly that of orbifolding [@Donagi:2008]. These are the first three-generation models of their kind. More analysis will be done in Sec. \[sec: Three\_Generation\_Models\_With\_a\_Geometric\_Interpretation\] regarding this new class of three-generation models. Other statistics for this data set are plotted in Figure \[fig: SU5\_NAHE\_O3L1\_Statistics\].
\
Pati-Salam Models
-----------------
Pati-Salam models consist of models with a gauge group $SO(6)\otimes SO(4)$, which is isomorphic to the gauge group $SU(4)\otimes SU(2)\otimes SU(2)$. The latter is the form of this gauge group which appears in WCFFHS models. The quark and lepton generations are in representations of (4,2,1), while the anti-quarks and anti-leptons are in representations of ($\overline{4}$,1,2). Because these data sets are formed by examining all permutations of possible observable sectors, the same statistics will emerge when the chiral generations and anti-generations are examined separately. As there are no three-generation models in this data set, this case will not be considered. The equation for the number of net chiral generations is $$|N_{(4,2,1)}-N_{(\overline{4},2,1)}|.\label{eqn: PS_Net_Generations}$$ There are 243 unique models containing this gauge group in the order-2 NAHE extensions, and there are 1,648 unique models with this gauge group in the order-3 NAHE extensions. The abundance of these models is expected; as they contain the most common non-Abelian gauge group, $SU(2)$, along with a group that the NAHE set model already has.
No order-2 models with the Pati-Salam gauge group have net chiral matter generations, but other statistics related to those models are presented, as are order-3 model statistics. For the order-2 models, the hidden sector gauge group content is given in Table \[tab: PS\_NAHE\_O2L1\_Gauge\_Groups\]. In this data set, both $SU(2)$ at KM-level 1 and $SU(2)^{(2)}$ at KM-level 2 are included as possibilities for the observable Pati-Salam gauge group. The number of charged exotics is presented in Figure \[fig: PS\_NAHE\_O2L1\_Exotics\]. The hidden sector gauge group content for the order-3 models is presented in Table \[tab: PS\_NAHE\_O3L1\_Gauge\_Groups\]. The number of chiral generations and observable sector charged exotics for these models are plotted in Figure \[fig: PS\_NAHE\_O3L1\_OS\_Statistics\].
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SO(5)$ 60 24.69%
$SO(8)$ 51 20.99%
$SO(10)$ 61 25.1%
$SO(12)$ 1 0.4115%
$SO(16)$ 80 32.92%
$SO(20)$ 2 0.823%
$E_7$ 86 35.39%
$E_8$ 75 30.86%
$U(1)$ 185 76.13%
: The hidden sector gauge group content in Pati-Salam models from the NAHE + O2L1 data set.
\[tab: PS\_NAHE\_O2L1\_Gauge\_Groups\]
\[ybar, ylabel = Number of Distinct Models, xlabel = Number of Charged Exotics\] coordinates[ (4,6)(6,8)(7,4)(8,52)(9,8)(10,42)(11,22)(12,140)(13,72)(14,144)(15,190)(16,174)(17,154)(18,228)(19,230)(20,212)(21,292)(22,248)(23,346)(24,254)(25,204)(26,258)(27,312)(28,210)(29,214)(30,120)(31,152)(32,144)(33,186)(34,140)(35,80)(36,48)(37,100)(38,60)(39,56)(40,68)(41,16)(42,42)(43,6)(44,62)(45,48)(46,24)(47,88)(48,28)(49,40)(50,36)(51,40)(52,22)(53,8)(54,52)(55,40)(56,24)(58,6)(59,62)(60,32)(61,16)(66,24)(70,24)]{};
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(3)$ 344 20.87%
$SU(5)$ 174 10.56%
$SU(6)$ 414 25.12%
$SU(7)$ 116 7.039%
$SU(8)$ 214 12.99%
$SU(9)$ 30 1.82%
$SU(10)$ 31 1.881%
$SU(11)$ 5 0.3034%
$SU(12)$ 24 1.456%
$SO(8)$ 422 25.61%
$SO(10)$ 332 20.15%
$SO(12)$ 255 15.47%
$SO(14)$ 163 9.891%
$SO(16)$ 106 6.432%
$SO(20)$ 33 2.002%
$E_6$ 81 4.915%
$E_7$ 56 3.398%
$E_8$ 15 0.9102%
$U(1)$ 1615 98%
: The hidden sector gauge group content of the Pati-Salam models in the NAHE + O3L1 data set.
\[tab: PS\_NAHE\_O3L1\_Gauge\_Groups\]
Note that the number of chiral generations is zero in most cases for the NAHE + O3L1 data set as well. This implies that the symmetry breaking to the two $SU(2)$ gauge groups splits the three distinct generations of the NAHE observable sector evenly. Thus, three generation Pati-Salam models likely require more complicated basis vector sets. The remaning statistics are presented in Figure \[fig: PS\_NAHE\_O2L1\_Statistics\]. Statistics for the NAHE + O3L1 data set are presented in Figure \[fig: PS\_NAHE\_O3L1\_Statistics\].
\
\
Left-Right Symmetric Models
---------------------------
The final GUT considered in this study is a derivative of the Pati-Salam GUT group referred to as the Left-Right Symmetric group. It retails the dual-$SU(2)$ nature of the Pati-Salam GUT, but the $SU(4)$ gauge group is broken into an $SU(3)$ group directly representing the strong force. The generations of quarks fit into a (3,2,1)-dimensional representation while the generations of anti-quarks fit into a ($\overline{3}$,1,2)-dimensional representation. The lepton and anti-lepton generations are placed in a (1,2,1) and (1,1,2) representation, respectively. As the quark generations are usually more constraining in WCFFHS models, the term chiral matter generation refers only to the quarks, while the term chiral anti-generation refers only to the anti-quarks. Lepton generations will need to be taken into account when considering a quasi-realistic model, but here statistics will be gathered only with respect to the quark generations for simplicity.
The number of net chiral (anti)generations is given by $$|N_{(3,2,1)} - N_{(\overline{3},2,1)}|.\label{eqn: LRSym_Net_Generations}$$ Since the statistics loop over all possible observable sector configurations, the statistical data on the net number of chiral generations and anti-generations are identical. There are no models with the gauge group in the NAHE + O2L1 data set, as that data set contains no models with $SU(3)$ gauge groups. There are 628 distinct models in the NAHE + O3L1 data set with this gauge group. The hidden sector gauge content of those models is presented in Table \[tab: LRSym\_NAHE\_O3L1\_Gauge\_Groups\]. The number of net chiral generations is presented in Figure \[fig: LRSym\_NAHE\_O3L1\_OS\_Statistics\] along with the number of observable sector charged exotics.
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(4)$ 344 54.78%
$SU(5)$ 264 42.04%
$SU(6)$ 112 17.83%
$SU(7)$ 149 23.73%
$SU(8)$ 84 13.38%
$SU(9)$ 29 4.618%
$SU(10)$ 10 1.592%
$SU(11)$ 17 2.707%
$SO(8)$ 89 14.17%
$SO(10)$ 95 15.13%
$SO(14)$ 53 8.439%
$E_6$ 41 6.529%
$U(1)$ 628 100%
: The hidden sector gauge group content of the Left-Right Symmetric models in the NAHE + O3L1 data set.
\[tab: LRSym\_NAHE\_O3L1\_Gauge\_Groups\]
Like the $SU(5)\otimes U(1)$ data set, there are three-generation models present here. There are 70 models with three net chiral generations. One such model will be presented as an example at the end of the paper. The remaining statistical information on these models is presented in Figure \[fig: LRSym\_NAHE\_O3L1\_Statistics\].
\
MSSM-like Models
----------------
The MSSM[^1] gauge group is $SU(3)\otimes SU(2)\otimes U(1)$. A generation of quarks fit in a (3,2) representation of these groups. The leptons fit in a (1,2) representation. The generations of antimatter are charged differently, as the antiparticles do not have isospin. A generation of antimatter consists of two ($\overline{3}$,1) representations, one for the “up"-type quarks and one for the “down"-type quarks. While the leptons fit into a (1,2) representation, the anti-leptons are (1,1) singlets. As was the case with Left-Right Symmetric models, the terms chiral generation and anti-generation refer only to the quarks. While the lepton generations must also be considered, statistics are only gathered for the quark generations, as they are more constraining.
The equation for the number of net chiral matter generations is $$|N_{(3,2)}-N_{(\overline{3},2)}|,\label{eqn: MSSM_Net_Generations}$$ while the number of net chiral antimatter generations is $$|N_{(3,1)}-N_{(\overline{3},1)}|.\label{eqn: MSSM_Net_Anti_Generations}$$ There are no models with this gauge group from the NAHE + O2L1 data set since there are no $SU(3)$ gauge groups. There are, however, 775 models in the NAHE + O3L1 data set with the MSSM group. The hidden sector gauge group content of models containing the MSSM gauge group is presented in Table \[tab: MSSM\_NAHE\_O3L1\_Gauge\_Groups\]. The number of net chiral generations and anti-generations are presented in Figure \[fig: MSSM\_NAHE\_O3L1\_OS\_Statistics\]. The number of observable sector charged exotics is also plotted in Figure \[fig: MSSM\_NAHE\_O3L1\_OS\_Statistics\].
Gauge Group Number of Unique Models % of Unique Models
------------- ------------------------- --------------------
$SU(4)$ 412 53.16%
$SU(5)$ 374 48.26%
$SU(6)$ 112 14.45%
$SU(7)$ 169 21.81%
$SU(8)$ 112 14.45%
$SU(9)$ 41 5.29%
$SU(10)$ 10 1.29%
$SU(11)$ 17 2.194%
$SO(8)$ 97 12.52%
$SO(10)$ 111 14.32%
$SO(12)$ 35 4.516%
$SO(14)$ 68 8.774%
$E_6$ 46 5.935%
: The hidden sector gauge group content of the MSSM models in the NAHE + O3L1 data set.
\[tab: MSSM\_NAHE\_O3L1\_Gauge\_Groups\]
\
There are models with three chiral matter generations, as well as models with three anti-generations. However, none of the models have three generations of quarks and anti-quarks. While these findings are still significant due to their novelty, they do not point towards phenomenologically realistic models as the $SU(5)\otimes U(1)$ and Left-Right Symmetric models do. The remaining statistics for these models are presented in Figure \[fig: MSSM\_NAHE\_O3L1\_Statistics\].
\
ST SUSYs
--------
There is a trend regarding the number of ST SUSYs in GUT models — that distributions of ST SUSYs for the most part do not change. Figure \[fig: NAHE\_O2L1\_All\_ST\_SUSYs\] contains the ST SUSY distributions for the full data set, the $SO(10)$ models, and the Pati-Salam models.
\
It is clear that the ST SUSY distributions are relatively even for each of the sample sets of models. The same can be said of order-3 models, whose ST SUSY distributions are presented in Figures \[fig: NAHE\_O3L1\_All\_ST\_SUSYs\] and \[fig: NAHE\_O3L1\_All\_ST\_SUSYs\_B\]. Only the $E_6$ models display any sort of statistical coupling to the number of ST SUSYs, having significantly more N=2 models than any other data set. The other samples, however, have nearly identical distributions, suggesting that the number of ST SUSYs is not statistically linked to the GUT group content.
\
\
\
\
Further investigations of these findings show several statistical couplings for higher ST SUSY models containing certain gauge group factors. The statistical test to be used invokes the Central Limit Theorem, which is applicable to populations which are well behaved, such as the number ST SUSY distributions discussed. We will average the number of ST SUSYs per model for models containing each of the gauge groups present. Random sample averages will be close to the average of the population; therefore any sample drawn based on gauge groups which has an average close to the population average indicates that the average number of ST SUSYs is not coupled to the gauge group. If the average number of ST SUSYs per model for a particular gauge group (for example, $E_6$) is higher than the population (and if the sample is large enough to be significant), then we can conclude that the gauge group content has an effect on the number of ST SUSYs for a significant percentage of models. These significances are plotted in for the NAHE + O2L1 data set and for the NAHE + O3L1 data set.
![The significance values for models in the NAHE + O2L1 data set. Any significance values greater than three indicate a strong statistical significance.[]{data-label="fig: NAHE_O2L1_ST_SUSY_Significances"}](NAHE_O2L1.pdf)
![The significance values for models in the NAHE + O3L1 data set. Any significance values greater than three indicate a strong statistical significance.[]{data-label="fig: NAHE_O3L1_ST_SUSY_Significances"}](NAHE_O3L1.pdf)
While there are no significant gauge groups in the NAHE + O2L1 data set, all three exceptional groups, as well as $SU(9)$ and $SU(12)$, have significant effects on the average number of ST SUSYs. This is likely due to the most common embeddings of the SUSY sectors causing additional roots in the gauge groups, promoting them from $SU(n+1)$ to $SO(2n)$ or $E_n$. Additional analysis will be needed to confirm the cause of this significance.
Three Generation Models With a Geometric Interpretation {#sec: Three_Generation_Models_With_a_Geometric_Interpretation}
=======================================================
Several models containing three net chiral matter generations with $SU(5)\otimes U(1)$ and Left-Right Symmetric GUT groups were found in the NAHE + O3L1 data set. As previously mentioned, this finding is novel because these models do not have rank-cuts, and thus have a geometric interpretation. The usual statistics will be reported for these models, and a potentially realistic model from each gauge group will be presented as an example. To determine the viability of these models, more phenomenology must be done. In particular, finding the $U(1)$ charges and the superpotential would be the first step, then the D- and F-flat directions can be found. If the flat directions can eliminate the observable sector charged exotic matter simultaneously with the anomalous $U(1)$ charge, then the model could be considered a quasi-realistic model.
A Three Generation $SU(5)\otimes U(1)$ Model
--------------------------------------------
Presented in this section is an explicit example of a NAHE based three-generation $SU(5)\otimes U(1)$ model with $N=1$ ST SUSY. The gauge group for this model is $SU(3)^2\otimes SU(4)\otimes SU(5)\otimes SU(6)\otimes U(1)^5$.
Sec $N_R$ $\psi$ $x^{12}$ $x^{34}$ $x^{56}$ $\overline{\psi}^{~1,...,5}$ $\overline{\eta}^{~1}$ $\overline{\eta}^{~2}$ $\overline{\eta}^{~3}$ $\overline{\phi}^{~1,...,8}$
----------- ------- -------- ---------- ---------- ---------- ------------------------------------- ------------------------ ------------------------ ------------------------ -----------------------------------
$\vec{v}$ 3 1 1 0 0 0,0,$\frac{2}{3}$,...,$\frac{2}{3}$ $\frac{2}{3}$ 0 $\frac{2}{3}$ 0,0,$\frac{2}{3},...,\frac{2}{3}$
: A basis vector and $k_{ij}$ matrix row which produces a three-generation $SU(5)\otimes U(1)$ model.
\
Sec O $y^{~1,2}w^{~5,6}||\overline{y}^{~1,2}\overline{w}^{~5,6}$ $y^{~3,...,6}||\overline{y}^{~3,...,6}$ $w^{~1,...,4}||\overline{w}^{~1,...,4}$
----------- --- ------------------------------------------------------------ ------------------------------------------------------------ ------------------------------------------------------------------
$\vec{v}$ 3 0,0,1,1$||$0,0,0,0 0,0,0,0$||\frac{2}{3},\frac{2}{3},\frac{2}{3},\frac{2}{3}$ 0,0,1,1$||$$\frac{2}{3}$,$\frac{2}{3}$,$\frac{2}{3},\frac{2}{3}$
: A basis vector and $k_{ij}$ matrix row which produces a three-generation $SU(5)\otimes U(1)$ model.
\
$k_{\vec{v},j}$ = (0, 1, 1, 0, 1) \[tab: 3G\_SU5\_NAHE\_O3L1\_Example\_BVs\]
Table \[tab: 3G\_SU5\_NAHE\_O3L1\_Example\_BVs\] shows the basis vectors, and Table \[tab: 3G\_SU5\_NAHE\_O3L1\_Example\_Particles\] shows the particle content. The observable sector matter is tabulated in Table \[tab: 3G\_SU5\_NAHE\_O3L1\_Example\_Observable\_Matter\].
**QTY** $SU(3)$ $SU(3)$ $SU(4)$ $SU(5)$ $SU(7)$
--------- ---------------- ---------------- ---------------- ----------------- -----------------
2 $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
3 $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $7_{ }$
3 $3_{ }$ $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $\overline{7}$
1 $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $7_{ }$
2 $1_{ }$ $\overline{3}$ $1_{ }$ $5_{ }$ $1_{ }$
6 $1_{ }$ $1_{ }$ $6_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $4_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $1_{ }$ $4_{ }$ $5_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $1_{ }$ $\overline{5}$ $1_{ }$
2 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $\overline{21}$
6 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $21_{ }$
3 $1_{ }$ $1_{ }$ $1_{ }$ $5_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $\overline{4}$ $1_{ }$ $1_{ }$
1 $1_{ }$ $1_{ }$ $\overline{4}$ $5_{ }$ $1_{ }$
2 $1_{ }$ $3_{ }$ $1_{ }$ $\overline{5}$ $1_{ }$
1 $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$ $\overline{7}$
2 $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
1 $\overline{3}$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
1 $\overline{3}$ $1_{ }$ $1_{ }$ $\overline{10}$ $1_{ }$
2 $\overline{3}$ $1_{ }$ $1_{ }$ $\overline{5}$ $1_{ }$
1 $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$ $\overline{7}$
2 $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
: Particle content for the three-generation $SU(5)\otimes U(1)$ model. This model also has five $U(1)$ groups and $N=1$ ST SUSY.
\[tab: 3G\_SU5\_NAHE\_O3L1\_Example\_Particles\]
**QTY** $SU(3)$ $SU(3)$ $SU(4)$ $SU(5)$ $SU(7)$
--------- ---------------- ---------------- ---------------- ----------------- ---------
2 $1$ $\overline{3}$ $1$ $5$ $1$
1 1 1 4 5 1
2 $1$ $1$ $1$ $\overline{5}$ $1$
3 $1$ $1$ $1$ $5$ $1$
1 $1$ $1$ $\overline{4}$ $5$ $1$
2 $1$ $3$ $1$ $\overline{5}$ $1$
1 $\overline{3}$ $1$ $1$ $\overline{10}$ $1$
2 $\overline{3}$ $1$ $1$ $\overline{5}$ $1$
: Observable sector matter states without hidden sector charges for the three-generation $SU(5)\otimes U(1)$ model.
\[tab: 3G\_SU5\_NAHE\_O3L1\_Example\_Observable\_Matter\]
There are no (10,$\overline{5}$) generations and three ($\overline{10}$,5) generations in this model, giving it three net chiral generations of matter[^2]. However, counting the hidden sector charges as duplicates, there are 14 extra 5’s and 8 extra $\overline{5}$’s. Because of the numerous $U(1)$ charges, this model is ideal for future $U(1)$ and flat direction analysis.
A Three Generation Left-Right Symmetric Model
---------------------------------------------
Presented in this section is an explicitly constructed three-generation Left-Right Symmetric NAHE based model. The gauge group for this model is $SU(2)^2\otimes SU(3)^2\otimes SU(5)\otimes SO(10)\otimes U(1)^7$, and it has N=1 ST SUSY. The basis vectors for this model are given in Table \[tab: 3G\_LRSym\_NAHE\_O3L1\_Example\_BVs\]. The particle content of this model is presented in Table \[tab: 3G\_LRSym\_NAHE\_O3L1\_Example\_Particles\], and the observable matter is presented in Table \[tab: 3G\_LRSym\_NAHE\_O3L1\_Example\_Observable\_Matter\].
Sec $N_R$ $\psi$ $x^{12}$ $x^{34}$ $x^{56}$ $\overline{\psi}^{~1,...,5}$ $\overline{\eta}^{~1}$ $\overline{\eta}^{~2}$ $\overline{\eta}^{~3}$ $\overline{\phi}^{~1,...,8}$
----------- ------- -------- ---------- ---------- ---------- --------------------------------- ------------------------ ------------------------ ------------------------ -----------------------------------------------------
$\vec{v}$ 3 1 0 0 0 $\frac{2}{3}$,...,$\frac{2}{3}$ $\frac{2}{3}$ $\frac{2}{3}$ $\frac{2}{3}$ 0,...,0,$\frac{2}{3}$, $\frac{2}{3}$, $\frac{2}{3}$
: A basis vector and $k_{ij}$ matrix row which produces a three-generation Left-Right Symmetric model.
\
Sec O $y^{~1,2}w^{~5,6}||\overline{y}^{~1,2}\overline{w}^{~5,6}$ $y^{~3,...,6}||\overline{y}^{~3,...,6}$ $w^{~1,...,4}||\overline{w}^{~1,...,4}$
----------- --- ------------------------------------------------------------ -------------------------------------------- --------------------------------------------------------------------
$\vec{v}$ 3 0,0,1,1$||$0,0,$\frac{2}{3}$, $\frac{2}{3}$ 0,0,0,0$||$0,0,$\frac{2}{3}$,$\frac{2}{3}$ 0,0,1,1$||$$\frac{2}{3}$,$\frac{2}{3}$,$\frac{2}{3}$,$\frac{2}{3}$
: A basis vector and $k_{ij}$ matrix row which produces a three-generation Left-Right Symmetric model.
\
$k_{\vec{v},j}$ = (0, 1, 0, 0, 0) \[tab: 3G\_LRSym\_NAHE\_O3L1\_Example\_BVs\]
**QTY** $SU(2)_L$ $SU(2)_R$ $SU(3)_C$ $SU(3)$ $SU(5)$ $SO(10)$
--------- ----------- ----------- ---------------- ---------------- ----------------- -----------------
1 $2_{ }$ $2_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
2 $2_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
2 $2_{ }$ $1_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$
3 $2_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
3 $2_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $5_{ }$ $1_{ }$
1 $2_{ }$ $1_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$
1 $2_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $2_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $2_{ }$ $1_{ }$ $1_{ }$ $\overline{5}$ $1_{ }$
3 $1_{ }$ $2_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $2_{ }$ $1_{ }$ $1_{ }$ $5_{ }$ $1_{ }$
1 $1_{ }$ $2_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$
3 $1_{ }$ $2_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $3_{ }$ $3_{ }$ $1_{ }$ $1_{ }$
4 $1_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $3_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$
3 $1_{ }$ $1_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $\overline{10}$ $1_{ }$
1 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $\overline{5}$ $1_{ }$
1 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $\overline{16}$
3 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $10_{ }$
9 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
3 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $16_{ }$
5 $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $5_{ }$ $1_{ }$
1 $1_{ }$ $1_{ }$ $1_{ }$ $\overline{3}$ $\overline{10}$ $1_{ }$
1 $1_{ }$ $1_{ }$ $\overline{3}$ $3_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $\overline{5}$ $1_{ }$
: The particle content of the three-generation Left-Right Symmetric Model. This model also has 7 $U(1)$’s and $N=1$ ST SUSY.
\[tab: 3G\_LRSym\_NAHE\_O3L1\_Example\_Particles\]
**QTY** $SU(2)_L$ $SU(2)_R$ $SU(3)_C$ $SU(3)$ $SU(5)$ $SO(10)$
--------- ----------- ----------- ---------------- ---------------- ---------------- ----------
1 $2_{ }$ $2_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
2 $2_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
2 $2_{ }$ $1_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$
3 $2_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
3 $2_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $5_{ }$ $1_{ }$
1 $2_{ }$ $1_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$
1 $2_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $2_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $2_{ }$ $1_{ }$ $1_{ }$ $\overline{5}$ $1_{ }$
3 $1_{ }$ $2_{ }$ $1_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $2_{ }$ $1_{ }$ $1_{ }$ $5_{ }$ $1_{ }$
1 $1_{ }$ $2_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$
3 $1_{ }$ $2_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $3_{ }$ $3_{ }$ $1_{ }$ $1_{ }$
4 $1_{ }$ $1_{ }$ $3_{ }$ $1_{ }$ $1_{ }$ $1_{ }$
2 $1_{ }$ $1_{ }$ $3_{ }$ $\overline{3}$ $1_{ }$ $1_{ }$
1 $1_{ }$ $1_{ }$ $\overline{3}$ $3_{ }$ $1_{ }$ $1_{ }$
1 $1_{ }$ $1_{ }$ $\overline{3}$ $1_{ }$ $\overline{5}$ $1_{ }$
: The observable matter content of the three-generation Left-Right Symmetric Model.
\[tab: 3G\_LRSym\_NAHE\_O3L1\_Example\_Observable\_Matter\]
This model has three net generations of quarks, but no net generations of anti-quarks. Additionally there are thirty left- and right-handed lepton doublets. Other exotics include a quark triplet with left- and right-handed isospin, eight quark and ten anti-quark triplets without isospin. Thus, this model is not a favorable candidate for a quasi-realistic three-generation model. It does serve as a proof of concept that three generation models can be built with single-layer extensions to the NAHE set, however.
Conclusions {#sec: NAHE_Conclusions}
===========
The statistics presented in this paper make it clear that the NAHE set does serve its intended purpose as a basis for quasi-realistic WCFFHS models at a statistical level. Three generation models were constructed from order-3 extensions to the NAHE set. A summary of the GUT group analysis is presented in Table \[tab: NAHE\_GUT\_Generations\].
GUT Net Chiral Generations? Three Generations?
-------------------------- ------------------------- --------------------
O2L1 $SO(10)$ Yes No
O2L1 Pati-Salam No No
O3L1 $E_6$ Yes No
O3L1 $SO(10)$ Yes No
O3L1 $SU(5)\otimes U(1)$ Yes Yes
O3L1 Pati-Salam Yes No
O3L1 L-R Symmetric Yes Yes
O3L1 MSSM Yes Yes
: A summary of the GUT group study with regard to the number of chiral fermion generations in the NAHE set investigation.
\[tab: NAHE\_GUT\_Generations\]
Two three-generation models were discussed - a flipped-$SU(5)$ model and a Left-Right Symmetric model. While they did have the requisite number of chiral matter generations, there were several unfavorable properties in both models that prevented them from being considered quasi-realistic. They are a proof that three-generation models with geometric interpretations can be built with order-3 basis vector extensions.
The distributions of ST SUSYs across the GUT group subsets remained largely the same, save the $E_6$ models, which displayed a greater statistical tendency for ST SUSY enhancements. Further examination revealed that all of the exceptional groups, as well as $SU(9)$ and $SU(12)$, display more enhanced ST SUSY models for the NAHE + O3L1 data set. It was also shown that the presence of the $\vec{S}$ did not significantly impact the gauge content. However, the matter content of the models without $\vec{S}$ is affected, as the $\vec{S}$ sector produces states other than SUSY partners.
Acknowledgements
================
This work was supported by funding from Baylor University.
[99]{} \#1\#2\#3[[*AIP Conf. Proc.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Eur. Phys. J. *]{} [**C\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Ann. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Fortsch. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*J. Geom. and Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Nucl. Phys.*]{} [**B\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Nucl. Phys.*]{} [[**B**]{} (Proc. Suppl.) [**\#1**]{}]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Lett.*]{} [**B\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Rev.*]{} [**D\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Rev. Lett.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Phys. Rep.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Prog. Theo. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Mod. Phys. Lett.*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Mod. Phys. Lett.*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*New J. Phys.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Int. J. Mod. Phys.*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Int. J. Mod. Phys.*]{} [**A\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*JHEP*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Nuovo Cimento*]{} [**\#1A**]{} (\#2) \#3]{} \#1\#2\#3[[*Rept. Prog. Phys.*]{} [**\#1**]{} (\#2) \#3]{}
[^1]: Here MSSM refers only to the gauge group content. Models with this gauge group may or may not have ST SUSY.
[^2]: Recall that the definition of barred and unbarred representations is arbitrary.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We show that the complex dynamics of the general Bianchi IX universe in the vicinity of the spacelike singularity can be approximated by a simplified system of equations. Our analysis is mainly based on numerical simulations. The properties of the solution space can be studied by using this simplified dynamics. Our results will be useful for the quantization of the general Bianchi IX model.'
author:
- Claus Kiefer
- Nick Kwidzinski
- 'W[ł]{}odzimierz Piechocki'
title: |
On the dynamics of the general Bianchi IX spacetime\
near the singularity
---
=1
Introduction
============
The problem of spacetime singularities is a central one in classical and quantum theories of gravity. Given some general conditions, it was proven that general relativity leads to singularities, among which special significance is attributed to big bang and black hole singularities [@HP96].
The occurrence of a singularity in a physical theory usually signals the breakdown of that theory. In the case of general relativity, the expectation is that its singularities will disappear after quantization. Although a theory of quantum gravity is not yet available in finite form, various approaches exist within which the question of singularity avoidance can be addressed [@oup]. Quantum cosmological examples for such an avoidance can be found, for example, in [@ABKMM; @KKK16; @Bergeron:2015lka; @Bergeron:2015ppa] and the references therein.
Independent of the quantum fate of singularities, the question of their exact nature in the classical theory, and in particular for cosmology, is of considerable interest and has a long history; see, for example, [@Berger] and [@Uggla] for recent reviews. This is also the topic of the present paper.
Already in the 1940s, Evgeny Lifshitz investigated the gravitational stability of non-stationary isotropic models of universes. He found that the isotropy of space cannot be retained in the evolution towards singularities [@Lif1] (see [@Lif2] for an extended physical interpretation). This motivated the activity of the Landau Institute in Moscow to examining the dynamics of homogeneous spacetimes [@Bel]. A group of relativists inspired by Lev Landau, including Belinski, Khalatnikov and Lifshitz (BKL), started to investigating the dynamics of the Bianchi VIII and IX models near the initial spacelike cosmological singularity [@BKL1]. After several years, they found that the dynamical behaviour can be generalized to a generic solution of general relativity [@BKL2]. They did not present a mathematically rigorous proof, but rather a conjecture based on deep analytical insight. It is called the BKL conjecture (or the BKL scenario if specialized to the Bianchi-type IX model). The BKL conjecture is a locality conjecture stating that terms with temporal derivatives dominate over terms with spatial derivatives when approaching the singularity (with the exception of possible ‘spikes’ [@Uggla; @Czuchry:2016rlo]). Consequently, points in space decouple and the dynamics then turn out to be effectively the same as those of the (non-diagonal) Bianchi IX universe. (In canonical gravity, this is referred to as the strong coupling limit, see e.g. [@oup], p. 127.)
The dynamics of the Bianchi IX towards the singularity are characterized by an infinite number of oscillations, which give rise to a chaotic character of the solutions (see e.g [@Cornish:1996hx]). Progress towards improving the mathematical rigour of the BKL conjecture has been made by several authors (see e.g. [@Heinzle_Uggla_Rohr_2009]), while numerical studies giving support to the conjecture have been performed (see e.g. [@Garfinkle_2004]).
The dynamics of the diagonal Bianchi IX model, in the Hamiltonian formulation, were studied independently from BKL by Misner [@Misner_1969a; @Misner_1969b]. Misner’s intention was to search for a possible solution to the horizon problem by a process that he called “mixing”.[^1] Ryan generalized Misner’s formalism to the non-diagonal case in [@Ryan_1971a; @Ryan_1971b]. A qualitative treatment of the dynamics for all the Bianchi models may be found in the review article by Jantzen [@Jantzen:2001me], and we make reference to it whenever we get similar results.
Part of the BKL conjecture is that non-gravitational (‘matter’) terms can be neglected when approaching the singularity. An important exception is the case of a massless scalar field, which has analogies with a stiff fluid (equation of state $p=\rho$) and, in Friedmann models, has the same dependence of the density on the scale factor as anisotropies ($\rho\propto a^{-6}$). As was rigorously shown in [@AR01], such a scalar field will suppress the BKL oscillations and thus [*is*]{} relevant during the evolution towards the singularity. Arguments for the importance of stiff matter in the early universe were already given by Barrow [@Barrow78].
In our present work, we shall mainly address the general (non-diagonal) Bianchi IX model near its singularity. Our main motivation is to provide support for a rather simple asymptotic form of the dynamics that can suitably model its exact complex dynamics. We expect this to be of relevance in the quantization of the general Bianchi IX model, which we plan to investigate in later papers; see, for example, [@AGWP]. Apart from a few particular solutions that form a set of measure zero in the solution space, no general analytic solutions to the classical equations of motion are known. Therefore, we will restrict ourselves to qualitative considerations which will be supported by numerical simulations. The examination of the non-diagonal dynamics presented in [@Bogo2], though it is mathematically satisfactory, is based on the qualitative theory of differential equations, which is of little use for our purpose.
Our paper is organized as follows. Section II contains the formalism and presents our main results for a general Bianchi IX model. We first specify the kinematics and dynamics. We then consider a matter field in the form of (tilted) dust. This is followed by investigating the asymptotic regime of the dynamics near the singularity. Our conclusions are presented in Sec. III. The numerical methods used in our numerical simulations are described in the Appendix.
The general Bianchi IX spacetime
================================
Kinematics
----------
The general non-diagonal case describes a universe with rotating principal axes. The metric in a synchronous frame can be given as follows (see for the following e.g. [@Ryan_Shepley] and [@Ryan_HC]): $${\textnormal{d}}s^2=-N^2 {\textnormal{d}}t^2 + h_{ij} \sigma^i \otimes\sigma^j ,
\label{eq:Non_diagonal_BIX_metric}$$ where $N$ is the lapse function. Spatial hypersurfaces in the spacetime are regarded topologically as $S^3$ (describing closed universes), which can be parametrized by using three angles $\left\{\bar{\theta},\bar{\phi},\bar{\psi} \right\}\in [0,\pi]\times[0,2\pi]\times [0,\pi]$. The basis one-forms read $$\begin{aligned}
\sigma^1 &= -\sin (\bar{\psi}) {\textnormal{d}}\bar{\theta} + \cos (\bar{\psi})\sin(\bar{\theta}) {\textnormal{d}}\bar{\phi}\, ,
\\
\sigma^2 &= \cos (\bar{\psi}) {\textnormal{d}}\bar{\theta} + \sin (\bar{\psi})\sin(\bar{\theta}) {\textnormal{d}}\bar{\phi}\, ,
\\
\sigma^3 &= \cos(\bar{\theta}){\textnormal{d}}\bar{\phi} + {\textnormal{d}}\bar{\psi}
\ .
\end{aligned}$$ The $\sigma^i$, $i=1,2,3$, are dual to the vector fields $$\begin{aligned}
X_1 &= -&\sin(\bar{\psi}) \partial_{\bar{\theta}} +
\frac{\cos(\bar{\psi})}{\sin (\bar{\theta})}
\left[
\partial_{\bar{\phi}}
+ \cos(\bar{\theta}) \partial_{\bar{\psi}}
\right]\, ,
\\
X_2 &= &\cos(\bar{\psi}) \partial_{\bar{\theta}} +
\frac{\sin(\bar{\psi})}{\sin (\bar{\theta})}
\left[
\partial_{\bar{\phi}}
+ \cos(\bar{\psi}) \partial_{\bar{\psi}} \right]\, ,
\\
X_3 &=& \partial_{\bar{\psi}}\, ,
\\
\end{aligned}$$ which together with $X_0=\partial_t$ form an invariant basis of the Bianchi IX spacetime. The $X_i$ are constructed from the Killing vectors that generate the isometry group $SO(3,\mathbb{R})$ (see [@Ryan_Shepley] for more details). The basis one-forms satisfy the relation $${\textnormal{d}}\sigma^i=-\frac{1}{2}C^i_{jk}\sigma^j \wedge \sigma^k \ ,$$ with $C^i_{jk}=\varepsilon_{ijk}$ being the structure constants of the Lie algebra $\mathfrak{so}(3, \mathbb{R})$. The $X_i$ obey the algebra $[X_i,X_j]=-C^{k}_{ij}X_k$. We parametrize the metric coefficients in this frame as follows $$h_{ij} = O_i{}^{k} O_j{}^{l} \bar{h}_{ k l }\, ,$$ where $$\label{bar-h}
\bar{h}\equiv\left\{
\bar{h}_{i j}
\right\} =
{\textnormal{e}}^{2\alpha}\text{diag}\left(
{\textnormal{e}}^{2\beta_+ + 2\sqrt{3}\beta_-},
{\textnormal{e}}^{2\beta_+ - 2\sqrt{3}\beta_-},
{\textnormal{e}}^{-4\beta_+ }
\right)\equiv\text{diag}\left(\Gamma_1 ,\Gamma_2 , \Gamma_3 \right) .$$ The variables $\alpha$, $\beta_+$, and $\beta_-$ are known as the Misner variables. The scale factor $\exp(\alpha)$ is related to the volume, while the anisotropy factors $\beta_+$ and $\beta_-$ describe the shape of this model universe. The variables $\Gamma_1$, $\Gamma_2$ and $\Gamma_3 $ were used by BKL in their original analysis [@bkl].
We introduced here a matrix $O\equiv\left\{O_i{}^{j}\right\}\equiv
O_{\theta}O_{\phi}O_{\psi}$ ($i$ corresponding to rows and $j$ corresponding to columns), which is an $SO(3,\mathbb{R})$ matrix that can be parametrized by another set of Euler angles, $\left\{ \theta,\phi,\psi \right\}\in
[0,\pi]\times[0,2\pi]\times [0,\pi]$. Explicitly, $$\label{matrixO}
\begin{aligned}
O_{\psi}=\left(
\begin{array}{ccc}
\cos ( \psi ) & \sin ( \psi ) & 0 \\
- \sin ( \psi ) & \cos ( \psi ) & 0 \\
0 & 0 & 1
\end{array}
\right) \ , \quad
O_{\theta}=\left(
\begin{array}{ccc}
1 & 0 & 0 \\
0 &\cos ( \theta ) & \sin ( \theta ) \\
0 &- \sin ( \theta ) & \cos ( \theta )
\end{array}
\right)\, ,
\\
O_{\phi}=\left(
\begin{array}{ccc}
\cos ( \phi ) & \sin ( \phi ) & 0 \\
- \sin ( \phi ) & \cos ( \phi ) & 0 \\
0 & 0 & 1
\end{array}
\right)\, .\qquad\qquad\quad\qquad
\end{aligned}$$ The Euler angles $\theta$, $\phi$, and $\psi$ are now dynamical quantities and describe nutation, precession, and pure rotation of the principal axes, respectively. In the case of Bianchi IX spacetime, the group $SO(3,\mathbb{R})$ is the canonical choice for the diagonalization of the metric coefficients. For a treatment of other Bianchi models, see [@Jantzen:2001me].
Dynamics
--------
In the following, we shall discuss the Hamiltonian formulation of this model. In order to keep track of the diffeomorphism (momentum) constraints, we replace the metric (\[eq:Non\_diagonal\_BIX\_metric\]) by the ansatz $$\begin{aligned}
&{\textnormal{d}}s^2=- N^2 {\textnormal{d}}t^2 +
h_{ij}\left(N^i{\textnormal{d}}t+\sigma^i \right)\otimes\left(N^j{\textnormal{d}}t+\sigma^j \right),
\end{aligned}$$ where $N^i$ are the shift functions. The Hamiltonian formulation was first derived in a series of papers by Ryan: the symmetric (non-tumbling) case obtained by constraining ${\psi}$, ${\phi}$ to be constant and keeping $\theta $ dynamical is discussed in [@Ryan_1971a], and the general case can be found in [@Ryan_1971b]. We write the Einstein-Hilbert action in the well known ADM form, $$S_{EH}=
\frac{1}{16\pi G}\int \sigma^1\wedge\sigma^2\wedge \sigma^3\int{\textnormal{d}}t\
N\sqrt{h}\left[
\left(h^{ik} h^{jl} -h^{ij}h^{kl}\right) K_{ij}K_{kl}
+{}^{(3)} R
\right],
\label{eq:SEH_BIX}$$ where $$K_{ij}=\frac{1}{2N}\left(\dot{h}_{ij}-2 D_{(i} N_{j}) \right)$$ is the extrinsic curvature, and $D_i$ is the spatial covariant derivative in the non-coordinate basis $\left\{ X_i \right\}$. We will set $\frac{3}{4\pi G}\int \sigma^1\wedge\sigma^2\wedge \sigma^3=1$ for simplicity. The three-dimensional curvature ${}^{(3)}R$ on spatial hypersurfaces of constant coordinate time is given by $${}^{(3)} R = -\frac{{\textnormal{e}}^{-2\alpha}}{2}
\left(
{\textnormal{e}}^{-8\beta_+}-4{\textnormal{e}}^{-2\beta_+} \cosh \left(2\sqrt{3}\beta_- \right)
+ 2{\textnormal{e}}^{4\beta_+}
\left[ \cosh\left(
4\sqrt{3}\beta_-
\right)
-1
\right]
\right) .
\label{eq:3R_BIX}$$ We now turn to the calculation of the kinetic term and the diffeomorphism constraints. For this purpose, we define an antisymmetric angular velocity tensor $\omega^{i}{}_{j}$ by the matrix equation $$\boldsymbol{\omega}=\left\{\omega^{i}{}_{j} \right\}=\left(\begin{array}{ccc}
0 & \omega^{1}{}_{2} & -\omega^{3}{}_{1} \\
-\omega^{1}{}_{2} & 0 & \omega^{2}{}_{3} \\
\omega^{3}{}_{1} & -\omega^{2}{}_{3} & 0
\end{array} \right) \equiv O^{T}\dot{O}.$$ An explicit calculation of the right-hand side gives, using , $$\begin{aligned}
\omega^{2}{}_{3} = & \cos(\psi) \dot{\phi} +\sin(\psi)\sin(\phi)\dot{\theta} \ ,
\\
\omega^{3}{}_{1} = & \sin (\psi)\dot{\phi}-\cos(\psi)\sin(\phi)\dot{\theta} \ ,
\\
\omega^{1}{}_{2} = & \dot{\psi}+\cos(\phi)\dot{\theta} \ .\end{aligned}$$ The Lagrangian in the gauge $N^i=0$ then takes the form $$\text{L}=N {\textnormal{e}}^{3\alpha}\left[\frac{-\dot{\alpha}^2+\dot{\beta}^2_+ + \dot{\beta}^2_-
+ I_1 \left(\omega^{2}{}_{3}\right)^2
+ I_2 \left(\omega^{3}{}_{1}\right)^2
+ I_3 \left(\omega^{1}{}_{2}\right)^2
}{2N^2}
+ \frac{{}^{(3)} R}{12}\right] ,$$ where the ‘moments of inertia’ are given by $$3 I_1 \equiv\sinh^2 \left( 3\beta_+ - \sqrt{3}\beta_- \right) , \quad
3 I_2 \equiv \sinh^2 \left( 3\beta_+ + \sqrt{3}\beta_- \right) , \quad
3 I_3 \equiv \sinh^2 \left( 2\sqrt{3}\beta_- \right) .$$ Note, in particular, that the term $\frac{1}{2}\left[ I_1 \left(\omega^{2}{}_{3}\right)^2
+ I_2 \left(\omega^{3}{}_{1}\right)^2
+ I_3 \left(\omega^{1}{}_{2}\right)^2 \right]$ would formally correspond to the rotational energy of a rigid body if the moments of inertia were constant. The canonical momenta conjugate to the Euler angles are given by $$\begin{aligned}
p_{\theta}
& =
\frac{{\textnormal{e}}^{3\alpha}}{N}\left[
I_1 \sin(\psi )\sin (\phi)\omega^{2}{}_{3}-I_2\cos(\psi )\sin (\phi)\omega^{3}{}_{1}
+I_3 \cos(\theta)\omega^{1}{}_{2}
\right],
\\
p_{\phi}
& =
\frac{{\textnormal{e}}^{3\alpha}}{N}\left[
I_1 \cos (\psi)\omega^{2}{}_{3}
+I_2 \sin (\psi)\omega^{3}{}_{1}
\right],
\\
p_{\psi}
& =
\frac{{\textnormal{e}}^{3\alpha}}{N}\ I_3 \omega^{1}{}_{2}.
\end{aligned}$$ It is convenient to introduce the following (non-canonical) angular momentum like variables: $$l_1 \equiv\frac{{\textnormal{e}}^{3\alpha}}{N}I_1 \omega^{2}{}_3 , \quad
l_2 \equiv\frac{{\textnormal{e}}^{3\alpha}}{N}I_2 \omega^{1}{}_3 \quad \text{and} \quad
l_3 \equiv\frac{{\textnormal{e}}^{3\alpha}}{N}I_3 \omega^{2}{}_1 .$$ The relation to the canonical momenta can now explicitely be given by $$\begin{gathered}
p_{\theta}
=
\sin (\psi)\sin(\phi) l_1 - \cos(\psi)\sin(\phi) l_2 +\cos(\phi) l_3 \ ,
\\
p_{\phi}
=
\cos(\psi) l_1 +\sin(\psi) l_2 \ ,
\\
p_{\psi}
=
l_3 \ .
\end{gathered}$$ It is readily shown that the variables $l_i$ obey the Poisson bracket algebra $
\left\{ l_i,l_j \right\} =-C^k_{ij}{}l_k
$. After the usual Legendre transform, we obtain the Hamiltonian constraint, $$\mathcal{H} =
\frac{{\textnormal{e}}^{-3\alpha}}{2}\left(-p_\alpha^2 +p_+^2 +p_-^2
+\frac{l_1^2}{I_1}+\frac{l_2^2}{I_2}+\frac{l_3^2}{I_3} -
\frac{{\textnormal{e}}^{6\alpha}}{6} {}^{(3)}R \right) .$$ From (\[eq:SEH\_BIX\]), we find that the diffeomorphism constraints ($\partial L/\partial N^i=0$) can be written as $$\mathcal{H}_i = 2C^{j}{}_{il} h_{jk}p^{kl} ,$$ where $$p^{ij}=\frac{\sqrt{h}}{24N}\left( h^{ik}h^{jl}-h^{ij}h^{kl}
\right)K_{kl}$$ is the ADM momentum. From this expression we can finally compute the diffeomorphism constraints in terms of the angular momentum-like variables and obtain $$\label{momentum-constraints}
\mathcal{H}_i=O_{i}{}^{j}l_j,$$ that is, we can identify the diffeomorphism constraints with a basis of the generators of $SO(3,\mathbb{R})$. The full gravitational Hamiltonian then reads $$H=N\mathcal{H}+N^i\mathcal{H}_i.$$ From the diffeomorphism constraints we conclude that in the vacuum case $l_i=0$ and that therefore no rotation is possible, that is, we recover the diagonal case. If we want to obtain a Bianchi IX universe with rotating principal axes, we are thus forced to add matter to the system. A formalism for obtaining equations of motion for general Bianchi class A models filled with fluid matter was developed by Ryan [@Ryan_HC]. For simplicity, we will only consider the case of dust as discussed by Kuchař and Brown in [@Kuchar_Brown_1995]. If we were, for example, interested in the study of the quantum version of this model, it would be desirable to introduce a fundamental matter field instead of an ideal fluid. Standard scalar fields alone cannot lead to a rotation for Bianchi IX models. The easiest way to achieve this is, to our knowledge, the introduction of a Dirac field [@Damour_Spindel_2011].
Adding dust to the system
-------------------------
The energy momentum tensor for dust reads $T_{\mu\nu} = \rho u_\mu u_\nu$. The local energy conservation $\nabla_\mu T^{\mu\nu}=0$ leads to a geodesic equation for the positions of the dust particles. Let us start therefore by considering the geodesic equation for a single dust particle, whose four-velocity we can express in the non-coordinate frame $\sigma^i$ used above by the Pfaffian form $$\mathbf{u}=u_0 {\textnormal{d}}t + u_i \sigma^i
\quad \text{with} \ \
\langle \mathbf{u},\mathbf{u} \rangle=-1\ .
\label{eq:BIX_dust_form1}$$ We partially fix the gauge by setting $N^i=0$. The normalization condition implies $$u_0=- N \sqrt{1+h^{ij}u_i u_j}.
\label{eq:u_0}$$ We have chosen here the minus sign because this guarantees that the proper time in the frame of the dust particle has the same orientation as the coordinate time $t$. The geodesic equation for the spatial components of the four velocity can then be written as $$u^0(\partial_t u_i) - C^k_{ij}u_k u^j = 0 \ .
\label{eq:u_geodesic}$$ The geodesic equation implies the existence of a constant of motion. To see this explicitly, we compute the expression $\sum\limits_{i=1,2,3}\dot{u}_i u_i$ and convince ourselves that it vanishes identically. Thus the Euclidean sum $$C^2\equiv(u_1)^2+(u_2)^2+(u_3)^2
\label{eq:u_geodesic_constant}$$ is a constant of motion. Defining $\vec{u}\equiv(u_1, \ u_2, \ u_3)^{T}$, the geodesic equation (\[eq:u\_geodesic\]) can be rewritten in vector notation, $$\partial_t \vec{u}= \frac{N \left[ \vec{u}\times (O \bar{h}^{-1} O^T\vec{u})\right] }{\sqrt{1+\vec{u}^T O \bar{h}^{-1} O^T \vec{u}}} \ ,$$ where “$\times$” denotes the usual cross product in the three dimensional Euclidean space. Defining for convenience $\vec{v}\equiv O^T \vec{u}/C$, we have $(v_1)^2+(v_2)^2+(v_3)^2 = 1$, and the geodesic equation simplifies to $$\left( \partial_t+\boldsymbol{\omega} \right)\vec{v} \ =\frac{ N C \left[ \vec{v}\times ({{\bar{h}}}^{-1}\vec{v} )\right]}{\sqrt{1+C^2 \vec{v}^T {{\bar{h}}}^{-1}\vec{v} }}
\label{eq:v_geodesic} .$$ Note that we can also write $\boldsymbol{\omega}\vec{v}=\vec{v}\times\vec{\omega}$, where $$\vec{\omega}\equiv\{ \omega^i \}= \frac{N}{{\textnormal{e}}^{3\alpha}} \left(
\frac{l_1}{I_1},\ \frac{l_2}{I_2} ,\ \frac{l_3}{I_3} \right)^T.$$ It will thus be possible to eliminate $\boldsymbol{\omega}$ from the geodesic equation by using the diffeomorphism constraints.
We now add homogeneous dust to the system. The formalism developed in [@Kuchar_Brown_1995] leads to the following form of the Hamiltonian and diffeomorphism constraints for dust in a Bianchi IX universe, $$\begin{aligned}
&\mathcal{H} + \mathcal{H}^{(m)}
=
\frac{{\textnormal{e}}^{-3\alpha}}{2}\left(-p_\alpha^2 +p_+^2 +p_-^2
+\frac{l_1^2}{I_1}+\frac{l_2^2}{I_2}+\frac{l_3^2}{I_3}
- \frac{{\textnormal{e}}^{6\alpha}}{6} {}^{(3)}R
+ 2{\textnormal{e}}^{3\alpha} p_T \sqrt{1+h^{ij}u_i u_j}
\right),
\\
&\mathcal{H}_i +\mathcal{H}^{(m)}_i
= O_i{}^j\left( l_j - C p_T v_j \right) ,
\end{aligned}
\label{eq:Hamiltonian+diffeo_constraints_dust}$$ where $p_T$ denotes the momentum canonically conjugate to $T$, where $T$ is the global ‘dust time’. Since the Hamiltonian does not explicitly depend on $T$, the momentum $p_T$ is a constant of motion. The fact that $l_1^2+l_2^2+l_3^2$ commutes with $\mathcal{H}$ implies that $l_1^2+l_2^2+l_3^2=(Cp_T)^2$ is a conserved quantity. This is consistent with . We note that a similar form of the constraints was already presented in [@Ryan_1971a; @Ryan_1971b]. The formalism is not entirely canonical and must be complemented by the geodesic equation (\[eq:u\_geodesic\]).
For our numerical purposes, it will be convenient to rewrite the equations of motion in the variables $\Gamma_i$ introduced in . We find that the choice of the variables $\log \Gamma_i$ allow for a better control over the error in the Hamiltonian constraint. Moreover, we pick the quasi-Gaussian gauge $N={\textnormal{e}}^{3\alpha}=\sqrt{\Gamma_1 \Gamma_2 \Gamma_3}$, $N^i=0$. Recall that the singularity is reached in a finite amount of comoving time (corresponding to the gauge $N=1$). The choice $N={\textnormal{e}}^{3\alpha}$ allows to resolve the oscillations in the approach towards the singularity. With these choices the Hamiltonian constraint becomes $$\begin{aligned}
&-(\log\Gamma_1)^{\cdot}(\log\Gamma_2)^{\cdot}-(\log\Gamma_2)^{\cdot}(\log\Gamma_3)^{\cdot}-(\log\Gamma_1)^{\cdot}(\log\Gamma_3)^{\cdot}
\\
&+\Gamma_1^2+\Gamma_2^2+\Gamma_3^2-2(\Gamma_1\Gamma_2+\Gamma_3\Gamma_1+\Gamma_2\Gamma_3)
\\
&+ 24 \left[\frac{l_1^2}{I_1}+\frac{l_2^2}{I_2}+\frac{l_3^2}{I_3}
+ 2 |p_T|\sqrt{\Gamma_1\Gamma_2\Gamma_3+ C^2 \left(\Gamma_2\Gamma_3 v_1^2+\Gamma_1\Gamma_3v_2^2+\Gamma_1\Gamma_2 v_3^2\right) } \right] = 0,
\end{aligned}$$ where the moments of inertia are $$I_1 = \frac{(\Gamma_3-\Gamma_2)^2}{12\Gamma_3\Gamma_2} ,
\quad
I_2 = \frac{(\Gamma_1-\Gamma_3)^2}{12\Gamma_1\Gamma_3} ,
\quad
I_3 = \frac{(\Gamma_1-\Gamma_2)^2}{12\Gamma_1\Gamma_2} .$$ The diffeomorphism constraints read $l_i=p_T C v_i $ and can be used to eliminate the angular momentum variables from the equations of motion. These equations can then be written as $$\begin{aligned}
(\log \Gamma_1)^{\cdot \cdot}
=&
(\Gamma_2-\Gamma_3)^2-\Gamma_1^2 +
2p_T'^2C^2\left[
\frac{\Gamma_1\Gamma_3(\Gamma_1+\Gamma_3)v_2^2}{(\Gamma_1-\Gamma_3)^3}
+\frac{\Gamma_1\Gamma_2(\Gamma_1+\Gamma_2)v_3^2}{(\Gamma_1-\Gamma_2)^3} \right]
\\
& +
\frac{p_T'(\Gamma_1 \Gamma_2 \Gamma_3+2 C^2 v_1^2 \Gamma_2\Gamma_3)}
{\sqrt{\Gamma_1\Gamma_2\Gamma_3+ C^2 \left(\Gamma_2\Gamma_3 v_1^2+\Gamma_1\Gamma_3v_2^2+\Gamma_1\Gamma_2 v_3^2\right) }},
\\
(\log \Gamma_2)^{\cdot \cdot}
= &
(\Gamma_3-\Gamma_1)^2-\Gamma_2^2 +
2p_T'^2C^2\left[
\frac{\Gamma_1\Gamma_2(\Gamma_1+\Gamma_2)v_3^2}{(\Gamma_2-\Gamma_1)^3}
+\frac{\Gamma_2\Gamma_3(\Gamma_2+\Gamma_3)v_1^2}{(\Gamma_2-\Gamma_3)^3} \right]
\\
& +
\frac{p_T'(\Gamma_1 \Gamma_2 \Gamma_3+2 C^2 v_2^2 \Gamma_1\Gamma_3)}
{\sqrt{\Gamma_1\Gamma_2\Gamma_3+ C^2 \left(\Gamma_2\Gamma_3 v_1^2+\Gamma_1\Gamma_3v_2^2+\Gamma_1\Gamma_2 v_3^2\right) }},
\\
(\log \Gamma_3)^{\cdot \cdot}
=&
(\Gamma_1-\Gamma_2)^2-\Gamma_3^2 +
2p_T'^2C^2\left[
\frac{\Gamma_1\Gamma_3(\Gamma_1+\Gamma_3)v_2^2}{(\Gamma_3-\Gamma_1)^3}
+\frac{\Gamma_3\Gamma_2(\Gamma_3+\Gamma_2)v_1^2}{(\Gamma_3-\Gamma_2)^3} \right]
\\
& +
\frac{p_T'(\Gamma_1 \Gamma_2 \Gamma_3+2 C^2 v_3^2 \Gamma_1\Gamma_2)}
{\sqrt{\Gamma_1\Gamma_2\Gamma_3+ C^2 \left(\Gamma_2\Gamma_3 v_1^2+\Gamma_1\Gamma_3v_2^2+\Gamma_1\Gamma_2 v_3^2\right) }},
\end{aligned}
\label{eq:BIX_nondiag_eoms_rot_dust}$$ where we have set $p_T'\equiv12p_T$ for convenience. Note that these equations are exact. (In [@BKL1], the matter terms were neglected.) We use the diffeomorphism constraints to eliminate $\vec{\omega}$ from the geodesic equation (\[eq:v\_geodesic\]).
If expressed in the gauge $N={\textnormal{e}}^{3\alpha}$ and using the $\Gamma_i$, the geodesic equation can be written as $$\begin{aligned}
\dot{\vec{v}}= & C\vec{v}\times \left( M\vec{v} \right) \quad \text{where}
\\
M = &\frac{\text{diag}\left(\Gamma_2 \Gamma_3, \ \Gamma_1 \Gamma_3,\ \Gamma_1 \Gamma_2\right)}{\sqrt{\Gamma_1\Gamma_2\Gamma_3+ C^2 \left(\Gamma_2\Gamma_3 v_1^2+\Gamma_1\Gamma_3v_2^2+\Gamma_1\Gamma_2 v_3^2\right) }} \\
&+ p_T'
\text{diag}\left( \frac{\Gamma_2\Gamma_3}{\left[\Gamma_2-\Gamma_3 \right]^2} ,
\frac{\Gamma_1\Gamma_3}{\left[\Gamma_3-\Gamma_1 \right]^2},
\frac{\Gamma_1\Gamma_2}{\left[\Gamma_1-\Gamma_2 \right]^2}\right) .
\end{aligned}
\label{eq:BIX_geodesic_final}$$ Together with the constraint $v_1^2+v_2^2+v_3^2=1$ this is all we need for numerical integration. Note that all dependence on the Euler angles and their momenta has dropped out from the equations of motion (\[eq:BIX\_nondiag\_eoms\_rot\_dust\],\[eq:BIX\_geodesic\_final\]). The numerical method we use is described in Appendix \[App:Numerical\_analysis\].
The tilted dust case
--------------------
A qualitative picture for the dynamics of the universe can be obtained by considering the Hamiltonian (\[eq:Hamiltonian+diffeo\_constraints\_dust\]) in the quasi-Gaussian gauge $N={\textnormal{e}}^{3\alpha}$, $N_i=0$. The Hamiltonian can then be interpreted as the “relativistic energy” of a point particle (called the universe point) with “spacetime coordinates” $(\alpha,\beta_+,\beta_-)$. The universe point is subject to the forces generated by the dynamical potential $$\frac{l_1^2}{I_1}+\frac{l_2^2}{I_2}+\frac{l_3^2}{I_3}
- \frac{{\textnormal{e}}^{6\alpha}}{6} {}^{(3)}R
+ 2{\textnormal{e}}^{3\alpha} p_T \sqrt{1+h^{ij}u_i u_j} \ ,$$ which is depicted in Fig. \[fig:BIX\_potential\_gen\]. The contour lines of the curvature potential $-\frac{{\textnormal{e}}^{6\alpha}}{12} {}^{(3)}R$ are represented by solid black lines. The curvature potential is exponentially steep and takes its minimum at the origin $\beta_\pm=0$. When the universe evolves towards the singularity ($\alpha\rightarrow-\infty$), the curvature potential walls move away from the origin while becoming [*effectively*]{} hard walls in the vicinity of the singularity. The term ${\textnormal{e}}^{3\alpha} p_T \sqrt{1+h^{ij}u_i u_j} $ can be interpreted as three rotational potential walls. These potentials are rather unimportant in the examination of the asymptotic dynamics, since they move away from the origin with unit speed. The term $\frac{l_1^2}{I_1}+\frac{l_2^2}{I_2}+\frac{l_3^2}{I_3}$ can be interpreted as three singular centrifugal potential walls. They are represented by the dashed red lines. Asymptotically close to the singularity, these walls are expected to become static. In general, however, the centrifugal potential walls are dynamical and change in a complicated manner dictated by the geodesic equation (\[eq:BIX\_geodesic\_final\]). The centrifugal walls will prevent the universe point from penetrating certain regions of the configuration space. Misner [@Misner_1969b] and Ryan [@Ryan_1971a; @Ryan_1971b] employed these facts to obtain approximate solutions in a diagrammatic form. The other Bianchi models can be treated in a similar way [@Jantzen:2001me].
![Potential picture for the tumbling case. The black contour lines represent the curvature potential while the dotted (blue) and dashed (red) lines represent the rotation and centrifugal walls, respectively.[]{data-label="fig:BIX_potential_gen"}](BIX_potential_gen.pdf){width="7.0cm"}
### Special classes of solutions
Before doing the numerics, we comment on particular classes of solutions: one class of solutions is obtained if we choose, for example, the initial conditions $v_1=0=v_2$ and $v_3 = 1$. The geodesic equation (\[eq:BIX\_geodesic\_final\]) implies now that the velocities stay constant in time. This implies that at all times $l_1= 0= l_2$ and $l_3 = p_\psi = p_T C$. This class of solutions is known as the non-tumbling case. Furthermore, there are classes of solutions which are rotating versions of the Taub solution. These solutions should be divided into two subclasses: one class that oscillates between the centrifugal walls and the curvature potential and one class that runs through the valley straight into the singularity. We set $$v_1=v_2=\frac{1}{2} \ , v_3=0 \quad \text{and} \quad \beta_- = 0 .$$ For the $\Gamma_i$ variables it means that $\Gamma_1={\textnormal{e}}^{2\alpha}{\textnormal{e}}^{2\beta_+}=\Gamma_2$ and $\Gamma_3={\textnormal{e}}^{2\alpha}{\textnormal{e}}^{-4\beta_+}$. With this choice we obtain $I_3=0$ and $3I_1=3I_2=\sinh^2\left(3\beta_+ \right)$. Most importantly, the geodesic equation (\[eq:BIX\_geodesic\_final\]) is trivially satisfied, that is, $v_1=v_2= 1/2$ and $v_3=0$ for all times. When setting $C=0$ we obtain the diagonal case which contains the isotropic case of a closed Friedmann universe. The simulation plotted in Fig. \[fig:BIX\_tumbling2\] was performed for the tumbling case, that is, the $v_i$ are chosen to be non-zero.
[0.4]{} ![Numerical simulation of the tumbling case. In the region where $0<t<5\times 10^5$ (arbitrary units) one can see a typical Kasner era. The solution bounces in the valley formed between the curvature and one of the centrifugal potential walls. Increasing $t$ corresponds to evolution towards the singularity.[]{data-label="fig:BIX_tumbling2"}](beta_plane.pdf "fig:"){width="8cm"}
[0.4]{} ![Numerical simulation of the tumbling case. In the region where $0<t<5\times 10^5$ (arbitrary units) one can see a typical Kasner era. The solution bounces in the valley formed between the curvature and one of the centrifugal potential walls. Increasing $t$ corresponds to evolution towards the singularity.[]{data-label="fig:BIX_tumbling2"}](bkl.pdf "fig:"){width="8cm"}
\
### The asymptotic regime close to the singularity
In order to simplify the dynamics of the general case, BKL made two assumptions based on qualitative considerations of the equations of motion. The first assumption states that anisotropy of space grows without bound. This means that the solution enters the regime $$\Gamma_1\gg \Gamma_2\gg \Gamma_3 .
\label{eq:BIX_Gamma_inequality}$$ The ordering of indices is irrelevant. In fact, there are six possible orderings of indices which each correspond to the universe point being constrained to one of the six regions bounded by the rotation and centrifugal walls sketched in Fig. \[fig:BIX\_potential\_gen\]. The region $\Gamma_1 > \Gamma_2> \Gamma_3$ corresponds to the right region above the line $\beta_-=0$ in Fig. \[fig:BIX\_potential\_gen\]. More precisely, the inequality (\[eq:BIX\_Gamma\_inequality\]) means that $$\Gamma_2 / \Gamma_1 \rightarrow 0 \quad \text{and}\quad
\Gamma_3 / \Gamma_2 \rightarrow 0 \ .$$ Our numerical simulations support the validity of this assumption (see the plot of the ratios $\Gamma_2 / \Gamma_1$ and $\Gamma_3 / \Gamma_2$ in Fig. \[fig:ratios\_velocities\]). We provide plots of the two ratios $\Gamma_2/\Gamma_1$, $\Gamma_3/\Gamma_2$ and the velocities $\vec{v}$ in order to provide a sanity check of the approximation we will perform later on.
[0.4]{} ![Plots of the ratios $\Gamma_2/\Gamma_1$, $\Gamma_3/\Gamma_2$ and the velocity $\vec{v}$. The plots correspond to the numerical solution presented in Fig. \[fig:BIX\_tumbling2\]. The peaks in the ratios appear during bounces of the universe point with the centrifugal walls. As we can see, the height of these peaks decreases in the evolution towards the singularity.[]{data-label="fig:ratios_velocities"}](ratios.pdf "fig:"){width="8cm"}
[0.4]{} ![Plots of the ratios $\Gamma_2/\Gamma_1$, $\Gamma_3/\Gamma_2$ and the velocity $\vec{v}$. The plots correspond to the numerical solution presented in Fig. \[fig:BIX\_tumbling2\]. The peaks in the ratios appear during bounces of the universe point with the centrifugal walls. As we can see, the height of these peaks decreases in the evolution towards the singularity.[]{data-label="fig:ratios_velocities"}](velocities.pdf "fig:"){width="8cm"}
The second assumption made by BKL states that the Euler angles assume constant values: $$(\theta,\phi,\psi)\rightarrow (\theta_0,\phi_0,\psi_0) \ ,$$ that is, the rotation of the principal axes stops for all practical purposes and the metric becomes [*effectively*]{} diagonal. The analysis of BKL [@bkl] supports the consistency of making both assumptions at the same time. [Similar heuristic considerations can possibly be applied to other Bianchi models as well [@Jantzen:2001me]]{}. In the dust model under consideration, this assumption is equivalent to the statement that the dust velocities $\vec{v}$ assume constant values $\vec{v}\rightarrow \vec{v}^{(0)}$. Our numerical results indicate that this is in fact true (see Fig. \[fig:ratios\_velocities\]). BKL then arrive at the simplified effective set of equations.
Let us now carry out the approximation and apply it to our equations of motion. The kinetic term stays untouched during the approximation. The first step in the approximation is to ignore the rotational potential. In view of the strong inequality (\[eq:BIX\_Gamma\_inequality\]), we approximate the curvature potential via $$\Gamma_1^2+\Gamma_2^2+\Gamma_3^2-2(\Gamma_1\Gamma_2+\Gamma_3\Gamma_1+\Gamma_2\Gamma_3)
\approx \Gamma_1^2 \ .$$ Furthermore, we approximate the centrifugal potential by $$\frac{l_1^2}{I_1}+\frac{l_2^2}{I_2}+\frac{l_3^2}{I_3} \approx 12C^2 p_T^2 \left[
\frac{\Gamma_3}{\Gamma_2}\left( v^{(0)}_1 \right)^2+\frac{\Gamma_2}{\Gamma_1}\left( v^{(0)}_3 \right)^2\right] \ .$$ Note that one centrifugal wall was ignored completely. Having Fig. \[fig:BIX\_potential\_gen\] in mind, this approximation is well motivated since only two of the centrifugal walls are expected to have a significant influence on the dynamics of the universe point. After defining the new variables $$a\equiv\Gamma_1\ , \quad
b\equiv2p_T'^2C^2\left( v^{(0)}_3\right)^2\Gamma_2 ,\quad
c\equiv4p_T'^4C^4\left( v^{(0)}_1 v^{(0)}_3\right)^2\Gamma_3 ,$$ we arrive at a simplified Hamiltonian constraint and equations of motion, $$\begin{aligned}
(\log a)^{\cdot}(\log b)^{\cdot }
+ (\log a)^{\cdot} (\log c)^{\cdot}
+ (\log b)^{\cdot} (\log c)^{\cdot} =
a^2 + b/a +c/b ,
\\
(\log a)^{\cdot \cdot}
= b /a - a^2 ,
\quad
(\log b)^{\cdot \cdot}
= a^2 - b/a + c/b ,
\quad
(\log c)^{\cdot \cdot} = a^2 - c/b ,
\end{aligned}
\label{eq:BIX_nondiag_asympt}$$ which coincides with the asymptotic form of equations obtained in [@bkl]. Equations (\[eq:BIX\_nondiag\_asympt\]) can now be treated by the numerical methods which we have used in the previous sections. One must ensure that initial conditions are chosen such that the simulation starts close to the asymptotic regime (\[eq:BIX\_Gamma\_inequality\]).
Conclusions
===========
The numerical simulations indicate that the non-diagonal Bianchi IX solutions, with tilted dust, evolve into the regime where $\Gamma_1 \gg \Gamma_2 \gg \Gamma_3$ and $v_i\approx$const. The results motivate us to formulate the conjecture:
[*Given a tumbling solution to the general Bianchi IX model filled with pressureless tilted matter, there exists $t_0\in \mathbb{R}$ such that the solution is well approximated by a solution to the asymptotic equations of motion for all times $t>t_0$ describing the vicinity of the singularity.*]{}
To make the notion of “approximation” mathematically more precise, a suitable measure of the “distance” on the set of solutions is needed. For this purpose, we propose to use the following simple measure: $$\Delta (t)\equiv\sqrt{\left(\log \Gamma_1 (t) - \log \bar{a} (t) \right)^2
+\left(\log \Gamma_2 (t) - \log \bar{b} (t)\right)^2
+\left(\log \Gamma_3 (t) - \log \bar{c} (t) \right)^2} ,
\label{eq:error}$$ where $\{\Gamma_1,\Gamma_2,\Gamma_3\} $ denotes the numerical solution to the exact equations of motion (\[eq:BIX\_nondiag\_eoms\_rot\_dust\])–(\[eq:BIX\_geodesic\_final\]), and $$a=\bar{a} \ , \quad
b=2p_T'^2C^2\left( v^{(0)}_3\right)^2\bar{b} \ ,\quad
c=4p_T'^4C^4\left( v^{(0)}_1 v^{(0)}_3\right)^2\bar{c} $$ denote the numerical solution to the asymptotic equations of motion (\[eq:BIX\_nondiag\_asympt\]).
We have evolved the exact system of equations from $t=0$ forward in time until $t=3 \times 10^6$. There we used the same initial conditions as the ones we used to obtain the solution shown in Fig. \[fig:BIX\_tumbling2\]. We then took the final state at $t=3 \times 10^6$ as an initial condition for the asymptotic system of equations and evolved it backwards in time towards the re-bounce until $t=-980$.
Fig. \[fig:BIX\_Error\] presents the measure (\[eq:error\]) as a function of time. We can see fast decrease of $\Delta$ with increasing time (evolution towards the singularity) and fast increase of $\Delta$ with decreasing time (evolution away from the singularity).
[0.45]{} ![The difference between the exact and the asymptotic solutions: (a) evolution towards the singularity, (b) evolution away from the singularity.[]{data-label="fig:BIX_Error"}](Error1.pdf "fig:"){width="8cm"}
[0.45]{} ![The difference between the exact and the asymptotic solutions: (a) evolution towards the singularity, (b) evolution away from the singularity.[]{data-label="fig:BIX_Error"}](Error2.pdf "fig:"){width="8cm"}
Our numerical simulations give strong support to the conjecture concerning the asymptotic dynamics of the general Bianchi IX spacetime put forward long ago by Belinski, Khalatnikov, and Ryan [@bkl]. We remark that approximating the diagonal or non-tumbling case by the asymptotic dynamics (\[eq:BIX\_nondiag\_asympt\]) is invalid (see [@ENW] for more details).
It is sometimes stated that “matter does not matter” in the asymptotic regime, but this does not mean that one is allowed to use the dynamics of the purely diagonal case. One only encounters an [ *effectively*]{} diagonal case, which is expressed in terms of the directional scale factors $\{a,b,c\}$. So there exist serious differences between the purely diagonal and effectively diagonal cases (see [@ENW] for more details).
Employing the asymptotic form of the equations of motion may enable one to study the chaotic behaviour and other properties of the solution space for the general model. This is also important for quantizing the general Bianchi IX model, where the quantization of the exact dynamics seems to be quite difficult, whereas the quantization of the asymptotic case seems to be feasible [@AGWP].
We are grateful to Vladimir Belinski and Claes Uggla for helpful discussions. This work was supported by the German-Polish bilateral project DAAD and MNiSW, No. 57391638, “Model of stellar collapse towards a singularity and its quantization”.
Numerical analysis {#App:Numerical_analysis}
==================
Numerical simulations of the diagonal Bianchi IX model were already carried out in the late 1980s and early 1990s (see e.g. [@Berger_1990; @Hobill_1991]; for a modern account, see [@Berger]). Our main interest here is in the nondiagonal case. With given initial conditions (respecting the constraints $\mathcal{H}=0$), the system (\[eq:BIX\_nondiag\_eoms\_rot\_dust\]) and (\[eq:BIX\_geodesic\_final\]) can be integrated by using a suitable numerical method. In this work we employ the MATLAB R2016b solver ode113 [@matlab_ode]. This code is an implementation of an Adams-Bashforth-Moulton method. It turns out to lead to the best results when compared to other MATLAB solvers. The relative error tolerance of the solver was chosen to be of the order $10^{-14}$. We can integrate the equations of motion together with the geodesic equation to obtain a numerical solution to the system. We set up initial conditions at $t=0$ and evolve the system forward in time towards the final singularity and away from the rebounce.
A major problem in numerical relativity is that the Hamiltonian constraint is not preserved exactly by the numerical procedure. Similar to [@Berger_1990; @Hobill_1991], we find that the error in the Hamiltonian constraint varies strongest after the start of the simulation. Furthermore, it varies strongly when the evolution of the universe approaches the point of maximal expansion. While approaching the singularity, the error approaches an approximately constant value. This can be seen in Fig. \[fig:error\_Hamiltonian\_constraint\]. Therefore we can minimize the error when we choose the initial conditions far away from the point of maximal expansion. Moreover, it turned out that the error can be further reduced when constraining the solver’s maximally allowed time step size. This time step size should, however, not be chosen too small since small time step sizes can drive the propagation of round off errors. Small step sizes are, of course, also numerically more expensive. By manually fine tuning the initial conditions and the maximally allowed time step size it was possible to keep the order of the error lower than $10^{-15}$.
[0.45]{} ![Subfigure (a) shows the error in the Hamiltonian constraint. Subfigure (b) shows the speed of the universe point in the beta plane as measured in “$\alpha$-time”. This plot can be viewed as another check of the numerics. Between two successive bounces from the potential walls this quantity should be close to one. If this ceased to be true it would indicate that the error in the Hamiltonian constraint becomes relevant and cannot be neglected in the approach towards the singularity (see [@Berger] for a more detailed discussion). Both plots correspond to the numerical solution shown in Fig. \[fig:BIX\_tumbling2\].[]{data-label="fig:error_Hamiltonian_constraint"}](H.pdf "fig:"){width="8cm"}
[0.45]{} ![Subfigure (a) shows the error in the Hamiltonian constraint. Subfigure (b) shows the speed of the universe point in the beta plane as measured in “$\alpha$-time”. This plot can be viewed as another check of the numerics. Between two successive bounces from the potential walls this quantity should be close to one. If this ceased to be true it would indicate that the error in the Hamiltonian constraint becomes relevant and cannot be neglected in the approach towards the singularity (see [@Berger] for a more detailed discussion). Both plots correspond to the numerical solution shown in Fig. \[fig:BIX\_tumbling2\].[]{data-label="fig:error_Hamiltonian_constraint"}](beta_speed.pdf "fig:"){width="8cm"}
Recall that the dynamics of Bianchi IX are chaotic, that is, slightly changing initial conditions have a large effect on the long time behaviour of solutions. Since the propagation of random numerical errors cannot be avoided, we will be dealing with a “butterfly effect” and it should in general not be expected that our numerical solution is an actual approximation of some exact solution of the equations of motion when considering large time intervals.
[10]{}
S. Hawking and R. Penrose, [*The Nature of Space and Time*]{} (Princeton University Press, Princeton, 1996).
C. Kiefer, [*Quantum Gravity*]{}, third edition (Oxford University Press, Oxford, 2012).
I. Albarran, M. Bouhmadi-López, C. Kiefer, J. Marto, and P. V. Moniz, “Classical and quantum cosmology of the little rip abrupt event”, Phys. Rev. D [**94**]{}, 063536 (2016).
A. Kamenshchik, C. Kiefer, and N. Kwidzinski, “Classical and quantum cosmology of Born-Infeld type models”, [*ibid.*]{} [**93**]{}, 083519 (2016).
H. Bergeron, E. Czuchry, J. P. Gazeau, P. Ma[ł]{}kiewicz, and W. Piechocki, “Smooth quantum dynamics of the mixmaster universe,” Phys. Rev. D [**92**]{} 061302 (2015).
H. Bergeron, E. Czuchry, J. P. Gazeau, P. Ma[ł]{}kiewicz, and W. Piechocki, “Singularity avoidance in a quantum model of the mixmaster universe,” Phys. Rev. D [**92**]{}, 124018 (2015).
B. K. Berger, “Singularities in cosmological spacetimes”, in: [*Springer Handbook of Spacetime*]{}, edited by A. Ashtekar and V. Petkov (Springer, Berlin, 2014), pp. 437–460.
C. Uggla, “Spacetime singularities: recent developments”, Int. J. Mod. Phys. D [**22**]{}, 1330002 (2013).
E. Czuchry, D. Garfinkle, J. R. Klauder, and W. Piechocki, “Do spikes persist in a quantum treatment of spacetime singularities?”, Phys. Rev. D [**95**]{}, 024014 (2017).
E. Lifshitz, “On the gravitational stability of the expanding universe”, J. Phys. (USSR) [**10**]{}, 116 (1946). Republished as a Golden Oldie in: Gen. Relativ. Grav. [ **49**]{}, 18 (2017), with an editorial note by G. F. R. Ellis.
E. M. Lifshitz and I. M. Khalatnikov, “Investigations in Relativistic Cosmology”, Adv. Phys. [**12**]{}, 185 (1963).
V. Belinski and M. Henneaux, [*The Cosmological Singularity*]{} (Cambridge University Press, Cambridge, 2017). See also V. A. Belinski, “On the cosmological singularity”, Int. J. Mod. Phys. D [ **23**]{}, 1430016 (2014) for a shorter review.
V. A. Belinskii, I. M. Khalatnikov, and E. M. Lifshitz, “Oscillatory approach to a singular point in the relativistic cosmology”, Adv. Phys. [**19**]{}, 525 (1970).
V. A. Belinskii, I. M. Khalatnikov, and E. M. Lifshitz, “A general solution of the Einstein equations with a time singularity”, Adv. Phys. [**31**]{}, 639 (1982).
N. J. Cornish and J. J. Levin, “The Mixmaster universe: A chaotic Farey tale,” Phys. Rev. D [**55**]{}, 7489 (1997).
J.M. Heinzle, C. Uggla, and N. Rohr, “The cosmological billiard attractor”, Adv. Theor. Math. Phys. [**13**]{}, 293 (2009).
D. Garfinkle, “Numerical simulations of generic singularities”, Phys. Rev. Lett. [**93**]{}, 161101 (2004).
C. W. Misner, “Mixmaster universe”, Phys. Rev. Lett. [**22**]{}, 1071 (1969).
C. W. Misner, “Quantum cosmology I”, Phys. Rev. [**186**]{}, 1319 (1969).
M. P. Ryan, “Qualitative cosmology: diagrammatic solutions for Bianchi type IX universes with expansion, rotation, and shear. I. The symmetric case”, Ann. Phys. (N.Y.) [**65**]{}, 506 (1971).
M. P. Ryan, “Qualitative cosmology: diagrammatic solutions for Bianchi type IX universes with expansion, rotation, and shear. II. The general case”, Ann. Phys. (N.Y.) [**68**]{}, 541 (1971).
R. T. Jantzen, “Spatially homogeneous dynamics: a unified picture”, arXiv:gr-qc/0102035. Originally published in the Proceedings of the International School Enrico Fermi, Course LXXXVI (1982) on [*Gamov Cosmology*]{}, edited by R. Ruffini and F. Melchiorri (North Holland, Amsterdam, 1987), pp. 61–147.
L. Andersson and A. D. Rendall, “Quiescent cosmological singularities”, Commun. Math. Phys. [ **218**]{}, 479 (2001).
J. D. Barrow, “Quiescent cosmology”, Nature [ **272**]{}, 211 (1978).
A. Góźdź, W. Piechocki, and G. Plewa, “Quantum Belinski-Khalatnikov-Lifshitz scenario”, in preparation.
O. I. Bogoyavlenskii, “Some properties of the type IX cosmological model with moving matter”, Sov. Phys. JETP [**43**]{}, 187 (1976).
M. P. Ryan and L. C. Shepley, [*Homogeneous Relativistic Cosmologies*]{} (Princeton University Press, Princeton, 1975).
M. P. Ryan, [*Hamiltonian Cosmology*]{} (Springer, Berlin, 1972).
J. D. Brown and K. V. Kuchar, “Dust as a standard of space and time in canonical quantum gravity”, Phys. Rev. D [**51**]{}, 5600 (1995).
T. Damour and P. Spindel, “Quantum Einstein-Dirac Bianchi universes”, Phys. Rev. D [**83**]{}, 123520 (2011).
V. A. Belinskii, I. M. Khalatnikov, and M. P. Ryan, “The oscillatory regime near the singularity in Bianchi-type IX universes”, Preprint [**469**]{} (1971), Landau Institute for Theoretical Physics, Moscow (unpublished); published as sections 1 and 2 in M. P. Ryan, Ann. Phys. (N.Y.) [**70**]{}, 301 (1971).
E. Czuchry, N. Kwidzinski, and W. Piechocki, “Comparing the dynamics of diagonal and general BIX spacetimes”, in preparation.
B. K. Berger, “Numerical study of initially expanding mixmaster universes”, Class. Quant. Grav. [**7**]{}, 203 (1990).
D. Hobill, D. Bernstein, M. Welge, and D. Simkins, “The Mixmaster cosmology as a dynamical system”, Class. Quant. Grav. [**8**]{}, 1155 (1991).
L. F. Shampine and M. W. Reichelt, “The MATLAB ODE Suite”, SIAM Journal on Scientific Computing [**18**]{}, 1 (1997).
[^1]: This is how the diagonal Bianchi IX model received the name mixmaster universe.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- '[Sophia Yakoubov]{}'
date: |
[`[email protected]`]{}\
[MIT]{}
title: |
\
[Pattern Avoidance in Extensions of Comb-Like Posets]{}
---
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Qing-Dong Jiang'
- 'Zhi-qiang Bao'
- 'Qing-Feng Sun'
- 'X. C. Xie'
date:
-
-
title: Theory for electric dipole superconductivity with an application for bilayer excitons
---
**Exciton superfluid is a macroscopic quantum phenomenon in which large quantities of excitons undergo the Bose-Einstein condensation. Recently, exciton superfluid has been widely studied in various bilayer systems. However, experimental measurements only provide indirect evidence for the existence of exciton superfluid. In this article, by viewing the exciton in a bilayer system as an electric dipole, we provide a general theory for the electric dipole superconductivity, and derive the London-type and Ginzburg-Landau-type equations for the electric dipole superconductors. By using these equations, we discover the Meissner-type effect and the electric dipole current Josephson effect. These effects can provide direct evidence for the formation of the exciton superfluid state in bilayer systems and pave new ways to drive an electric dipole current.**\
Since the idea of excitonic condensation was proposed about fifty years ago[@Keldysh1965; @Lozovik1975], exciton systems have attracted a lot of interest. With the development of micromachining technology in the last two decades, high-quality bilayer exciton systems can be fabricated in the laboratories, in which one layer hosts electrons and the other layer hosts holes.[@science1368; @eisenstein2004] Many new interaction phenomena have been experimentally reported in the bilayer exciton systems,[@kellogg2004vanishing; @Naturephy799; @spielman2000resonantly; @tutuc2004counterflow; @nandi2012exciton; @gorbachev2012strong; @Science99; @titov2013giant; @Nature409; @Nature584] including the vanishing Hall resistance for each layer[@kellogg2004vanishing], the resonantly enhanced zero-bias inter-layer tunneling phenomenon[@spielman2000resonantly], the large bilayer counterflow conductivity[@tutuc2004counterflow], the Coulomb drag[@nandi2012exciton; @gorbachev2012strong; @Science99; @titov2013giant], etc. These phenomena strongly imply the formation of the exciton condensate superfluid state, in which many excitons crowd into the ground state. However, because the exciton is charge neutral, there are still no effective methods to directly confirm the formation of the superfluid state. Thus, whether the superfluid state really forms is still unclear.
Before any further discussion, we need first to point out the specificity of excitons in bilayer systems. Because the electrons and holes are separated in space and bound with each other by the Coulomb interaction, the exciton in a bilayer system can be seen as a charge neutral electric dipole (as shown in Fig. \[fig:1\]a). On the other hand, superconductivity has been one of the central subjects in physics. The superconductor state has several fascinating properties, such as zero resistance[@onnes], the Meissner effect[@meissner], the Josephson effect[@Josephson1962], and so on, which have many applications nowadays[@gennes1989superconductivity]. It is now well known that the superconductor is the condensate superfluid state of the Cooper pairs[@bardeen1957theory], which can be viewed as electric monopoles. In other words, the superconductor state is the electric monopole condensated superfluid state. Thus, it is natural to ask whether the electric dipole superfluid state possesses many similar fascinating properties, just like its counterpart, the electric monopole superfluid state.
In this article, we will construct a general theory of electric dipole superconductivity under an external electromagnetic field, and apply this theory to the bilayer exciton systems, revealing the basic characteristics of the electric dipole superconductors. Apart from the bilayer exciton systems, the electric dipole superconductor may also exist in other two or three dimensional systems, e.g., the Bose-Einstein condensate of ultracold polar molecules[@phystoday64-27; @nature464-1324; @science322-231]. In fact, the ultracold polar molecules have been successfully produced in the laboratory over the past decade[@phystoday64-27; @nature464-1324; @science322-231]. In addition, a new quantum state was proposed recently[@Sun2011; @Sun2013; @Bao2013], a magnetic dipole superconductor named the spin superconductor. Both electric and magnetic dipole superconductors contain some similar properties. Below we first derive the London-type and Ginzburg-Landau-type (GL-type) equations of the electric dipole superconductor. These equations can be applied to all electric dipole superconductors independent of specific systems and we apply them to study various physical properties of the electric dipole superconductor. By using these equations, we find that the Meissner-type effect and the Josephson effect of the electric dipole current. With the Meissner-type effect, a non-uniform external magnetic field can cause a super electric dipole current in an electric dipole superconductor, and a super electric dipole current can generate a magnetic field that is against the spatial variation of the external magnetic field. Considering the bilayer exciton systems, we show that the magnetic field induced by the super dipole current is measurable by today’s technology. We also show that the frequency in the AC Josephson effect of the electric dipole current is equal to that of the AC Josephson effect in the normal superconductor. These new effects discovered in this work can not only provide direct evidence for the existence of the exciton condensate superfluid state in the bilayer systems, but also pave new ways to drive an electric dipole current.
![ **A side view of the exciton in bilayer system and the induced supercurrent by a non-uniform magnetic field.** (**a**) The top and bottom layers host holes and electrons respectively, and the middle blue block stands for the interlayer barrier which prevents tunneling between the two layers. (**b**) The left (right) panel shows the induced super electric dipole current for $\partial B_z/\partial z<0$ ($\partial B_z/\partial
z>0$). The arrows on the blue lines denote the direction of positive charge flow in each layer.\[fig:1\]](Figure-1.eps){height="5.4cm" width="8cm"}
**Results**
**London-type equations of the electric dipole superconductor**
Considering a bosonic electric dipole condensate superfluid state, namely the electric dipole superconductor, under an external electric field, the force on the electric dipole $\bold{p}_0$ is $$\begin{aligned}
\label{eq:1}
\bold{F}=(\bold{p}_0 \cdot\nabla)\bold{E}=m^*\frac{d\bold v}{dt}\end{aligned}$$ which accelerates the dipole. Here, $\bold E$ is the electric field, $m^*$ is the effective mass of the dipole and $\bold v$ is its velocity. The moving electric dipole induces an electric dipole current, which can be described by a tensor $\mathbb{J}_{\bold
p}=n\bold{p_0}\bold{v}$, where $n$ is the dipole density. It should be pointed out that, in this work the direction of the electric dipole vector is fixed in the $z$-direction (as is the case in a bilayer exciton system), thus we only need one vector $\bold
{J_p}=np_0\bold{v}$ to describe the super dipole current density. The derivative of $\bold{J_p}$ with respect to time is $$\begin{aligned}
\label{eq:2}
\frac{d{\bold {J_p}}}{dt}=\beta(\hat p_0 \cdot \nabla )\bold E,\end{aligned}$$ where $\beta=n p_0^2/m^*$ and $\hat p_0 ={\bold e}_z$ is the unit vector in the direction of $\bold p_0$. We can see that, just as $\bold E$ accelerates the super electric current[@gennes1989superconductivity], $(\hat
p_0\cdot\nabla)\bold E$ accelerates the super electric dipole current. Taking curl on both sides of the equation (\[eq:2\]) and using the Maxwell equation[@griffiths] $\nabla\times\bold
E=-\partial\bold B/ \partial t$, we have $\frac{\partial}{\partial
t}(\nabla\times\bold J_p)=-\frac{\partial}{\partial t}[\beta(\hat
p_0\cdot\nabla)\bold B]$. Integrating over the time $t$, we get the London-type equation for $\bold J_p$, $$\begin{aligned}
\label{eq:3}
\nabla \times{\bold J_p}=-\beta (\hat p_0\cdot \bold\nabla)\bold B,\end{aligned}$$ where the integral constant is taken to be zero due to the requirement of thermodynamic equilibrium. Equations (\[eq:2\]) and (\[eq:3\]) play similar roles as the London equations for normal superconductors[@gennes1989superconductivity], so we call them the first and the second London-type equations for the electric dipole superconductor.
Equation (\[eq:3\]) implies that the gradient of a magnetic field ${\bold B}$ will induce a super electric dipole current. As is shown in Fig. \[fig:1\]b, if the gradient of magnetic field $\partial
B_z/\partial z<0$, the super dipole current flows in the counterclockwise direction (left panel of Fig.\[fig:1\]b); if $\partial B_z/\partial z>0$, the super dipole current flows in the clockwise direction (right panel of Fig.\[fig:1\]b). In addition, the super dipole current can also have a feedback for an external magnetic field. The magnetic field generated by a moving electric dipole $\bold p_0$ with velocity $\bold v$ is equivalent to that generated by a static magnetic moment $\bold m=-\bold v\times\bold
p_0$.[@griffiths; @Magneticdipole] As a consequence, the magnetic field induced by the super dipole current $\bold J_p$ is equivalent to that induced by the static magnetic moment distribution (magnetization) $\bold M=n \bold m=-\bold J_p\times \hat p_0$. In materials, the last Maxwell equation takes the form $\nabla\times\bold B=\mu_0(\bold j_f+\nabla\times\bold M+{\partial
\bold D}/{\partial t})$,[@griffiths] where $\bold j_f$ stands for the free electric current and $\bold D$ is the effective electric field. In the equilibrium case, ${\partial \bold
D}/{\partial t} =0$ and with no free electric current present, only the super dipole current exists, so we obtain the magnetic field equation in the electric dipole superconductor: $$\begin{aligned}
\label{eq:4}
\nabla\times\bold B=-\mu_0 \nabla\times(\bold J_p\times \hat p_0).\end{aligned}$$
The London-type equation (\[eq:3\]) and the magnetic field equation (\[eq:4\]) govern the magnetic field and the super dipole current in an electric dipole superconductor. An alternative set of equations that are equivalent to equations (\[eq:3\]) and (\[eq:4\]) are given in Supplementary Note 1. From equations (\[eq:3\]) and (\[eq:4\]), we can obtain the Meissner-type effect against the spatial variation of an external magnetic field, which will be studied below. We can see the effect of equation (\[eq:3\]) by considering a massless Dirac particle for which the factor $\beta\rightarrow\infty$. In this case, the spatial variation of the total magnetic field $\partial_z B_z$ has to vanish everywhere inside an electric dipole superconductor in order to satisfy the equation (\[eq:3\]). This means that $\partial_z B_z$ is completely screened out.
**Ginzburg-Landau-type equations of the electric dipole superconductor**
Since the electric dipole condensate is a macroscopical quantum state, we can use a quasi-wave function (or the order parameter) $\psi(\bold r)$ to describe it. Then its free energy can be written as $F_s$=$\int_V f_s d\bold r$, where $f_s$ is the free energy density. In analogy with the superconductor, $f_s$ can be expressed as: $$\begin{gathered}
\label{eq:5}
f_s =f_n +\alpha(T)|\psi(\bold r)|^2+\frac{\beta(T)}{2}|\psi(\bold r)|^4+ \\
\frac{|(\hat{\bold p} + \bold p_0\times \bold B)\psi(\bold r)|^2}{2m^*}+\frac{|\bold B|^2}{2\mu_0} ,\end{gathered}$$ where $f_n$ is the density of free energy in normal state and the momentum operator ${\bold {\hat p}}=-i\hbar\nabla$. The two terms $\alpha(T)|\psi(\bold r)|^2$ and $\beta(T)|\psi(\bold r)|^4/2$ are the lower order terms in the series expansion of the free energy $f_s$, which have similar meanings as those in the normal superconductor.[@gennes1989superconductivity; @GLequation] Particularly, the gauge invariant term $|(\bold p + \bold
p_0\times\bold B)\psi(\bold r)|^2/2m^*$ can be viewed as the kinetic energy of the electric dipole superconductor (see Supplementary Note 2). Substitute ${\mathbf B}=\nabla\times{\mathbf A}$ and minimize the free energy with respect to $\psi^*$ and the magnetic vector potential $\bold A$ respectively, we get (see Supplementary Note 3) $$\begin{aligned}
&&\mbox{}\hspace{-10mm}\alpha(T)\psi+\beta(T)|\psi|^2\psi+
\frac{[\hat{\bold p}+\bold p_0\times \bold B]^2\psi}{2m^*}=0, \label{eq:6}\\
&&\mbox{}\hspace{-10mm}\nabla\times\bold B=-\mu_0 \nabla\times(\bold J_p\times \hat p_0),\label{eq:7}\end{aligned}$$ where $$\begin{aligned}
\label{eq:8}
\bold J_p=\frac{p_0}{2
m^*}[i\hbar(\psi\nabla\psi^*-\psi^*\nabla\psi)+2\bold p_0\times\bold
B\,|\psi|^2] .\end{aligned}$$ Equations (\[eq:6\]) and (\[eq:8\]) are the first and the second GL-type equations, respectively. They provide a full phenomenological description of the dipole superconductor. Since the velocity operator is $\bold {\hat v}=(\bold {\hat p}+\bold
p_0\times\bold B)/m^*$, the super electric dipole current can be expressed as $p_0 Re(\psi^*\bold{\hat v}\psi)=\frac{p_0}{2
m^*}[i\hbar(\psi\nabla\psi^*-\psi^*\nabla\psi)+2\bold p_0\times\bold
B\,|\psi|^2]$. Comparing the expression $p_0 Re(\psi^*\bold{\hat
v}\psi)$ and equation (\[eq:8\]), we find that $\bold J_p$ is exactly the super dipole current density. It should be noted that equation (\[eq:7\]) is the same as the magnetic field equation (\[eq:4\]). It indicates that the GL-type theory gives a more general result. Next we derive the London-type equations from the GL-type equations. The order parameter $\psi(\bold r)$ can be written as $|\psi(\bold r)|e^{i\theta(\bold r)}$, where $|\psi(\bold
r)|^2$ is proportional to the density of dipoles $n$ and $\theta$ represents the phase. For simplicity, we assume the amplitude $|\psi|$ is the same everywhere in the dipole superconductor, whereas the phase $\theta(\bold r)$ are allowed to change in order to account for the super dipole current. Substitute $\psi=|\psi(\bold r)|e^{i\theta(\bold r)}$ into equation (\[eq:8\]), we get $\bold J_p=\frac{p_0 n}{m^*}(\hbar
\nabla\theta+\bold p_0\times\bold B)$. Furthermore, if we take curl on both sides, the London-type equation (\[eq:3\]) is recovered. It indicates that the London-type equations can be obtained from the GL-type equations, which shows the validity and consistency of our theory.
{height="3.9cm" width="16.9cm"}
**Meissner-type effect of the electric dipole superconductor**
In the following, we will use the London-type equations to analyse the Meissner-type effect. To begin with, we consider a two-dimensional circular dipole superconductor with the radius $r_{out}$ located in a non-uniform external magnetic field $\bold B$ created by a cylindrical hollow conductor with the inner (outer) radius $R_{in}$ ($R_{out}$) and the height $h$ (shown in Fig. \[fig:2\]a). The distance between the cylindrical hollow conductor and the dipole superconductor is $t$. Fig. \[fig:2\]b depicts the cross-section of the device. A uniform electric current along the azimuthal direction in the hollow conductor creates a non-uniform magnetic field with a gradient $\partial_{z}B_z^{ext}$ (see Supplementary Note 4), which can induce a super dipole current in the electric dipole superconductor. Substitute $\partial_z B_z^{ext}$ into the London-type equation (\[eq:3\]), considering the rotational symmetry of the whole device and $\nabla \cdot {\bold
J}_p =0$, we can obtain the super electric dipole current density $J_p$ (flowing in the azimuthal direction): $$\begin{aligned}
\label{eq:9}
J_p(r)=\frac{-\beta}{r}\int_0^r r^{\prime}dr^{\prime}[\partial_z B_z^{ext}(r^{\prime},z)].\end{aligned}$$
In the following, we consider that the electric dipole superconductor is the bilayer exciton system (see Fig. \[fig:1\]a). Then the super dipole current can be viewed as a counterflow electric current in the bilayer as shown in Fig.\[fig:1\]b and thereby induces a magnetic field $\bold
B^{ind}$. Since the counterflow current in bilayer and the rotational symmetry about the $z$ axis, the induced magnetic field $\bold B^{ind}$ in middle plane $m$ only has the nonzero r-component $B^{ind}_r$. Although the z-component $B^{ind}_z =0$, the gradient $\partial_z B_z^{ind}$ does not vanish. Fig. 2d and 2e show $\partial_z B_z^{ind}$ and $B^{ind}_r$ in plane $m$ which can easily be calculated from the Biot-Savart law (see Supplementary Note 5).
In the calculation, we take cylindrical hollow conductor sizes as $R_{in}=1mm$, $R_{out}=1\,cm$ and $h=1.5\,cm$. The current density in conductor $j=10^8 A/m^2$, which generates the non-uniform external magnetic field ${\bold B}^{ext}$. The bilayer exciton specimen, the electric dipole superconductor, is below the conductor with $t=0.1\, mm$ and the radius $r_{out}=1mm$. Here the specimen is just in the hollow region of the conductor, because $\partial_z
B_z^{ext}$ is relatively large there (see Supplementary Note 4). The two-dimensional carrier density in each layer $n$ is chosen $10^{12}cm^{-2}$ and the effective mass of exciton $m^*=0.01 m_e$ with the electron mass $m_e$. Fig. \[fig:2\]c, \[fig:2\]d and \[fig:2\]e show respectively the induced super dipole current density $J_p$, the induced magnetic field variation $\partial_z B_z^{ind}$ and $B_r^{ind}$ versus radius $r$ for bilayer thickness $d=3nm$ and $d=10 nm$. A quite large $J_p$ is induced near the edge of the specimen, in which the corresponding electric current density in each layer near the edge is about $15 A/m$. From Fig. \[fig:2\]d, we find that $\partial_z B_z^{ind}(r)$ counteracts the variation $\partial_z B_z^{ext}(r)$. This is a Meissner-type effect in the dipole superconductor against a spatial variation of a magnetic field. Notice that it is not against the magnetic field. This is the main difference between the dipole superconductor and (monopole) superconductor. Also notice in Fig. \[fig:2\]d, $\partial_z B_z^{ind}(r)$ is much smaller than $\partial_z B_z^{ext}(r)$ because the thickness $d$ of the dipole superconductor is very small now. If for the thick dipole superconductor or for very small $m^*$, $\partial_z B_z^{ind}(r)$ can almost be of the same value as $\partial_z B_z^{ext}(r)$, then the spatial variation of the total magnetic field vanishes inside of the dipole superconductor. Fig. \[fig:2\]e shows the induced magnetic field $B_r^{ind}$ by the super dipole current, which can reach about 0.05 Gauss. This magnetic field can be accurately detected by the today’s technology. In addition, a recent work has successfully used the SQUID to detect a tiny edge current (around $0.5 A/m$) in the Hall specimen[@nowack2013imaging]. In our case, the edge current density is around $10 A/m$, so it should be detectable using the same method.
{height="3.9cm" width="16.9cm"}
**The detection of the zero dipole resistance**
The most remarkable phenomenon of the Bose-Einstein condensate macroscopic quantum system is superfluid, e.g. the zero resistance phenomenon of the (monopole) superconductor[@onnes; @gennes1989superconductivity]. For the dipole superconductor, the dipole resistance is zero, i.e. the electric dipole can flow without dissipation. In the following, we suggest a method to detect the zero dipole resistance.
From the first London-type equation (\[eq:2\]), we know that a variation of an electric field $\partial_z \bold E$ can excite a super dipole current. This excited super dipole current will maintain for a very long time if the dipole resistance is zero. Now, consider an annular dipole superconductor specimen. This annular specimen is placed below the cylindrical hollow conductor (see Fig. \[fig:3\]a and Fig. \[fig:3\]b), and there $\partial_z
B_z^{ext}$ is relatively small and $\partial_z A_{\theta}^{ext}$ is quite large[@simpson] (see Supplementary Note 4). First, let the hollow conductor have an azimuthal electric current $j$ and then cool the specimen into the dipole superconductor state. Next, we abruptly turn off the current $j$ in the conductor. In this process, an azimuthal dipole current ${J_p}$ will be excited. From the equation (\[eq:2\]) and $\bold E=-\frac{\partial \bold
A}{\partial t}$, we have $d J_p/dt = \beta
\partial_z E_{\theta}^{ext} = -\frac{\partial}{\partial t}[\beta \partial_z
A_{\theta}^{ext}]$. Integrating over the time $t$, we obtain the excited super dipole current: $$\begin{aligned}
\label{eq:10}
J_p = \beta \partial_z A_{\theta}^{ext},\end{aligned}$$ where the vector potential $A_{\theta}^{ext}$ is that before the current in the conductor was turned off. Here we have used $A_{\theta}^{ext}=0$ after the current turned off and $J_p$ is almost zero at the beginning. Fig. \[fig:3\]c and \[fig:3\]d show the excited super dipole current $J_p$ and the magnetic field $B_r^{ind}$ induced by $J_p$. From which we find that $J_p$ and $B_r^{ind}$ are quite large. Moreover, this excited super dipole current $J_p$ does not decay for a very long time because of the zero dipole resistance. So one can measure the non-decayness of $J_p$ to confirm the zero dipole resistance.
{height="3cm" width="12.9cm"}
**DC and AC electric dipole current Josephson effect**
The Josephson effect is another highlight of the superconductor[@Josephson1962]. Below we use the GL-type equations of the electric dipole superconductor to discuss the dipole current Josephson effect in a dipole superconductor-insulator-dipole superconductor junction. From the GL-type equations (\[eq:6\]) and (\[eq:8\]) we can see that if ${\mathbf B}=0$, they are the same as the GL equations of the superconductor when ${\mathbf A}=0$. Therefore, the DC Josephson effect of the dipole superconductor and the (monopole) superconductor are similar, [*i.e.*]{}, the super electric dipole current is $j=j_{0}\sin\gamma_{0}$, where $\gamma_{0}=\gamma_{1}-\gamma_{2}$. $j_{0}$ is the Josephson critical super dipole current, and $\gamma_{1}$, $\gamma_{2}$ are phases of the dipole superconductors.
Now we study the AC dipole current Josephson effect and consider an electric field variation $\partial_z E_x$. From the first GL-type equation (\[eq:6\]), we can get the change of the phase when the super dipole current (in $x$-direction) passes through the Josephson junction. Its expression is $\gamma=\gamma_{0}+\frac{1}{\hbar}\int_{1}^{2}({\mathbf
p}_{0}\times{\mathbf B})\cdot{\mathbf e}_{x}dx$. Taking the derivative with respect to time, we get $\frac{\partial\gamma}{\partial
t}=\frac{p_{0}}{\hbar}\int_{1}^{2}({\mathbf
e}_{z}\times\frac{\partial{\mathbf B}}{\partial t})\cdot{\mathbf
e}_{x}dx$. Substituting $\frac{\partial{\mathbf B}}{\partial
t}=-\nabla\times{\mathbf E}$, we have $\gamma=\gamma_0 -\omega_0 t$ with $\omega_0 =-\frac{p_0 }{\hbar}\int_{1}^{2}\partial_z E_xdx$. As a result, the super dipole current can be written as $j=j_0 \sin
(\gamma-\omega_0 t)$. It shows that the dipole current is an alternating current, although the electric field spatial variation is time-independent. We can compare this with the (monopole) superconductor. For the superconductor, a time-independent electric field (or bias) can drive an AC Josephson current. Now a spatial variation of an electric field drives an AC dipole Josephson current.
Next, we consider that the dipole superconductor is the bilayer exciton system. Fig. \[fig:4\]a shows the schematic diagram for the device of dipole current Josephson junction. A thin wire connects the left sides of the two layers, which enables the current to flow between them. Then, if we apply the voltages $-V_2$ and $V_2$ to the right sides between the bilayer, it establishes an electric field in $-x$ ($x$) direction in the bottom (top) layer (see Fig. \[fig:4\]a). This means that a spatial variation of electric field, $\partial_z E_x$, is added on junction, so an AC dipole Josephson current is driven and an alternating electric current emerges in the external circuit, although only a DC bias is added. For the bilayer exciton system, $p_0=ed$ and $\partial_z E_x
= 2E_x/d$, so the frequency $\omega_0 =-\frac{2 e d
}{d\hbar}\int_{1}^{2} E_xdx = \frac{2e}{\hbar} (V_2-V_1)$. This frequency is the same with that of the AC Josephson effect of the superconductor[@Josephson1962]. The reason is as follows. In the electric current Josephson effect, two electrons form a Cooper pair, and the Cooper pair moves in response to external voltages. In the dipole current Josephson effect in the bilayer system, however, a pair consists of an electron and a hole, which moves in response to the counter voltages. The difference between them is shown in Fig. \[fig:4\]b. It can be seen that the pairs feel the same electric forces in the superconductor and the dipole superconductor. Thus the frequencies of the alternating current should be the same in the two cases. The fact that the results are indeed exactly the same shows that these results of the dipole superconductor are reasonable and credible. In addition, since its frequency is the same as that of the superconductor, it is not difficult to measure in an experiment because the latter has been observed for a long time. As a result, detecting the electric dipole current Josephson effect is another feasible method to verify the formation of the dipole superconductor.
**Discussion**
In conclusion, we view the exciton condensate superfluid state in bilayer electron system as an electric dipole superconductor state. Then, from the properties of the electric dipoles in an external electromagnetic field, we derive the London-type and the GL-type equations for an electric dipole superconductor. These equations are universal to all dipole superconductors and can also be used to study various properties of a dipole superconductor. By using these equations, we discover the Meissner-type effect against the spatial variation of the magnetic field and the DC and AC dipole current Josephson effects, and also suggest a method to detect the zero dipole resistance. These new effects discovered in this work can not only provide direct evidence for the existence of the exciton superfluid in bilayer electron systems, but also pave new ways to drive an electric dipole current.
**References**
Keldysh, L. V. & Kopaev, Y. V. Possible instability of semimetallic state toward Coulomb interaction. *Sov. Phys. Solid State* **6**, 2219 (1965).
Lozovik, Y. E. & Yudson, V. I. Feasibility of superfluidity of paired spatially separated electrons and holes - new superconductivity mechanism. *JETP Lett.* **22**, 274 (1975).
Snoke, D. Spontaneous Bose Coherence of Excitons and Polaritons. *Science* **298**, 1368-1372 (2012).
Eisenstein, J. P. & MacDonald, A. H. Bose-Einstein condensation of excitons in bilayer electron systems. *Nature* **432**, 691-694 (2004).
Kellogg, M., Eisenstein, J. P., Pfeiffer, L. N. & West, K. W. Vanishing Hall resistance at high magnetic field in a double-layer two-dimensional electron system. *Phys. Rev. Lett.* **93**, 036801 (2004).
Su, J.-J. & Macdonaild, A. H. How to make a bilayer exciton condensate flow. *Nature Physics* **4**, 799-802 (2008).
Spielman, I. B., Eisenstein, J. P., Pfeiffer, L. N. & West, K. W. Resonantly enhanced tunneling in a double layer quantum Hall ferromagnet. *Phys. Rev. Lett.* **84**, 5808 (2000).
Tutuc, E., Shayegan, M. & Huse, D. A. Counterflow measurements in strongly correlated GaAs hole bilayers: evidence for electron-hole pairing. *Phys. Rev. Lett.* **93**, 036802 (2004).
Nandi, D., Finck, A. D. K., Eisenstein, J. P., Pfeiffer, L. N. & West, K. W. Exciton condensation and perfect Coulomb drag. *Nature* **488**, 481-484 (2012).
Gorbachev, R. V. *et al.* Strong Coulomb drag and broken symmetry in double-layer graphene. *Nature Physics* **8**, 896-901 (2012).
Price, A. S., Savchenko, A. K., Narozhny, B. N., Allison, G. & Ritchie, D. A. Giant Fluctuations of Coulomb Drag in a Bilayer System. *Science* **316**, 99-102 (2007).
Titov, M. *et al.* Giant Magnetodrag in Graphene at Charge Neutrality. *Phys. Rev. Lett.* **111**, 166601 (2013).
Kasprzak, J. *et al.* Bose-Einstein condensation of exciton polaritons. *Nature* **443**, 409-414 (2006).
High, A. A. *et al.* Spontaneous coherence in a cold exciton gas. *Nature* **483**, 584-588 (2012).
Onnes, H. K. The superconductivity of mercury. *Comm. Phy. Lab. Univ. Leiden* **122**, 122-124 (1991).
Meissner, W. & Ochsenfeld, R. Ein neuer Effekt bei Eintritt der Supraleitf$\ddot{a}$higkeit. *Naturwiss* **21**, 787 (1933).
Josephson, B. D. Possible new effects in superconductive tunnelling. *Phys. Lett.* **1**, 251 (1962).
de Gennes, P. G. *Superconductivity of Metals and Alloys* (Westview Press, Boulder, 1999).
Bardeen, J., Cooper, L. N. & Schrieffer, J. R. Theory of superconductivity. *Phys. Rev.* **108**, 1175 (1957).
Jin, D. S. & Ye, J. Polar molecules in the quantum regime. *Phys. Today* **64**, 27-31 (2011).
Ni, K.-K. *et al.* Dipolar collisions of polar molecules in the quantum regime. *Nature* **464**, 1324-1328 (2010).
Ni, K.-K. *et al.* A high phase-space-density gas of polar molecules. *Science* **322**, 231-235 (2008).
Sun, Q.-F., Jiang, Z. T., Yu, Y. & Xie, X. C. Spin superconductor in ferromagnetic graphene. *Phys. Rev. B* **84**, 214501 (2011).
Sun, Q.-F. & Xie, X. C. The spin-polarized $\nu=0$ state of graphene: a spin superconductor. *Phys. Rev. B* **87**, 245427 (2013).
Bao, Z-Q., Xie, X. C. & Sun, Q.-F. Ginzburg-Landau-type theory of spin superconductivity. *Nature Communications* **4**, 2951 [2013]{}.
Griffiths D. J. *Introduction to Electrodynamics, 4th ed.* (Pearson, London, 2013).
Hnizdo V. Magnetic dipole moment of a moving electric dipole. *Am. J. Phys.* **80**, 625 (2012).
Ginzburg, V. L. & Landau, L. D On the theory of superconductivity. *Zh. Eksp. Teor. Fiz.* **20**, 1064-1082 (1950).
Nowack, K. C. *et al.* Imaging currents in HgTe quantum wells in the quantum spin Hall regime. *Nature Materials* **12**, 787-791 (2013).
Simpsons, J., Lane, J., Immer, C. & Youngquist, R. Simple Analytic Expressions for the Mangetic Field of a Circular Current Loop. *Tech. Rep. NASA*, (2011).
**Acknowledgments**
This work was financially supported by NBRP of China (2012CB921303 and 2012CB821402) and NSF-China under Grants Nos. 11274364 and 91221302.
**Author contributions**
Q.D.J., Z.Q.B., Q.F.S., and X.C.X. performed the calculations, discussed the results and wrote this manuscript together. Q.D.J. and Z.Q.B. contributed equally to this work.
**Additional information**
**Supplementary Information** accompanies this paper at http://www.nature.com/naturecommunications
**Correspondence and requests** for materials should be addressed to Q.F.S. or X.C.X.
**Competing financial interests:** The authors declare no competing financial interests.
Supplementary Information for “Theory for electric dipole superconductivity with an application for bilayer excitons"
=====================================================================================================================
Qing-Dong Jiang, Zhi-qiang Bao, Qing-feng Sun, and X. C. Xie
**Supplementary Note 1: An alternative set of equations for the electric dipole superconductor.**
Starting from Eqs. (3) and (4) in the main text, we can get an alternative set of equations to describe the electric dipole superconductor in the steady state. First of all, make some simplification on Eq. (4) in the main text, i.e, $$\begin{aligned}
\label{eq:s1}
\begin{split}
\nabla\times\bold B=&-\mu_0\nabla\times(\bold J_p\times\hat p_0)\\=&-\mu_0(\hat p_0\cdot\nabla)\bold J_p-\mu_0\hat p_0(\nabla\cdot\bold J_p)\\=&-\mu_0(\hat p_0\cdot\nabla)\bold J_p.
\end{split}\end{aligned}$$ The term $\nabla\cdot\bold J_p$ is taken to be zero in the steady state because of the conservation of the super electric dipole current in material. By taking curl on both sides of the London-type equation (3) in the main text, one has $\nabla\times(\nabla\times\bold
J_p)=-\beta(\hat p_0\cdot\nabla)(\nabla\times\bold B)$. Combining this equation and Eq. (\[eq:s1\]), we obtain the equation for the super electric dipole current density $$\begin{aligned}
\label{eq:s2}
\nabla\times(\nabla\times\bold J_p)=\mu_0 \beta (\hat p_0\cdot\nabla)^2\bold J_p\label{eq:s2},\end{aligned}$$ where the characteristic dimensionless ratio $\mu_0 \beta=\mu_0 n
p_0^2/m^*$ governs the screening strength. If we take curl on both sides of Eq. (\[eq:s1\]), it transforms into $\nabla\times(\nabla\times\bold B)=-\mu_0(\hat
p_0\cdot\nabla)(\nabla\times\bold J_p)$. Combine this equation and the London-type equation (3) in the main text, then we get the equation for magnetic field $$\begin{aligned}
\label{eq:s3}
\nabla\times(\nabla\times\bold B)=\mu_0 \beta (\hat p_0\cdot\nabla)^2\bold B.\end{aligned}$$ Now, we have obtained an alternative set of Eqs. (\[eq:s2\]) and (\[eq:s3\]) describing the magnetic field and the super electric dipole current separately in the electric dipole superconductor.
**Supplementary Note 2: The Hamiltonian of a moving electric dipole and the kinetic energy term of the electric dipole superconductor**
An electric dipole moving with velocity $\bold v$ in magnetic field $\bold B$ can feel an electric field $\bold E^{\prime}=\bold
v\times\bold B$, and the corresponding energy is $-\bold
p_0\cdot\bold E^{\prime}=-\bold p_0\cdot(\bold v\times\bold B)=\bold
v\cdot(\bold p_0\times \bold B)$.[@griffiths; @Magneticdipole] Therefore, the Lagrangian of this electric dipole is $\mathscr{L}=\frac{1}{2}m^*\bold v^2-\bold v\cdot(\bold p_0\times
\bold B)$, where $m^*$ is the effect mass of the electric dipole. The canonical momentum is $\bold p=\frac{\partial \mathscr
L}{\partial \bold v}=m^*\bold v-\bold p_0\times \bold B$. Thus the Hamiltonian of a moving electric dipole in a magnetic field is $$\begin{aligned}
\label{eq:s4}
\mathscr H =\bold {\hat p}\cdot\bold v-\mathscr L=\frac{[\hat{\bold p} + \bold p_0\times \bold B]^2}{2m^*}.\end{aligned}$$ The term $\bold p_0\times \bold B$ is analogous to the term $e\bold
A/c$ for an electron in a magnetic field. So the kinetic energy term of the electric dipole superconductor can be written as: [$\frac{|(\hat{\bold p} + \bold p_0\times \bold B)\psi(\bold
r)|^2}{2m^*}$]{}.
**Supplementary Note 3: The derivation of the Ginzburg-Landau-type equations of the electric dipole superconductor**
First of all, we minimize the free energy shown in Eq. (5) in the main text with respect to the complex conjugate of the order parameter $\psi^*$. For the second and third terms of Eq. (5) in the main text, we have $$\begin{aligned}
\label{eq:s5}
\delta\int_V d\bold r\left\lbrace\alpha(T)|\psi(\bold r)|^2+\frac{\beta(T)}{2}|\psi(\bold r)|^4\right\rbrace =\int_V d\bold r \left\lbrace\left[\alpha(T)\psi(\bold r)+\beta(T) |\psi(\bold r)|^2\psi(\bold r)\right]\delta \psi^*(\bold r)\right\rbrace.\end{aligned}$$ For the fourth term, we get $$\begin{aligned}
\label{eq:s6}
\delta\int_V d\bold r\frac{|(\hat {\bold p} + \bold p_0\times \bold B)\psi(\bold r)|^2}{2m^*}&=&\int_V d\bold r\left\lbrace[\frac{i\hbar}{2 m^*}\nabla\delta \psi^*(\bold r)]\cdot[(-i\hbar\nabla+\bold p_0\times\bold B)\psi(\bold r)]\right.\nonumber\\&&\left.+\frac{1}{2 m^*}[(\bold p_0\times\bold B)\delta \psi^*(\bold r)]\cdot[(-i\hbar\nabla+\bold p_0\times\bold B)\psi(\bold r)]\right\rbrace.\end{aligned}$$ It should be noted that $$\begin{aligned}
\label{eq:s7}
\int_Vd\bold r\left\lbrace\frac{i\hbar}{2 m^*}\nabla\delta \psi^*(\bold r)
\cdot [(-i\hbar\nabla+\bold p_0\times\bold B)\psi(\bold r)]\right\rbrace &=&\frac{i\hbar}{2 m^*}\oint d\bold S\cdot\lbrace\delta \psi^*(\bold r)(-i\hbar\nabla+\bold p_0\times\bold B)\psi(\bold r)\rbrace \nonumber\\& &-\frac{i\hbar}{2 m^*}\int_V d\bold r\lbrace\delta \psi^*\nabla\cdot[(-i\hbar\nabla+\bold p_0\times\bold B)\psi(\bold r)]\rbrace.\end{aligned}$$ Combining Eqs. (\[eq:s5\]), (\[eq:s6\]) and (\[eq:s7\]), we can obtain: $$\begin{aligned}
\label{eq:s8}
%\begin{split}
\alpha(T)\psi(\bold r)+\beta(T)|\psi(\bold r)|^2\psi(\bold r)
+ \frac{[\hat{\bold p}+\bold p_0\times \bold B]^2\psi(\bold r)}{2m^*}=0 ,\end{aligned}$$ and $$\begin{aligned}
\label{eq:s9}
[-i\hbar\nabla+\bold p_0\times\bold B]_n\psi(\bold r)=0.\end{aligned}$$
Eq. (\[eq:s8\]) is the first Ginzburg-Landau-type equation and Eq. (\[eq:s9\]) is the boundary condition for the first Ginzburg-Landau-type equation, where the subscript n stands for the component perpendicular to the surface. Here, we emphasize that this boundary condition is actually the requirement of the variational principle. In fact, if we substitute Eq. (\[eq:s9\]) into Eq. (8) in the main text, we can get $\bold {J_p}_n=0$, which means that there is no electric dipole current entering or leaving the electric dipole superconductor. Similar discussions on boundary condition of the (monopole) superconductor can be found in the original paper written by Ginzburg and Landau[@GLequation].
Next, we minimize the free energy with respect to the vector potential $\bold A$. For the fourth term of Eq. (5) in the main text, we have $$\begin{aligned}
\label{eq:s10}
\delta\int_Vd\bold r\frac{|(\hat{\bold p} + \bold p_0\times \bold B)\psi(\bold r)|^2}{2m^*} &
=&\frac{1}{2m^{\ast}}\int_V d\bold r[\bold p_0\times\delta(\nabla\times\bold A)]\psi\cdot[(i\hbar\nabla
+\bold p_0\times(\nabla\times\bold A))\psi^*]+c.c \nonumber\\
&=&\frac{1}{2m^{\ast}}\int_V d\bold r\delta(\nabla\times\bold A)\cdot \lbrace [\psi(i\hbar\nabla
+\bold p_0\times(\nabla\times \bold A))\psi^*]\times\bold p_0\rbrace+c.c \nonumber\\
&=&\frac{1}{2m^{\ast}}\int_V\,d\bold r\,\delta \bold A\cdot\lbrace\nabla\times\lbrace[\psi(i\hbar\nabla
+\bold p_0\times(\nabla\times \bold A))\psi^*]\times\bold p_0\rbrace\rbrace \nonumber\\
&&\left.+\frac{1}{2m^{\ast}}\oint d\bold S\cdot\lbrace \delta\bold A\times[[\psi(i\hbar\nabla\right.
+\bold p_0\times(\nabla\times \bold A))\psi^*]\times\bold p_0]\rbrace+c.c.\end{aligned}$$
If we variate the last term of Eq. (5) in the main text with respect to vector $\bold A$, we get $$\begin{aligned}
\label{eq:s11}
\delta\int_V d\bold r\frac{|\nabla\times\bold A |^2}{2\mu_0}
=\int_V\, d\bold r\delta \bold A\cdot\frac{\nabla\times(\nabla\times\bold A)}{\mu_0}
+\oint d\bold S\cdot\frac{[\delta \bold A\times(\nabla\times\bold A)]}{\mu_0}.\end{aligned}$$
Combining Eqs. (\[eq:s10\]) and (\[eq:s11\]), we can get the second Ginzburg-Landau-type equation, i.e., $$\begin{aligned}
\label{eq:s12}
\nabla\times\bold B=-\mu_0 \nabla\times(\bold J_p\times \hat p_0) ,\end{aligned}$$ where $$\begin{aligned}
\label{eq:s13}
\bold J_p \equiv \frac{p_0}{2 m^*}[i\hbar(\psi\nabla\psi^*-\psi^*\nabla\psi)+2\bold p_0\times\bold B\,|\psi|^2].\end{aligned}$$ It should be noted that in the derivation the surface integral vanishes due to the requirement of free energy minimization [@gennes1989superconductivity; @GLequation].
![ **The spatial variations of the magnetic field and the vector potential created by the cylindrical hollow conductor.** (**a**) The spatial variation of the magnetic field $\partial_z B_z^{ext}$ created by the cylindrical hollow conductor versus the coordinate $r$ in the specimen plane $m$. (**b**) $\partial_z A_{\theta}^{ext}$ versus the coordinate $r$ in the specimen plane $m$. This magnetic field will be applied on the electric dipole superconductor specimen. The parameters are same as in Fig. 2 in the main text. \[fig:S1\]](supfig-1.eps "fig:"){height="3.9cm" width="12cm"}\
**Supplementary Note 4: The non-uniform magnetic field created by the electric current in the cylindrical hollow conductor**
In the article, we consider that a non-uniform magnetic field is created by a cylindrical hollow conductor. The inner radius of the cylindrical hollow conductor is $R_{in}$, the outer radius is $R_{out}$ and the length is $h$ (see Fig. 2a and 2b in the main text). A uniform electric current density $j$ along the azimuthal direction is applied in this conductor. In the following, we calculate the non-uniform magnetic field caused by this electric current in the cylindrical hollow conductor. This non-uniform magnetic field is the external magnetic field applied on the electric dipole superconductor specimen.
To deduce the magnetic field caused by the electric current in the hollow conductor, we start from the analytical expression of the magnetic field induced by a circular current loop of radius $R$ carrying current $I$. The circular current loop lies in the $z=z_0$ plane, and is centered at $z$ axis. Then the magnetic field $\bold
B$ and magnetic vector potential $\bold A$ at $(r, \theta, z)$ (in cylindrical coordinate) can be immediately calculated through Biot-Savart law[@griffiths; @simpson]. The results are $$\begin{aligned}
\label{eq:s14}
B_r(r, \theta, z)=&\frac{\mu_0 I (z-z_0)}{2 \pi \xi^2\eta r}[(r^2+R^2
+(z-z_0)^2)E(k)-\xi^2 K(k)],\end{aligned}$$ $$\begin{aligned}
\label{eq:s15}
B_{\theta}(r, \theta, z)=0,\end{aligned}$$ $$\begin{aligned}
\label{eq:s16}
B_z(r, \theta, z)=&-\frac{\mu_0}{2 \pi \xi^2\eta}[(r^2-R^2
+(z-z_0)^2)E(k)-\xi^2 K(k)],\end{aligned}$$ $$\begin{aligned}
\label{eq:s17}
\begin{split}
\partial_z B_z (r, \theta, z)=&\frac{\mu_0 I (z-z_0)}{2 \pi \xi^4\eta^3}[(6 R^2(r^2-(z-z_0)^2)\\
&-7 R^4+(r^2+(z-z_0)^2)^2)E(k)\\
&-\xi^2(r^2-R^2+(z-z_0)^2)K(k)],
\end{split}\end{aligned}$$ $$\begin{aligned}
\label{eq:s18}
\partial_z A_{\theta}(r, \theta, z)=-B_r(r, \theta, z).\end{aligned}$$ In the above analytical expressions, $\xi=\sqrt{R^2+r^2+(z-z_0)^2-2
r R}$, $\eta=\sqrt{R^2+r^2+(z-z_0)^2+2 r R}$, $k=1-\xi^2/\eta^2$, and $K(k)$ and $E(k)$ are the first kind and the second kind of the complete elliptic integral. By using the Biot-Savart law and the principle of superposition, we calculate the total external magnetic field gradient $\partial_z B_z^{ext}$ and the gradient of magnetic vector potential $\partial_z A_{\theta}^{ext}$. It gives $$\begin{aligned}
\label{eq:s19}
\partial_z B_z^{ext} &= &\frac{\mu_0 j}{2 \pi}\int_{-h/2}^{h/2}dz_0\int_{R_{in}}^{R_{out}}dR
\left\lbrace\frac{z-z_0}{\xi^4\eta^3}\left\lbrace\left[6 R^2(r^2-(z-z_0)^2)-7 R^4 \right.\right.\right. \nonumber\\
&&\left.\left.\left. + (r^2+(z-z_0)^2)^2\right]E(k)-\xi^2[r^2-R^2
+(z-z_0)^2]K(k)\right\rbrace\right\rbrace\end{aligned}$$ and $$\begin{aligned}
\label{eq:s20}
\partial_z A_{\theta}^{ext}=-B_r=-\frac{\mu_0 j}{2 \pi}\int_{-h/2}^{h/2}dz_0\int_{R_{in}}^{R_{out}}dR \left\lbrace\frac{z-z_0}{\xi^2\eta r}\left[(r^2+R^2+(z-z_0)^2)E(k)-\xi^2 K(k)\right]\right\rbrace.\end{aligned}$$ Due to the rotational symmetry of the cylindrical hollow conductor, $\partial_z B_z^{ext}$ and $\partial_z A_{\theta}^{ext}$ are independent of the angle $\theta$.
Fig. \[fig:S1\]a and \[fig:S1\]b in the Supplementary Note 4 show respectively the spatial variation of the magnetic field $\partial_z B_z^{ext}$ and $\partial_z A_{\theta}^{ext}$ versus the coordinate $r$ in the specimen plane $m$. Here the cylindrical hollow conductor sizes are $R_{in}=1mm$, $R_{out}=1\,cm$ and $h=1.5\,cm$. The current density in conductor $j=10^8 A/m^2$. We can see that $\partial_z B_z^{ext}$ is quite large while $r<R_{in}$. So we suggest to put the electric dipole superconductor specimen in the hollow ($r<R_{in}$) if to investigate the Meissner-type effect of the electric dipole superconductor. In addition, Fig. \[fig:S1\]a and \[fig:S1\]b in the Supplementary Note 4 also show that $\partial_z B_z^{ext}$ is relatively small and $\partial_z A_{\theta}^{ext}$ is large when below the cylindrical hollow conductor with $7mm<r<9mm$ (see the yellow color shadow region), and that is why we choose to put the annular dipole superconductor specimen in there for the investigation of the zero electric dipole resistance.
**Supplementary Note 5: The magnetic field induced by the super electric dipole current**
In this section, we calculate the magnetic field induced by the super electric dipole current in the bilayer exciton system. In the bilayer exciton system, the electric dipole current can be viewed as counter-flow electric currents. Since the bilayer system has the reflection symmetry about the middle plane $m$ and the rotational symmetry about the $z$ axis, the induced magnetic field in plane $m$ only has the nonzero r-component $B_r$. Although the $z$-component of the induced magnetic field vanishes in plane $m$, the spatial variation $\partial_z B_z^{ind}$ does not vanish. Moreover, we note that the induced magnetic field $B_r^{ind}$ and its spatial variation $\partial_z B_z^{ind}$ are twice of that induced by the electric current in one layer. We will calculate the magnetic field induced by the electric current in one layer, and then times $2$. At last, we obtain $\partial_z B_z^{ind}$ and $B_r^{ind}$ in plane $m$: $$\begin{aligned}
\label{eq:s21}
\partial_z B_z^{ind}(r) &= &\frac{\mu_0}{2 \pi}\int_{0}^{r_{out}}d R
\left\lbrace\frac{J_{p}}{\xi_0^4\eta_0^3}[(6 R^2(r^2-(d/2)^2)-7 R^4 \right.\nonumber\\
&&+ \left.(r^2+(d/2)^2)^2)E(k_0)-\xi_0^2(r^2-R^2+(d/2)^2)K(k_0)]\right\rbrace\end{aligned}$$ $$\begin{aligned}
\label{eq:s22}
B_r^{ind}(r)=-\frac{\mu_0}{2 \pi}\int_0^{r_{out}}d R
\left\lbrace\frac{J_{p}}{\xi_0^2\eta_0
r}[(r^2+R^2+(d/2)^2)E(k_0)-\xi_0^2 K(k_0)]\right\rbrace\end{aligned}$$ where $\xi_0=\sqrt{R^2+r^2+(d/2)^2-2 r R}$, $\eta_0=\sqrt{R^2+r^2+(d/2)^2+2 r R}$ and $k_0=1-\xi_0^2/\eta_0^2$.\
**Supplementary References**
Griffiths D. J. *Introduction to Electrodynamics 4th ed.* (Pearson, London, 2013).
Hnizdo V. Magnetic dipole moment of a moving electric dipole. *Am. J. Phys.* **80**, 625 (2012).
Ginzburg, V. L.& Landau, L. D. On the theory of superconductivity. *Zh. Eksp. Teor. Fiz.* **20**, 1064-1082 (1950).
de Gennes, P. G. *Superconductivity of Metals and Alloys* (Westview Press, Boulder, 1999).
Simpsons, J., Lane, J., Immer, C. & Youngquist, R. Simple Analytic Expressions for the Mangetic Field of a Circular Current Loop. *Tech. Rep. NASA*, (2011).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Measuring the vulnerability of communities in complex network has become an important topic in the research of complex system. Numerous existing vulnerability measures have been proposed to solve such problems, however, most of these methods have their own shortcomings and limitations. Therefore, a new entropy-based approach is proposed in this paper to address such problems. This measure combines the internal factors and external factors for each communities which can give the quantitative description of vulnerability of community. The internal factors contain the complexity degree of community and the number of edges inside the community, and the external factors contain the similarity degree between chosen community and other communities and the number of nodes outside the community. Considering community vulnerability from the perspective of entropy provides a new solution to such problem. Due to sufficient consideration of community information, more reasonable vulnerability result can be obtained. In order to show the performance and effectiveness of this proposed method, one example network and three real-world complex network is used to compare with some exiting methods, and the sensitivity of weight factors is analysed by Sobol’ indices. The experiment results demonstrate the reasonableness and superiority of this proposed method.'
address: 'Institute of Fundamental and Frontier Science, University of Electronic Science and Technology of China, Chengdu, 610054, China'
author:
- Tao Wen
- Yong Deng
bibliography:
- 'myreference.bib'
title: 'The vulnerability of communities in complex network: An entropy approach'
---
Complex network, Community vulnerability, Entropy
Introduction
============
Recently, cyber-physical system (CPS) has attracted much attention in numerous fields, such as microgrid [@Guan2010Microgrid], smart city [@BRUNEO201912service], internet of things [@Dautov2019Hierarchical; @Dautov2018Metropolitan], and so on. Meanwhile, how to model CPS into specific physical models to analyze their performance and property has become the focus of research [@Lee2019cascading]. Therefore, complex network has been applied in this field to better understand the performance of CPS [@Cheng2019Smart; @Guo2019network], because nodes in the network represent individuals in the system, and edges would show the relationship between these individuals. Most of previous researches focus on the structure and topological property, which can quantify the characteristic and performance of network. Particularly, the community structure has received increasing attention, because it can reveal human dynamics [@wang2018exploiting; @wang2017onymity], inference reliable links [@Ma2019Reliable], and identify influential nodes [@ZHANG2019249; @wentao2019nodes]. The community structure in the network demonstrates a higher density of nodes and edges, which can cause critical influence on the function and structure of subnetwork, improve the reliability of system [@DUI2019162System; @LEVITIN2019289Optimization], and counteract the aging effect [@LEVITIN2019397Dynamic; @LEVITIN201963Joint].
There are several different problems about the research of community structure, which can be divided into two issues. The first one is about the structure of community, such as dividing network’s community structure [@Zhu2019Community; @Orman2015Interpreting; @Rocco2017Effects], detecting overlapping community [@Orman2015Overlapping], and dynamic changes of community in evolving network [@ORMAN2017375]. Another one is the property of community, including measuring the reliance of community [@Ramirez2018Quantifying; @Zhang2018uncertainty; @Cerqueti2019resilience], reconfiguring network [@Zhang2019Reconfiguration], quantifying the reliability of community[@Ramirez2016Robustness; @Zhang2017Reliability], and measuring community vulnerability [@wentao2018evaluating]. The vulnerability of community in network gradually aroused researchers’ interest recently. For example, Rocco et al. [@Rocco2011Vulnerability] defined vulnerability set and value for different community, and proposed relative vulnerability value to compare with remaining communities. Wei et al. [@Wei2018Measuring] proposed a measure which consider more information about community itself, and used non-linear weighted function to combine these factors. Aniko et al. [@Kovacs2019vulnerability] proposed a topological index (distance-based fragmentation) to quantify the structural vulnerability in plant-visitation network. Alim et al. [@Alim2016Structural] assessed the community vulnerability through social-based forwarding and routing methods in opportunistic networks, which shows significant contribution about some devices on the performance of entire network. Che et al. [@Yanbo2019Vulnerability] modified original evolution method, and proposed a nondimensionalized scoring standard to form a complete assessment system to measure the vulnerability of urban power grid. Chen et al. [@CHEN2019CORRELATION] explored the relationship between vulnerability of complex network and fractal dimension. These methods have their own limitations, like computational complexity, inaccurate measurements, and not suitable for certain scenarios.
Since entropy is an useful tool to measure the uncertain of information [@Dengwei2018; @wang2016statistical], it has been wildly in the network theory, like dimension presentation [@wb2019; @WU2018HADAMARD; @wentao2018information], evidence theory [@jiang2018Correlation; @Jiang2019Znetwork], influential nodes identification [@li2018evidential; @wang2018amodified], time series prediction [@xu2018visibility; @Fan2019timeseries]. In addition, the structure and property of communities can be expressed by probability sets, entropy-based method has gradually been an reasonable and effective method to quantify the property of network [@wentao2019similar; @wentao2019structure]. Therefore, an entropy approach is applied in this paper to measure the vulnerability of community which can overcome the shortcomings and limitations of previous method.
In this paper, an entropy-based measure is proposed to quantify the vulnerability degree of community structure. This proposed method can combine two parts of information, i.e., internal factors and external factors, which can consider more information about community and give a reasonable vulnerability result of each community. The internal factors contains the number of edges inside the community and the complexity degree of community which is measured by Tsallis structure entropy, and the external factors contains the number of edges outside the community and the similarity degree between chosen community and other communities which is measured by relative entropy. These two entropy can quantify the property of community more reasonable and effective. Finally, the vulnerability and relative vulnerability result can be obtained by this proposed method to quantitatively describe vulnerability of different community. In order to show the performance and effectiveness of this proposed method, one example network and three real-world complex network are applied in this paper. In addition, the sensitive of four weight factors are analysed by Sobol’ indices in Manzi network, and the vulnerability order obtained by different method are compared in Italian 380KV power grid network. The experiments results show the superiority and reasonableness of this proposed method, meanwhile, this proposed method can overcome the shortcomings and limitations of previous method,
The organization of the rest of this paper is as follows. Section 2 presents some basic properties about node and detecting community methods. This novel entropy-based method is proposed in Section 3 to measure the vulnerability of community. Meanwhile, numerical experiments are performed to illustrate the reasonableness and effectiveness of the proposed method in Section 4. Conclusion is conducted in Section 5.
Preliminaries
=============
In this section, some basic concepts about complex network are introduced. In addition, a community detection algorithm and classical community vulnerability measure are described in this section.
Node properties in network
--------------------------
A given complex network can be denoted as $G(N,E)$, where $N = (1,2, \cdots ,n)$ and $E = (1,2, \cdots ,m)$ is the set of nodes and edges respectively, and $n$ and $m$ is the number of nodes and edges in the complex network respectively. $A$ is the adjacency matrix of complex network whose size is $n \times n$, where ${a_{ij}} = 1$ represents there is an edge between node $i$ and node $j$, and ${a_{ij}} = 0$ is the opposite.
(Node Degree). The degree of node $i$ in the complex network is denoted as ${d_i}$ and defined as follows, $$\label{E_degree}
{d_i} = \sum\limits_{j = 1}^n {{a_{ij}}}$$ where ${a_{ij}}$ is one element of adjacency matrix $A$. The degree distribution of node $i$ is defined as follows, $$\label{E_deg_dis}
{p_i} = \frac{{{d_i}}}{{\sum\limits_{i \in N} {{d_i}} }}$$
(Node Betweenness). The betweenness distribution of node $i$ in the complex network is denoted as ${p_i}^\prime $ and defined as follows, $$\label{E_betweenness}
{p_i}^\prime = \sum\limits_{s,e \ne i} {\frac{{{g_{se}}(i)}}{{{g_{se}}}}}$$ where ${{g_{se}}}$ is the total number of shortest paths between node $s$ and node $e$, and ${{g_{se}}(i)}$ is the number of shortest paths between node $s$ and node $e$ which pass through node $i$.
Community detection algorithm
-----------------------------
Lots of measures have been proposed to detect the community structure in complex network. In order to find the community structure of network, Newman’s modularity method [@Newman2004Fast] is applied in this paper.
(Newman’s modularity). For a given complex network $G$ with $k$ communities, the modularity is denoted as $Q$ and defined as follows, $$\label{E_Newman}
Q = \sum\limits_{{c_k} = 1}^k {\left( {\frac{{{e_{{c_k}}}}}{m} - {{\left( {\frac{{De{g_{{c_k}}}}}{{2m}}} \right)}^2}} \right)}$$ where $k$ is the number of communities, $m$ is the total number of edges in complex network, ${{e_{{c_k}}}}$ is the number of edges in community ${c_k}$, and ${De{g_{{c_k}}}}$ is the total degree of nodes in community ${c_k}$ which is defined as follows, $$\label{E_Community_degree}
De{g_{{c_k}}} = \sum\limits_{i \in {c_k}} {{d_i}}$$ where ${c_k}$ is the set of nodes in community, and ${{d_i}}$ is the degree of node $i$.
The value of $Q$ can measure the difference of different communities which can show the presence of community structure in complex network. Different value of $Q$ represents different situation. $Q = 0$ means all of the nodes in the network are in one single community and there is no community structure in the network. $Q > 0$ represents there are some kinds of community structure, and $Q = 1$ means the community structure is strong in the network. Meanwhile, Newman and Girvan [@Newman2004Finding] suggested the value of $Q$ should fall in the range $0.2 \sim 0.7$, and this value of $Q$ would show the existence of community structures.
The main idea of this method is to find the changes in $Q$, and the step to detect community structure is shown as follows,
**Step 1**: Each node in the complex network is divided in a single community.
**Step 2**: Every two communities are integrated into one community in turn, and the value of modularity change $\Delta {Q_{ij}}$ can obtained from the community structure.
**Step 3**: Community $i$ and community $j$ are integrated into one community with the highest $\Delta {Q_{ij}}$.
**Step 4**: Repeat Step 2 and Step 3 until $\Delta {Q_{ij}} < 0$.
Classical community vulnerability measure
-----------------------------------------
To measure the vulnerability of community, lots of measures have been proposed. One classical measure is introduced in this section.
(Community vulnerability measure). The vulnerability of community $x$ is denoted as ${v_x}$ and defined as follows, $$\label{E_vx}
{v_x} = \frac{1}{{\left| {{V_x}} \right|}},\forall {V_x} \ne \emptyset$$ where ${V_x}$ is the set of communities which are connected with community $x$, and $\left| {{V_x}} \right|$ is the number of links which are connected with community $x$.
The relative vulnerability of community $x$ is denoted as ${R_x}$ and defined as follows, $$\label{E_Rx}
{R_x} = \frac{{{v_x}}}{v},v = \mathop {\min }\limits_y ({v_y})$$
The proposed method
===================
Basic method
------------
In this section, a novel method is proposed to measure community vulnerability via entropy approach. This proposed method focuses two parts of information which can consider more details in the network, including internal factors and external factors. The internal factors include the complexity degree of community and number of edges within the community, and external factors include the similarity degree and number of edges between chosen community and other communities. The complexity and similarity would be obtained by entropy method which would overcome the shortcomings and limitations of previous method. The flow chart of this proposed method is shown in Fig. \[fig\_flow chart\].
![**The flow chart of this proposed method.**[]{data-label="fig_flow chart"}](flowchart.eps "fig:"){width="15cm"}\
### Complexity measure
Firstly, the complexity degree of community is measured by Tsallis structure entropy which combines the degree distribution and betweenness distribution. Because degree distribution focuses on the local topological information of central node and betweenness distribution considers the global topological information, Tsallis structure entropy which combines these two topological information can give a reasonable measure for community complexity.
\[Def\_complexity\] (Complexity measure based on Tsallis structure entropy). The complexity degree of community ${c_k}$ is denoted as ${T_{{c_k}}}$ and defined as follows, $$\label{E_complexity}
{T_{{c_k}}} = \sum\limits_{i = 1}^{\left| {{c_k}} \right|} {\frac{{{p_i}^{{q_i}} - {p_i}}}{{1 - {q_i}}}}$$ where $\left| {{c_k}} \right|$ is the number of nodes in community ${c_k}$, ${p_i}$ is the degree distribution node $i$ which can be obtained by Eq. (\[E\_deg\_dis\]), ${q_i}$ can be obtained from betweenness distribution, and the relationship between ${q_i}$ and ${p_i}^\prime $ is shown as follows, $$\label{E_pq}
{q_i} = 1 + ({p_{\max }}^\prime - {p_i}^\prime )$$ where ${p_i}^\prime $ can be obtained from Eq. (\[E\_betweenness\]), ${p_{\max }}^\prime $ is the maximum value of betweenness ${p_i}^\prime $. The purpose of Eq. (\[E\_pq\]) is to make the index ${q_i}$ bigger than 1 which can show the influence of subnetwork to community ${c_k}$.
When each node’s ${q_i}$ equals to 1, Tsallis entropy would degenerate to Shannon entropy which is shown as follows, $$\label{E_Shannon}
{T_{{c_k}}} = - \sum\limits_{i = 1}^{\left| {{c_k}} \right|} {{p_i}\log {p_i}}$$ This form of entropy would only focuses on the local topological structure information.
The degree distribution is based on the local topological structure around central node $i$. The betweenness distribution focuses on the whole topological structure which can describe the global properties of community. Most of time, ${q_i}$ would be bigger than 1 which can show the influence of subnetwork. When each nodes’s ${q_i}$ equals to 1, the Tsallis entropy would degenerate to Shannon entropy based on degree distribution which only focuses on the local structure. Using betweenness distribution to replace the constant parameter $q$ can describe the information about the network itself which is more reasonable for measuring the communities’ complexity. This method’s property also obeys the classical Tsallis entropy.
### Similarity measure
Then, relative entropy is used in this section to obtain the similarity degree of chosen community and other communities. The relative entropy (Kullback$ - $Leibler divergence) was wildly used in information theory and probability theory which is proposed by Kullback and Leibler et al. [@Kullback1951ON]. In general, the relative entropy is used to measure the difference between two probability set. In this section, the relative entropy is based on Shannon entropy and degree distribution which can measure the similarity between two community structure.
For two community ${c_i}$ and ${c_j}$, the community structure is denoted as ${L_{{c_i}}}({N_i},{D_i})$ and ${L_{{c_j}}}({N_j},{D_j})$ respectively, where ${N_i}$ and ${D_i}$ is the set of nodes and set of degree of nodes in community ${c_i}$. $\left| {{N_i}} \right|$ is the number of nodes in community ${c_i}$ and $\max \left| {{N_i}} \right|$ is the maximum size of community in the network. The probability set of community $i$ is denoted as $P(i)$ and obtained by degree distribution. The scale of every probability set $s$ would be same which equals to $\max \left| {{N_i}} \right|$. So the probability set of community ${c_i}$ can be shown as follows, $$\label{E_probability_set}
P(i) = [p(i,1),p(i,2), \cdots ,p(i,s)]$$ The element in probability set is based on degree distribution. When the size $\left| {{N_i}} \right|$ of community ${c_i}$ equals to $\max \left| {{N_i}} \right|$, all of the elements would be obtained by nodes’ degree, but when $\left| {{N_i}} \right| < \max \left| {{N_i}} \right|$, some elements would equal to zero to make the probability set complete. The detail of $p(i,k)$ is defined as follows, $$\label{E_element}
p(i,k) = \left\{ {\begin{array}{*{20}{c}}
{\frac{{{d_k}}}{{\sum\limits_{k = 1}^{\left| {{N_i}} \right|} {{d_k}} }}}&{k \le \left| {{N_i}} \right|}\\
0&{k > \left| {{N_i}} \right|}
\end{array}} \right.$$ where ${{d_k}}$ is the degree of node $k$, ${\left| {{N_i}} \right|}$ is the number of nods in community ${c_i}$.
To measure the similarity between community ${c_i}$ and ${c_j}$, the relative entropy is used in this section and it is defined as follows,
\[Def\_similarity\] (Similarity measure based on relative entropy). The difference between two communities is obtained by relative entropy ${R_{ij}}$ and defined as follows, $$\label{E_similarity}
{R_{ij}}(P'(i)||P'(j)) = \sum\limits_{k = 1}^{s'} {p'(i,k)\log \frac{{p'(i,k)}}{{p'(j,k)}}}$$ Because the order of element would affect the relative entropy and similarity result, ${p'(i,k)}$ and ${p'(j,k)}$ are the decreasing order of ${p(i,k)}$ and ${p(j,k)}$ in Eq. (\[E\_probability\_set\]). ${s'}$ can be obtained as follows, $$\label{E_s_pie}
s' = \min (\left| {{N_i}} \right|,\left| {{N_j}} \right|)$$ The adjustment of $s'$ is to avoid ${\frac{{p'(i,k)}}{{p'(j,k)}}}$ being 0 or positive infinity, which would be beneficial for calculation. The relative entropy’ property is not symmetry, so the following changes are needed to make it symmetrical, $$\label{E_r_ij}
{r_{ij}} = {R_{ij}}(P'(i)||P'(j)) + {R_{ji}}(P'(j)||P'(i))$$ Thus, ${r_{ij}} = {r_{ji}}$ holds, and the relative entropy between two communities are symmetry. Because the relative entropy measure the difference between two probability set, the difference between two communities are obtained in this situation. The bigger ${r_{ij}}$, the greater the difference between two communities structure is. So the similarity index is obtained based on relative entropy to show the similarity between two communities, and it is denoted as ${s_{ij}}$ and defined as follows, $$\label{E_s_ij}
{s_{ij}} = 1 - \frac{{{r_{ij}}}}{{\max ({r_{ij}})}}$$ where ${s_{ij}}$ is also symmetry, and shows the similarity between two communities structure. The more similar the two communities, the less the difference between them is, the closer ${r_{ij}}$ is to ${\max ({r_{ij}})}$ and the closer ${s_{ij}}$ is to zero.
So the similarity between two communities structure can be measured by the relative entropy, which can give a novel approach to this problem. The relative entropy focuses on the local structure topological information in the community structure, which is more reasonable.
### Edges in the network
The number of edges inside and outside the community are also important for community vulnerability measuring. In this section, the number of edges is considered.
\[Def\_inside\_edge\] (Number of edges inside the community). The number of edges inside the community ${c_k}$ is donated as $D_{{c_k}}^{in}$ and defined as follows, $$\label{E_edges_in}
D_{{c_k}}^{in} = \sum\limits_{i \in {c_k}} {\sum\limits_{j \in {c_k}} {{a_{ij}}} }$$ where node $i$ and node $j$ are within the community ${c_k}$, ${a_{ij}}$ is the element of adjacency matrix $A$. Thus, ${a_{ij}}$ is entirely inside the community.
\[Def\_outside\_edge\] (Number of edges outside the community). The number of edges outside the community ${c_k}$ is donated as $D_{{c_k}}^{out}$ and defined as follows, $$\label{E_edges_out}
D_{{c_k}}^{out} = \sum\limits_{i \in {c_k}} {\sum\limits_{j \notin {c_k}} {{a_{ij}}} }$$ where node $i$ is within the community ${c_k}$, and node $j$ is outside the community ${c_k}$, ${a_{ij}}$ is the element of adjacency matrix $A$. Thus, ${a_{ij}}$ connects the chosen community and other communities which can show the relationship between them.
### Community vulnerability measure
Lastly, all of the factors defined in Definition \[Def\_complexity\] to \[Def\_outside\_edge\] are considered in the vulnerability measuring model. This proposed method would consider the internal factors and external factors which consider more details of community, and is defined as follows,
(Proposed community vulnerability measure). The vulnerability of community $x$ is donated as $Vu{l_x}$ and defined as follows, $$\label{E_pro_vul}
Vu{l_x} = \frac{{{{\left( {{S_x}} \right)}^\alpha }}}{{{{\left( {D_x^{out}} \right)}^\beta }}}\frac{1}{{{{\left( {D_x^{in}} \right)}^\lambda }{{\left( {{T_x}} \right)}^\eta }}}$$ where $\alpha ,\beta ,\lambda ,\eta $ are the weight factors of different parameters, and all of them are bigger than zero. ${D_x^{in}}$ and ${D_x^{out}}$ is the number of edges inside and outside the community $x$ respectively, ${{T_x}}$ is the complexity degree of community $x$, and ${{S_x}}$ represents the similarity degree between community $x$ and other communities (exclude community $x$ itself) which can be shown as follows, $$\label{E_S_x}
{S_x} = \sum\limits_{j = 1}^k {{s_{xj}}},j \ne x$$ where ${{s_{xj}}}$ can be obtained by Eq. (\[E\_s\_ij\])
The relative vulnerability of community $x$ is denoted as $R{V_x}$ and defined as follows, $$\label{E_pro_rel_vul}
R{V_x} = \frac{{Vu{l_x}}}{{Vul}},Vul = \mathop {\min }\limits_y (Vu{l_y})$$
In order for these parameters to be considered on the same scale, all of these four parameters ${s_{xj}},D_x^{out},D_x^{in},{T_x}$ are normalized firstly. The weight factors $\alpha ,\beta ,\lambda ,\eta $ can give the weight to consider different parameters, which can be adjusted in different situation. This setting of weight factor makes this proposed method more reasonable. Some special cases of this proposed method $Vu{l_x}$ are shown as follows,
**1)** When $\alpha {\rm{ = }}\beta {\rm{ = }}\lambda {\rm{ = }}\eta $, this four parameters are considered equally.
**2)** When $\beta {\rm{ = }}1$, and $\alpha {\rm{ = }}\lambda {\rm{ = }}\eta {\rm{ = }}0$, this proposed method $Vu{l_x}$ would degenerate to the classical vulnerability measure ${v_x}$ in Eq. (\[E\_vx\]).
**3)** When $\alpha {\rm{ = }}\beta {\rm{ = 0}}$, this proposed method $Vu{l_x}$ would consider the external factors, which is the communities connected with chosen community.
**4)** When $\lambda {\rm{ = }}\eta {\rm{ = }}0$, this proposed method $Vu{l_x}$ would only consider the internal factors, i.e., the chosen community.
Sensitive analysis
------------------
Because these four weight factors ($\alpha ,\beta ,\lambda ,\eta $) are important for community vulnerability measuring, and the vulnerability result would have a related changes as weight factors change, thus, how to determine factors has been a problem in this model. In this section, the sensitive of these weight factors are analysed. In general, the global sensitivity analysis ia a useful tool to obtain the influence of inputs on the output variability in mathematical and physical model, and Sobol’ indices based on variance decomposition is applied in this paper. The first-order Sobol’ index $SI({X_i})$ and total effect index $ST({X_i})$ are defined as follows respectively, $$\label{E_first_order}
SI({X_i}) = \frac{{Va{r_{{X_i}}}({E_{{X_{ \sim i}}}}(Y\left| {{X_i}} \right.))}}{{Var(Y)}}$$ $$\label{E_total_effect}
ST({X_i}) = \frac{{{E_{{X_{ \sim i}}}}(Va{r_{{X_i}}}(Y\left| {{X_{ \sim i}}} \right.))}}{{Var(Y)}}$$ where $Y$ represents the output of system, ${{X_i}}$ is the *$i$*th independent input $X$, ${{X_{ \sim i}}}$ is all of the inputs exclude ${{X_i}}$, ${Var(Y)}$ is the variance which change with these inputs. The first-order Sobol’ index $SI(i)$ can get the contribution of ${{X_i}}$ to $Y$, and total effect index $ST(i)$ can get the contribution to the variance of $Y$ by the variability of each input ${{X_i}}$, which considering its individual effects and the interaction with other variables.
Each weight factors are randomly generated 10000 times by Monte Carlo method, and the range fall into \[0.2, 5\]. The vulnerability result would be obtained by these random factor combinations, and the contribution of different weight factors can be obtained by first-order Sobol’ index and total effect index.
An illustrative example
-----------------------
In this section, an example network is given to show the difference between this proposed method $Vu{l_x}$ and classical measure ${v_x}$. The network structure is shown in Fig. \[fig\_example\_network\]. Observing from Fig. \[fig\_example\_network\], this network has 9 nodes and 14 edges, and the community structure of network is detected by Newman’s modularity in Eq. (\[E\_Newman\]). The network is divided into three communities ($Q = 0.2857$) and each of the community is a fully-connected subnetwork. All of these four weight factors $\alpha ,\beta ,\lambda ,\eta $ equal to one which make four parameters equally important. According to this proposed method in Eq. (\[E\_pro\_vul\]), (\[E\_pro\_rel\_vul\]) and classical measure in Eq. (\[E\_vx\]), (\[E\_Rx\]), four parameters and the vulnerability of three communities are shown in Table \[table\_example\_network\].
![**An example network with 9 nodes.**[]{data-label="fig_example_network"}](example.eps "fig:"){width="7cm"}\
Community ${{S_x}}$ ${{T_x}}$ ${D_x^{in}}$ ${D_x^{out}}$ ${\left| {{V_x}} \right|}$ $Vu{l_x}$ $R{V_x}$ ${v_x}$ ${R_x}$
------------- ----------- ----------- -------------- --------------- ---------------------------- ----------- ---------- --------- ---------
Community A 1 0.5 0.1667 0.5 0.5 24 27.6264 2 2
Community B 0.4298 0.7924 0.5 0.5 0.5 2.1696 2.4975 2 2
Community C 0.8687 1 1 1 1 0.8687 1 1 1
: **The vulnerability of three communities in example network in Fig. \[fig\_example\_network\].**
\[table\_example\_network\]
From Table \[table\_example\_network\] when $\alpha {\rm{ = }}\beta {\rm{ = }}\lambda {\rm{ = }}\eta {\rm{ = 1}}$, it can be found that the vulnerability of community C is the lowest, which is the same as the classical measure, but the vulnerability of community A and B is different with classical measure. It can be found from Table \[table\_example\_network\] that the classical measure of the vulnerability ${v_x}$ of community A and B is same and it is 2 because of the same number of edges outside of chosen community. However, this classical method is not reasonable, because the vulnerability of one chosen community is determined not only by external factors but also by internal factors. From Fig. \[fig\_example\_network\], community A is a fully-connected subnetwork with only two nodes, but community B is a fully-connected subnetwork with three nodes. When the network’ structure is similar, i.e., fully-connected, the network with more nodes would be more robust, so community B is more robust than community A. The similar vulnerability result can be obtained by this proposed method ($Vu{l_B} = 2.1696 < Vu{l_C} = 24$), which is more reasonable for real-world application. The relative vulnerability $R{V_x}$ can be more obvious to show the vulnerability difference between different community. From the comparison result of example network, this proposed vulnerability measure $Vu{l_x}$ outperforms classical method, and can distinguish the vulnerability level of community that the classical method cannot distinguish.
Experimental study
==================
In this section, three real-world complex networks are applied to show the performance and effectiveness of this proposed method. These three networks are namely as Manzi network [@Manzi2001Fishman], Karate network [@Zachary1977Information], Italian power network [@Crucitti2005Locating] respectively. The topological properties of these three networks are shown in Table \[table\_network\_property\]. Observing from Table \[table\_network\_property\], $n$ and $m$ is the number of nodes and edges respectively. $\left\langle k \right\rangle $ and ${k_{\max }}$ is the average and maximum value of degree respectively, and $\left\langle d \right\rangle $ and ${d_{\max }}$ is the average and maximum value of shortest distance respectively in the network.
Network $n$ $m$ $\left\langle k \right\rangle $ ${k_{\max }}$ $\left\langle d \right\rangle $ ${d_{\max }}$
--------- ----- ----- --------------------------------- --------------- --------------------------------- ---------------
Manzi 52 76 2.8077 5 5.5000 13
Karate 34 78 4.5882 17 2.4082 5
Italian 127 171 2.6929 7 8.5682 25
: **The topological properties of real-world complex networks.**
\[table\_network\_property\]
Manzi network
-------------
Firstly, the telephone network in Belgium [@Manzi2001Fishman] which was analyzed for reliability purposes is used in this section. The topological structure of this network and the community structure obtained by Newman’s modularity [@Newman2004Fast] is shown in Fig. \[fig\_Manzi\_network\]. Observing from Fig. \[fig\_Manzi\_network\], Manzi network is divided into seven communities ($Q = 0.6316$), and the detail nodes in each community is shown in Table \[table\_Manzi\].
![**Manzi et al. network [@Manzi2001Fishman].**[]{data-label="fig_Manzi_network"}](Manzi.eps "fig:"){width="15cm"}\
The vulnerability of each community can guide the identification of critical community in the network. The vulnerability $Vu{l_x}$ and relative vulnerability $R{V_x}$ of each community are given in Table \[table\_vulnerability\_Manzi\], the classical comparing vulnerability measure ${v_x}$ and ${R_x}$ are also shown in Table \[table\_vulnerability\_Manzi\]. Observing classical measure ${R_x}$ in Table \[table\_vulnerability\_Manzi\], community 7 is the most vulnerable community, but community 3 is the most robust community in the network. The vulnerability of the rest of community (community 1, 2, 4, 5, 6) cannot be identified by the classical measure because of the same value of ${R_x}$. Thus, this novel method is proposed based on two parts of information, including internal factors and external factors. The result of this proposed method ($Vu{l_x}$, $R{V_x}$) are shown in Table \[table\_vulnerability\_Manzi\]. It can be found that community 3 is also the most robust community in the network which is same as the classical method. The value of $R{V_x}$ in community 4 and community 7 is close, which are more vulnerable than other communities. The rest of communities can get close but different $R{V_x}$, which can give a vulnerable order of this community (community 6 $>$ community 5 $>$ community 2 $>$ community 1). The vulnerable order of community in the network is community 4 $>$ community 7 $>$ community 6 $>$ community 5 $>$ community 2 $>$ community 1 $>$ community 3. So this proposed method can consider more information in the network and give a detail vulnerable order for these communities which can overcome some limitations of classical method.
community $k$ Nodes in community $k$
--------------- ----------------------------------------
1 3,6,7,8,9,10,13
2 1,2,4,5,11,12,14
3 15,16,17,18,20,21,24,25,29,31
4 19,22,26
5 23,38,30,32,36,37,41,43
6 27,33,34,35,38,39,40,42,45,46,47,49,50
7 44,48,51,52
: **The community details of Manzi network.**
\[table\_Manzi\]
\[table\_vulnerability\_Manzi\]
Then, Sobol’ indices introduced in Section 3.2 is used in this section to analysis the global sensitivity of these four weight factor $\alpha ,\beta ,\lambda ,\eta $. These weight factors can adjust the consideration of different parameters which can give a different vulnerability result. The sensitivity analysis result for the vulnerability of different community with different weight factors are shown in Table \[table\_sensitivity\]. Some conclusions can be obtained as follows,
**1)** The value of first-order Sobol’ index can show the sensitivity of different wight factors. For instance, the vulnerability of community 1 is most sensitive to weight factor $\alpha $, followed by $\lambda$, the other two factors $\beta $ and $\eta$ are less sensitive.
**2)** When the parameters of community equals to one, the first-order Sobol’ index and total effect index would equal to 1. That is because no matter how weight factor change, the influence parameter would remain the same, i.e., equal to 1. For example, ${{T_x}}$ and ${D_x^{in}}$ equal to 1 in community 6, so $SI(\eta )$, $ST(\eta )$, $SI(\lambda )$ and $ST(\lambda )$ equal to 0. Thus, the variability of these two weight factors would not affect the vulnerability measure of community 6. The same situation can occur in ${{S_x}}$ of community 4 and ${D_x^{out}}$ of community 3.
**3)** In most of communities, the first-order Sobol’ index $SI(\beta )$ and $SI(\eta )$ are smaller than $SI(\alpha )$ and $SI(\lambda )$, which means the vulnerability measure of community is more sensitive with $\alpha$ and $\lambda$. The similarity degree and the number of edges within community are more influential to the vulnerability results.
**4)** The first-order Sobol’ index $SI(\beta )$, $SI(\lambda )$, $SI(\eta )$ would be smaller when the value of parameters ${{T_x}}$, ${D_x^{in}}$, and ${D_x^{out}}$ are bigger, and $SI(\alpha )$ is different situation which is bigger with bigger parameters ${{S_x}}$. The situation occurs because $Vu{l_x}$ is positively correlated with ${{S_x}}$, but negatively correlated with ${{T_x}}$, ${D_x^{in}}$, and ${D_x^{out}}$. Hence the value trend of these parameters would have different impact on the vulnerability of communities. These patterns can be observed from all of these communities vulnerability sensitivity analysis in Table \[table\_sensitivity\].
**5)** The sum of the first-order Sobol’ index over these four weight factors in different communities are less than 1, which means there is an interaction between these four parameters. But this situation does not occur in total effect index.
**6)** Because of the interaction between these parameters, there would be a huge difference between first-order Sobol’ index and total effect index. But it is interesting to find that the order of total effect index would be the same as first-order Sobol’ index. For instance, the order of first-order Sobol’ index in community 1 is $SI(\eta ) < SI(\beta ) < SI(\lambda ) < SI(\alpha )$, the order of total effect index is $ST(\eta ) < ST(\beta ) < ST(\lambda ) < ST(\alpha )$, which is same as previous order.
\[table\_sensitivity\]
Because Sobol’ indices is convenient to obtained, and similar results can be obtained from differen network, we only analyse the sensitivity of weight factors in Manzi network and don’t analyse the subsequent network.
Karate network
--------------
Next, a social network is used in this section to show the performance of this proposed method. This social network is named as Karate club network, which describes the relationship between 34 members of one club in US university [@Zachary1977Information]. The topological structure and community structure divided by Newman’s modularity is shown in Fig. \[fig\_Karate\_network\]. Every nodes in the network denote a member in karate club, including the instructors and administrators, and the edges in the network represent the relationship between two members beyond their normal activities in the club. The network is divided into two communities ($Q = 0.38$), and it is same as the well-known community structure result because there has been disagreement between administrators and instructors [@Zachary1977Information]. The detail member in each communities is shown in Table \[table\_Karate\], and it can be found that the members are divided equally and each community have 17 members. The leader in each community is node 1 and node 34 respectively because of their largest degree.
![**Karate club network [@Zachary1977Information].**[]{data-label="fig_Karate_network"}](Karate.eps "fig:"){width="10cm"}\
community $k$ Nodes in community $k$
--------------- -------------------------------------------------------------------
1 1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 12, 13, 14, 17, 18, 20, 22
2 9, 15, 16, 19, 21, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34
: **The community details of Karate network.**
\[table\_Karate\]
The vulnerability measure result of Karate network is shown in Table \[table\_vulnerability\_Karate\]. Because there is only two communities, it can be found that the external factors (${{S_x}}$, ${D_x^{out}}$) are determined by each other, and they are the same. Thus, the classical measure ${R_x}$ would obtain same result and cannot identify the vulnerability degree of each community. But the internal factors are determined by community itself, these parameters can get a different result. It is interesting to find that the number of edges inside the community ${D_x^{in}}$ is also same and it is 34, but the edges between nodes are different which result in a different complexity degree. The complexity degree ${{T_x}}$ of two communities are 0.7060 and 1 respectively, which would get a different vulnerability measure for different communities. The relative vulnerability $R{V_x}$ of each communities are 1.4162 and 1 respectively, which can get conclusion that community 1 is more vulnerable than community 2. The main reason for their different vulnerability result is the complexity degree of each community, and the initial reason is the topological structure of each community. From this case, we can find that the vulnerability cannot be distinguished when the number of communities is too small. The topological structure inside the community is also important for the vulnerability result, and more factors should be considered to make a accurate identification for their vulnerability. So this proposed method can get a reasonable vulnerability comparison result in Karate club network, whereas, the classical method ${R_x}$ can only get the same vulnerability for two communities.
Community ${{S_x}}$ ${{T_x}}$ ${D_x^{in}}$ ${D_x^{out}}$ ${\left| {{V_x}} \right|}$ $Vu{l_x}$ $R{V_x}$ ${v_x}$ ${R_x}$
----------- ----------- ----------- -------------- --------------- ---------------------------- ----------- ---------- --------- ---------
1 1 0.7060 1 1 1 1.4162 1.4162 1 1
2 1 1 1 1 1 1 1 1 1
: **The vulnerability of communities in Karate network.**
\[table\_vulnerability\_Karate\]
Italian 380KV power grid
------------------------
Lastly, the Italian 380KV power transmission grid network [@Crucitti2005Locating] is used in this section. This network has been frequently used to analyse the network vulnerability performance. The topological structure and community structure is shown in Fig. \[fig\_Italian\_network\]. Observing from Fig. \[fig\_Italian\_network\], this network is divided into 10 communities ($Q = 0.7596$), and different communities have different number of components and the detail is shown in Table \[table\_Italian\]. It can be found that community 9 and community 7 have the maximum and minimum number of components respectively.
![**Italian 380KV power grid network [@Crucitti2005Locating].**[]{data-label="fig_Italian_network"}](Italian.eps "fig:"){width="15cm"}\
\[table\_Italian\]
The vulnerability results are shown in Table \[table\_vulnerability\_Italian\]. In this network, a novel method which is proposed by Wei et al. [@Wei2018Measuring] is used as a comparing method, this method is also modified from Ref [@Rocco2011Vulnerability]. The detail parameters and vulnerability results of communities are shown in Table \[table\_vulnerability\_Italian\]. Observing from Table \[table\_vulnerability\_Italian\], community 5 is the most robust community from $R{V_x}$ and ${R_x}$, but ${{R'}_x}$ give a result that community 7 is the most robust. Community 7 only have 6 nodes which is the minimum number of nodes, so this conclusion (community 7 is the most robust) have low credibility, and it can be also seen from these four parameters (${{S_x}}$, ${{T_x}}$, ${D_x^{in}}$, ${D_x^{out}}$) in Table \[table\_vulnerability\_Italian\]. Community 10 is consider to be the most vulnerable from these three methods at the same time. This proposed method would magnify the vulnerability of community, like the relative vulnerability of propose method $R{V_x} = 43.8623$ and other method ${R_x} = 8$, ${{R'}_x} = 11.4347$. The $R{V_x}$ of most vulnerable community would be much bigger than other methods, which is convenient to find the vulnerability of community. Meanwhile, the classical method cannot identify some communities’ vulnerability, such as community 2, 4, 6 and community 1, 3, 7, 8, because of the same ${R_x}$. This proposed method and Wei et al. method can get a certain vulnerability order for these communities (2, 4, 6 and 1, 3, 7, 8). The detail vulnerability orders of these communities obtained by different methods are shown in Table \[table\_Italian\_order\]. From these orders, it can be found that the order obtained by this proposed method is more similar with classical method than Wei method. Specifically, community 9 is considered as the second to last vulnerable community in classical method and proposed method, but it is considered as second vulnerable community by Wei method. Community 3, 7, 8 is considered as the second vulnerable community by classical method at the same time, and this proposed method give a conclusion that community 7, 8, 3 is the second, third, forth vulnerable community respectively which is similar with classical method, but Wei method gives a dissimilar order. Other more detail information about the vulnerability order of communities can be obtained from Table \[table\_Italian\_order\]. So this proposed method would consider more information of community and give a certain vulnerability order, and is more reasonable than other methods.
\[table\_vulnerability\_Italian\]
Method Vulnerability order
------------------------------------- --------------------------------------------------------------------
Classical [@Rocco2011Vulnerability] Community 5 $<$ 9 $<$ 2 $=$ 4 $=$ 6 $<$ 1 $=$ 3 $=$ 7 $=$ 8 $<$ 10
Proposed Community 5 $<$ 9 $<$ 4 $<$ 1 $<$ 2 $<$ 6 $<$ 3 $<$ 8 $<$ 7 $<$ 10
Wei et al. [@Wei2018Measuring] Community 7 $<$ 6 $<$ 5 $<$ 2 $<$ 4 $<$ 3 $<$ 8 $<$ 1 $<$ 9 $<$ 10
: **The vulnerability order of Italian 380KV power network by different method.**
\[table\_Italian\_order\]
Conclusion
==========
The vulnerability measuring of community has already been a hot topic in the study of network theory. In this paper, a new entropy-based method is proposed to measure the vulnerability of communities which can overcome the shortcomings and limitations of previous methods. Different with previous method, this proposed method combines the internal factors and external factors of community which give sufficient consideration of community information. Thus, reasonable vulnerability result can be obtained by this proposed method. The internal factors contain the number of edges inside community and the complexity degree of community measured by Tsallis structure entropy, and the external factors contain the number of edges outside community and the similarity degree between chosen community and other communities measured by relative entropy. The vulnerability and relative vulnerability of community are obtained to give the quantitative description of vulnerability of community eventually. In order to show the performance and effectiveness of this proposed method, one example network and three real-world complex network are applied. Through the vulnerability order obtained by different methods, the rationality of this method is demonstrated. In addition, the sensitivity of weight factors are analysed by Sobol’ indices, the important parameters considered in this model can be obtained. The experiment results show the superiority and reasonableness of this propose method.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors thank Prof. Claudio Rocco for providing us with some network data. The work is partially supported by National Natural Science Foundation of China (Grant Nos. 61573290, 61503237).
Reference {#reference .unnumbered}
=========
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Natural language offers an intuitive and flexible means for humans to communicate with the robots that we will increasingly work alongside in our homes and workplaces. Recent advancements have given rise to robots that are able to interpret natural language manipulation and navigation commands, but these methods require a prior map of the robot’s environment. In this paper, we propose a novel learning framework that enables robots to successfully follow natural language route directions without any previous knowledge of the environment. The algorithm utilizes spatial and semantic information that the human conveys through the command to learn a distribution over the metric and semantic properties of spatially extended environments. Our method uses this distribution in place of the latent world model and interprets the natural language instruction as a distribution over the intended behavior. A novel belief space planner reasons directly over the map and behavior distributions to solve for a policy using imitation learning. We evaluate our framework on a voice-commandable wheelchair. The results demonstrate that by learning and performing inference over a latent environment model, the algorithm is able to successfully follow natural language route directions within novel, extended environments.'
author:
- |
Sachithra Hemachandra$^*$ Felix Duvallet$^*$ Thomas M. Howard\
Nicholas Roy Anthony Stentz Matthew R. Walter[^1][^2][^3][^4] [^5]
bibliography:
- 'references.bib'
title: |
**Learning Models for Following Natural Language Directions\
in Unknown Environments**
---
=1
Introduction {#sec:introduction}
============
Over the past decade, robots have moved out of controlled isolation and into our homes and workplaces, where they coexist with people in domains that include healthcare and manufacturing. One long-standing challenge to realizing robots that behave effectively as our partners is to develop command and control mechanisms that are both intuitive and efficient. Natural language offers a flexible medium through which people can communicate with robots, without requiring specialized interfaces or significant prior training. For example, a voice-commandable wheelchair [@hemachandra11] allows the mobility-impaired to independently and safely navigate their surroundings simply by speaking to the chair, without the need for traditional head-actuated switches or sip-and-puff arrays. Recognizing these advantages, much attention has been paid of late to developing algorithms that enable robots to interpret natural language expressions that provide route directions [@macmahon06; @kollar10; @chen11; @matuszek12], that command manipulation [@tellex11; @howard14a], and that convey environment knowledge [@walter13; @hemachandra14].
![Our goal is to enable robots to autonomously follow natural language commands without any prior knowledge of their environment.[]{data-label="fig:go_to_kitchen_down_hallway"}](./graphics/wheelchair-commanding2.pdf){width="1.0\linewidth"}
Natural language interpretation becomes particularly challenging when the expression references areas in the environment unknown to the robot. Consider an example in which a user directs the voice-commandable to “go to the kitchen that is down the hallway,” when the wheelchair is in an unknown environment and the hallway and kitchen are outside the field-of-view of its sensors (Fig. \[fig:go\_to\_kitchen\_down\_hallway\]). Unable to associate the hallway and kitchen with specific locations, most existing solutions to language understanding would result in the robot exploring until it happens upon a kitchen. By reasoning over the spatial and semantic environment information that the command conveys, however, the robot would be able to follow the spoken directions more efficiently.
In this paper, we propose a framework that follows natural language route directions within unknown environments by exploiting spatial and semantic knowledge implicit in the commands. There are three algorithmic contributions that are integral to our approach. The first is a learned language understanding model that efficiently infers environment annotations and desired behaviors from the user’s command. The second is an estimation-theoretic algorithm that learns a distribution over hypothesized world models by treating the inferred annotations as observations of the environment and fusing them as observations from the robot’s sensor streams (Fig. \[fig:belief\_world\]). The third is a belief space policy learned from human demonstrations that reasons directly over the world model distribution to identify suitable navigation actions.
This paper generalizes previous work by the authors [@duvallet14], which was limited to object-relative navigation within small, open environments. The novel contributions of this work enable robots to follow natural language route directions in large, complex environments. They include: a hierarchical framework that learns a compact probabilistic graphical model for language understanding; a semantic map inference algorithm that hypothesizes the existence and location of regions in spatially extended environments; and a belief space policy learned from human demonstrations that considers spatial relationships with respect to a hypothesized map distribution. We demonstrate these advantages through simulations and experiments with a voice-commandable wheelchair in an office-like environment.
Related Work {#sec:related_work}
============
Recent advancements in language understanding have enabled robots to understand free-form commands that instruct them to manipulate objects [@tellex11; @howard14a] or navigate through environments using route directions [@macmahon06; @kollar10; @chen11; @howard14a; @matuszek12a]. With few exceptions, most of these techniques require a priori knowledge of location, geometry, colloquial name, and type of all objects and regions within the environment [@kollar10; @howard14a; @tellex11]. Without known world models, however, interpreting free-form commands becomes much more difficult. Existing methods have dealt with this by learning a parser that maps the natural language command directly to plans [@macmahon06; @chen11; @matuszek12a]. Alternatively, Duvallet et al. [@duvallet2013] use imitation learning to train a policy that reasons about uncertainty in the grounding and that is able to backtrack as necessary. However, none of these approaches explicitly utilize the knowledge that the instruction conveys to influence their models of the environment, nor do they reason about its uncertainty. Instead, our framework treats language as an additional, albeit noisy, sensor that we use to learn a distribution over hypothesized world models, by taking advantage of information implicitly contained in a given command.
Related to our algorithm’s ability to learn world models, state-of-the-art semantic mapping frameworks exist that focus on using the robot’s sensor observations to update its representation of the world [@zender08; @pronobis10]. Some methods additionally incorporate natural language descriptions in order to improve the learned world models [@walter13; @hemachandra14]. These techniques, however, only use language to update regions of the environment that the robot has observed and are not able to extend the maps based on natural language. Our approach treats natural language as another sensor and uses it to extend the spatial representation by adding both topological and metric information regarding hypothesized regions in the environment, which is then used for planning. Williams et al. [@Williams2013] use a cognitive architecture to add unvisited locations to a partial map. However, they only reason about topological relationships to unknown places, do not maintain multiple hypotheses, and make strong assumptions about the environment that limit the applicability to real systems. In contrast, our approach reasons both topologically and metrically about regions, and can deal with ambiguity, which allows us to operate in challenging environments.
Approach Overview {#sec:overview}
=================
We define natural language direction following as one of inferring the robot’s trajectory $x_{t+1:T}$ that is most likely for a given command $\Lambda^t$: $${\ensuremath{\underset{x_{t+1:T} \, \in \, \Re^{n}}{\arg \! \max}\;}}
p\left(x_{t+1:T} | \Lambda^t, z^t, u^t \right),
\label{eqn:problem-statement}$$ where $z^t$ and $u^t$ are the history of sensor observations and odometry data, respectively. Traditionally, this problem has been solved by also conditioning the distribution over a known world model. Without any a priori knowledge of the environment, we treat this world model as a latent variable $S_t$. We then interpret the natural language command in terms of the latent world model, which results in a distribution over behaviors $\beta_t$. We then solve the inference problem by marginalizing over the latent world model and behaviors: $$\begin{split}
{\ensuremath{\underset{x_{t+1:T} \, \in \, \Re^{n}}{\arg \! \max}\;}}
\int\displaylimits_{\beta_t} \int\displaylimits_{S_{t}} p(x_{t+1:T}
&\vert \beta_t, S_{t}, \Lambda^t) \cdot
p(\beta_t \vert S_t, \Lambda^t)\\
&\cdot p(S_{t} \vert \Lambda^t) \, dS_{t} \, d\beta_t,
\end{split}\label{eqn:marginalization}$$ where we have omitted the measurement $z^t$ and odometry $u^t$ histories for lack of space.
By structuring the problem in this way, we are able to treat inference as three coupled learning problems. The framework (Fig. \[fig:framework\]) first converts the natural language direction into a set of environment annotations using learned language grounding models. It then treats these annotations as observations of the environment (i.e., the existence, name, and relative location of rooms) that it uses together with data from the robot’s onboard sensors to learn a distribution over possible world models (third factor in Eqn. \[eqn:marginalization\]). Our framework then infers a distribution over behaviors conditioned upon the world model and the command (second factor). We then solve for the navigation actions that are consistent with this behavior distribution (first factor) using a learned belief space policy that commands a single action to the robot. As the robot executes this action, we update the world model distribution based upon new utterances and sensor observations, and subsequently select an updated action according to the policy. This process repeats as the robot navigates.
The rest of this paper details each of these components in turn. We then demonstrate our approach to following natural language directions through large unstructured indoor environments on the robot shown in \[fig:go\_to\_kitchen\_down\_hallway\] as well as simulated experiments. We additionally evaluate our approach to learning belief space policies on a corpus of natural language directions through one floor of an indoor building.
(annotation-inference) at (0,0) [annotation inference]{}; (semantic-mapping) at (0,-2) [semantic mapping]{}; (behavior-inference) at (3.5,0) [behavior inference]{}; (policy-planner) at (3.5,-2) [policy planner]{}; (robot) at (-3.0,-2.0) [![Outline of the framework.[]{data-label="fig:framework"}](graphics/robot.png "fig:"){width="1.5cm"}]{}; (annotation-inference) – (semantic-mapping); (semantic-mapping) –(1.75,-2) – (1.75,0) – (behavior-inference); (semantic-mapping) – (policy-planner); (behavior-inference) – (policy-planner); (observationlabel) at (2.325,-1.0) [map distribution]{}; (annotationdistributionlabel) at (0.625,-1.0) [annotation distribution]{}; (behaviordistributionlabel) at (4.125,-1.0) [behavior distribution]{}; (speechlabel) at (-3.0,0) [“go to the kitchen that is down the hallway”]{}; (robot) – (speechlabel); (robot) – (semantic-mapping); (speechlabel) – (-3.0,1.0) – (0.0,1.0) – (annotation-inference); (speechlabel) – (-3.0,1.0) – (3.5,1.0) – (behavior-inference); (policy-planner) – (3.5,-3.5) – (-3.0,-3.5) – (robot); (observationslabel) at (-1.5,-1.75) [observations]{}; (parsetreelabel) at (0,1.25) [parse tree(s)]{}; (actionlabel) at (0,-3.25) [action]{};
Natural Language Understanding {#sec:nlu}
==============================
Our framework relies on learned models to identify the existence of annotations and behaviors conveyed by free-form language and to convert these into a form suitable for semantic mapping and the belief space planner. This is a challenge because of the diversity of natural language directions, annotations, and behaviors. We perform this translation using the Hierarchical Distributed Correspondence Graph (HDCG) model [@howard14b], which is a more efficient extension of the Distributed Correspondence Graph (DCG) [@howard14a]. The DCG exploits the grammatical structure of language to formulate a probabilistic graphical model that expresses the correspondence $\phi \in \Phi$ between linguistic elements from the command and their corresponding constituents (*groundings*) $\gamma \in \Gamma$. The factors $f$ in the DCG are represented by log-linear models with feature weights that are learned from a training corpus. The task of grounding a given expression then becomes a problem of inference on the DCG model.
The HDCG model employs DCG models in a hierarchical fashion, by inferring rules $\mathtt{R}$ to construct the space of groundings for lower levels in the hierarchy. At any one level, the algorithm constructs the space of groundings based upon a distribution over the rules from the previous level: $$\Gamma \rightarrow \Gamma\left(\mathtt{R}\right).$$ The HDCG model treats these rules and, in turn, the structure of the graph, as latent variables. Language understanding then proceeds by performing inference on the marginalized models: $$\begin{aligned}
&\operatorname*{arg\,max}_\Phi \int_{\mathtt{R}} p\left(\Phi\vert\mathtt{R},\Gamma\left(\mathtt{R}\right),\Lambda,\Psi\right) p\left(\mathtt{R} \vert \Gamma\left(\mathtt{R}\right),\Lambda,\Psi\right)\\
&\operatorname*{arg\,max}_\Phi \int_{\mathtt{R}} \prod_{i} \prod_{j} f\left(\Phi_{i_{j}},\Gamma_{i_{j}}\left(\mathtt{R}\right),\Lambda_{i},\Psi,\mathtt{R}\right)\times\\
&\qquad \qquad \qquad \prod_{i} \prod_{j} f\left(\mathtt{R},\Lambda_{i},\Psi,\Gamma_{i_{j}}\left(\mathtt{R}\right)\right).\nonumber\end{aligned}$$ We now describe how the HDCG model infers annotations (representing our knowledge of the environment inferred from the language) and behaviors (representing the intent of the command) to understand the natural language command given by the user.
Annotation Inference
--------------------
An annotation is a set of object types and subspaces. A subspace is defined here as a spatial relationship (e.g., down, left, right) with respect to an object type. In the experiments described in Section \[sec:results\] we assume 17 object types and 12 spatial relationships. We also permit object types to express a spatial relationship with another object type. We denote object types by their physical type (e.g., kitchen, hallway), subspaces as the relationship type with an object type argument (e.g., down(kitchen), left(hallway)), and object types with spatial relationships as an object type with a subspace argument (e.g., kitchen(down(hallway))). Since the number of possible combinations of annotations is equal to the power set of the number of symbols, $2^{3,485}$ annotations can be expressed by an instruction.[^6] The HDCG model infers a distribution of graphical models to efficiently generate annotations by assuming conditional independence of constituents and eliminating symbols that are learned to be irrelevant to the utterance. For example, Figure \[fig:annotation-inference\] illustrates the model for the direction “go to the kitchen that is down the hall.” In this example only 4 of the 3,485 symbols (two object types, one subspace, and one object type with a spatial relationship) are active in this model. Note that all factors with inactive correspondence variables are not illustrated in Figures \[fig:annotation-inference\] and \[fig:behavior-inference\]. At the root of the sentence the symbols for an object type (kitchen) and an object type with a spatial relationship (kitchen(down(hallway))) are sent to the semantic map to fuse with other observations.
Behavior Inference
------------------
A behavior is a set of objects, subspaces, actions, objectives, and constraints. Behavior inference differs from annotation inference by considering objects from the semantic map and subspaces defined with respect to objects from the semantic map instead of only object types. We denote actions by their type and an object or subspace argument (e.g., navigate(hallway)), objectives by their type (e.g., quickly, safely), and constraints as objects with spatial relationship from the semantic map (e.g., $o_{4}$(down($o_{3}$))). In the experiments presented in Section \[sec:results\] we assume 4 action types, 3 objectives, and 12 spatial relations. Just as with annotation inference, the HDCG model eliminates irrelevant action types, objective types, objects, and spatial relationships to efficiently infer behaviors. Figure \[fig:behavior-inference\] illustrates the model for the direction “go to the kitchen that is down the hall” in the context of an inferred map. In this example a *navigate* action with a goal relative to $o_{1}$ would be inferred as the most likely behavior for the policy planner.
Semantic Mapping
================
We represent the world model as a modified *semantic map* [@walter13] $S_t = \{G_t, X_t\}$, a hybrid metric and topological representation of the environment. The topology $G_t$ consists of nodes $n_i$ that denote locations in the environment, edges that denote inter-node connections, and non-overlapping regions $R_\alpha = \{n_1, n_2, \ldots, n_m\}$ that represent spatially coherent areas compatible with a human’s decomposition of space (e.g., rooms and hallways). We associate a pose $x_i$ with each node $n_i$, the vector of which constitutes the metric map $X_t$. Each region is also labeled according to its type (e.g., kitchen, hallway). An edge connects two regions that the robot has transitioned between or for which language indicates the existence of an inter-region spatial relation (e.g., that the kitchen is “down” the hallway).
Annotations extracted from a given command provide information regarding the existence, relative location, and type of regions[^7] in the environment. We learn a distribution over world models consistent with these annotations by treating them as observations $\alpha_t$ in a filtering framework. We combine these observations with those from other sensors onboard the robot (LIDAR and region appearance observations) $z_t$ to maintain a distribution over the semantic map:
$$\begin{aligned}
p(S_t \vert \Lambda^t, z^t, u^t)\! &\approx p(S_t \vert
\alpha^t, z^t, u^t)\\
\!&= p(G_t, X_t, \vert \alpha^t, z^t, u^t)\\
\!&= p(X_t \vert G_t, \alpha^t, z^t, u^t) p(G_t \vert \alpha^t, z^t, u^t),
\end{aligned}$$
where we assume that an utterance $\Lambda^t$ provides a set of annotations $\alpha_t$. The factorization within the last line models the metric map induced by the topology, as with pose graph representations [@kaess08]. We maintain this distribution over time using a Rao-Blackwellized particle filter (RBPF) [@doucet00], with a sample-based approximation of the distribution over the topology, and a Gaussian distribution over metric poses.
The robot observes transitions between environment regions and the semantic label of its current region. As scene understanding is not the focus of this work, we use AprilTag fiducials [@Olson2011] placed in each region that denotes its label. Unlike our earlier work [@hemachandra14] in which we segment regions based only on their spatial coherence using spatial clustering, here we additionally use the presence of conflicting spatial appearance tags to also segment the region. As such, we assume that we are aware of the segmentation of the space immediately, which is not possible with a purely spectral clustering based approach, allowing us to immediately evaluate each particle’s likelihood based on the observation of region appearance. In turn, we can down-weight particles that are inconsistent with the actual layout of the world sooner, reducing the number of actions the robot must take to satisfy the command.
We maintain each particle through the three steps of the RPBF. First, we propagate the topology by sampling modifications to the graph when the robot receives new sensor observations or annotations. Second, we perform a Bayesian update to the pose distribution based upon the sampled modifications to the underlying graph. Third, we update the weight of each particle based on the likelihood of generating the given observations, and resample as needed to avoid particle depletion. We now outline this process in more detail.
During the proposal step, we first add an additional node $n_t$ and edge to each particle’s topology that model the robot’s motion $u_t$, yielding a new topology $S_t^{(i)-}$. We then sample modifications to the topology $\Delta_t^{(i)} = \{\Delta_{\alpha_t}^{(i)}, \Delta_{z_t}^{(i)}\}$ based on the most recent annotations $\alpha_t$ and sensor observations $z_t$: $$\begin{gathered}
\label{eq:sample_graph}
p(S_t^{(i)} | S_{t-1}^{(i)}, \alpha_t, z_t, u_t)=p(\Delta_{\alpha_t}^{(i)} | S_{t}^{(i)-}, \alpha_t)\, \\
p(\Delta_{z_t}^{(i)} | S_{t}^{(i)-}, z_t)\,
p(S_t^{(i)-} | S_{t-1}^{(i)}, u_t).\end{gathered}$$ This updates the proposed graph topology $S_{t}^{(i)-}$ with the graph modifications $\Delta_t^{(i)}$ to yield the new semantic map $S_t^{(i)}$. The updates can include the addition and deletion of nodes and regions from the graph that represent newly hypothesized or observed regions, and edges that express express spatial relations inferred from observations or annotations.
We sample graph modifications from two independent proposal distributions for annotations $\alpha_t$ and robot observations $z_t$. This is done by sampling a grounding for each observation and modifying the graph according to the implied grounding.
Graph modifications based on natural language
---------------------------------------------
Given a set of annotations , we sample modifications to the graph for each particle. An annotation $\alpha_{t,j}$ contains a spatial relation and figure when the language describes one region (e.g., “go to the elevator lobby”), and an additional landmark when the language describes the relation between two regions (e.g., “go to the lobby through the hallway”). We use a likelihood model over the spatial relation to sample landmark and figure pairs for the grounding. This model employs a Dirichlet process prior that accounts for the fact that the annotation may refer to regions that exist in the map or to unknown regions. If either the landmark or the figure are sampled as new regions, we add them to the graph and create an edge between them. We also sample the metric constraint associated with this edge based on the spatial relation. The spatial relation models employ features that describe the locations of the regions, their boundaries, and robot’s location at the time of the utterance, and are trained based upon a natural language corpus [@tellex11].
Graph modifications based on robot observations
-----------------------------------------------
If the robot does not observe a region transition (i.e. the robot is in the same region as before), the algorithm adds the new node $n_t$ to the current region and modifies its spatial extent. If there are any edges denoting spatial relations to hypothesized regions, the algorithm resamples their constraint if its likelihood changes significantly due to the modified spatial extent of the current region. Alternatively, if the robot observes a region transition, the new node $n_t$ is assigned to a new or existing region as follows. First, the algorithm checks if the robot is in a previously visited region, based on spatial proximity, in which case it will add $n_t$ to that region. Otherwise, it will create a new region and check whether it matches a region that was previously hypothesized based on an annotation (for example, a newly-visited kitchen can be the same as a hypothesized kitchen described with language). We do so by sampling a grounding to any unobserved regions in the topology using a Dirichlet process prior. If this process results in a grounding to an existing hypothesized region, we remove the hypothesized region and adjust the topology accordingly, resampling any edges to yet-unobserved regions. For example, if an annotation suggested the existence of a “kitchen down the hallway,” and we grounded the robot’s current region to the hypothesized hallway, we would reevaluate the “down” relation for the hypothesized kitchen with respect to this detected hallway.
Re-weighting particles and resampling
-------------------------------------
After modifying each particle’s topology, we perform a Bayesian update to its Gaussian distribution. We then re-weight each particle according to the likelihood of generating language annotations and region appearance observations: $$\label{eq:weight_update_all}
w_t^{(i)}\!\!=\! p(z_t, \alpha_t | S_{t-1}^{(i)}) w_{t-1}^{(i)} \!\!=\! p(\alpha_t| S_{t-1}^{(i)}) p(z_t| S_{t-1}^{(i)}) w_{t-1}^{(i)}.$$ When calculating the likelihood of each region appearance observation, we consider the current node’s region type and calculate the likelihood of generating this observation given the topology. In effect, this down-weights any particle with a sampled region of a particular type existing on top of a known traversed region of a different type. We use a likelihood model that describes the observation of a region’s type, with a latent binary variable $v$ that denotes whether or not the observation is valid. We marginalize over $v$ to arrive at the likelihood of generating the given observation, where $R_u$ is the set of unobserved regions in particle $S_{t-1}^{(i)}$: $$\label{eq:likelihood_appearance}
p(z^t\vert{S_{t-1}^{(i)}}) = \prod_{R_i\in {R_u}}\left(\sum\limits_{v\in {1,0}}{p(z^t\vert{v,R_i}) \times p(v\vert{R_i})}\right).$$ For annotations, we use the language grounding likelihood under the map at the previous time step. As such, a particle with an existing pair of regions conforming to a specified language constraint will be weighted higher than one without. When the particle weights fall below a threshold, we resample particles to avoid particle depletion [@doucet00].
Reasoning and Learning in Belief Space {#sec:planning}
======================================
Searching for the complete trajectory that is optimal in the distribution of maps would be intractable. Instead, we treat direction following as sequential decision making under uncertainty, where a policy $\pi$ minimizes a single step of the cost function $c$ over the available actions $a \in A_t$ from state $x$: $$\label{eqnPolicy}
\pi {\left( x, {S_t}\right)} = {\ensuremath{\underset{a \in A_t}{\arg \! \min}\;}} c{\left( x, a, {S_t}\right)}.$$ After executing the action and updating the map distribution, we repeat this process until the policy declares it has completed following the direction using a separate stop action.
As the robot travels in the environment, it keeps track of the nodes in the topological graph $G_t$ it has visited (${\mathcal{V}}$) and frontiers (${\mathcal{F}}$) that lie at the edge of explored space. The action set $A_t$ consists of paths to nodes in the graph. An additional action ${a_\textrm{stop}}$ declares that the policy has completed following the direction. Intuitively, an action represents a single step along the path that takes the robot towards its destination. Each action may explore new parts of the environment (for example continuing to travel down a hallway) or backtrack if the policy has made a mistake (for example, traveling to a room in a different part of the environment). The following sections explain how the policy reasons in belief space, and the novel imitation learning formulation to train the policy from demonstrations of correct behavior.
Belief Space Reasoning using Distribution Embedding
---------------------------------------------------
The semantic map ${S_t}$ provides a distribution over the possible locations of the landmarks relevant to the command the robot is following. As such, the policy $\pi$ must reason about a *distribution* of action features when computing the cost of any action $a$. We accomplish this by embedding the action feature distribution in a Reproducing Kernel Hilbert Space (RKHS), using the mean feature map [@smola2007] consisting of the first $K$ moments of the features computed with respect to each map sample ${S_t^{{\left( i \right)}}}$ (and its likelihood): $$\begin{aligned}
\label{eqnPolicyDecomposition}
{\hat{\Phi}_1}{\left( x, a, {S_t}\right)} & = \sum_{{S_t^{{\left( i \right)}}}} p({S_t^{{\left( i \right)}}}) \: {\phi {\left( x, a, {S_t^{{\left( i \right)}}}\right)}}\\
{\hat{\Phi}_2}{\left( x, a, {S_t}\right)} & = \sum_{{S_t^{{\left( i \right)}}}} p({S_t^{{\left( i \right)}}}) \: {\left( {\phi {\left( x, a, {S_t^{{\left( i \right)}}}\right)}}- {\hat{\Phi}_1}\right)}^2\\
& \ldots \nonumber \\
{\hat{\Phi}_k}{\left( x, a, {S_t}\right)} & = \sum_{{S_t^{{\left( i \right)}}}} p({S_t^{{\left( i \right)}}}) \: {\left( {\phi {\left( x, a, {S_t^{{\left( i \right)}}}\right)}}- {\hat{\Phi}_1}\right)}^k\end{aligned}$$ Intuitively, this formulation computes features for the action and all hypothesized landmarks individually, aggregates these feature vectors, and then computes moments of the feature vector distribution (mean, variance, and higher order statistics). A simplified illustration, shown in , shows how our approach computes belief space features for two actions with a hypothesized kitchen (with two possible locations).
The cost function in can now be rewritten as a weighted sum of the first $K$ moments of the feature distribution: $$\label{eqnBeliefPolicySumK}
{c{\left( x, a, {S_t}\right)}}= \sum_{i=1}^K {w_i^T}\; {\hat{\Phi}_i}{{\left( x, a, {S_t}\right)}}.$$ By concatenating the weights and moments into respective column vectors $W := [w_1; \ldots; w_k]$ and $F := [{\hat{\Phi}_1}; \ldots; {\hat{\Phi}_k}]$, we can rewrite the policy in as minimizing a weighted sum of the feature moments $F_a$ for action $a$ : $$\label{eqnBeliefPolicyVectorK}
\pi {\left( x, {S_t}\right)} = {\ensuremath{\underset{a \in A_t}{\arg \! \min}\;}} W^T F_a.
$$
The vector $\phi (x, a, {S_t^{{\left( i \right)}}})$ are features of the action and a *single* landmark in ${S_t^{{\left( i \right)}}}$. It contains geometric features describing the shape of the action (e.g., the cumulative change in angle), the geometry of the landmark (e.g., the area of the landmark), and the relationship between the action and landmark (e.g., the difference between the ending and starting distances to the landmark). See [@duvallet2013] for more details.
Imitation Learning Formulation
------------------------------
We use imitation learning to train the policy by treating action prediction as a multi-class classification problem: given an expert demonstration, we wish to correctly predict their action among all possible actions for the same state. Although prior work introduced imitation learning for training a direction following policy, it operated in partially known environments [@duvallet2013]. Instead, we train a belief space policy that reasons in a *distribution* of hypothesized maps.
We assume the expert’s policy ${\pi^*}$ minimizes the unknown immediate cost $C(x, a^*, {S_t})$ of performing the demonstrated action $a^*$ from state $x$, under the map distribution ${S_t}$. However, since we cannot directly observe the true costs of the expert’s policy, we must instead minimize a surrogate loss that penalizes disagreements between the expert’s action $a^*$ and the policy’s action $a$, using the multi-class hinge loss [@Crammer2002]: $$\label{eqnSVMloss}
\ell {\left( x, {{a^{*}}}\!, c, {S_t}\right)} \! = \! \max \! {\left( \! 0, 1 \! + \! {c{\left( x, {{a^{*}}}\negmedspace, {S_t}\right)}}\! - \! \min_{a \ne {{a^{*}}}} { \left[ {c{\left( x, a, {S_t}\right)}}\right] } \! \right)}.$$ The minimum of this loss occurs when the cost of the expert’s action is lower than the cost of all other actions, with a margin of one. This loss can be re-written and combined with to yield: $$\label{eqLossAugmentation}
\ell {\left( x, {{a^{*}}}, W, {S_t}\right)} = {W^T}F_{{a^{*}}}- \min_a { \left[ {W^T}F_a - l_{xa} \right] },$$ where the margin $l_{xa} = 0$ if $a={{a^{*}}}$ and $1$ otherwise. This ensures that the expert’s action is better than all other actions by a margin [@Ratliff2006]. Adding a regularization term $\lambda$ to yields our complete optimization loss: $$\label{eqnOptimizationLoss}
\ell {\left( x, {{a^{*}}}, W, {S_t}\right)} \! = \! \frac{\lambda}{2} \Vert W \Vert^2 + {W^T}F_{{a^{*}}}- \min_{a} \left[ {W^T}F_a - l_{xa} \right].$$
Although this loss function is convex, it is not differentiable. However, we can optimize it efficiently by taking the subgradient of and computing action predictions for the loss-augmented policy [@Ratliff2006]: $$\begin{aligned}
{\frac{\partial \ell}{\partial W}}& = \lambda W + F_{{a^{*}}}- F_{a'} \\
a' & = {\ensuremath{\underset{a}{\arg \! \min}\;}} \left[ {W^T}F_a - l_{xa} \right].\end{aligned}$$ Note that $a'$ (the best loss-augmented action) is simply the solution to our policy using a loss-augmented cost. This leads to the update rule for the weights $W$: $$\label{eqnUpdateRule}
W_{t+1} \gets W_t - \alpha \; {\frac{\partial \ell}{\partial W}}$$ with a learning rate $\alpha \propto 1/t^\gamma$. Intuitively, if the current policy disagrees with the expert’s demonstration, decreases the weight (and thus the cost) for the features of the demonstrated action $F_{{a^{*}}}$, and increases the weight for the features of the planned action $F_{a'}$. If the policy produces actions that agree with the expert’s demonstration, the update will only be for the regularization term. As in our prior work, we train the policy using the [<span style="font-variant:small-caps;">DAgger</span>]{} (Dataset Aggregation) algorithm [@Ross2011], which learns a policy by iterating between collecting data (using the current policy) and applying expert corrections on all states visited by the policy (using the expert’s demonstrated policy).
Treating direction following in the space of possible semantic maps as a problem of sequential decision making under uncertainty provides an efficient approximate solution to the belief space planning problem. By using a kernel embedding of the distribution of features for a given action, our approach can learn a policy that reasons about the distribution of semantic maps.
Results {#sec:results}
=======
We implemented the algorithm on our voice-commandable wheelchair (Fig. \[fig:go\_to\_kitchen\_down\_hallway\]), which is equipped with three forward-facing cameras with a collective field-of-view of 120 degrees, and forward- and rearward-facing LIDARs. We set up an experiment in which the wheelchair was placed in a lobby within MIT’s Stata Center, with several hallways, offices, and lab spaces, as well as a kitchen on the same floor. As scene understanding is not the focus of this paper, we placed AprilTag fiducials [@Olson2011] to identify the existence and semantic type of regions in the environment. We trained the HDCG models from a parallel corpus of 54 fully-labeled examples. We then directed the wheelchair to execute the novel instruction “go to the kitchen that is down the hallway.”
[0.9]{}[Scccc]{} & &\
Algorithm & Mean & Std Dev & Mean & Std Dev\
Known Map & 13.10 & 0.67 & 62.48 & 16.61\
With Language & 12.62 & 0.62 & 122.14 & 32.48\
Without Language & 24.91 & 13.55 & 210.35 & 97.73\
\[tab:results\_robot\]
[0.9]{}[Scccc]{} & &\
Algorithm & Mean & Std Dev & Mean & Std Dev\
Known Map & 12.88 & 0.06 & 18.32 & 3.54\
With Language & 16.64 & 6.84 & 82.78 & 10.56\
Without Language & 25.28 & 12.99 & 85.57 & 17.80\
\[tab:results\_sim\]
We compare our framework against two other methods. The first emulates the previous state-of-the-art and uses a known map of the environment in order to infer the actions consistent with the route direction. The second assumes no prior knowledge of the environment (as with ours) and opportunistically grounds the command in the map, but does not use language to modify the map. We performed six experiments with our algorithm, three with the known map method, and five with the method that does not use language, all of which were successful (the robot reached the kitchen). compares the total distance traveled and execution time for the three methods. Our algorithm resulted in paths with lengths close to those of the known map, and significantly outperformed the method that did not use language. Our framework did require significantly more time to follow the directions than the known map case, due to the fact that it repeats the three steps of the algorithm when new sensor data arrives. shows a visualization of the semantic maps over several time steps for one successful run on the robot.
We performed a similar evaluation in a simulated environment comprised of an office, hallway, and kitchen. With the robot starting in the office, we ran ten simulations of each method. As with the physical experiment, our method resulted in an average length closer to that of the known map case, but with a longer average run time ().
To evaluate the performance of the learned belief space policy in isolation on a larger corpus of natural language directions (with more verbs, spatial relations, and landmarks), we performed cross-validation trials of the policy operating in a simplified simulated map. We evaluated the policy using a corpus of 55 multi-step natural language directions, some of which refer to navigation landmarks (for example, the direction shown in \[figFullDirectionAmbiguousEight\]). These directions are similar to those in our prior work [@duvallet2013]. For this cross-validation evaluation, we trained the policy on 28 randomly-sampled directions then evaluated the learned policy on the remaining 27 directions (measuring the average ending distance error across the held out directions). The results of this experiment, shown in \[figAmbiguousDistanceBoxPlots\], demonstrate the benefit of using the additional information available in the direction to infer a distribution of possible environment models. By contrast, our prior approach (without belief space reasoning) ignores this information which results in larger ending distance errors.
Conclusions
===========
Robots that can understand and follow natural language directions in unknown environments are one step towards intuitive human-robot interaction. Reasoning about parts of the environment that have not yet been detected would help enable seamless coordination in human-robot teams.
We have generalized our prior work to move beyond object-relative navigation in small, open environments. The primary contributions of this work include:
- a hierarchical framework that learns a compact probabilistic graphical model for language understanding;
- a semantic map inference algorithm that hypothesizes the existence and location of spatially coherent regions in large environments; and
- a belief space policy that reasons directly over the hypothesized map distribution and is trained based on expert demonstrations.
Together, these algorithms are integral to efficiently interpreting and following natural language route directions in unknown, spatially extended, and complex environments. We evaluated our algorithm through a series of simulations as well as demonstrations on a voice-commandable autonomous wheelchair tasked with following natural language route instructions in an office-like environment.
In the future, we plan to carry out experiments on a more diverse set of commands. Other future work will focus on handling sequences of commands, as well as streams of command that are given *during* execution to change the behavior of the robot.
Acknowledgments {#acknowledgments .unnumbered}
===============
[^1]: $^*$The first two authors contributed equally to this paper.
[^2]: S. Hemachandra and N. Roy are with the Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA USA [ {sachih,tmhoward,nickroy}@csail.mit.edu]{}
[^3]: F. Duvallet and A. Stentz are with the Robotics Institute, Carnegie Mellon University, Pittsburgh, PA USA [ {felixd,tony}@cmu.edu]{}
[^4]: T.M. Howard is with the University of Rochester, Rochester, NY USA [[email protected]]{}
[^5]: M.R. Walter is with the Toyota Technological Institute at Chicago, Chicago, IL USA [[email protected]]{}
[^6]: 3,485 symbols = 17 object types, 204 subspaces, and 3,264 object types with spatial relationships (we exclude object types with spatial relationships to the same object type)
[^7]: Regions as defined by the mapping framework are also considered as objects for the purpose of natural language understanding.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
We study a many-body system of interacting spin-1 particles in the context of ultracold atomic gases. Its description requires eight parameters among which there are three components of magnetization and five parameters associated with quadrupole degrees of freedom. Based on the symmetry considerations, we construct a many-body interaction Hamiltonian that includes eight generators of the SU(3) group related to the above description parameters. This Hamiltonian can be generalized to spin-S systems by considering the generators of the SU(2S+1) group. We apply the Hamiltonian to study the ferromagnetic phase of a weakly interacting gas of spin-1 atoms with Bose-Einstein condensate. It is shown how the quadrupole degrees of freedom entering the Hamiltonian modify the ground state and single-particle excitation spectra in comparison with those obtained from the Hamiltonian containing the spin operators only. We discuss the issue of taking into account the local character of interaction to obtain the correct spectra of single-particle excitations.
[**Keywords:**]{} high spin magnets; Ultracold gases, Bose-Einstein condensate; Ferromagnetic phase; Single-particle excitations
author:
- 'A.S. Peletminskii'
- 'S.V. Peletminskii'
- 'Yu.V. Slyusarenko'
title: 'SU(3) symmetry in theory of a weakly interacting gas of spin-1 atoms with Bose-Einstein condensate'
---
Introduction
============
The magnetic properties of spin-1/2 crystalline systems are well understood and presented in the literature [@ABP; @GM]. If one ignores the relativistic effects associated with the interaction of the electron magnetic moments, then the description is based on the Heisenberg model Hamiltonian and depending on the sign of the exchange integral, the system exhibits the ferromagnetic or antiferromagnetic ordering. The physical nature of the exchange interaction can be explained by considering the electrostatic interaction of two electrons between themselves and protons in a hydrogen molecule or by employing only the symmetry considerations which require the Hamiltonian to be invariant with respect to spatial rotations [@ABP]. The resulting Hamiltonian is expressed in terms of the Pauli matrices representing the generators of the SU(2) group. It takes into account purely quantum (exchange) effects originating from the Fermi-Dirac statistics for electrons.
The interaction in high-spin ($S>1/2$) crystalline systems has a more complicated character that goes beyond the usual Heisenberg model, while their phase diagram exhibits a more rich structure [@Nagaev]. In particular, for spin-1 systems with bilinear and biquadratic exchange interactions, the exotic orderings, such as nematic [@Nagaev; @Matveev; @Andreev; @Papa] and semi-ordered [@Papa] phases, may exist along with the traditional ferromagnetic and antiferromagnetic phases. Moreover, the non-Heisenberg structure of the spin-spin interaction affects even the traditional phases. In recent years, there have been intensive studies of unconventional orderings in magnets with $S\geqslant 1$ [@Toth; @BH; @NC; @Fridman; @Kosm; @Kolezhuk; @Kov1].
Nowadays, the interest in high-spin systems is attracted by the studies of ultracold atomic gases providing remarkable opportunities to examine and model various effects and phenomena in quantum many-body systems in a well controlled manner. In particular, quantum gases loaded in an optical lattice represent an artificial but effective simulator of magnetic phenomena in crystalline systems [@Sotnikov1; @Sotnikov2]. The first theoretical studies of magnetic phases and corresponding excitations in dilute Bose gases with condensate [@Ohmi; @APS; @Ho] were stimulated by experiments on optical trapping a condensate of ${}^{23}$Na spin-1 atoms [@Stamper]. A comprehensive study of the so-called spinor Bose gases, including those of spin-2 and spin-3 atoms, has been reviewed in Refs. [@Ueda; @Kurn], where the interaction between the atomic degrees of freedom was taken as a bilinear form in spin operators, like in the usual Heisenberg model. However, as was mentioned, such a form of interaction energy is not sufficient to describe properly the magnetic properties of high-spin systems.
In this paper, proceeding solely from the symmetry considerations, we propose a general recipe for obtaining the many-body Hamiltonian describing a system of interacting spin-S particles. Such systems are characterized by $(2S+1)^{2}-1$ parameters among which there are three components of the magnetization vector and the rest can be treated as the multipole degrees of freedom. These additional parameters are induced in a many-body system by the spin of the structural constituents of matter (particles) and appear at the macroscopic level. The spin degrees of freedom of a particle generate also the SU(2S+1) symmetry of the many-body Hamiltonian in the absence of external fields. As the most intriguing case with a view to physics of ultracold Bose gases, we present a detailed study of the SU(3) symmetric Hamiltonian describing a system of spin-1 particles. Then we apply it to examine the ferromagnetic phase of a weakly interacting Bose gas with condensate in a magnetic field. It is shown that the constructed Hamiltonian modifies the ground state properties and single-particle excitation spectra of the system.
Formulation of the problem and many-body interaction Hamiltonian of internal degrees of freedom
===============================================================================================
Consider a many-body system of spin-S particles whose reduced description is performed in terms of the single-particle density matrix $f_{\alpha\beta}({\bf p})= {\rm Tr}\,\varrho a^{\dagger}_{{\bf p}\beta}a_{{\bf p}\alpha}$, where $\varrho$ can be either an equilibrium or non-equilibrium statistical operator, $a^{\dagger}_{{\bf p}\alpha}$, $a_{{\bf p}\alpha}$ are the creation and annihilation operators, with index $\alpha$ running $2S+1$ values. Depending on the spin value, these operators meet the following bosonic commutation (integer spin) or fermionic anticommutation (half-integer spin) relations: $$\begin{gathered}
[a_{{\bf p}\alpha},a^{\dagger}_{{\bf p}'\alpha'}]_{B}=\delta_{{\bf p}{\bf p}'}\delta_{\alpha\alpha'}, \quad [a_{{\bf p}\alpha},a_{{\bf p}'\alpha'}]_{B}=0, \label{eq:2.1} \\
\{a_{{\bf p}\alpha},a^{\dagger}_{{\bf p}'\alpha'}\}_{F}=\delta_{{\bf p}{\bf p}'}\delta_{\alpha\alpha'}, \quad \{a_{{\bf p }\alpha},a_{{\bf p}'\alpha'}\}_{F}=0. \label{eq:2.2}\end{gathered}$$ Since below we study a homogeneous interacting Bose gas, we use the momentum ${\bf p}$ to specify the individual state of a particle. For the lattice models, one should consider a lattice site index instead of ${\bf p}$.
In the case of spin-1/2 system, the density matrix $f_{\alpha\beta}({\bf p})$, being a square matrix of the second order, can be written as a linear combination of the Pauli matrices $\sigma^{i}_{\alpha\beta}$ and unit matrix $I$ which form a basis for the vector space of $2\times 2$ matrices. The scalar part of the single-particle density matrix in such a decomposition defines the density of the system, whereas its vectorial part specifies three components of the magnetization vector. The latter quantity is induced by the spin of the microscopic constituents of matter. The many-body Hamiltonian of two-particle interaction includes the so-called spin-spin interaction given by $$\label{eq:2.3}
V={1\over 2\mathcal{V}}\sum_{{\bf p}_{1},\ldots{\bf p}_{4}}I({\bf p}_{1}-{\bf p}_{3})a^{\dagger}_{{\bf p}_{1}\alpha}a^{\dagger}_{{\bf p}_{2}\beta}S^{i}_{\alpha\gamma}S^{i}_{\beta\delta}a_{{\bf p}_{3}\gamma}a_{{\bf p}_{4}\delta}\,\delta_{{\bf p}_{1}+{\bf p}_{2},\,{\bf p}_{3}+{\bf p}_{4}}, \quad S^{i}_{\alpha\beta}={1\over 2 }\sigma^{i}_{\alpha\beta},$$ where $a_{{\bf p}\alpha}$, $a^{\dagger}_{{\bf p}\alpha}$ satisfy the permutation relations given by Eqs. (\[eq:2.2\]) and $I({\bf p}_{1}-{\bf p}_{3})$ denotes the exchange interaction. Here and below, the summation over the repeated indices related to internal symmetry is assumed. The above Hamiltonian commutes with the spin operator of a many-body system, $$S^{i}=\sum_{\bf p}a^{\dagger}_{{\bf p}\alpha}S^{i}_{\alpha\beta}a_{{\bf p}\beta}$$ and, consequently, the later represents the conserved quantity or the integral of motion. It is related to magnetization vector $M^{i}$ by $M^{i}=2\mu_{0}S^{i}$, where $\mu_{0}=e\hbar/2mc$ is the Bohr magneton. Note that three components of a particle spin (microscopic characteristic) generate the same number of macroscopic parameters necessary to describe a many-body system of spin-1/2 particles. The Hamiltonian given by Eq. (\[eq:2.3\]) can be applied to describe a gas of spin-1/2 atoms or interacting electron gas embedded in a solid state system. Note that in the lattice models, the creation and annihilation operators carry lattice site index instead of momentum ${\bf p}$.
Now we address the description of a many-body system of spin-1 particles. In this case, the single-particle density matrix, being a reduced description parameter, can be written as a linear combination of the unit $3\times 3$ matrix $I_{\alpha\beta}$ and the Gell-Mann linearly independent traceless Hermitian matrices $\lambda^{a}_{\alpha\beta}$ (see Appenix): $$\label{eq:2.4}
f_{\alpha\beta}({\bf p})=f^{0}({\bf p})I_{\alpha\beta}+f^{a}({\bf p})\lambda_{\alpha\beta}^{a}, \quad a=1,\dots,8.$$ The scalar $f^{0}({\bf p})$ and vectorial $f^{a}({\bf p})$ coefficients are given by $$\label{eq:2.5}
f^{0}={1\over 3}{\rm Tr}\,f({\bf p}), \quad f^{a}({\bf p})={1\over 2}{\rm Tr}\,f({\bf p})\lambda^a.$$ In contrast to spin-1/2 systems, we see that three components of spin are insufficient to describe the many-body states of the system. Indeed, according to Eqs. (\[eq:2.4\]), (\[eq:2.5\]), the states are specified by eight independent parameters determined by the generators $\lambda^{a}$ of the SU(3) group.
To clarify the physical meaning of eight parameters associated with internal symmetry, consider the realization of spin-1 operators in the vector (Cartesian) basis $|x\rangle$, $|y\rangle$, $|z\rangle$ instead of the usual canonical (irreducible) basis $|S,m\rangle$ with $S=1$ and $m=-1,0,1$. These two are related by (see e.g. [@Papa]): $$|x\rangle=\sqrt{1/2}\left(|1,1\rangle+|1,-1\rangle\right), \quad |y\rangle=-i|1,0\rangle,
\quad |z\rangle=i\sqrt{1/2}\left(|1,1\rangle-|1,-1\rangle\right).$$ In the vector basis we have $$\label{eq:2.6}
\langle i|k\rangle=\delta_{ik}, \quad S^{i}|k\rangle=i\varepsilon_{ikl}|l\rangle,$$ so that $S^{i}$ meet the usual commutation relations for spin operators, $$[S^{i},S^{k}]=i\varepsilon_{ikl}S^{l}.$$ From Eq. (\[eq:2.6\]), one finds the matrix elements for the corresponding spin operators, $$\langle k|S^{i}|l\rangle\equiv (S^{i})_{kl}=-i\varepsilon_{ikl},$$ whence $$\label{eq:2.7}
S^{x}=\left(
\begin{array}{ccc}
0 & 0 & 0 \\ [-2ex]
0 & 0 & -i \\ [-2ex]
0 & i & 0 \\
\end{array}
\right), \quad
S^{y}=\left(
\begin{array}{ccc}
0 & 0 & i \\[-2ex]
0 & 0 & 0 \\[-2ex]
-i & 0 & 0 \\
\end{array}
\right), \quad
S^{z}=\left(
\begin{array}{ccc}
0 & -i & 0 \\ [-2ex]
i & 0 & 0 \\ [-2ex]
0 & 0 & 0 \\
\end{array}
\right).$$ One can easily seen that $S^{x}=\lambda^{7}$, $S^{y}=-\lambda^{5}$, and $S^{z}=\lambda^{2}$ (see Eqs. (\[eq:A1\])), so that subalgebra of these matrices generates an SU(2) subgroup of SU(3) group. The remaining five Gell-Mann matrices, due to their properties given by Eq. (\[eq:A6\]), can be expressed in terms of the quadratic combinations of spin operators: $$\begin{gathered}
\lambda^{1}=-\{S^{x},S^{y}\}, \quad \lambda^{3}=(S^{y})^{2}-(S^{x})^{2}, \quad \lambda^{4}=-\{S^{x},S^{z}\}, \nonumber \\
\lambda^{6}=-\{S^{y},S^{z}\}, \quad \lambda^{8}=\sqrt{3}(S^{z})^{2}-{2\over\sqrt{3}}\,I, \label{eq:2.8}\end{gathered}$$ where $\{\ldots,\ldots\}$ denotes an anticommutator and $I$ is the unit $3\times3$ matrix. Since the traceless quadrupole matrix $Q^{ik}\equiv S^{i}S^{k}+S^{k}S^{i}-(4/3)\,\delta_{ik}$ is determined by the above five independent components, $$Q^{ik}={1\over\sqrt{3}}\left(
\begin{array}{ccc}
-\lambda^{8}-\sqrt{3}\lambda^{3} & -\lambda^{1} & -\lambda^{4} \\ [-1.5ex]
-\lambda^{1} & -\lambda^{8}+\sqrt{3}\lambda^{3} & -\lambda^{6} \\ [-1.5ex]
-\lambda^{4} & -\lambda^{6} & 2\lambda^{8} \\
\end{array}
\right),$$ we call them the quadrupole operators. These operators can be considered as the components of a single vector $q^{b}=(-\lambda^{1},-\lambda^{3},-\lambda^{4},-\lambda^{6},\lambda^{8})$. Therefore, if the microscopic constituents of a many-body system have a unit spin, then its macroscopic state is described by the eight parameters originating from the generators of the SU(3) group: $$\label{eq:2.9}
\Lambda^{a}=\sum_{\bf p}a^{\dagger}_{{\bf p}\alpha}\lambda^{a}_{\alpha\beta}a_{{\bf p}\beta}, \quad a=1,\ldots 8, \quad \alpha,\, \beta=x,\,y,\,z$$ that can be splitted into the spin and quadrupole operators, $$\label{eq:2.10}
S^{i}=\sum_{\bf p}a^{\dagger}_{{\bf p}\alpha}S^{i}_{\alpha\beta}a_{{\bf p}\beta} \quad (i=x,\,y,\,z), \quad Q^{b}=\sum_{\bf p}a^{\dagger}_{{\bf p}\alpha}q^{b}_{\alpha\beta}a_{{\bf p}\beta} \quad (b=1,\,3,\,4,\,6,\,8).$$
In the second quantization method the general two-body operator (or binary operator) $A^{(2)}$ can be expressed in the form (see, e.g., [@MethStatPhys]) $$A^{(2)}={1\over 4}\sum_{\alpha_{1},\ldots \alpha_{4}}a^{\dagger}_{\alpha_{1}}a^{\dagger}_{\alpha_{2}}A_{\alpha_{1}\alpha_{2};\alpha_{3}\alpha_{4}} a_{\alpha_{3}}a_{\alpha_{4}}, \quad A_{\alpha_{1}\alpha_{2};\alpha_{3}\alpha_{4}}=\langle \alpha_{1},\alpha_{2}|A^{(2)}|\alpha_{3},\alpha_{4}\rangle.$$ Therefore, one can write the two-body Hamiltonian describing the interaction of internal degrees of freedom in a many-body system of spin-1 particles, $$\label{eq:2.11}
V_{\lambda}={1\over 2\mathcal{V}}\sum_{{\bf p}_{1},\ldots{\bf p}_{4}}J({\bf p}_{1}-{\bf p}_{3})a^{\dagger}_{{\bf p}_{1}\alpha}a^{\dagger}_{{\bf p}_{2}\beta}\lambda^{a}_{\alpha \gamma}\lambda^{a}_{\beta\delta}a_{{\bf p}_{3}\gamma}a_{{\bf p}_{4}\delta}\,\delta_{{\bf p}_{1}+{\bf p}_{2},\,{\bf p}_{3}+{\bf p}_{4}},$$ where ${\cal V}$ is the volume of the system. The structure of the obtained Hamiltonian can also be justified within a phenomenological quasiparticle theory, where the energy of the system is considered to be a functional of the single-particle density matrix, like in the normal Fermi-liquid theory [@Landau; @Silin]. For a not dense system, one can restrict ourselves by the energy functional quadratic in the single-particle density matrix [@FNT]. Since each density matrix represents the eight-component vector, they should enter the energy functional as a scalar product. On the other hand, it is clear, that such a functional should be obtained from the microscopic Hamiltonian given by Eq. (\[eq:2.11\]) by statistical averaging and using the Bloch–De Dominicis (or Wick’s) theorem [@BogBog].
Taking into account the commutation relations for the creation and annihilation operators as well as the properties of the structure constants $f^{abc}$ (see Eqs. (\[eq:2.1\]), (\[eq:A4\])), one can show that $[V_{\lambda},\Lambda^{a}]=0$ and, consequently, $\Lambda_{a}$ is the integral of motion. Note that the above Hamiltonian and the corresponding macroscopic state are SU(3) symmetric. This symmetry, however, is broken if the coupling between the spin and magnetic field (the Zeeman term) is taken into account. The role of SU(3) symmetry has been discussed when studying the dynamics and relaxation in high-spin magnets [@Kov2; @Kov3; @Ivanov].
Truncated Hamiltonian for a weakly interacting Bose gas with internal degrees of freedom
========================================================================================
As we noted, ultracold quantum gases provide a powerful tool to study various effects and phenomena in quantum many-body systems. Therefore, we apply the obtained Hamiltonian to study the ferromagnetic phase of a weakly interacting Bose gas of spin-1 atoms with Bose-Einstein condensate. To this end, we employ the Bogoliubov model [@Bogoliubov] based on $c$-number treatment of creation and annihilation operators for condensate particles. Our starting point is the following Hamiltonian consisting of the kinetic energy term $H_{0}$ and the terms corresponding to the potential interaction $V_{p}$ as well as the interaction between the internal degrees of freedom $V_{\lambda}$: $$\label{eq:3.1}
H=H_{0}+V_{\rm p}+V_{\lambda},$$ where $$\begin{gathered}
H_{0}=\sum_{\bf p}a^{\dagger}_{{\bf p}\alpha}\left[\varepsilon_{\bf p}\delta_{\alpha\beta}- hS^{z}_{\alpha\beta}\right]a_{{\bf p}\beta}, \quad S^{z}_{\alpha\beta}\equiv\lambda^{2}_{\alpha\beta}, \label{eq:3.2} \\
V_{\rm p}={1\over 2\cal{V}}\sum_{{\bf p}_{1},\ldots{\bf p}_{4}}U({\bf p}_{1}-{\bf p}_{3})a^{\dagger}_{{\bf p}_{1}\alpha}a^{\dagger}_{{\bf p}_{2}\beta}a_{{\bf p}_{3}\alpha}a_{{\bf p}_{4}\beta}\,\delta_{{\bf p}_{1}+{\bf p}_{2},\,{\bf p}_{3}+{\bf p}_{4}}, \label{eq:3.3}\end{gathered}$$ and $V_{\lambda}$ is given by Eq. (\[eq:2.11\]). Here $\varepsilon_{\bf p}=p^{2}/2m$ is the kinetic energy of a particle, $a^{\dagger}_{{\bf p}\alpha}$ and $a_{{\bf p}\alpha}$ are the bosonic creation and annihilation operators satisfying Eqs. (\[eq:2.1\]), $h=g\mu_{B}H$ with $g$, $\mu_{B}$, and $H$ being, respectively, the Landé hyperfine factor [@Ueda], the Bohr magneton, and external magnetic field directed along $z$-axis. Note that usually the interaction Hamiltonian is written in terms of the corresponding scattering lengths describing the low energy collisions of atoms at ultra low temperature [@Ohmi; @Ho; @Ueda; @Kurn]. However, such parametrization of interaction does not take into account the local character of interaction and the divergencies appear when computing the ground state energy or chemical potential so that it is necessary to use the renormalization of the coupling constant [@Pethick; @Stringari; @JPhys2017]. Moreover, as we see below, it may lead to an incomplete structure of the spectrum of single-particle excitations. Therefore, the interaction given by Eqs. (\[eq:2.11\]), (\[eq:3.3\]) is characterized by the corresponding functions $U({\bf p})$ and $J({\bf p})$.
Since the number of Bose condensed atoms is a macroscopic value proportional to the volume of the system ${\cal V}$, the next step, according to the Bogoliubov model [@Bogoliubov], is to replace the creation and annihilation operators of condensed atoms with zero momentum by $c$-numbers, $a_{0}^{\dagger}\to\sqrt{\cal V}\Psi^{*}_{\alpha}$ and $a_{0}\to\sqrt{\cal V}\Psi_{\alpha}$, in all operators of relevant physical quantities, where $\Psi_{\alpha}$ represents the condensate wave function. This procedure has been proved to be exact in the thermodynamic limit [@Ginibre]. The $c$-number terms in the Hamiltonian and those that are quadratic in creation and annihilation operators allow to define the ground state and the corresponding spectrum of single-particle excitations (quasiparticles), while the higher order terms in creation and annihilation operators are relevant when describing the interaction effects between the quasiparticles themselves. Therefore, performing the above replacement with $\Psi_{\alpha}$ being a variational parameter and neglecting the terms of the third and fourth order, one can obtain the Hamiltonian truncated up to quadratic terms in the creation and annihilation operators: $$\label{eq:3.5}
H(\Psi)\simeq H^{(0)}(\Psi)+H^{(2)}(\Psi),$$ where $H^{(0)}(\Psi)$ is the $c$-number part of the truncated Hamiltonian given by $$\label{eq:3.6}
{1\over{\cal V}}H^{(0)}(\Psi)={U(0)\over 2}(\Psi^{*}\Psi)^{2}+{J(0)\over 2}(\Psi^{*}\lambda^{a}\Psi)^{2}-h(\Psi^{*}\lambda^{2}\Psi).$$ The quadratic part reads $$\label{eq:3.6'}
H^{(2)}(\Psi)=H_{0}^{(2)}(\Psi)+V^{(2)}_{\rm p}(\Psi)+V^{(2)}_{\lambda}(\Psi),$$ where $H_{0}^{(2)}(\Psi)$ does not include the interparticle interactions, $$\label{eq:3.7}
H_{0}^{(2)}(\Psi)=\sum_{{\bf p}\neq 0}\varepsilon_{\bf p}(a^{\dagger}_{\bf p}a_{\bf p})-h\sum_{{\bf p}\neq 0}(a^{\dagger}_{\bf p}\lambda^{2}a_{\bf p}).$$ Two other terms describing the interaction have the form $$\label{eq:3.8}
V_{\rm p}^{(2)}(\Psi)=U(0)\sum_{{\bf p}\neq 0}(\Psi^{*}\Psi)(a^{\dagger}_{\bf p}a_{\bf p})+{1\over 2}\sum_{{\bf p}\neq 0}U({\bf p})\left[(\Psi a^{\dagger}_{\bf p})(\Psi^{*}a_{\bf p})+(\Psi a^{\dagger}_{\bf p})(\Psi a^{\dagger}_{-{\bf p}})+{\rm h.c.}\right]$$ and $$\begin{gathered}
V_{\lambda}^{(2)}(\Psi)=J(0)\sum_{{\bf p}\neq 0}(\Psi^{*}\lambda^{a}\Psi)(a^{\dagger}_{{\bf p}}\lambda^{a}a_{\bf p}) \nonumber\\
+{1\over 2}\sum_{{\bf p}\neq 0}J({\bf p})\left[(a^{\dagger}_{\bf p}\lambda^{a}\Psi)(\Psi^{*}\lambda^{a}a_{\bf p})+(a^{\dagger}_{\bf p}\lambda^{a}\Psi)(a^{\dagger}_{-{\bf p}}\lambda^{a}\Psi)+{\rm h.c.}\right], \label{eq:3.9}\end{gathered}$$ where we use the following notations $(\Psi^{*}\Psi)\equiv\Psi_{\alpha}^{*}\Psi_{\alpha}$, $(a^{\dagger}_{\bf p}a_{\bf p})\equiv a^{\dagger}_{{\bf p}\alpha}a_{{\bf p}\alpha}$, $(\Psi^{*}\lambda^{a}\Psi)\equiv \Psi_{\alpha}^{*}\lambda^{a}_{\alpha\beta}\Psi_{\beta}$, and so on, assuming matrix multiplication. Note that the replacement of creation and annihilation operators by $c$-numbers implies the gauge symmetry breaking and leads to non-conservation of the total number of atoms. Therefore, the problem should be considered in the grand canonical ensemble, where the chemical potential $\mu$, being a Lagrange multiplier, reflects the conservation of the total number of atoms $N=\sum_{\bf p}(a^{\dagger}_{\bf p}a_{\bf p})$. The corresponding Gibbs statistical operator for the above truncated Hamiltonian reads $$\label{eq:3.10}
w(\Psi)\simeq\exp\left[\Omega-\beta\left({\cal H}^{(0)}(\Psi)+{\cal H}^{(2)}(\Psi)\right)\right],$$ where $$\begin{gathered}
{\cal H}^{(0)}(\Psi)=H^{(0)}(\Psi)-\mu{\cal V}\left(\Psi^{*}\Psi\right), \label{eq:3.11} \\
{\cal H}^{(2)}(\Psi)=H^{(2)}(\Psi)-\mu\sum_{{\bf p}\neq 0}\left(a^{\dagger}_{\bf p}a_{\bf p}\right). \label{eq:3.12}\end{gathered}$$ The grand thermodynamic potential $\Omega$ as a function of reciprocal temperature $\beta={1/T}$, chemical potential $\mu$, and variational parameter $\Psi_{\alpha}$ is found from the normalization condition ${\rm Tr}\,w=1$, $$\label{eq:3.13}
\Omega=\beta{\cal H}^{(0)}(\Psi)-\ln{\rm Tr}\left[ \exp(-\beta{\cal H}^{(2)}(\Psi))\right],$$ where the trace is taken in the space of occupation numbers of bosons with ${\bf p}\neq 0$. In the standard Bogoliubov approach the relation between the condensate wave function and chemical potential is determined by the $c$-number part of the thermodynamic potential assuming that it represents the leading term, $$\label{eq:3.14}
\omega^{(0)}={U(0)\over 2}(\Psi^{*}\Psi)^{2}+{J(0)\over 2}(\Psi^{*}\lambda^{a}\Psi)^{2}-h(\Psi^{*}\lambda^{2}\Psi)-\mu(\Psi^{*}\Psi),$$ where we introduced the thermodynamic potential density $\omega=\Omega/\beta{\cal V}$. The later, up to a sign, coincides with pressure $P$, $\omega=-P$ and it is employed when studying macroscopic dynamics of superfluid systems, both classical and relativistic [@JPhysA; @TMF]. The variation of Eq. (\[eq:3.14\]) over $\Psi^{*}_{\alpha}$ yields $$\label{eq:3.15}
\mu\Psi_{\alpha}-U(0)(\Psi^{*}\Psi)\Psi_{\alpha}-J(0)(\Psi^{*}\lambda^{a}\Psi)\lambda^{a}_{\alpha\beta} \Psi_{\beta}+h\lambda^{2}_{\alpha\beta}\Psi_{\beta}=0$$ (we do not write the complex conjugate equation). This equation ensures the minimum of the thermodynamic potential and gives a relation between the chemical potential and condensate wave function. The contribution of the quadratic terms in creation and annihilation operators to Eqs. (\[eq:3.14\]), (\[eq:3.15\]) is examined in Refs. [@Tolmachev; @JPhysB] for atoms with zero spin.
Ferromagnetic ground state and excitations
==========================================
Now we use the obtained equations to study the ground state properties and corresponding single-particle excitations of a weakly interacting Bose gas of spin-1 atoms. In order to introduce the condensate density $n_{0}$, consider the normalized state vector $\zeta_{\alpha}$, $$\label{eq:4.1}
\Psi_{\alpha}=\sqrt{n_{0}}\zeta_{\alpha}, \quad \zeta^{*}_{\alpha}\zeta_{\alpha}\equiv(\zeta^{*}\zeta)=1.$$ In the Cartesian basis, the ferromagnetic ordering is specified by the following vector [@Ohmi]: $$\label{eq:4.2}
\zeta={1\over\sqrt{2}}(1,\,i,\,0).$$ As we have shown, the description of a many-body system of spin-1 constituents requires the introduction of additional parameters along with the ordinary magnetization vector. These parameters are determined by Eqs. (\[eq:2.9\]), (\[eq:2.10\]). In the problem under consideration, the above ferromagnetic state vector generates the ordinary magnetization along $z$–direction, $$\label{eq:4.3}
\langle S^{i}\rangle=\Psi^{*}S^{i}\Psi=n_{0}\delta_{iz}, \quad S^{i}= (S^{x}\equiv\lambda^{7},\,S^{y}\equiv-\lambda^{5},\,S^{z}\equiv\lambda^{2})$$ and one more parameter associated with the quadrupole degrees of freedom, $$\label{eq:4.4}
\langle Q^{b}\rangle=\Psi^{*}q^{b}\Psi={n_{0}\over\sqrt{3}}\delta_{b8}.$$ The quadrupole tensor for the ferromagnetic state becomes $$\label{eq:4.5}
\langle Q^{ik}\rangle=\Psi^{*}Q^{ik}\Psi=n_{0}\left(
\begin{array}{ccc}
-1/3 & 0 & 0 \\ [-1.5ex]
0 & -1/3 & 0 \\ [-1.5ex]
0 & 0 & 2/3 \\
\end{array}
\right).$$ Since $\langle Q^{xx}\rangle=\langle Q^{yy}\rangle$, the order parameter $\zeta$ is invariant with respect to rotations about $z$-axis, as it should be in the ferromagnetic state. Next, multiplying Eq. (\[eq:3.15\]) by $\Psi_{\alpha}^{*}$ and performing the summation over $\alpha$, one can obtain the relation between the chemical potential and condensate density for the above state vector $\zeta$: $$\label{eq:4.6}
\mu=n_{0}\left(U(0)+{4\over 3}J(0)\right)-h.$$ In a similar manner, the thermodynamic potential density determined by Eq. (\[eq:3.14\]) is written as $$\omega^{(0)}={n_{0}^{2}\over 2}\left(U(0)+{4\over 3}J(0)\right)-n_{0}(h+\mu),$$ or eliminating the condensate density by using Eq. (\[eq:4.6\]), one finds $$\label{eq:4.7}
\omega^{(0)}=-{1\over 2}{(\mu+h)^{2}\over{U(0)+(4/3)J(0)}}.$$ In order for the equilibrium state to be stable, the thermodynamic potential density $\omega^{(0)}$ must be negative (the pressure is positive) that implies $U(0)+(4/3)J(0)>0$. Since the necessary stability condition for a condensed gas of spinless particles requires the predominance of repulsive forces over attractive forces, we also consider $U(0)>0$. Therefore, from Eq. (\[eq:4.7\]), it becomes evident that the condition $J(0)<0$ must be satisfied for the stability of ferromagnetic state under consideration.
Having defined the ferromagnetic ground state, we now address the issue of single-particle excitations. To obtain the corresponding spectra, let us return to the Hamiltonian given by Eqs. (\[eq:3.6’\])-(\[eq:3.9\]), (\[eq:3.12\]). Eliminating the chemical potential by using Eq. (\[eq:4.6\]) and taking into account the explicit form for the ground state vector and for matrices $\lambda^{a}$ (see Eqs. (\[eq:4.1\]), (\[eq:4.2\]), (\[eq:A1\])), the quadratic Hamiltonian is reduced to $$\label{eq:4.8}
{\cal H}^{(2)}(n_{0})={\cal H}^{(2)}_{1}(n_{0})+{\cal H}^{(2)}_{2}(n_{0}),$$ where $$\label{eq:4.9}
{\cal H}^{(2)}_{1}(n_{0})=\sum_{{\bf p}\neq 0}\left[\varepsilon_{\bf p}+h+2n_{0}J({\bf p})-2n_{0}J(0)\right]a^{\dagger}_{{\bf p}z}a_{{\bf p}z},$$ and $$\label{eq:4.10}
{\cal H}^{(2)}_{2}(n_{0})=\sum_{{\bf p}\neq 0}a^{\dagger}_{{\bf p}\alpha}A_{\alpha\beta}({\bf p})a_{{\bf p}\beta}+{1\over 2}\sum_{{\bf p}\neq 0}a^{\dagger}_{{\bf p}\alpha}B_{\alpha\beta}({\bf p})a^{\dagger}_{-{\bf p}\beta}+{1\over 2}\sum_{{\bf p}\neq 0}a_{{\bf p}\alpha}B^{*}_{\alpha\beta}({\bf p})a^{\dagger}_{-{\bf p}\beta}, \quad \alpha,\,\beta=x,\,y.$$ Here $A=A^{\dagger}$ and $B=B^{T}$ are the Hermitian and symmetric matrices, respectively, $$\label{eq:4.11}
A_{\alpha\beta}({\bf p})=\left(
\begin{array}{ccc}
A({\bf p}) & i{\cal A}({\bf p}) \\ [-1.0ex]
-i{\cal A}({\bf p}) & A({\bf p}) \\
\end{array}
\right), \quad
B_{\alpha\beta}({\bf p})=\left(
\begin{array}{ccc}
B({\bf p}) & iB({\bf p}) \\ [-1.0ex]
iB({\bf p}) & -B({\bf p}) \\
\end{array}
\right)$$ with the following matrix elements: $$\begin{gathered}
A({\bf p})=\varepsilon_{\bf p}+h+{5\over 3} n_{0}J({\bf p})-n_{0}J(0)+{1\over 2}{n_{0}U({\bf p})}, \nonumber \\
{\cal A}({\bf p})={1\over 3}n_{0}J({\bf p})-{1\over 2}n_{0}U({\bf p})-n_{0}J(0)+h, \nonumber \\
B({\bf p})={2\over 3}n_{0}J({\bf p})+{1\over 2}n_{0}U({\bf p}). \label{eq:4.12}\end{gathered}$$ The first part ${\cal H}_{1}^{(2)}(n_{0})$ of the total quadratic Hamiltonian has already diagonal form with the following spectrum of single-particle excitations: $$\label{eq:4.13}
\omega^{({\rm I})}_{{\bf p}z}=\varepsilon_{\bf p}+h+2n_{0}\left[J({\bf p})-J(0)\right],$$ while ${\cal H}_{2}^{(2)}(n_{0})$ should be diagonalized in creation and annihilation operators. Note that since ${\cal H}_{2}^{(2)}(n_{0})$ commutes with ${\cal H}_{1}^{(2)}(n_{0})$, it can be diagonalized independently. To this end, we apply the Bogoliubov canonical transformation method which allows to reduce the general Hermitian quadratic form in bosonic operators to a diagonal structure [@BogBog]. Therefore, let us introduce the unitary operator $U$ mixing up $a_{{\bf p}\lambda}$ and $a^{\dagger}_{{\bf p}-\lambda}$: $$\begin{gathered}
Ua_{{\bf p}\alpha}U^{\dagger}=\sum_{\lambda=x,y}\left[u_{\alpha\lambda}({\bf p})a_{{\bf p}\lambda}+v^{*}_{\alpha\lambda}({\bf p})a^{\dagger}_{-{\bf p}\lambda}\right], \nonumber \\
Ua^{\dagger}_{{\bf p}\alpha}U^{\dagger}=\sum_{\lambda=x,y}\left[u^{*}_{\alpha\lambda}({\bf p })a^{\dagger}_{{\bf p}\lambda}+v_{\alpha\lambda}({\bf p})a_{-{\bf p}\lambda}\right] \label{eq:4.14}\end{gathered}$$ and transforming ${\cal H}_{2}^{(2)}(n_{0})$ to the diagonal form $$\label{eq:4.15}
U{\cal H}_{2}^{(2)}(n_{0})U^{\dagger}=\sum_{{\bf p}\neq 0}\sum_{\lambda=x,y}\omega_{{\bf p}\lambda}a^{\dagger}_{{\bf p }\lambda}a_{{\bf p}\lambda}+{\cal E}_{0},$$ where $\omega_{{\bf p}\lambda}$ are the spectra of single-particle excitations and ${\cal E}_{0}$ redefines the vacuum energy or the ground state thermodynamic potential. The creation and annihilation operators $Ua^{\dagger}_{{\bf p}\alpha}U^{\dagger}$ and $Ua_{{\bf p}\alpha}U^{\dagger}$ given by Eqs. (\[eq:4.14\]) must satisfy the same bosonic commutation relations as the operators $a_{{\bf p}\alpha}$ and $a^{\dagger}_{{\bf p}\alpha}$. This requirement results in the following normalization and orthogonality conditions for functions $u_{\alpha\lambda}({\bf p})$ and $v_{\alpha\lambda}(\bf p)$: $$\begin{gathered}
\sum_{\lambda=x,y}\left[u_{\alpha\lambda}({\bf p})u^{*}_{\beta\lambda}({\bf p})-v^{*}_{\alpha\lambda}({\bf p})v_{\beta\lambda}({\bf p})\right]=\delta_{\alpha\beta}, \nonumber \\
\sum_{\lambda=x,y}\left[u_{\alpha\lambda}({\bf p})v^{*}_{\beta\lambda}({\bf p})-v^{*}_{\alpha\lambda}({\bf p})u_{\beta\lambda}({\bf p})\right]=0. \label{eq:4.16}\end{gathered}$$ Note that $u({\bf p})$ and $v({\bf p})$ are constructed from the same quantities as the matrices $A_{\alpha\beta}({\bf p})$ and $B_{\alpha\beta}({\bf p})$ and, therefore, they can be considered as even functions of momentum. The energies of single-particle excitations (or quasiparticles) $\omega_{{\bf p}\lambda}$ satisfy the following eigenvalue equations [@BogBog]: $$\begin{gathered}
\sum_{\lambda=x,y}\left[A_{\alpha\lambda}({\bf p })u_{\lambda\gamma}({\bf p })+B_{\alpha\lambda}({\bf p })v_{\lambda\gamma}({\bf p})\right]=\omega_{{\bf p}\gamma}u_{\alpha\gamma}({\bf p}), \nonumber \\
\sum_{\lambda=x,y}\left[A^{*}_{\alpha\lambda}({\bf p })v_{\lambda\gamma}({\bf p })+B^{*}_{\alpha\lambda}({\bf p })u_{\lambda\gamma}({\bf p})\right]= -\omega_{{\bf p}\gamma}v_{\alpha\gamma}({\bf p}). \label{eq:4.17}\end{gathered}$$ This system of homogeneous liner equations has non-zero solution when the corresponding determinant turns to zero. Therefore, taking into account Eqs. (\[eq:4.11\]), one finds the equation for $\omega_{{\bf p}x}$: $$(A^{2}-4B^{2}-2A{\cal A}+{\cal A}^{2}-\omega_{x}^{2})((A+{\cal A})^{2}-\omega_{x}^{2})=0,$$ Next, taking into account Eqs. (\[eq:4.12\]), one obtains two different excitation spectra, $$\label{eq:4.18}
\omega_{{\bf p}x}^{(\rm II)}=\varepsilon_{\bf p}+2h+2n_{0}(J({\bf p})-J(0))=\omega_{{\bf p}z}^{\rm (I)}+h$$ and $$\label{eq:4.19}
\omega_{{\bf p}x}^{({\rm III})}=\left[\varepsilon_{\bf p}^{2}+2\varepsilon_{\bf p}\left(n_{0}U({\bf p})+{4\over 3}n_{0}J({\bf p })\right)\right]^{1/2}.$$ Note that $\omega_{{\bf p}y}$ satisfies exactly the same equation as $\omega_{{\bf p}x}$ and, consequently, the corresponding spectra are identical or degenerate [@Ohmi]. Therefore, the ferromagnetic phase of a weakly interacting Bose gas with condensate is characterized by three types of excitations with the dispersion laws given by Eqs. (\[eq:4.13\]), (\[eq:4.18\]), (\[eq:4.19\]) and any of $\omega_{{\bf p}x}$ can be related to operators $a^{\dagger}_{{\bf p}y}$, $a_{{\bf p}y}$ in the Hamiltonian determined by Eq. (\[eq:4.15\]).
The spectrum given by Eq. (\[eq:4.19\]) is independent of magnetic field and represents the gapless Bogoliubov mode modified by the interaction of internal degrees of freedom. At small momenta, it represents the phonon excitations, $$\omega_{{\bf p}x}^{\rm (II)}\approx cp, \quad c=\left[{n_{0}\over m}\left(U(0)+{4\over 3}J(0)\right)\right]^{1/2},$$ where $c$ is a speed of sound. The requirement for the speed of sound to be real leads to the stability condition $U(0)+(4/3)J(0)>0$ obtained above.
Two other spectra $\omega_{{\bf p}z}^{\rm ({\rm I})}$ and $\omega_{{\bf p}x}^{\rm ({\rm II})}$ describe the excitations related to the internal degrees of freedom, or “spin-quadrupole” waves. When the applied magnetic field is zero ($h=0$), both spectra become identical so that the system is described by two types of single-particle excitations. Note that the quadratic form determined by Eq. (\[eq:4.10\]) must be positive definite. This requirement implies $J({\bf p})-J(0)>0$ at any momentum ${\bf p}$, moreover $J(0)<0$ for ferromagnetic ordering. It is worth stressing that in contrast to the previous studies of spin-1 Bose-Einstein condensates [@Ohmi; @Ho; @Ueda; @Kurn], both spectra depends on the interaction parameter that is absolutely clear for the system of interacting atoms. This is due to the fact that we do not parameterize the interaction by the corresponding scattering lengths. Indeed, in doing so, $J({\bf p})=J(0)={4\pi\hbar^{2}a_{s}/m}$, where $a_{s}$ is the corresponding scattering length and the spectra become independent of the interaction parameters. Therefore, the description of the interaction effects in ultracold gases by the scattering length represents a sufficiently rough approximation that does not take into account the local character of interaction. The role of nonlocal interaction was recently discussed for ultracold Bose [@JPhysB; @JPhysB2] and Fermi [@Ferm-nonloc1; @Fermi-nonloc2] gases.
Finally, if the interaction Hamiltonian is SU(2) symmetric so that it contains the spin operators $S^{i}$ only, then the ferromagnetic state of spin-1 condensate is characterized by the following spectra of single-particle excitations [@APS; @PelPelSl]: $$\begin{gathered}
\omega_{{\bf p}}^{(1)}=\varepsilon_{\bf p}-2J(0)n+2h, \\
\omega_{{\bf p}}^{(2)}=\varepsilon_{\bf p}-n_{0}(J({\bf p})-J(0))+h, \\
\omega_{\bf p}^{(3)}=\left[\varepsilon_{\bf p}^{2}+2\varepsilon_{\bf p}n_{0}\left(U({\bf p})+J({\bf p })\right)\right]^{1/2}\end{gathered}$$ which are in agreement with other studies [@Ho; @Ohmi; @Ueda] if the interaction is taken to be of the contact type, $$U({\bf p})=U(0)={{g_{0}+2g_{2}}\over 3}, \quad J({\bf p})=J(0)={{g_{2}-g_{0}}\over 3},$$ with $g_{0}$ and $g_{2}$ being related to the s-wave scattering lengths of the total spin-1 channel [@Ueda].
Therefore, the extension of the Hamiltonian to SU(3) symmetry leads to the appearance of new description parameters (along with the magnetization vector) and changes the ground state and single-particle excitation spectra of ferromagnetic Bose-Einstein condensate.
Concluding remarks
==================
We have proposed a general approach for describing a many-body system of interacting spin-1 objects. It has been shown that the microscopic spin-1 objects inevitably induce both SU(3) symmetry of a many-body system and appearance of additional description parameters along with the ordinary magnetization vector. These parameters are related to the quadrupole degrees of freedom. From the symmetry considerations, we have found the many-body SU(3) symmetric Hamiltonian describing the interaction of internal degrees of freedom. It includes eight generators of the SU(3) group, among which there are three spin and five quadrupole operators. The Hamiltonian has been applied to study the ferromagnetic state of a weakly interacting Bose gas of spin-1 atoms with Bose-Einstein condensate. It has been shown that the ground state is specified by non-zero magnetization vector and additional quadrupole parameter. In addition, due to the quadrupole degrees of freedom entering the Hamiltonian, the single-particle excitation spectra are modified in comparison with those obtained from the usually employed Hamiltonian containing only the spin operators. It is also worth noting that the nonlocal character of interaction essentially affects the structure of the spectra so that the parametrization of interaction by the scattering length is a sufficiently rough approximation. Finally, the studied Hamiltonian can be easily generalized to spin-S systems by considering the generators $T^{a}$ of the SU(N) group with $N=2S+1$. The resulting Hamiltonian can be applied to study the magnetic properties of high-spin dilute Fermi and Bose gases.
The properties of Gell-Mann matrices
====================================
The Gell-Mann matrices, being the generators of the SU(3) group, are defined as follows: $$\begin{gathered}
\lambda^{1}=\left(
\begin{array}{ccc}
0 & 1 & 0 \\ [-2ex]
1 & 0 & 0 \\ [-2ex]
0 & 0 & 0 \\
\end{array}
\right), \quad
\lambda^{2}=\left(
\begin{array}{ccc}
0 & -i & 0 \\[-2ex]
i & 0 & 0 \\[-2ex]
0 & 0 & 0 \\
\end{array}
\right), \quad
\lambda^{3}=\left(
\begin{array}{ccc}
1 & 0 & 0 \\ [-2ex]
0 & -1 & 0 \\ [-2ex]
0 & 0 & 0 \\
\end{array}
\right), \quad
\lambda^{4}=\left(
\begin{array}{ccc}
0 & 0 & 1 \\[-2ex]
0 & 0 & 0 \\[-2ex]
1 & 0 & 0 \\
\end{array}
\right), \nonumber \\
\lambda^{5}=\left(
\begin{array}{ccc}
0 & 0 & -i \\ [-2ex]
0 & 0 & 0 \\ [-2ex]
i & 0 & 0 \\
\end{array}
\right), \quad
\lambda^{6}=\left(
\begin{array}{ccc}
0 & 0 & 0 \\[-2ex]
0 & 0 & 1 \\[-2ex]
0 & 1 & 0 \\
\end{array}
\right), \quad
\lambda^{7}=\left(
\begin{array}{ccc}
0 & 0 & 0 \\ [-2ex]
0 & 0 & -i \\ [-2ex]
0 & i & 0 \\
\end{array}
\right), \quad
\lambda^{8}={1\over\sqrt{3}}\left(
\begin{array}{ccc}
0 & 0 & 1 \\[-2ex]
0 & 0 & 0 \\[-2ex]
1 & 0 & 0 \\
\end{array}
\right).\label{eq:A1}\end{gathered}$$ They have the following property: $$\label{eq:A2}
{\rm Sp}\lambda^{a}\lambda^{b}=2\delta_{ab}$$ and satisfy the commutation relations, $$\label{eq:A3}
[\lambda^{a},\lambda^{b}]=2if^{abc}\lambda^{c}.$$ The structure constants $f^{abc}$ of the SU(3) group, according to Eq. (\[eq:A2\]), are found to be $$f^{abc}=-{i\over 4}{\rm Sp}\,\lambda^{c}[\lambda^{a},\lambda^{b}],$$ whence $$\label{eq:A4}
f^{abc}=-f^{bac}=f^{bca}.$$ One can easily find the numerical values of the structure constants, $$\label{eq:A5}
f^{123}=1, \quad f^{147}=-f_{156}=f^{246}=f^{257}=f^{345}=-f^{367}={1\over 2}, \quad f^{456}=f^{678}={\sqrt{3}\over 2}.$$ All other numerical values of $f^{abc}$ not related to the indicated above by permutation are zero. The anticommutator of the Gell-Mann matrices, as well as the commutator, is linear in $\lambda_{a}$: $$\label{eq:A6}
\{\lambda^{a},\lambda^{b}\}={4\over 3}\delta_{ab}+2d^{abc}\lambda^{c}.$$ The coefficients $d^{abc}$, symmetric over all indices, are given by $$d^{abc}={1\over 4}{\rm Sp}\,\lambda^{c}\{\lambda^{a},\lambda^{b}\}.$$ The following their values are different from zero: $$\begin{gathered}
d^{118}=d^{228}=d^{338}=-d^{888}={1\over\sqrt{3}}, \nonumber \\
d^{146}=d^{157}=d^{256}=d^{344}=d^{355}=-d^{247}=-d^{366}=-d^{377}={1\over 2}, \nonumber \\
d^{448}=d^{558}=d^{668}=d^{778}=-{1\over2\sqrt{3}}. \label{eq:A8}\end{gathered}$$
[99]{}
Akhiezer, Bar’yakhtar, and Peletminskii, [*Spin waves*]{} (North-Holland Publishing Company, Amsterdam, 1968).
A.G. Gurevich and G.A. Melkov, [*Magnetization oscillations and waves*]{} (CRC press, Boca Raton, 1996).
E.L. Nagaev, Usp. Fiz. Nauk [**136**]{}, 61 (1982) \[Sov. Phys. Usp. [**25**]{}, 31 (1982)\].
V. M. Matveev, Zh. Eksp. Teor. Fiz. [**65**]{}, 1626 (1973) \[Sov. Phys.-JETP. [**38**]{}, 813 (1974)\].
A.F. Andreev and I.A. Grishchuck, Sov. Phys. JETP [**60**]{}, 267 (1984).
N. Papanicolaou, Nucl. Phys. B [**305**]{}, 367 (1988).
T.A. Tóth, A.M. Läuchli, F. Mila, and K. Penc, Phys. Rev. B [**85**]{}, 140403(R) (2012).
J. Bernatska and P. Holod, J. Phys. A [**42**]{}, 075401 (2009).
I. Niesen and P. Corboz, Phys. Rev. B [**97**]{}, 245146 (2018).
Yu.A. Fridman, O.A. Kosmachev, A.K. Kolezhuk, and B.A. Ivanov, Phys. Rev. Lett. [**106**]{}, 097202 (2011).
O.A. Kosmachev, Yu.A. Fridman, E.G. Galkina, and B.A. Ivanov, Zh. Eksp. Teor. Fiz. [**147**]{}, 320 (2015) \[JETP [**120**]{}, 281 (2015)\].
A. Kolezhuk, Phys. Rev. B [**78**]{}, 144428 (2008).
M.Yu. Kovalevsky and A.V. Glushchenko, Ann. Phys. [**349**]{}, 55 (2014).
A. Sotnikov and W. Hofstetter, Phys. Rev. A [**89**]{}, 063601 (2014).
A. Sotnikov, Phys. Rev. A [**92**]{}, 023633 (2015).
T. Ohmi and K. Machida, J. Phys. Soc. Jpn. [**67**]{}, 1822 (1998).
A.I. Akhiezer, S.V. Peletminskii, and Yu.V. Slyusarenko, Zh. Eksp. Teor. Fiz. [**113**]{}, 918 (1998) \[J. Exp. Theor. Phys. [**86**]{}, 501 (1998)\].
T.-L. Ho, Phys. Rev. Lett. [**81**]{}, 742 (1998).
D.M. Stamper-Kurn, M.R. Andrews, A.P. Chikkatur, S. Inouye, H.-J. Miesner, J. Stenger, and W. Ketterle, Phys. Rev. Lett. [**80**]{}, 2027 (1998).
Y. Kawaguchi and M. Ueda, Phys. Rep. [**520**]{}, 253 (2012).
D.M. Stamper-Kurn and M. Ueda, Rev. Mod. Phys. [**85**]{}, 1191 (2013).
A.I. Akhiezer and S.V. Peletminskii, [*Methods of statistical physics*]{} (Pergamon Press, Oxford, 1981).
L.D. Landau, Zh. Eksp. Teor. Fiz. [**30**]{}, 1058 (1956) \[Sov. Phys. JETP [**3**]{}, 920 (1957)\].
V.P. Silin, Zh. Eksp. Teor. Fiz. [**33**]{}, 495 (1957) \[Sov. Phys. JETP [**6**]{}, 387 (1958)\].
A.S. Peletminskii, S.V. Peletminskii, Yu.V. Slyusarenko, Fiz. Nizk. Temp. [**25**]{}, 417 (1999) \[Low. Temp. Phys. [**25**]{}, 303 (1999)\]; Fiz. Nizk. Temp. [**25**]{}, 211 (1999) \[Low. Temp. Phys. [**25**]{}, 153 (1999)\].
N.N. Bogolubov, N.N. Bogolubov Jr., [*Introduction to quantum statistical mechanics*]{} (World Scientific Publishing Company, 2009).
M.Yu. Kovalevsky and A.V. Glushchenko, J. Magn. Magn. Mat. [**355**]{}, 192 (2014).
M.Yu. Kovalevsky and Glushchenko, Fiz. Nizk. Temp. [**40**]{}, 560 (2014) \[Low Temp. Phys. [**41**]{}, 845 (2014)\].
V.G. Bar’yakhtar, V.I. Butrim, A.K. Kolezhuk, and B.A. Ivanov, Phys. Rev. B [**87**]{}, 224407 (2013).
N.N. Bogoliubov, J. Phys. USSR [**11**]{}, 23 (1947).
C.J. Pethick and H. Smith, [*Bose-Einstein Condensation in Dilute Gases*]{} (Cambridge University Press, Cambridge, 2002).
L. Pitaevskii and S. Stringari, [*Bose-Einstein Condensation*]{} (Oxford University Press, Oxford, 2003).
A.S. Peletminskii, S.V. Peletminskii, and Yu.V. Slyusarenko, J. Phys. B [**50**]{}, 145301 (2017).
J. Ginibre, Commun. Math. Phys. [**8**]{}, 26 (1968).
A.S. Peletminskii, J. Phys. A [**42**]{}, 045501 (2009); Phys. Lett A [**373**]{}, 3369 (2009).
A.S. Peletminskii and S.V. Peletminskii, Teor. Mat. Fiz. [**160**]{}, 333 (2009) \[Theor. Math. Phys. [**160**]{} 1146 (2009)\].
V.V. Tolmachev, [*Theory of a Bose Gas*]{} (Moscow: Moscow University Press, 1969) \[in Russian\].
M.S. Bulakhov, A.S. Peletminskii, S.V. Peletminskii, Yu.V. Slyusarenko, and A.G. Sotnikov, J. Phys. B [**51**]{}, 205302 (2018).
F. Haas and B. Eliasson, J. Phys. B [**51**]{}, 175302 (2018).
S. Simonucci, G. Garberoglio, and S. Taioli, Phys. Rev. A [**84**]{}, 043639 (2011).
S.F. Caballero-Benítez, R. Paredes, and V. Romero-Rochín, Phys. Lett. A [**377**]{}, 1756 (2013).
A.S. Peletminskii, S.V. Peletminskii, and Yu.V. Slyusarenko, Physica A [**380**]{}, 202 (2007).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- |
T. Bergmann$^{1}$, R. Engel$^{1}$, D. Heck$^{1}$, N. N. Kalmykov$^{2}$,\
S. Ostapchenko[^1] $^{1,2}$, T. Pierog$^{1}$, T. Thouw$^{1}$, K. Werner$^{3}$\
*$^{1}$Forschungszentrum Karlsruhe, Institut für Kernphysik, 76021 Karlsruhe, Germany*\
*$^{2}$D.V. Skobeltsyn Institute of Nuclear Physics, Moscow State University, 119992 Moscow, Russia*\
*$^{3}$SUBATECH, Université de Nantes – IN2P3/CNRS – Ecole des Mines, Nantes, France*\
title: ' **One-dimensional Hybrid Approach to Extensive Air Shower Simulation**\'
---
**Abstract**
An efficient scheme for one-dimensional extensive air shower simulation and its implementation in the program [conex]{} are presented. Explicit Monte Carlo simulation of the high-energy part of hadronic and electromagnetic cascades in the atmosphere is combined with a numeric solution of cascade equations for smaller energy sub-showers to obtain accurate shower predictions. The developed scheme allows us to calculate not only observables related to the number of particles (shower size) but also ionization energy deposit profiles which are needed for the interpretation of data of experiments employing the fluorescence light technique. We discuss in detail the basic algorithms developed and illustrate the power of the method. It is shown that Monte Carlo, numerical, and hybrid air shower calculations give consistent results which agree very well with those obtained within the <span style="font-variant:small-caps;">corsika</span> program.
Introduction\[intro.sec\]
==========================
Monte Carlo (MC) simulation of extensive air showers (EAS) is the most common method to calculate detailed theoretical predictions needed for interpreting experimental data of air shower arrays or fluorescence light detectors. However, for primary particles of very high energy, straight-forward MC simulation is not a viable option because of the unreasonably large computing time required. The situation can be improved by applying some weighted sampling algorithms, like the so-called “thinning” method [@Hillas81a], i.e. treating explicitly only a small portion of all shower particles and assigning each particle a corresponding weight factor. Although this approach allows the reduction of EAS calculation times to practically affordable values, it comes soon to its limits. The summation of particle contributions with very large weights creates significant artificial fluctuations for EAS observables of interest [@Kobal99a; @Risse01a; @Kobal:2001jx]. Imposing maximum weight limitations to ensure high simulation quality [@Kobal:2001jx], on the other hand, prevents one from using less detailed sampling and correspondingly from further speeding-up the calculation process. A possible alternative procedure is to describe EAS development numerically, based on the solution of the corresponding cascade equations [@Dedenko65a; @Hillas65a; @Bossard:2000jh]. Combining this with an explicit MC simulation of the most high-energy part of an air shower allows one to obtain accurate results both for average EAS characteristics and for their fluctuations [@KaM13].
In this article we describe a new EAS simulation program of such a hybrid type, called [<span style="font-variant:small-caps;">conex</span>]{}. In [<span style="font-variant:small-caps;">conex</span>]{} the MC treatment of above-threshold particle cascading is realized in the standard way and does not differ significantly from the implementation in e.g.<span style="font-variant:small-caps;">corsika</span> [@Heck98a]. On the other hand, the numerical description of lower energy sub-cascades is based on the solution of hadronic cascade equations, using an updated algorithm of Ref. [@Bossard:2000jh], and a newly developed procedure for the solution of electro-magnetic (e/m) cascade equations. The corresponding algorithms are characterized by high efficiency and good accuracy even if a comparatively crude binning with respect to particle energy and depth position is used. Furthermore, by accounting for neutrino production in addition to the typically considered particles, [<span style="font-variant:small-caps;">conex</span>]{} can also be used for the calculation of ionization energy deposit profiles.
The numeric solution of cascade equations can be replaced, in principle, by a pre-tabulation of the characteristics of secondary sub-cascades, obtained via an iterative MC procedure [@Stanev:1994pf; @Gaisser97a; @Alvarez-Muniz:2002ne]. An example of a combined approach, extended to a three-dimensional EAS simulation, is described in Ref. [@Drescher:2002cr], where hadronic sub-cascades are treated using the method of Ref. [@Bossard:2000jh] and e/m sub-cascades are tabulated using the <span style="font-variant:small-caps;">egs4</span> MC program [@egs4].
The approach presented in the current work does not require any pre-tabulation of particle cascades and is characterized by high efficiency and large flexibility. It can be applied to various initial conditions, i.e.a wide range of energies and angles of incidence of a primary particle, including the case of upward-going showers, as well as arbitrary parameterizations of the atmosphere of the Earth. These features make [<span style="font-variant:small-caps;">conex</span>]{} ideally suited for applications related to the event-by-event based analysis of EAS data, in particular, of fluorescence light based measurements.
This is the first paper in a series in which we will investigate various features of EAS and their relation to the characteristics of hadronic multiparticle production. In this work, we shall present the hybrid simulation scheme in detail, leaving the study of shower predictions to a forthcoming article.
The outline of the paper is as follows. Section \[scheme\] describes the calculation scheme and its basic procedures. In Section \[eas-char\] we show some examples of calculated EAS characteristics and investigate the accuracy and efficiency of the method comparing the hybrid approach with pure MC or numerical procedures. The reliability of the predictions is checked by a detailed comparison with <span style="font-variant:small-caps;">corsika</span> results. Finally, a summary is given in Section \[sec:Summary\] and both potential applications of the program and the prospects for its further development are discussed.
Calculation scheme\[scheme\]
============================
Physics overview
----------------
The calculation scheme consists of two main stages: an explicit MC simulation of the cascade for particles with energies above some chosen threshold $E_{\rm thr}$ (being a free parameter of the scheme) and a solution of nuclear-electro-magnetic cascade equations for sub-cascades of smaller energies. Both MC and numerical parts are characterized by the same physics content, as described below.
In the hadronic cascade one follows the propagation, interaction and decay (where applicable) of (anti-) nucleons, charged pions, charged and neutral kaons; all other types of hadrons produced in interactions and decays are assumed to decay immediately. Particle interactions in the MC part are treated within a chosen high energy hadronic interaction model (implemented are <span style="font-variant:small-caps;">nexus 3.97</span> [@Drescher:2000ha; @Pierog:2002gj; @Werner:2003], <span style="font-variant:small-caps;">qgsjet 01</span> [@Kalmykov:1994ys; @Kalmykov:1997te] and <span style="font-variant:small-caps;">ii</span> [@Ostapchenko:2004ss; @Ostapchenko:2006vr; @Ostapchenko:2005yj], and <span style="font-variant:small-caps;">sibyll 2.1</span> [@Engel:1992vf; @Fletcher:1994bd; @Engel99a]), decays are simulated using the corresponding routines of the <span style="font-variant:small-caps;">nexus</span> model. The same models are used to pre-calculate secondary particle spectra for later use in the numerical treatment of hadronic cascade equations. Optionally, below some energy $E_{\rm low}^{\rm had}\sim100$ GeV, <span style="font-variant:small-caps;">gheisha</span> [@Fesefeldt85a] is employed as low-energy hadronic interaction model.
The MC treatment of the e/m cascade is realized by means of the <span style="font-variant:small-caps;">egs4</span> code [@egs4], supplemented by an account of the Landau-Pomeranchuk-Migdal (LPM) effect [@Migdal:1956tc; @Konishi:1990ya; @Heck:1998gr] for ultra-high energy electrons (positrons) and photons. The simulation of photonuclear interactions and muon pair production was added to the <span style="font-variant:small-caps;">egs</span> package closely following the <span style="font-variant:small-caps;">corsika</span> implementation [@Heck98a]. The system of coupled e/m cascade equations is based on the same interaction processes as implemented in the MC, using Bethe-Heitler cross sections for the bremsstrahlung and pair production with energy-dependent correction factors according to Storm and Israel [@Storm70a], the Klein-Nishina formula for the Compton process, and accounting for Moeller and Bhabha processes as well as for positron-electron annihilation (see, for example [@egs4; @Gaisser90a]). Both LPM suppression and photo-effect are neglected in the e/m cascade Eqs., as the latter are employed in the energy ranges where these processes are not important. Ionization losses of electrons and positrons are described by the Bethe-Bloch formula with corrections to account for the density effect [@Sternheimer:1983mb].
High energy interactions of muons (bremsstrahlung, pair production and muon-nuclear interaction [@Bottai:2000en; @Andreev:1997pf; @Bezrukov:1981ci; @Bilokon:1990nc; @Lohmann:1985qg]) are taken into account in the MC part but are neglected in the cascade Eqs.
In general, an individual shower is simulated as follows. One starts with the primary particle of given energy, direction and initial position in the atmosphere (by default, at 100 km above sea level, if no special geometry is required, e.g. up-going showers). The initial particle direction thus defines the position of the “true” shower axis, in the following referred to as the “shower trajectory”. For a hadron as primary particle, one simulates the hadronic cascade explicitly, recording all secondary particles at a number of pre-chosen depth levels and energy intervals, until all produced secondaries have an energy lower than the threshold $E_{\rm thr}$. The levels are defined with respect to the projected depth $X$, i.e. the slant depth for the particle position projected to the initial shower axis (shower trajectory), as described in more detail in the Appendix. All sub-threshold hadrons/muons and e/m particles are filled into energy-depth tables that form the “source terms” for the cascade equations. In parallel, the above-threshold e/m particles are transferred to <span style="font-variant:small-caps;">egs4</span> for simulating the e/m particle cascade in a similar way, with all sub-threshold e/m particles being added to the e/m source terms.
In the next step the hadronic cascade at energies below $E_{\rm thr}$ is calculated numerically for the first depth level using the corresponding cascade equations and initial conditions specified by the source terms. As the result, one obtains discretized energy spectra of hadrons of different types at the next depth level. All sub-threshold e/m particles produced at this stage are added to the e/m source term. Then sub-threshold e/m cascades are calculated by solving the corresponding e/m cascade equations for the given initial conditions. Hadrons due to photonuclear interaction and pair-produced muons that are generated in the numerical solution of the e/m cascade Eqs. are added to the hadronic source term of the next slant depth level. This procedure is repeated for the following depth levels, each time using the hadronic and e/m source terms of the previous level.
Ultra-high energy e/m particles can undergo geomagnetic pair production and bremsstrahlung well above the atmosphere of the Earth [@Erber:1966vv; @Stanev:1996ux; @Vankov:2002cb]. Therefore, in case of the primary particle being a photon or an electron, the simulation process starts with the calculation of possible interactions with the geomagnetic field using the [<span style="font-variant:small-caps;">preshower</span>]{} code [@Homola:2003ru] and the above described procedure is applied to the secondary particles.
Hadronic cascade equations\[sub:Hadronic-cascade-equations\]
------------------------------------------------------------
The backbone of a hadron-initiated extensive air shower is the hadronic cascade which develops via particle propagation, decay, and interaction with air nuclei of both the initial particle and of produced secondary hadrons. The corresponding integro-differential equations are given by [@Bossard:2000jh] (see also [@Drescher:2002cr]) $$\begin{aligned}
\frac{\partial\left.h_{a}(E,X)\right|_{T}}{{\partial
X}}&=&-\frac{\left.h_{a}(E,X)\right|_{T}}{\lambda_{a}(E)}-\left.h_{a}(E,X)
\right|_{T}\frac{\left|\frac{dL}{dX}\right|_{T}}{\tau_{a}(E)\,
c}
+\frac{\partial}{\partial E}\left(\beta_{a}^{{\rm
ion}}(E) \left.h_{a}(E,X)\right|_{T}\right)\nonumber \\
& & +\sum_{d}\int_{E}^{E_{\max}}\!
dE'\;\left.h_{d}(E',X)\right|_{T}\;\left[\frac{W_{d\rightarrow
a}(E',E)}{\lambda_{d}(E')}+D_{d\rightarrow
a}(E',E)\frac{\left|\frac{dL}{dX}\right|_{T}}{\tau_{d}(E')\,
c}\right]\nonumber \\
& & +\left.S_{a}^{{\rm had}}(E,X)\right|_{T},
\label{sys1-had}\end{aligned}$$ where $\left.h_{a}(E,X)\right|_{T}$ are the differential energy spectra of hadrons of type $a$ with energy $E$ at depth position $X$ along a given straight line trajectory $T$ (in the following the $T$-symbol will be omitted), $\beta_{a}^{{\rm ion}}(E)=-dE_{a}/dX$ is the ionization energy loss of particle $a$ per depth unit. A muon is treated like a hadron, but without interaction term.
The first term in the r.h.s. of Eq. (\[sys1-had\]) represents the decrease of hadron number due to interactions with air nuclei $$\frac{dh_{a}}{dX}=-\frac{h_{a}}{\lambda_{a}},$$ with the corresponding mean free path $\lambda_{a}=m_{\mathrm{air}}/\sigma_{\mathrm{inel}}^{a-{\rm air}}$, where $m_{{\rm air}}$ is the average mass of air molecules and $\sigma_{\mathrm{inel}}^{a-{\rm air}}$ is the hadron $a$ - air nucleus inelastic cross section.
The second term describes particle decay, with the decay rate on a path $dL$ being $$dh_{a}=-h_{a}\:\frac{dL}{\tau_{a}\, c}\label{dhdt},$$ where $\tau_{a}$ is the life time of hadron $a$ in the lab. system, related to the proper life time $\tau_{a}^{(0)}$ by $\tau_{a}=\tau_{a}^{(0)}\, E/m_{a}$, with $m_{a}$ being the hadron mass and $c$ the velocity of light. From the definition of slant depth (\[depth\]) follows $$\left|\frac{dL}{dX}\right|=\frac{1}{\rho_{{\rm air}}(X)}\label{dldx}.$$
The third term in Eq. (\[sys1-had\]) takes into account particle ionization energy losses and the integral term in Eq. (\[sys1-had\]) represents the production of particles of type $a$ in interactions and decays of higher energy parents of type $d$, with $W_{d\rightarrow a}(E',E)$, $D_{d\rightarrow a}(E',E)$ being the corresponding inclusive spectra of secondaries.
Finally, the so-called source term $S_{a}^{{\rm had}}(E,X)$ defines the initial conditions and is determined during the MC simulation of above-threshold particle cascading. It consists of contributions of all sub-threshold hadrons produced at that stage $$S_{a}^{{\rm had}}(E,X)=S_{a}^{{\rm MC\rightarrow
had}}(E,X)=\sum_{i=1}^{N_{{\rm source}}^{{\rm
had}}}\delta_{d_{i}}^{a}\,\delta(E-E_{i})\,\delta(X-X_{i}),\label{source1}$$ with $d_{i}$, $E_{i}$, $X_{i}$ being type, energy, and depth position of the source particles.
The numerical method of solving the hadronic cascade equations is similar to the approach of [@Bossard:2000jh] and is summarized in Appendix \[sub:Numerical-had\].
Electro-magnetic cascade equations
----------------------------------
The e/m cascade development can be described by the following system of integro-differential equations (see, for example, [@Gaisser90a]) $$\begin{aligned}
\frac{\partial l_{e^{-}}\!(E,X)}{{\partial
X}}&=&-\sigma_{e^{-}}\!(E)\: l_{e^{-}}\!(E,X)
+\frac{\partial}{\partial E}\left( \beta_{e^{-}}^{{\rm ion}}(E) l_{e^{-}}(E,X)\right)
\label{sys1-el}
\nonumber \\
& & +\int_{E}^{E_{\max}}\!
dE'\,\left[l_{e^{-}}\!(E',X)\, W_{e^{-}\rightarrow
e^{-}}\!(E',E)+l_{e^{+}}\!(E',X)\right.
\nonumber \\
& & \times\left.\,
W_{e^{+}\rightarrow e^{-}}\!(E',E)+l_{\gamma}\!(E',X)\,
W_{\gamma\rightarrow e^{-}}\!(E',E)\right]+S_{e^{-}}^{{\rm
e/m}}\!(E,X)
\\
\nonumber \\
\frac{\partial
l_{e^{+}}\!(E,X)}{{\partial X}}&=&-\sigma_{e^{+}}\!(E)\:
l_{e^{+}}\!(E,X)
+\frac{\partial}{\partial E}\left(\beta_{e^{+}}^{{\rm ion}}(E)
l_{e^{+}}(E,X)\right)
\nonumber \\
& & +\int_{E}^{E_{\max}}\!
dE'\,\left[l_{e^{+}}\!(E',X)\, W_{e^{+}\rightarrow
e^{+}}\!(E',E)\right.
\nonumber \\
& & +\left.l_{\gamma}\!(E',X)\,
W_{\gamma\rightarrow e^{+}}\!(E',E)\right]+S_{e^{+}}^{{\rm
e/m}}\!(E,X)
\\
\frac{\partial
l_{\gamma}\!(E,X)}{{\partial X}}&=&-\sigma_{\gamma}\!(E)\:
l_{\gamma}\!(E,X)+\int_{E}^{E_{\max}}\! dE'\label{sys1-gam}
\nonumber \\
& & \times\left[l_{\gamma}\!(E',X)W_{\gamma\rightarrow\gamma}
\!(E',E)+l_{e^{-}}\!(E',X)\,
W_{e^{-}\rightarrow\gamma}\!(E',E)\right.
\nonumber \\ & &
+\left.l_{e^{+}}\!(E',X)\,
W_{e^{+}\rightarrow\gamma}\!(E',E)\right]+S_{\gamma}^{{\rm
e/m}}\!(E,X),\end{aligned}$$ where $l_{a}\!(E,X)$ ($a=e^{-},\, e^{+},\,\gamma$) are energy spectra of electrons, positrons, and photons at depth[^2] $X$, $\sigma_{a}\!(E)$ are interaction cross sections (in units area/mass, see Sec. \[sub:Hadronic-cascade-equations\]) $$\begin{aligned}
\sigma_{e^{-}}&=&\sigma_{\rm (bremsstrahlung)}
+\sigma_{\rm (Moeller)}
\nonumber \\
\sigma_{e^{+}}&=&\sigma_{\rm (bremsstrahlung)}
+\sigma_{\rm (Bhabha)} +\sigma_{\rm (annihilation)}
\nonumber \\
\sigma_{\gamma}&=&\sigma_{\rm (pair~production)} +\sigma_{\rm (Compton)}
+\sigma_{\rm (photonuclear)} +\sigma_{\rm (muon~pair)},\end{aligned}$$ and $W_{d\rightarrow a}(E',E)$ are corresponding differential energy spectra of secondary particles $$\begin{aligned}
& & W_{e^{-}\rightarrow e^{-}}\!(E',E)=
W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E',E)+W_{e^{-}\rightarrow e^{-}}^{{\rm Moeller}}\!(E',E)
\nonumber \\
& & W_{e^{+}\rightarrow e^{+}}\!(E',E)=
W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E',E)+W_{e^{+}\rightarrow e^{+}}^{{\rm Bhabha}}\!(E',E)
\nonumber \\
& & W_{\gamma\rightarrow\gamma}\!(E',E)=
W_{\gamma\rightarrow\gamma}^{{\rm Compton}}\!(E',E)
\nonumber \\
& & W_{e^{-}\rightarrow\gamma}\!(E',E)=
W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E',E'-E)
\nonumber \\
& & W_{e^{+}\rightarrow\gamma}\!(E',E)=
W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E',E'-E)+W_{e^{+}\rightarrow\gamma}^{{\rm annih}}\!(E',E)
\nonumber \\
& & W_{e^{+}\rightarrow e^{-}}\!(E',E)=
W_{e^{+}\rightarrow e^{+}}^{{\rm Bhabha}}\!(E',E'-E)
\nonumber \\
& & W_{\gamma\rightarrow e^{-}}\!(E',E)=
W_{\gamma\rightarrow e^{-}}^{{\rm pair}}\!(E',E)+W_{\gamma\rightarrow\gamma}^{{\rm Compton}}\!(E',E'-E)
\nonumber \\
& & W_{\gamma\rightarrow e^{+}}\!(E',E)=
W_{\gamma\rightarrow e^{-}}^{{\rm pair}}\!(E',E) .
\label{W-def}\end{aligned}$$ Here $\sigma_{\rm (Moeller)}$ and $\sigma_{\rm (Bhabha)}$ correspond to the process of $\delta$-electron knock out above some energy threshold $E_{\min}^{{\rm e/m}}$ $$\sigma_{e^{-/+}}^{{\rm Moeller/Bhabha}}\!(E)=
\int_{E_{\min}^{{\rm e/m}}}^{E}\! dE'\, W_{e^{-/+}\rightarrow e^{-/+}}^{{\rm Moeller/Bhabha}}\!(E,E'),
\label{moeller}$$ whereas the contribution of those processes below $E_{\min}^{{\rm e/m}}$ is treated as continuous energy losses and constitutes a part of $\beta_{{\rm ion}}^{e^{\pm}}\!(E)$ $$\beta_{e^{\pm}}^{{\rm ion}}\!(E)=
\left.-\frac{dE_{e^{\pm}}}{dX}\right|_{{\rm ionization}+\delta(<E_{\min}^{{\rm e/m}})}.
\label{beta-ion}$$
The bremsstrahlung cross section diverges due to the characteristic infra-red singular behavior $1/E'$ of the secondary photon spectrum $W_{e^{-}\rightarrow\gamma}^{{\rm brems}}\!(E,E')
=W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E,E-E')$ and normaly requires to introduce some low energy cutoff. The sub-cutoff photon emission could be treated as continuous “radiation” energy losses as is done in [<span style="font-variant:small-caps;">egs4</span>]{} [@egs4]. However, this is not necessary in our calculation scheme as is shown in Appendix \[sub:Numerical-em\], where the numerical solution is presented.
Photonuclear effect and muon pair production
--------------------------------------------
To take into account hadron and muon production by photons, two additional terms are introduced that couple e/m and hadronic cascade equations via source terms. Photoproduction of hadrons, i.e. photonuclear interaction, is implemented using the cross section of Ref. [@Stanev:1986cy]. Particle production distributions are approximated by those of $\pi^0$-air interactions. Then the corresponding source term can be written as $$S_{a}^{{\rm em\rightarrow had}}(E,X)=\int_{E}^{E_{\max}}\! dE'\;
l_{\gamma}(E',X)\; W_{\pi^{0}\rightarrow
a}(E',E)\,\sigma_{\gamma}^{\mathrm{photonuc}}(E'),$$ where $W_{\pi^{0}\rightarrow a}(E',E)$ is defined in analogy to Eq. (\[W-def\]). With a small cross section, a photon can also produce a $\mu^{+}\mu^{-}$ pair. This gives another contribution to the hadronic source term $$S_{\mu}^{{\rm em\rightarrow\mu}}(E,X)=\int_{E}^{E_{\max}}\! dE'\;
l_{\gamma}(E',X)\;
W_{\gamma\rightarrow\mu}(E',E)\,\sigma_{\gamma}^{\mathrm{mu-pair}}(E').$$ Finally the source term defined in Eq. (\[source1\]) becomes $$S_{a}^{{\rm had}}(E,X)=S_{a}^{{\rm MC\rightarrow had}}(E,X)
+S_{a}^{{\rm em\rightarrow had}}(E,X)
+S_{\mu}^{{\rm em\rightarrow\mu}}(E,X)\;\delta _a^{\mu}\,.$$
Applications\[eas-char\]
========================
In the following we demonstrate the reliability of the [<span style="font-variant:small-caps;">conex</span>]{} code by comparing predictions for shower observables calculated with cascade Eqs. and the full hybrid scheme to that of MC simulations. As [<span style="font-variant:small-caps;">conex</span>]{} can also run in pure MC mode, both [<span style="font-variant:small-caps;">conex</span>]{} and [<span style="font-variant:small-caps;">corsika</span>]{} are used for calculating the MC predictions.
If not specified otherwise, the calculations were performed using the <span style="font-variant:small-caps;">qgsjet</span> model for hadronic interactions at energies $E>E_{{\rm low}}^{{\rm had}}=80$ GeV and <span style="font-variant:small-caps;">gheisha</span> model for $E<E_{{\rm low}}^{{\rm had}}$. Our default choice for the cutoff between the MC and the numerical parts is $E_{{\rm thr}}=10^{-2}E_{0}$, $E_{0}$ being the energy of the primary particle. The default energy grid for solving the cascade Eqs. is 30 bins per energy decade ($d_{E}=30$) for hadrons and e/m particles and the slant depth binning has a 5 g/cm$^{2}$ elementary step ($\Delta X$). When applying the hybrid scheme, high energy particles are treated in MC. As a consequence the energy transfer from hadronic to e/m particles is more precise, allowing us to use larger bins, i.e. 20 bins per energy decade and a 10 g/cm$^{2}$ slant depth step size.
Hadronic shower component
-------------------------
In Fig. \[cap:Hadron-stability-16\] we investigate the stability of our scheme and compare both longitudinal profiles of nucleons and charged pions and their energy spectra at 500 g/cm$^{2}$ for different choices of energy and depth discretization. Results only change significantly for very large discretization intervals.
![Average hadronic shower size profiles (left panel) and energy spectra at $X= 500$ g/cm$^{2}$ (right panel) of nucleons and charged pions for proton-initiated vertical ($\theta=0^{{\rm o}}$) showers of $10^{16}$ eV. Compared are the results of solving numerically the system of cascade equations (CE) with different discretization bin sizes in energy and depth. \[cap:Hadron-stability-16\]](fig1-1.eps){width="1.0\columnwidth"}
In Fig. \[cap:Longitudinal-profiles-of-hadron-18\] we plot similar characteristics of charged pions and muons for $10^{18}$ eV proton-initiated showers simulated with [<span style="font-variant:small-caps;">qgsjet 01</span>]{} at high energy and [<span style="font-variant:small-caps;">gheisha</span>]{} at low energy. The results are compared to [<span style="font-variant:small-caps;">corsika</span>]{} predictions. The agreement between the results from the different [<span style="font-variant:small-caps;">conex</span>]{} calculation methods as well as <span style="font-variant:small-caps;">corsika</span> simulations is very good.
![Average longitudinal shower size profiles (left panel) and energy spectra (right panel) of charged pions and muons with energies above 1 GeV. The calculations were done for proton-initiated vertical showers of $10^{18}$ eV. Compared are the predictions obtained with [<span style="font-variant:small-caps;">conex</span>]{} applying the hybrid (dashed line), pure MC (points) and numerical calculation (full line) schemes. In addition CORSIKA predictions are shown as symbols (stars). \[cap:Longitudinal-profiles-of-hadron-18\]](fig2.eps){width="1.0\columnwidth"}
Electromagnetic shower component
--------------------------------
The longitudinal profiles of electrons, positrons, and photons for a $10^{14}$ eV vertical photon-initiated shower are shown in Fig. \[cap:Longitudinal-profiles-of-gamma-14-nm\] (left panel). The shower size profiles are given for the cutoff energies $E_{\min}^{{\rm e/m}}=$1 MeV, 1000 MeV using again the hybrid, pure MC, and cascade equation approaches.
While for a large cutoff energy, for example 1000 MeV, the agreement between the different methods is good we notice systematically larger particle numbers in the hybrid and numerical calculations for $E_{\min}^{{\rm e/m}}=$1 MeV. The corresponding difference is clearly visible in the particle energy spectra and is related to spatial effects in the shower development. Low energy electrons (positrons) undergo significant angular deflections due mainly to multiple Coulomb scattering. In turn low energy bremsstrahlung photons produced by such deflected particles also have significant directional deviations from the initial shower axis. If only the track projected to the shower axis is considered, this leads to an apparently faster absorption of low energy particles (higher interaction rate and ionization energy loss) compared to that expected for particles traveling along the shower axis only (see also discussion in [@Alvarez-Muniz:2003my]). Although a full account of this effect requires a three-dimensional treatment of the particle cascade at MeV energies a reasonable improvement can be achieved by introducing an “average angular deflection”. As the effect is only important for low energy leptons which anyway lose their energy quite fast we may estimate the corresponding average scattering angle of an electron (positron) as $$\langle\theta^{2}\rangle\sim\frac{E_{s}^{2}}{E^{2}}L\!(E),
\label{theta-scat0}$$ where $E_{s}\simeq21$ MeV and $L\!(E)$ is the average travel distance of an electron of energy $E$ in units of radiation length. With $L\!(E)\simeq E/E_{{\rm crit}}$ and $E_{{\rm crit}}\simeq81$MeV, we have $$\langle\theta^{2}\rangle\sim\frac{E_{s}^{2}}{ E\, E_{{\rm crit}}}\,.
\label{theta-scat}$$ For numerical applications, expression (\[theta-scat\]) has to be modified to satisfy the boundary condition $\theta \le \pi/2$. In the following we chose the ansatz $$\langle\theta^{2}\rangle=
\left(\frac{\pi}{2}\right)^{2}\left[1-\exp\!\left(-\frac{E_{e^{\pm}}^{{\rm eff}}}{E}\right)\right],
\label{theta-scat-final}$$ which reduces to the functional form of (\[theta-scat\]) for large $E$ and approaches $(\pi/2)^{2}$ in the small $E$ limit. The same kind of formula is used for photons but with a different parameter $E_{\gamma}^{{\rm eff}}$, as photons themselves do not undergo multiple scattering. Good agreement between fully three-dimensional MC simulations and cascade equation calculations is obtained for $E_{e^{\pm}}^{{\rm eff}}=9.5\cdot10^{-4}$ GeV and $E_{\gamma}^{{\rm eff}}=5\cdot10^{-4}$ GeV. This is shown in Fig. \[cap:Longitudinal-profiles-of-gamma-14\] where the shower size profiles and energy spectra of the two approaches are compared for e/m showers of $10^{14}$ eV and $10^{16}$ eV.
![Left panel: Average longitudinal shower size profiles of charged particles for $10^{14}$ eV (top) and $10^{16}$ eV (bottom) photon-initiated showers and $E_{\min}^{{\rm e/m}}=1$ MeV. Right panel: Particle energy spectra at $X = 700$ g/cm$^2$. Shown are the results of MC simulations with [<span style="font-variant:small-caps;">conex</span>]{} and [<span style="font-variant:small-caps;">corsika</span>]{} and of the hybrid and cascade equation schemes. Here an “average angular deflection” is used in the numerical and hybrid schemes. \[cap:Longitudinal-profiles-of-gamma-14\]](fig5-1.eps "fig:"){width="1.0\columnwidth"}\
![Left panel: Average longitudinal shower size profiles of charged particles for $10^{14}$ eV (top) and $10^{16}$ eV (bottom) photon-initiated showers and $E_{\min}^{{\rm e/m}}=1$ MeV. Right panel: Particle energy spectra at $X = 700$ g/cm$^2$. Shown are the results of MC simulations with [<span style="font-variant:small-caps;">conex</span>]{} and [<span style="font-variant:small-caps;">corsika</span>]{} and of the hybrid and cascade equation schemes. Here an “average angular deflection” is used in the numerical and hybrid schemes. \[cap:Longitudinal-profiles-of-gamma-14\]](fig5-2.eps "fig:"){width="1.0\columnwidth"}
![ Left panel: Average longitudinal profiles of charged particles and photons of energies above 1 MeV for proton-initiated vertical showers of $E_{0}=10^{18}$ eV. Right panel: Particle energy spectra of photons, electrons, and positrons for the atm. depths $X= 700$ and 1000 g/cm$^{2}$. Shown are the results from the hybrid calculation (dashed line), pure MC simulation (points), and numerical cascade Eqs. solution (full line). In addition <span style="font-variant:small-caps;">corsika</span> predictions are given by stars. \[cap:p-charged-18\] ](fig6-1.eps){width="1.0\columnwidth"}
![ Left panel: Average longitudinal profiles of charged particles and photons of energies above 1 MeV for proton-initiated vertical showers of $E_{0}=10^{18}$ eV. Right panel: Particle energy spectra of photons, electrons, and positrons for the atm. depths $X= 700$ and 1000 g/cm$^{2}$. Shown are the results from the hybrid calculation (dashed line), pure MC simulation (points), and numerical cascade Eqs. solution (full line). In addition <span style="font-variant:small-caps;">corsika</span> predictions are given by stars. \[cap:p-charged-18\] ](fig6-2.eps){width="0.94\columnwidth"}
After having shown that both hadronic and e/m showers can be well described by the presented cascade equations and the corresponding hybrid simulation scheme, we test the coupling between the e/m and hadronic shower components. In Fig. \[cap:p-charged-18\] we compare both longitudinal profiles and energy spectra of e/m particles for vertical proton-induced showers of $10^{18}$ eV. Very good agreement is found between the results of the different calculation methods and also with the <span style="font-variant:small-caps;">corsika</span> predictions. The simulations were performed using <span style="font-variant:small-caps;">qgsjet</span> 01 as high-energy interaction model and <span style="font-variant:small-caps;">gheisha</span> for hadronic interactions at low energy.
Energy deposit
--------------
The longitudinal profile of simulated EAS depends on the e/m low-energy cut-off $E_{\rm min}^{\rm e/m}$ used in the calculation (in this work 1 MeV). Lowering this threshold to, for example, 50 keV would increase the number of charged particles at the shower maximum by a few percent [@Nerling:2006yt]. In the case of the photon number, the dependence of the simulation results on the low-energy cut-off is much stronger. In fact, the number of photons diverges if the simulation cut-off is set to 0. Problems of this kind can be avoided if, instead of the secondary particle profile, the ionization energy deposit profile of showers is considered. Another advantage of the energy deposit profile is its direct relation to the light curve measured in air fluorescence experiments. Current measurements support the theoretical expectation that the fluorescence yield is proportional to the ionization energy deposit [@Arciprete:2006pb].
Therefore, the hybrid simulation scheme implemented in <span style="font-variant:small-caps;">conex</span> has been extended to also allow the calculation of energy deposit profiles as described below.
The calculation of the energy deposit profile in the MC part of <span style="font-variant:small-caps;">conex</span> is very similar to that in <span style="font-variant:small-caps;">corsika</span>, see [@Risse:2003fw]. The deposited ionization energy is counted for each particle and traversed slant depth bin. If a particle reaches the cut-off energy of the calculation, depending on the particle type, either all its energy or a part of it is assumed to be deposited locally. For example, a positron is expected to annihilate in the end and, therefore, all its energy plus electron mass are assumed to be deposited. The different energy contributions taken to be deposited locally are given in Table \[cap:Deposited-energy\].
particle $\gamma$ $e^{-}$ $e^{+}$ $\mu^{\pm}$ $\nu$ protons neutrons $\pi^{\pm}$,$\mathrm{K^{\pm}}$
---------------- --------------- --------------- --------------------- ----------------- ------- --------------- ---------- --------------------------------
energy deposit $E_{\rm tot}$ $E_{\rm kin}$ $E_{\rm tot}+m_{e}$ $E_{\rm tot}/3$ 0 $E_{\rm kin}$ 0 $E_{\rm tot}/4$
: Energy deposited locally for particles with $E_{\rm kin}<E_{\rm min}$ ($\mathrm{m_{e}}$ denotes the electron mass and $m_{N}$ is the nucleon mass). \[cap:Deposited-energy\]
particle $\mathrm{K^{0}}$ baryons anti-baryons nuclei anti-nuclei
---------------- ------------------ --------------------- --------------------- ---------------------------- ----------------------------
energy deposit $E_{\rm tot}/2$ $E_{\rm tot}-m_{N}$ $E_{\rm tot}+m_{N}$ $E_{\rm tot}-A\cdot m_{N}$ $E_{\rm tot}+A\cdot m_{N}$
: Energy deposited locally for particles with $E_{\rm kin}<E_{\rm min}$ ($\mathrm{m_{e}}$ denotes the electron mass and $m_{N}$ is the nucleon mass). \[cap:Deposited-energy\]
Solving cascade equations numerically one can calculate the energy deposit of particles with $E>E_{\rm min}$ explicitly. In addition the number of newly created particles at each step in atmospheric depth is given by the source functions. However, the energy carried by particles falling below the cut-off energy threshold and that of neutrinos is not directly calculated.
For the e/m cascade equations, the total energy deposit per slant depth bin can be estimated from energy conservation. The deposited energy follows from the difference between the total kinetic energy at slant depths $X_m$ and $X_{m-1}$, including a correction for positrons similar to the MC treatment $$\begin{aligned}
E_{\rm dep}^{\rm e/m}(X_{m}) & = & \sum_{i=i_{\rm min}}^{i_{\rm max}}
\left(E_{i}^{\mathrm{e/m}}\sum_{a=e^{\pm},\gamma}l_{a}^{i}(X_{m})
+2\, m_{\mathrm{e}}\;l_{e^{+}}^{i}(X_{m})\right)
\nonumber\\
& &- \sum_{i=i_{min}}^{i_{max}}
\left(E_{i}^{\mathrm{e/m}} \sum_{a=e^{\pm},\gamma} l_{a}^{i}(X_{m-1})
+2\, m_{\mathrm{e}}\;l_{e^{+}}^{i}(X_{m-1})\right),\end{aligned}$$ where $m_{e}$ is electron mass.
A similar energy conservation-based method can be applied to the hadronic shower component, however, the deposited energy and the energy going into neutrino production have to be distinguished. Since we know the ionization energy deposited by hadrons and can calculate the number of neutrinos produced within a given slant depth bin, we can obtain the energy of all the particles falling below the low-energy threshold from $$E_{\rm cut}(X_{m})=E_{\rm bal}(X_{m})-E_{\rm ion}(X_{m})-E_{\nu}(X_{m}),$$ where $$\begin{aligned}
E_{\rm bal}(X_{m}) & = & \sum_{i=i_{min}}^{i_{max}}\left(E_{i}^{\mathrm{had}}
\sum _a h_{a}^{i}(X_{m})+\sum_{a}m_{a}\; h_{a}^{i}(X_{m})\right)\nonumber\\
& - & \sum_{i=i_{min}}^{i_{max}}\left(E_{i}^{\mathrm{had}}
\sum _a h_{a}^{i}(X_{m-1})+\sum_{a}m_{a}\; h_{a}^{i}(X_{m-1})\right)\\
E_{\rm ion}(X_{m}) & = & \sum_{i=i_{min}}^{i_{max}}\sum_{a}\, h_{a}^{i}(X_{m-1})\:
\left(1-\exp\left[-\frac{\beta_{a}^{{\rm ion}}(E_{i}^{{\rm had}})}{
E_{i}^{{\rm had}}-E_{i-1}^{{\rm had}}}\Delta X\right]\right)\nonumber\\
& \times & \exp\!\left[-\frac{1-W_{a\rightarrow a}^{ii}}{\lambda_{a}(E_{i}^{{\rm had}})}\Delta X
-\frac{m_{a}\left|L(X_{m})-L(X_{m-1})\right|}{c\tau_{a}^{(0)}E_{i}^{{\rm had}}}\right]\\
E_{\nu}(X_{m}) & = & \int_{X_{m-1}}^{X_{m}}\! d X^\prime \sum_{d}\sum_{j=i+1}^{i_{\max}}\,
h_{d}^{j}(X')\, D_{d\rightarrow\nu}^{ji}\frac{m_{d}/(c\tau_{d}^{(0)})}{\rho_{{\rm air}}(X')}.\end{aligned}$$ To account for the fact that only a part of the energy carried by hadrons falling below the low-energy threshold is deposited as ionization energy, a factor $f_{\mathrm{had}/\mu}=0.45$ is introduced [@Barbosa:2003dc]. Thus, the energy deposit for the hadronic part of the system of cascade Eqs. is given by $$E_{\rm dep}^{\rm had}(X_{m})=E_{\rm ion}(X_{m})+f_{\mathrm{had}/\mu}\; E_{\rm cut}(X_{m}).$$
Summing the different contributions, the longitudinal energy deposit profile can be calculated. A comparison of the results of different calculation methods within [<span style="font-variant:small-caps;">conex</span>]{} with [<span style="font-variant:small-caps;">corsika</span>]{} predictions is shown in Fig. \[Longitudinal-Energy-Deposit\] (top panel). Good agreement is found.
An interesting consistency check of the energy deposit calculation scheme is the investigation of the dependence of corresponding results on the low-energy threshold used in the e/m cascade Eqs. In Fig. \[Longitudinal-Energy-Deposit\] (bottom panel), the mean longitudinal energy deposit profiles of iron-initiated showers of $\theta = 60^\circ$ and $E = 10^{19}$eV are shown for different low-energy thresholds. The profiles are independent of this threshold over a wide energy range – the approximation of local energy deposit breaks down only for a threshold energy of 100 MeV and higher.
![ Upper panel: Average longitudinal energy deposit profile for vertical proton-initiated showers of $10^{18}$eV. The results obtained with the e/m energy cutoff $E_{\rm min}^{{\rm e/m}}=1$ MeV, using hybrid (dashed line), MC (points), or numerical approaches (full line), are compared to [<span style="font-variant:small-caps;">corsika</span>]{} predictions (stars). Lower panel: Energy deposit profiles for $10^{19}$ eV inclined ($\theta=60^{{\rm o}}$) iron-initiated showers for the e/m energy cutoff $E_{\min}^{{\rm e/m}}=$1 (full line), 5 (dashed line), 10 (dotted line) and 100 MeV (dashed-dotted line). \[Longitudinal-Energy-Deposit\]](fig9.eps){width="0.55\columnwidth"}
Shower fluctuations
-------------------
So far we have only studied mean shower observables, i.e. ones averaged over many showers. As a proper description of shower fluctuations is of central importance for the analysis of experimental data, we will also discuss the treatment of shower fluctuations.
Running [<span style="font-variant:small-caps;">conex</span>]{} in hybrid mode allows us to benefit from the fast numerical solution of cascade Eqs. and, at the same time, to obtain a good description of shower-to-shower fluctuations. Here, the key parameter of the method is the energy threshold that separates the explicit MC simulation from the application of cascade Eqs. By default, this energy threshold is set to $E_{\rm thr} = 0.01\cdot E_0$ for all particles in [<span style="font-variant:small-caps;">conex</span>]{}. In principle, it can be chosen differently for e/m and hadronic particles to further reduce the simulation time needed for high-energy showers.
In Fig. \[Xmax18-fluctuations\] the distribution of the depth of shower maximum is shown for proton- and iron-initiated showers. Within the statistical uncertainties, the [<span style="font-variant:small-caps;">conex</span>]{} results agree very well with that obtained with [<span style="font-variant:small-caps;">corsika</span>]{} MC simulations. Not only the fluctuations but also the mean depth of shower maximum obtained with the two codes are the same (not shown).
The dependence of the simulated fluctuations on the energy threshold is shown in Fig. \[Xmax18-threshold\]. Even a threshold as large as $E_{\rm thr} = 0.99\cdot E_0$ is sufficient to reproduce almost the full $X_{\rm max}$ distribution. This comparison demonstrates that the fluctuations of particle production in the first interaction of such high-energy showers determine almost the entire shower profile.
Summary\[sec:Summary\]
======================
We have developed a fast and efficient one-dimensional hybrid simulation scheme for ultra-high energy air showers. It combines explicit MC simulation of high-energy particle interaction, propagation and decay with the numerical solution of a system of cascade equations for calculating the low-energy part of the particle cascade.
The presented hybrid simulation scheme is implemented in the code [<span style="font-variant:small-caps;">conex</span>]{}[^3]. Several high- and low-energy hadronic interaction models are available within [<span style="font-variant:small-caps;">conex</span>]{} to study theoretical predictions and the model-dependence of data analyses.
All relevant interaction and decay processes are considered in both the MC and the cascade equation parts of [<span style="font-variant:small-caps;">conex</span>]{}. These processes also include muon pair production and photonuclear interactions of muons. At ultra-high energy, the LPM effect and possible e/m preshowering in the geo-magnetic field are simulated.
The hybrid simulation scheme has been extended to include the simultaneous calculation of both shower size profiles of various particles and the generation of ionization energy deposit profiles. The latter are independent of the low-energy cut-off that has to be applied in all shower simulations. Knowing both the shower size profile (with an arbitrary low-energy cut-off) and the energy deposit profile allows us to simulate directly the fluorescence and Cherenkov light signal of air showers. Together with the fully three-dimensional implementation of the shower axis geometry, this makes [<span style="font-variant:small-caps;">conex</span>]{} ideally suited for event simulation and data analysis of fluorescence light experiments such as HiRes [@Abbasi:2004nz], Auger [@Abraham:2004dt], TA [@Kasahara:2005pk], and EUSO [@Pallavicini:2003we].
In developing [<span style="font-variant:small-caps;">conex</span>]{}, particular emphasis is put on the accuracy and reliability of the shower simulation to make the code directly applicable to data analysis of air shower experiments. Extensive comparisons with [<span style="font-variant:small-caps;">corsika</span>]{} simulations show that all shower distributions agree very well. Both mean shower profiles and energy distributions as well as their fluctuations were compared, only a small fraction of which could be shown in this paper.
In a forthcoming work we will study the influence of different hadronic interaction models on air shower predictions. In particular we will investigate the total calorimetric energy deposited by a shower in air. First results of this work have been presented in [@Pierog:2005aa].
Appendix
========
Geometry
--------
Using a one-dimensional treatment of EAS development, a given shower trajectory may be characterized by a single parameter - its distance to the center of the Earth $O$: $R_{\perp}^{T}=\left|P_{l}^{T}O\right|$ - see Fig. \[traj\], where $P_{l}^{T}$ is the lowest trajectory point.
In case the observer is positioned at some height $h_{{\rm obs}}$ above sea level and $R_{{\rm Earth}}+h_{{\rm obs}}>R_{\perp}^{T}$ ($R_{{\rm Earth}}$ being the Earth radius) the observed shower inclination is $$\theta=\arcsin\frac{R_{\perp}^{T}}{R_{{\rm Earth}}+h_{{\rm obs}}}.
\label{theta}$$
The position of a particle moving along a given trajectory may be then characterized by its local azimuthal angle $\theta_{P}^{T}$ with respect to the vertical direction, or alternatively, by its height $H_{P}^{T}$, or by the distance $L_{P}^{T}=\left|P_{l}^{T}P\right|$ to the lowest trajectory point $$\begin{aligned}
& & H_{P}^{T}=\frac{R_{\perp}^{T}}{\sin\theta_{P}^{T}}
\label{height}\\
& & L_{P}^{T}=\frac{R_{\perp}^{T}}{\cos\theta_{P}^{T}}.
\label{dist}\end{aligned}$$
For the solution of cascade equations, it is more convenient to use instead the slant grammage $X_{P}^{T}$, i.e. the integral over the atmospheric density $\rho_{{\rm air}}(h)$ along the trajectory $T$ from the point $P$ to infinity $$X_{P}^{T}=\int_{L_{P}^{T}}^{\infty}\! dl'\;\rho_{{\rm air}}\left(H(l')\right),
\label{depth}$$ and characterize an arbitrary particle position by two variables, $R_{\perp}^{T}$ and $X$.
Numerical treatment of hadronic cascade equations\[sub:Numerical-had\]
----------------------------------------------------------------------
In order to reduce the problem to the solution of a system of ordinary differential equations, we perform a discretization of particle energy spectra. Introducing an energy binning as $E_{i}^{{\rm had}}=E_{\min}^{{\rm had}}\; C^{i-1}$, $C=\,10^{1/d_{E}}$ ($E_{\min}^{{\rm had}}=1$ GeV, $d_{E}=10\div30$) and replacing the smooth particle spectra $h_{a}\!(E,X)$ by discrete contributions $h_{a}^{i}\!(X)$ of representative particles of energies $E_{i}^{{\rm had}}$ ($i=1,...,i_{\max}$) we may get instead of (\[sys1-had\]) $$\begin{aligned}
\frac{dh_{a}^{i}(X)}{dX}&=&
-h_{a}^{i}(X)\left[\frac{1}{\lambda_{a}(E_{i}^{{\rm had}})}
+\left|\frac{dL}{dX}\right|\frac{m_{a}}{c\,\tau_{a}^{(0)}\:E_{i}^{{\rm had}}}+
\frac{\beta_{a}^{{\rm ion}}(E_{i}^{{\rm had}})}{E_{i}^{\mathrm{had}}-E_{i-1}^{\mathrm{had}}}\right]
\nonumber\\
& & +\sum_{d}\sum_{j=i}^{i_{\max}}\; h_{d}^{j}(X)\;
\left[\frac{W_{d\rightarrow a}^{ji}}{\lambda_{d}(E_{j}^{{\rm had}})}+
D_{d\rightarrow a}^{ji}\frac{m_{d}/(c\,\tau_{d}^{(0)})}
{E_{j}^{{\rm had}}\;\rho_{{\rm air}}(X)}\right]
\nonumber \\
& & +h_{a}^{i+1}(X)\,\frac{\beta_{a}^{{\rm ion}}(E_{i+1}^{{\rm had}})}
{E_{i+1}^{\mathrm{had}}-E_{i}^{\mathrm{had}}}+S_{ai}^{{\rm had}}(X),
\label{sys2-had}\end{aligned}$$ where we used (\[dldx\]), replaced the integral over parent particle energies $\int\! dE'$ by the discrete sum $\sum_{j}$, introduced the discretized source term $S_{ai}^{{\rm had}}(X)$ and the discrete interaction (decay) spectra $W_{d\rightarrow a}^{ij}$ ($D_{d\rightarrow a}^{ij}$) via $$\begin{aligned}
S_{ai}^{{\rm had}}(X)&=&\int_{\max\left[E_{i-1}^{{\rm had}},E_{\min}^{{\rm had}}
\right]}^{E_{i}^{{\rm had}}}\! dE'\;
S_{a}^{{\rm had}}(E',X)\; K(E'/E_{i}^{{\rm had}})
\nonumber\\
& & +\int_{E_{i}^{{\rm had}}}^{\min\left[E_{i+1}^{{\rm had}},
E_{{\rm thr}}\right]}\! dE'\; S_{a}^{{\rm had}}(E',X)\,\left[1-K(E'/E_{i+1}^{{\rm had}})\right]
\label{source2}\\
W_{d\rightarrow a}^{ij}&=&\int_{\max\left[E_{i-1}^{{\rm had}},E_{\min}^{{\rm had}}
\right]}^{E_{i}^{{\rm had}}}\! dE'\; W_{d\rightarrow a}(E',E_{j}^{{\rm had}})\;
K(E'/E_{i}^{{\rm had}})\nonumber\\
& & +\int_{E_{i}^{{\rm had}}}^{\min\left[E_{i+1}^{{\rm had}},E_{{\rm thr}}\right]}
\! dE'\; W_{d\rightarrow a}(E',E_{j}^{{\rm had}})\,\left[1-K(E'/E_{i+1}^{{\rm had}})\right],
\label{wij2}\end{aligned}$$ and the discrete energy loss term via :$$-\beta_{a}^{{\rm ion}}(E_{i}^{{\rm had}})\:\frac{h_{a}^{i}(X)}
{E_{i}^{\mathrm{had}}-E_{i-1}^{\mathrm{had}}}
+\beta_{a}^{{\rm ion}}(E_{i+1}^{{\rm had}})\:\frac{h_{a}^{i+1}(X)}
{E_{i+1}^{\mathrm{had}}-E_{i}^{\mathrm{had}}}\,.$$ Here the condition of both energy and particle number conservation gives $$K(\varepsilon)=\frac{C\:\varepsilon-1}{C-1}\,.
\label{keps}$$
The solution of the homogeneous part of Eq. (\[sys2-had\]) is $$h_{a}^{i}(X) = h_{a}^{i}(X_{0})\:
\exp\!\left[-\left(\frac{1-W_{a\rightarrow a}^{ii}}{\lambda_{a}(E_{i}^{{\rm had}})}
+\frac{\beta_{a}^{{\rm ion}}(E_{i}^{{\rm had}})}
{E_{i}^{{\rm had}}-E_{i-1}^{{\rm had}}}\right)(X-X_{0})
-\frac{m_{a}\left|L(X)-L(X_{0})\right|}{c\tau_{a}^{(0)}E_{i}^{{\rm had}}}\right].
\label{hhomo}$$
Correspondingly the solution of the full equation can be obtained in an iterative way $$\begin{aligned}
h_{a}^{i}(X) & = & h_{a}^{i}(X_{0})\:\exp\!\left[-\left(\frac{1-W_{a\rightarrow a}^{ii}}{\lambda_{a}(E_{i}^{{\rm had}})}+\frac{\beta_{a}^{{\rm ion}}(E_{i}^{{\rm had}})}{E_{i}^{{\rm had}}-E_{i-1}^{{\rm had}}}\right)(X-X_{0})-\frac{m_{a}\left|L(X)-L(X_{0})\right|}{c\tau_{a}^{(0)}E_{i}^{{\rm had}}}\right]
\nonumber\\
& & + \int_{X_{0}}^{X}\! dX'\left\{ \sum_{d}\sum_{j=i+1}^{i_{\max}}h_{d}^{j}(X')\,\left[\frac{W_{d\rightarrow a}^{ji}}{\lambda_{d}(E_{j}^{{\rm had}})}+D_{d\rightarrow a}^{ji}\frac{m_{d}/(c\tau_{d}^{(0)})}{E_{j}^{{\rm had}}\rho_{{\rm air}}(X')}\right]\right.
\nonumber \\
& & \left.+h_{a}^{i+1}(X')\frac{\beta_{a}^{{\rm ion}}(E_{i+1}^{{\rm had}})}{E_{i+1}^{\mathrm{had}}-E_{i}^{\mathrm{had}}}+S_{ai}^{{\rm had}}(X')\right\}
\nonumber \\
& & \times\exp\!\left[-\left(\frac{1-W_{a\rightarrow a}^{ii}}
{\lambda_{a}(E_{i}^{{\rm had}})}+\frac{\beta_{a}^{{\rm ion}}(E_{i}^{{\rm had}})}
{E_{i}^{{\rm had}}-E_{i-1}^{{\rm had}}}\right)(X-X')
-\frac{m_{a}\left|L(X)-L(X')\right|}{c\tau_{a}^{(0)}E_{i}^{{\rm had}}}\right].
\label{hfull} \end{aligned}$$
Discretizing depth positions as $X_{m}=m\,\Delta\! X$, $m=1,...,m_{\max}$, $X_{m_{\max}}$ corresponding to the observation level for the given shower trajectory, the formulas (\[hhomo\]-\[hfull\]) can be used to calculate spectra of different hadrons at all depths $X_{m}$, starting from the initial condition $h_{a}^{i}(X_{m})$ for $m=1$, calculating first $h_{a}^{i_{\max}}(X_{m+1})$, using Eq. (\[hhomo\]), and $h_{a}^{i_{\max}-1}(X_{m+1})$, $h_{a}^{i_{\max}-2}(X_{m+1})$, ... , $h_{a}^{1}(X_{m+1})$, using Eq. (\[hfull\]), etc. for $m=2,...,m_{\max}$. The depth integral in Eq. (\[hfull\]) is taken using the Simpson formula; particle spectra values $h_{a}^{i}(X_{m+1/2})$ at mids of the bins $X_{m+1/2}=(X_{m}+X_{m+1})/2$ are obtained via a logarithmic interpolation between previously calculated $h_{a}^{i}(X_{m})$, $h_{a}^{i}(X_{m+1})$.
As for neutral pions, we assume them to decay at the place, calculate the number of $\pi^{0}$s produced at depth $X_{m}$ as $$\begin{aligned}
h_{\pi^{0}}^{i}(X_{m})&=&\int_{X_{m-1}}^{X_{m}}\! dX'\sum_{d}\sum_{j=i+1}^{i_{\max}}
\, h_{d}^{j}(X')
\nonumber\\
& & \times\left[\frac{W_{d\rightarrow\pi^{0}}^{ji}}{\lambda_{d}(E_{j}^{{\rm had}})}
+D_{d\rightarrow\pi^{0}}^{ji}\frac{m_{d}/(c\,\tau_{d}^{(0)})}{E_{j}^{{\rm had}}\;
\rho_{{\rm air}}(X')}\right]+\int_{X_{m-1}}^{X_{m}}\! dX'\: S_{\pi^{0}i}^{{\rm had}}(X')
\label{pi0}\end{aligned}$$ and add photons resulting from $\pi^{0}$ decay to the e/m source function.
Numerical treatment of e/m cascade equations\[sub:Numerical-em\]
----------------------------------------------------------------
To reduce the problem to the solution of a system of ordinary differential equations we again, like in case of hadron cascade, discretize particle energy spectra $l_{a}\!(E,X)$ using an energy grid $E_{i}^{{\rm e/m}}=E_{\min}^{{\rm e/m}}\, C^{i-1}$ ($E_{\min}^{{\rm e/m}}=0.1\div1$ MeV), replace the integral over parent particle energy $E'$ in (\[sys1-el\]-\[sys1-gam\]) by a sum over discrete energies, and discretize the source term $S_{a}^{{\rm e/m}}\!(E,X)$ and the differential energy spectra $W_{d\rightarrow a}(E',E)$ according to (\[source2\]-\[wij2\]) (with $E_{i}^{{\rm had}}$, $E_{\min}^{{\rm had}}$, $S_{a}^{{\rm had}}\!(E',X)$ being replaced by $E_{i}^{{\rm e/m}}$, $E_{\min}^{{\rm e/m}}$, $S_{a}^{{\rm e/m}}\!(E',X)$). Then Eqs. (\[sys1-el\]-\[sys1-gam\]) are transformed to $$\begin{aligned}
& & \frac{dl_{a}^{i}(X)}{dX}=\sum_{d}\sum_{j=i}^{i_{\max}}
\bar{W}_{d\rightarrow a}^{ji}\; l_{d}^{j}(X)+S_{ai}^{{\rm e/m}}\!(X),
\label{sys2-em}\end{aligned}$$ where we replaced the continuous energy loss term $\frac{\partial}{\partial E}\left( \beta_{e^{\pm}}^{{\rm ion}}(E)
l_{e^{\pm}}(E,X)\right)$ by $$-\beta_{e^{\pm}}^{{\rm ion}}\!(E_{i}^{{\rm e/m}})\;
\frac{l_{e^{\pm}}^{i}\!(X)}{E_{i}^{{\rm e/m}}-E_{i-1}^{{\rm e/m}}}
+\beta_{e^{\pm}}^{{\rm ion}}\!(E_{i+1}^{{\rm e/m}})\;
\frac{l_{e^{\pm}}^{i+1}\!(X)}{E_{i+1}^{{\rm e/m}}-E_{i}^{{\rm e/m}}}$$ and included the two terms, together with the interaction cross sections, into the discrete particle production spectra $\bar{W}_{d\rightarrow a}^{ji}$, defined as follows $$\begin{aligned}
& & \bar{W}_{e^{\pm}\rightarrow e^{\pm}}^{ii}=W_{e^{\pm}\rightarrow e^{\pm}}^{ii}-\sigma_{e^{\pm}}\!(E_{i}^{{\rm e/m}})-\frac{\beta_{e^{\pm}}^{{\rm ion}}\!(E_{i}^{{\rm e/m}})}{(C-1)E_{i}^{{\rm e/m}}}
\label{wii-ep}\\
& & \bar{W}_{e^{\pm}\rightarrow e^{\pm}}^{i,i-1}=W_{e^{\pm}\rightarrow e^{\pm}}^{i,i-1}+\frac{\beta_{e^{\pm}}^{{\rm ion}}\!(E_{i}^{{\rm e/m}})}{(C-1)E_{i}^{{\rm e/m}}}
\label{wii1-ep}\\
& & \bar{W}_{\gamma\rightarrow\gamma}^{ii}=W_{\gamma\rightarrow\gamma}^{ii}-\sigma_{\gamma}\!(E_{i}^{{\rm e/m}}),\label{wiigg}\end{aligned}$$ and $\bar{W}_{d\rightarrow a}^{ij}=W_{d\rightarrow a}^{ij}$ for all other combinations of $i$, $j$, $a$, $d$.
It is worth verifying that all elements of the matrixes $\bar{W}_{d\rightarrow a}^{ij}$ are free of singularities. Indeed, singular terms appear in $\sigma_{e^{\pm}}^{{\rm brems}}\!(E_{i}^{{\rm e/m}})=
\int_{0}^{E_{i}^{{\rm e/m}}}\! dE'\; W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E_{i}^{{\rm e/m}},E')$ and in $W_{e^{\pm}\rightarrow e^{\pm}}^{ii}$ (defined by Eq. (\[wij2\])) due to the characteristic $1/(E-E')$ dependence of $W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E,E')$. Nevertheless, in $\bar{W}_{e^{\pm}\rightarrow e^{\pm}}^{ii}$ defined by Eq. (\[wii-ep\]) the corresponding singularities are canceled against each other $$\begin{aligned}
\bar{W}_{e^{\pm}\rightarrow e^{\pm}}^{({\rm brems})ii}&=&
W_{e^{\pm}\rightarrow e^{\pm}}^{({\rm brems})ii}-\sigma_{e^{\pm}}^{{\rm brems}}\!(E_{i}^{{\rm e/m}})\\
&=&\int_{E_{i-1}^{{\rm e/m}}}^{E_{i}^{{\rm e/m}}}\! dE'\;
W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E_{i}^{{\rm e/m}},E')\;
\frac{C\: E'/E_{i}^{{\rm e/m}}-1}{C-1}\nonumber\\
& & -\int_{0}^{E_{i}^{{\rm e/m}}}\! dE'\;
W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E_{i}^{{\rm e/m}},E')\\
&=& -\frac{C}{E_{i}^{{\rm e/m}}\:(C-1)}
\int_{E_{i-1}^{{\rm e/m}}}^{E_{i}^{{\rm e/m}}}\! dE'\;
(E_{i}^{{\rm e/m}}-E')\; W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E_{i}^{{\rm e/m}},E')\nonumber\\
& & -\int_{0}^{E_{i-1}^{{\rm e/m}}}\! dE'\; W_{e^{-}\rightarrow e^{-}}^{{\rm brems}}\!(E_{i}^{{\rm e/m}},E'),\end{aligned}$$ where the last two integrals are finite. Similarly one can check the finiteness of $\bar{W}_{e^{\pm}\rightarrow e^{\pm}}^{i,i-1}$.
To solve the system (\[sys2-em\]), we first find the solution of the homogeneous equation system $$\begin{aligned}
& & \frac{dl_{a}^{i}(X)}{dX}=\sum_{d}\bar{W}_{d\rightarrow a}^{ii}\;
l_{d}^{i}(X),
\label{sys-emhomo}\end{aligned}$$ which is given as $$\begin{aligned}
& & l_{a}^{i}(X)=\sum_{I=1}^{3}D_{aI}^{i}(X_{0})\; e^{\Lambda_{I}^{i}\:(X-X_{0})}
\label{lept-homo}\end{aligned}$$ Here $\Lambda_{I}^{i}$ and $D_{aI}^{i}(X_{0})$ are correspondingly the eigenvalues and eigenvectors of the system of linear algebraic equations obtained by inserting (\[lept-homo\]) into (\[sys-emhomo\]):$$\Lambda^{i}\; D_{a}^{i}(X_{0})-\sum_{d}\bar{W}_{d\rightarrow a}^{ii}\;
D_{d}^{i}(X_{0})=0,$$ with the normalization fixed by the initial conditions at $X=X_{0}$: $\sum_{I=1}^{3}D_{aI}^{i}(X_{0})=l_{a}^{i}(X_{0})$.
Then the solution of the full equation system (\[sys2-em\]) may be given in a recursive form $$\begin{aligned}
l_{a}^{i}(X)&=&\sum_{I=1}^{3}D_{aI}^{i}(X_{0})\:
\left[e^{\Lambda_{I}^{i}\:(X-X_{0})}+\sum_{d}F_{Id}^{i}(X_{0})
\int_{X_{0}}^{X}\! dX'\right.
\nonumber\\
& & \left.\times\: e^{\Lambda_{I}^{i}\:(X-X')}\:\left[S_{di}^{{\rm e/m}}\!(X')
+\sum_{j=i+1}^{i_{\max}}\sum_{g}\bar{W}_{g\rightarrow d}^{ji}\;
l_{g}^{j}(X')\right]\right],
\label{lept-full}\end{aligned}$$ where the matrix $F_{Id}^{i}$ is inverse with respect to $D_{aI}^{i}$ $$\sum_{I}D_{aI}^{i}(X_{0})\; F_{Id}^{i}(X_{0})=\delta_{a}^{d}\,.
\label{inver}$$
Equations (\[lept-homo\]-\[lept-full\]) can be used to calculate the discrete spectra of e/m particles $l_{a}^{i}(X_{m})$ at all depths $X_{m}$ in the same way as in case of the hadronic cascade.
#### Acknowledgments {#acknowledgments .unnumbered}
The authors thank Hans-Joachim Drescher, Michael Unger, and Ralf Ulrich for fruitful discussions. N.N.K. acknowledges the financial support of the Russian Foundation for Basic Research (RFBR, grant 05-02-16401). The work of one of the authors, S.O., has been supported in part by the German Ministry for Education and Research (BMBF, Grant 05 CU1VK1/9).
[10]{}
A. M. Hillas, in Proc. 17th Int. Cosmic Ray Conf., [**v. 8**]{}, p. 193, Paris, France, 1981.
M. Kobal, A. Filipčič, and D. Zavrtanik, Proc. of 26th Int. Cosmic Ray Conf., [**v. 1**]{}, p. 490, Salt Lake City, Utah, USA, 1999.
M. Risse, D. Heck, J. Knapp, and S. S. Ostapchenko, Proc. 27th Int. Cosmic Ray Conf., [**v. 2**]{}, p. 522, Hamburg, Germany, 2001.
M. Kobal, Astropart. Phys. [**15**]{}, 259 (2001).
L. G. Dedenko, Proc. of 9th Int. Cosmic Ray Conf., [**v. 1**]{}, p. 662, London, UK, 1965.
M. Hillas, Proc. of 9th Int. Cosmic Ray Conf. [**v. 1**]{}, p. 758, London, UK, 1965.
G. Bossard, H. J. Drescher, N. N. Kalmykov, S. Ostapchenko, A. I. Pavlov, T. Pierog, E. A. Vishnevskaya, and K. Werner, Phys. Rev. [**D63**]{}, 054030 (2001), hep-ph/0009119.
N. N. Kalmykov and M. V. Motova, Yadernaya Fizika (Rus) [**43**]{}, 630 (1986).
D. Heck, J. Knapp, J. Capdevielle, G. Schatz, and T. Thouw, Wissenschaftliche Berichte FZKA 6019, Forschungszentrum Karlsruhe, 1998.
T. Stanev and H. P. Vankov, Astropart. Phys. [**2**]{}, 35 (1994).
T. K. Gaisser, P. Lipari, and T. Stanev, in Proc. 25th Int. Cosmic Ray Conf., [**v. 6**]{}, p. 281, Durban, South Africa, 1997.
J. Alvarez-Muniz, R. Engel, T. K. Gaisser, J. A. Ortiz, and T. Stanev, Phys. Rev. [**D66**]{}, 033011 (2002), astro-ph/0205302.
H. J. Drescher and G. R. Farrar, Phys. Rev. [**D67**]{}, 116001 (2003), astro-ph/0212018.
W. R. Nelson, H. Hirayama, and D. W. O. Rogers, SLAC-265, Stanford Linear Accelerator Center, 1985.
H. J. Drescher, M. Hladik, S. Ostapchenko, T. Pierog, and K. Werner, Phys. Rept. [**350**]{}, 93 (2001), hep-ph/0007198.
T. Pierog, H. J. Drescher, F. M. Liu, S. Ostapchenko, and K. Werner, Nucl. Phys. [**A715**]{}, 895 (2003), hep-ph/0211202.
K. Werner, F. M. Liu, S. Ostapchenko, and T. Pierog, J. Phys. G: Nucl. Part. Phys. [**30**]{}, S211 (2003), hep-ph/0306152.
N. N. Kalmykov, S. S. Ostapchenko, and A. I. Pavlov, Bull. Russ. Acad. Sci. Phys. [**58**]{}, 1966 (1994).
N. N. Kalmykov, S. S. Ostapchenko, and A. I. Pavlov, Nucl. Phys. Proc. Suppl. [**52B**]{}, 17 (1997).
S. Ostapchenko, Nucl. Phys. Proc. Suppl. [**151**]{} (2006), 143, hep-ph/0412332.
S. Ostapchenko, Phys. Rev. D (2006), to be published, hep-ph/0505259.
S. Ostapchenko, to appear in Proc. of INFN Eloisatron Project 44th Workshop on QCD at Cosmic Energies, Erice, Italy, 2004, hep-ph/0501093.
R. Engel, T. K. Gaisser, T. Stanev, and P. Lipari, Phys. Rev. [**D46**]{}, 5013 (1992).
R. S. Fletcher, T. K. Gaisser, P. Lipari, and T. Stanev, Phys. Rev. [**D50**]{}, 5710 (1994).
R. Engel, T. K. Gaisser, P. Lipari, and T. Stanev, Proc. of 26th Int. Cosmic Ray Conf., [**v. 1**]{}, p. 415, Salt Lake City, Utah, USA, 1999.
H. Fesefeldt, preprint PITHA-85/02, RWTH Aachen, 1985.
A. B. Migdal, Phys. Rev. [**103**]{}, 1811 (1956).
E. Konishi, A. Adachi, N. Takahashi, and A. Misaki, J. Phys. G: Nucl. Part. Phys. [**17**]{}, 719 (1991).
D. Heck and J. Knapp, Wissenschaftliche Berichte FZKA 6097, Forschungszentrum Karlsruhe, 1998.
E. Storm and H. I. Israel, Nuclear Data Tables [**A7**]{}, 565 (1970).
T. K. Gaisser, , Cambridge Univ. Press, Cambridge, 1990.
R. M. Sternheimer, M. J. Berger, and S. M. Seltzer, At. Data Nucl. Data Tabl. [**30**]{}, 261 (1984).
S. Bottai and L. Perrone, Nucl. Instrum. Meth. [**A459**]{}, 319 (2001), hep-ex/0001018.
Y. M. Andreev and E. V. Bugaev, Phys. Rev. [**D55**]{}, 1233 (1997).
L. B. Bezrukov and E. V. Bugaev, Yadernaya Fizika (Rus) [**33**]{}, 1195 (1981).
H. Bilokon [*et al.*]{}, Nucl. Instrum. Meth. [**A303**]{}, 381 (1991).
W. Lohmann, R. Kopp, and R. Voss, CERN-85-03, 1985.
T. Erber, Rev. Mod. Phys. [**38**]{}, 626 (1966).
T. Stanev and H. P. Vankov, Phys. Rev. [**D55**]{}, 1365 (1997), astro-ph/9607011.
H. P. Vankov, N. Inoue, and K. Shinozaki, Phys. Rev. [**D67**]{}, 043002 (2003), astro-ph/0211051.
P. Homola [*et al.*]{}, Comput. Phys. Commun. [**173**]{}, 71 (2005), astro-ph/0311442.
T. Stanev, T. K. Gaisser, and F. Halzen, Phys. Rev. [**D32**]{}, 1244 (1985).
J. Alvarez-Muñiz, E. Marques, R. A. Vazquez, and E. Zas, Phys. Rev. [**D67**]{}, 101303 (2003), astro-ph/0302491.
F. Nerling, J. Blümer, R. Engel, and M. Risse, Astropart. Phys. [**24**]{}, 421 (2006).
AirFly Collab., F. Arciprete [*et al.*]{}, Nucl. Phys. Proc. Suppl. [**150**]{}, 186 (2006).
M. Risse and D. Heck, Astropart. Phys. [**20**]{}, 661 (2004), astro-ph/0308158.
H. M. J. Barbosa, F. Catalani, J. A. Chinellato, and C. Dobrigkeit, Astropart. Phys. [**22**]{}, 159 (2004), astro-ph/0310234.
HiRes Collab., R. U. Abbasi [*et al.*]{}, Astrophys. J. [**622**]{}, 910 (2005), astro-ph/0407622.
Pierre Auger Collab., J. Abraham [*et al.*]{}, Nucl. Instrum. Meth. [**A523**]{}, 50 (2004).
TA Collab., K. Kasahara [*et al.*]{}, astro-ph/0511177, 2005.
EUSO Collab., M. Pallavicini, Nucl. Instrum. Meth. [**A502**]{}, 155 (2003).
T. Pierog [*et al.*]{}, Proc. 29-th Int. Cosmic Ray Conf., [**v. 7**]{}, p. 103, Pune, India, 2005.
[^1]: corresponding author, e-mail: [email protected]
[^2]: In the absence of particle decays there is no dependence on a particular shower trajectory, apart from the density effect correction.
[^3]: The [<span style="font-variant:small-caps;">conex</span>]{} code is available upon request from [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a sparse estimation and dictionary learning framework for compressed fiber sensing based on a probabilistic hierarchical sparse model. To handle severe dictionary coherence, selective shrinkage is achieved using a Weibull prior, which can be related to non-convex optimization with $\ell_p$-norm constraints for $0\!<\!p\!<\!1$. In addition, we leverage the specific dictionary structure to promote collective shrinkage based on a local similarity model. This is incorporated in form of a kernel function in the joint prior density of the sparse coefficients, thereby establishing a Markov random field-relation. Approximate inference is accomplished using a hybrid technique that combines Hamilton Monte Carlo and Gibbs sampling. To estimate the dictionary parameter, we pursue two strategies, relying on either a deterministic or a probabilistic model for the dictionary parameter. In the first strategy, the parameter is estimated based on alternating estimation. In the second strategy, it is jointly estimated along with the sparse coefficients. The performance is evaluated in comparison to an existing method in various scenarios using simulations and experimental data.'
author:
- Christian Weiss
- ' Abdelhak M. Zoubir'
bibliography:
- './bibliography.bib'
title: Dictionary Learning Strategies for Compressed Fiber Sensing Using a Probabilistic Sparse Model
---
Introduction
============
Fiber sensors are versatile devices with broad applicability [@Kersey1997; @Culshaw2008; @Nakazaki2009; @Yamashita2009]. They are of high interest in smart structures to sense and react to the environment [@Measures1992; @Udd1996]. For quasi-distributed sensing based on wavelength-division fiber Bragg grating (FBG) sensors are often employed due to their sensitivity to strain or An FBG describes a local variation of the refractive index and reflects light at a certain wavelength, called *Bragg wavelength*. Typically, a number of detuned FBGs is imprinted into the core of an optical fiber. Fiber interrogation is performed using broadband light sources or wavelength-tunable lasers. The latter feature higher local signal-to-noise ratios However, in order to monitor time-varying perturbations, the laser has to sweep quickly through the tuning range. This requires high-speed analog-to-digital converters (ADCs) and produces large amounts of data.\
*Compressed sensing* (CS) [@Baraniuk2007; @Candes2008; @Eldar2012] can help to alleviate these problems by taking samples in form of projections into a low-dimensional subspace. The original signal can be reconstructed by exploiting the sparsity of the signal with respect to an adequate dictionary [@Candes2005; @Candes2008a]. This task strongly resembles the *sparse synthesis* problem with redundant dictionaries Besides greedy methods, such as Orthogonal Matching Pursuit (OMP) [@Pati1993], $\ell_1$-minimization is a popular method to solve the sparse reconstruction problem [@Tibshirani1994; @Donoho2003; @Candes2011]. It relies on the *restricted isometry property* (RIP), which essentially states that unique sparse solutions can be recovered by restricting the $\ell_1$-norm instead of the $\ell_0$-norm [@Donoho2001; @Candes2006]. Redundant dictionaries can yield highly sparse representations, that allow for estimating quantities at high resolution directly in the sparse domain [@Donoho2003; @Malioutov2005]. However, redundancy causes inter-column coherence and it is likely that the required RIP conditions are no longer fulfilled [@Donoho2003; @Rauhut2008; @Candes2011]. The with $0\!<\!p\!<\!1$, offers a trade-off to avoid an NP-hard combinatorical problem imposed by the $\ell_0$-norm, while a unique solution might still be retrieved [@Chartrand2007; @Chartrand2008].\
Dictionaries can be classified as parametric or non-parametric. Non-parametric dictionaries are typically learned from training data and often used if no analytical model is While they can yield sparser representations of certain data realizations [@Mairal2008a], non-parametric dictionaries usually lack an interpretable structure and are inefficient in terms of Parametric dictionaries, in turn, rely on an analytical model for the observed signal. Their analytic form offers an efficient implementation and a means to obtain optimality proofs and error bounds [@Rubinstein2010]. They are also favorable in terms of scalability and storage-. *Translation-invariant* dictionaries represent an important sub-class of parametric dictionaries, that can be used to estimate the translation coefficients of localized Nonetheless, due to the complexity of natural signals, some model parameters might be unknown or contain uncertainty. Parametric *Dictionary* *Learning* (DL) addresses this problem with the aim of estimating these parameters from the measured data. Herein, statistical DL methods, such as maximum or maximum *a posteriori* (MAP) estimation, are commonly employed [@Rubinstein2010]. In order to solve the resulting optimization problem, *alternating estimation* (AE) is a frequently -optimal paradigm, that iteratively optimizes a local objective In a Bayesian setting, the Expectation Maximization (EM) algorithm is a popular variant of AE-based estimation [@Rubinstein2010].\
A model for the sparse coefficients can be of deterministic or probabilistic nature. While the deterministic case is often assumed in sparse estimation [@Donoho2003; @Candes2011], a probabilistic model offers high flexibility to take model deviations and measurement errors into account. Moreover, a hierarchical structure can be used to incorporate additional uncertainty in prior assumptions. Sparsity can either be promoted by continuous distributions, resulting in *weakly sparse* models, or by discrete mixtures, leading to *strongly sparse* A prominent example of discrete mixtures are *Spike & Slab* models [@Ishwaran2005a]. They are based on binary activations and yield strongly sparse representations. Continuous sparse priors, such as a Gaussian or double-exponential (Laplace) prior, feature high excess kurtosis with heavy tails and a narrow peak around Besides sparsity, additional knowledge of the signal, e.g. correlation, can be incorporated [@Eldar2010; @Zhang2011a].\
For many practical models, evaluating the posterior distribution is not feasible and approximate methods, such as *Markov Chain Monte Carlo (MCMC)* or variational Bayes methods, have to be used to accomplish Variational methods use rather simple analytic functions to approximate the posterior distribution by factorization, which is favorable in terms of scalability and computational costs but leads to a deterministic approximation [@Seeger2008; @Bishop2006]. MCMC methods attempt to sample the posterior distribution, where subsequent samples form a *Markov chain* [@Bishop2006]. The method is a powerful technique, that is especially suitable for sampling high-dimensional spaces in the presence of correlation [@Neal2011]. However, MCMC performance is generally limited by the available computation time, thereby relying on a stochastic approximation. Another application of MCMC is found in non-convex optimization, where Stochastic MCMC has gained popularity for large-scale Bayesian learning [@Chen2014; @Chen2015; @Chen2016].\
In the present work, we consider the problem of *Compressed Fiber Sensing* (CFS) with highly coherent translation-invariant dictionaries and imperfectly known parameters. For the sparse coefficients, a weakly sparse hierarchical model is considered. We also establish a relation between this model and non-convex optimization with $\ell_p$-norm constraints for $0\!\!<\!\!p\!\!<\!\!1$. In order to alleviate the problem of dictionary coherence, we leverage additional structure of the dictionary and achieve augmented sparsity by establishing a *Markov random* relation among the sparse coefficients. For dictionary learning, we pursue two different strategies: In the first we consider a deterministic dictionary parameter, that is estimated using a Monte Carlo EM algorithm. In the second a probabilistic hierarchical model for the dictionary parameter is considered, leading to a full Bayesian formulation and joint estimation of the sparse coefficients and the dictionary parameter. In both strategies, approximate inference is accomplished using a hybrid MCMC method based on Gibbs sampling and HMC. Finally, we use simulations and real data to compare the proposed methods to previous work where a deterministic model is considered for the sparse coefficients and the dictionary parameter. For the deterministic case, we derive the Cram[é]{}r-Rao bound (CRB) to assess the performance gain achieved by a probabilistic model.
Contributions
-------------
- We propose a probabilistic model for the sparse coefficients, where a Weibull prior is used to promote (weak) sparsity. Additional collective shrinkage is achieved by establishing an MRF-relation among the sparse coefficients based on a bivariate kernel function in the joint prior density. This helps to moderate the impact of severe dictionary coherence and can be used in general sparse synthesis problems with similar dictionary structure. We also establish a relation to non-convex optimization with constraints on the $\ell_p$-norm for $0<p<1$.
- For dictionary learning, we investigate two conceptually different strategies, assuming either a deterministic (***S1***) or a stochastic (***S2***) dictionary parameter. In both strategies, the noise level can be jointly estimated along with the sparse coefficients. We further highlight advantages, disadvantages and limitations to offer support in choosing an adequate method for practical systems.
- To accomplish inference in these models, we use a hybrid MCMC method, combining HMC and Gibbs sampling, We show its applicability and efficacy in the considered sampling problem for CFS.
- We use simulations to evaluate the performance of the proposed sparse estimation and DL methods for various scenarios of different CS sample sizes, SNRs and CS matrices. These results are compared to an existing method in [@Weiss2016], where the sparse coefficients and the dictionary parameter are assumed to be deterministic. In addition, we provide a real-data example to verify the practical applicability of ***S1*** and ***S2***.
- We derive the Cram[é]{}r-Rao bound for jointly estimating the sparse coefficients and the dictionary parameter in the deterministic case. It is a valid bound for the competing method in [@Weiss2016], and serves to assess the achieved performance gain of our probabilistic approach.
Related Work
============
There exists little work addressing the combination of CS and DL for the application of WDM-based distributed fiber-optic In [@Weiss2013], a model for the received sensor signal is presented, from which a redundant shift-invariant parametric dictionary is created. The works focus on the aspect of CS and sparse estimation in the case of uncertain dictionary parameters. The authors use AE-based estimation to determine the dictionary parameters, where a pre-processing routine accounts for severe dictionary coherence. Unlike our approach, these works use a deterministic model for the sparse coefficients and dictionary parameters.\
Weakly sparse models have been widely used in the literature. A comprehensive analysis of different hierachical sparse prior models is provided in [@Mohammad-Djafari2012]. The general problem of choosing the prior in weakly sparse models for sparse regression is addressed in [@Polson2010], where the authors describe various properties of different shrinkage priors and illuminate the selection problem from two perspectives: prior distributions and penalty functions. The work in [@Mohamed2012] also investigates Bayesian methods with different sparse models in comparison to classical $\ell_1$-minimization. Seeger [@Seeger2008] found that the Laplace prior is able to shrink most components close to zero, while allowing for selected components to become sufficiently large. This effect, termed *selective shrinkage* is most noticeable for heavy-tailed priors, e.g. the Student’s $t$-prior [@Seeger2008] or the *horseshoe* prior in [@Carvalho2010; @Polson2010]. Based on these findings, we select a sparsity prior that resembles a positive version of the horseshoe prior. Other works, that focus on the perspective of penalized regression, report higher sparsity levels by penalizing the $\ell_p$-norm with $0<p<1$ instead of the $\ell_1$-norm [@Gupta2013]. The authors in [@Chartrand2007] show that the RIP requirements for the dictionary can be relaxed in this case. It is also pointed out in [@Chartrand2007; @Chartrand2008] that non-convex CS with $\ell_p$-norm penalization requires less measurements than standard CS, which is based on the $\ell_1$-norm. We rely on these results and show a relation between the considered sparsity prior and non-convex optimization with $\ell_p$-norm constraints.\
There exist several approaches to exploit additional structure of the signal. One example is *block sparsity* [@Eldar2010]. A block sparse Bayesian learning framework is proposed in [@Zhang2011a], pointing out how correlation can be exploited in regularization algorithms. Wakin *et* introduce the concept of *joint sparsity* for signal recovery in distributed CS theory. In [@Malioutov2005], temporal correlation across subsequent CS measurements is considered, while the authors in [@Chen2013a] use correlation to achieve smoothness. Another related concept is proposed in [@Altmann2015], where a truncated multivariate Ising MRF model is used to describe the correlation between adjacent pixels for image processing. Different from these works, we use the ideas of MRFs [@Murphy2012] and exploit correlation to achieve collective shrinkage among the sparse coefficients.\
A comparative analysis in [@Mohamed2012] suggests that MCMC methods are powerful for inference in sparse models. In [@Neal2011], the benefits of HMC and Gibbs sampling in hierarchical models are outlined. It is also shown, that HMC can be more effective than a Gibbs sampler for sampling high-dimensional spaces in the presence of correlation. According to these results, we consider a hybrid MCMC method that combines HMC and Gibbs sampling for inference in our hierarchical model, where the sparse coefficients are high-dimensional and correlated. For parametric DL, the Monte Carlo EM algorithm in ***S1*** represents one variant of the frequently applied AE-based estimation Comparable to ***S2*** is the Bayesian framework for sparse estimation and DL in [@Hansen2014]. However, the authors use a Gaussian prior without correlation.
Outline
-------
In , the signal model for CFS is introduced, and in , the CRB for joint estimation of the deterministic sparse coefficients and dictionary parameters is derived. details the sparsity and local similarity model, while Section \[sec:approx\_inference\] describes the hybrid MCMC method for approximate inference in this model. The parametric DL strategies ***S1*** and ***S2*** are described in Section \[sec:PDL\]. Section \[sec:performance\] shows the working principle along with a performance analysis of the proposed and an existing method based on simulations and experimental data. A discussion of the results and findings is given in Section \[sec:discussion\]. Section \[sec:conclusion\] concludes this work.
Signal Model {#sec:problem_and_model}
============
In order to determine the quantity and nature of impairments at the FBGs in a WDM-based fiber sensor, the time delays of the reflections from the individual FBGs need to be estimated. We adopt the model in [@Weiss2013; @Weiss2015; @Weiss2016], where CS-based acquisition is employed to reduce the number of samples to be stored and processed. The CS measurements are described by $$\label{eq:basic_model}
\mathbf{y} = \boldsymbol{\Phi}\mathbf{A}(\theta)\mathbf{x} + \mathbf{n}\,,$$ where $\boldsymbol{\Phi} \in \mathbb{R}^{M\times L}$ is the CS sampling matrix and $\mathbf{n}\in \mathbb{R}^{M}$ is a Gaussian noise component with independent and identically distributed (i.i.d.) entries, $n_m \sim \mathcal{N}(0,\sigma_n^2)$, $m=1,\dots,M$. The vector $\mathbf{x} \in \mathbb{R}^N$ is sparse with $K$ significant components, and $\theta\!\in\!\mathbb{R}$ is a scalar dictionary parameter. The matrix $\mathbf{A}(\theta)$ represents a redundant shift-invariant dictionary and its columns, called *atoms*, represent FBG reflections on a dense grid of delays. The indices of the $K$ significant components in $\x$ indicate the desired reflection delays. They are collected in the set $\mathcal{S}\! =\! \{i_1,\dots,i_K\}$. We can write the full data likelihood function for this model by $$\label{eq:Gauss_likelihood}
\hspace{0.05cm} {p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\theta \hspace{0.8pt} )} = (\sqrt{2\pi}\sigma_n)^{-M}{\mathrm{exp}\text{$\left( \!-\frac{1}{2\sigma_n^2}\|\y-\boldsymbol{\Phi}\mathbf{A}(\theta)\x\|_2^2 \right)$} }\!.\hspace{-0.3cm}$$ The $i$-th dictionary atom, $i\!=\!1,\dots,\!N\!$, is defined $$\label{eq:dict_atoms_elements}
[\mathbf{a}_i]_l(\theta) = r(lT_d - i\delta t, \theta),\ \ l=1,\dots,L \,,$$ where the generating function, $r(lT_d - i\delta t, \theta)$, describes the reflection from a single FBG, incrementally shifted by $\delta t$ and sampled with a design sampling period, $T_d$. In order to specify the dictionary parameter in CFS according to [@Weiss2016], we write $$\label{eq:sensor_signa_IFT_model}
r(t,\theta) = \int_{-\infty}^{\infty} \text{e}^{j2\pi f t} H_{\text{\tiny LP}}(f,\theta)\,i_{\text{ph}}(f)\, \text{d}f \,.$$ Herein, $i_{\text{ph}}(f)$ is the received photocurrent in the frequency domain, and $H_{\text{\tiny LP}}(f,\theta)$ is the transfer function of a lowpass filter, that models a limited *effective bandwidth* of the receiver circuitry. This bandwidth is described in terms of a positive dictionary parameter, $\theta\in\mathbb{R}_+$. As an auxiliary parameter, it accounts for different indistinguishable sources of uncertainty, that all contribute to the broadening in the temporal response of the FBG reflections. A detailed model for $i_{\text{ph}}(f)$ is
The CRB for joint estimation of ($\x,\theta$) in CFS {#sec:CRB}
====================================================
We derive the CRB for jointly estimating the deterministic parameters ($\x,\theta$). This is a valid bound for the model considered in [@Weiss2016] and can be used to assess the relative performance gain achieved by the proposed probabilistic sparse model and DL strategies. Although the Bayesian CRB in [@Bobrovsky1987] can be empirically determined, we found that this bound is very lose, due to the high information content in the considered sparsity prior. Therefore, and in regard of the comparative analysis with the deterministic case in , the non-Bayesian CRB is more useful in this case.\
The constrained CRB for estimating $\x$ with sparsity constraints has been derived in [@Ben-Haim2010]. However, this derivation does not assume uncertainty in the dictionary. It is based on locally balanced sets and involves the projection of the *Fisher Information matrix* (FIM), $\boldsymbol{\mathcal{I}}(\x)$, onto a low-dimensional subspace spanned by the so-called *feasible directions*. Any estimator, $\hat{\x}$, for which the constrained CRB is a valid lower bound, must be unbiased with respect to these directions. The projection matrix can be created from the unit vectors corresponding to the non-zero coefficients in $\x$, that is $\mathbf{U}=[\mathbf{e}_{i_1},\dots,\mathbf{e}_{i_K}]$ with $i_k\in \mathcal{S}$, For a Gaussian likelihood as in (\[eq:Gauss\_likelihood\]), the FIM can be derived from the expected value of the Hessian matrix of the log-likelihood function, i.e. [@Kay1993; @Ben-Haim2010] $$\boldsymbol{\mathcal{I}}(\x) = - {\mathbb{E}_{\y}\hspace{4.0pt} }{\nabla_{\hspace{-1.5pt}\x}}^2 \log {p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\theta \hspace{0.8pt} )} = \frac{1}{\sigma_n^2}\mathbf{B}^{\!\top}\mathbf{B},$$ with $\mathbf{B} = \boldsymbol{\PHI}\mathbf{A}$. Further, we define the reduced FIM by . Then, given that $\x$ is exactly $K$-sparse, the constrained CRB for a *known* dictionary becomes [@Ben-Haim2010] $$\text{Cov}(\hat{\x}) \succeq \mathbf{U}\,\boldsymbol{\mathcal{I}}_K^{-1}\mathbf{U}^{\!\top},\quad \|\x\|_0 = K.$$ Based on these results, we derive the CRB for the joint parameters $\boldsymbol{\gamma} = (\x,\theta)$. First, we derive the Fisher information for $\theta$, given that $\x$ is known. It is given by $$\begin{aligned}
\nonumber \mathcal{I}(\theta) &=& - {\mathbb{E}_{\y}\hspace{4.0pt} } {\frac{\partial^2 }{\partial \theta^2}}\log {p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\theta \hspace{0.8pt} )}\\[0.0cm]
\nonumber &=& {\mathbb{E}_{\y}\hspace{4.0pt} } {\frac{\partial^2 }{\partial \theta^2}}\frac{1}{2\sigma_n^2} (\y-\PHI\mathbf{A}(\theta)\x)^{\!\top} ( \y-\PHI\mathbf{A}(\theta)\x)\\[0.0cm]
\label{eq:Fisher_theta} &=& \frac{1}{\sigma_n^2}\,\x^{\!\top}\mathbf{A}'(\theta)^{\!\top}\PHI^{\!\top}\PHI\mathbf{A}'(\theta)\x .\end{aligned}$$ Herein, $\mathbf{A}'(\theta)$ denotes the (element-wise) derivative of $\mathbf{A}(\theta)$ with respect to $\theta$. Next, we have to take into account that $\x$ and $\theta$ share some mutual information. Therefore, we define the combined FIM: $$\boldsymbol{\mathcal{I}}(\boldsymbol{\gamma}) = \left(
\begin{matrix}
\boldsymbol{\mathcal{I}}(\x) & -{\mathbb{E}_{\y}\hspace{4.0pt} }\mathbf{u}\\[0.1cm]
-{\mathbb{E}_{\y}\hspace{4.0pt} }\mathbf{u}^{\!\top} & \mathcal{I}(\theta)
\end{matrix} \right),$$ where $\mathbf{u}=[u_1,\dots,u_N]^T$ and $u_i = {\frac{\partial }{\partial x_i}}{\frac{\partial }{\partial \theta}}\log {p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\theta \hspace{0.8pt} )}$, Since the partial derivatives can be interchanged, the off-diagonal elements are identical. In order to complete the definition of $\boldsymbol{\mathcal{I}}(\boldsymbol{\gamma})$, we determine $$\begin{aligned}
\nonumber -{\mathbb{E}_{\y}\hspace{4.0pt} \!u_i } &=& -{\mathbb{E}_{\y}\hspace{4.0pt} }\frac{\partial^2}{\partial x_i\partial\theta} \log{p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\theta \hspace{0.8pt} )}\\
&=& \frac{1}{\sigma_n^2}\x^{\!\top}\mathbf{A}'(\theta)^{\!\top}\PHI^{\!\top}\PHI\,\mathbf{a}_i(\theta).\end{aligned}$$ The reduced FIM is obtained by appending the set of feasible directions, such that the coordinate $\theta$ is included, i.e. Hence, $\boldsymbol{\mathcal{I}}_{K+1}=\mathbf{\tilde{U}}^{\!\top}\boldsymbol{\mathcal{I}}(\boldsymbol{\gamma})\mathbf{\tilde{U}}$. To obtain the inverse, we apply twice the matrix inversion lemma [@Higham2002] $$\boldsymbol{\mathcal{I}}_{K+1}^{-1} =
\left(\begin{matrix}
\left(\boldsymbol{\mathcal{I}}_K-\frac{\mathbf{v}\mathbf{v}^{\!\top}}{\mathcal{I}(\theta)}\right)^{-1} &
-\frac{1}{\breve{b}}\boldsymbol{\mathcal{I}}_K^{-1}\mathbf{v} \\[0.2cm]
-\frac{1}{\breve{b}}\mathbf{v}^{\!\top}\boldsymbol{\mathcal{I}}_K^{-1} &
\frac{1}{\breve{b}}
\end{matrix}\right),$$ where $\breve{b} = \mathcal{I}(\theta)-\mathbf{v}^{\!\top}\boldsymbol{\mathcal{I}}_K^{-1}\mathbf{v}$, and $$\left(\boldsymbol{\mathcal{I}}_K-\frac{\mathbf{v}\mathbf{v}^{\!\top}}{\mathcal{I}(\theta)}\right)^{-1} = \boldsymbol{\mathcal{I}}_K^{-1}+\frac{1}{\breve{b}}\boldsymbol{\mathcal{I}}_K^{-1}\mathbf{v}\mathbf{v}^{\!\top}\boldsymbol{\mathcal{I}}_K^{-1}\,.$$ The constrained CRB for the joint parameters in $\boldsymbol{\gamma}$ becomes $$\text{Cov}(\boldsymbol{\gamma})\ \succeq\ \mathbf{\tilde{U}}\boldsymbol{\mathcal{I}}_{K+1}^{-1}\mathbf{\tilde{U}}^{\!\top},\quad \|\x\|_0 = K\,.$$ Finally, a lower bound for the *mean squared error (MSE)* in the joint setting is obtained by updating the individual estimation errors to account for the information shared between $\x$ and $\theta$: $$\begin{aligned}
\hspace{-0.4cm} {\text{MSE}\hspace{1.0pt}(\hat{\x})} &\!\!\! \geq&\!\!\! ({\text{Tr}\hspace{3.0pt}\ \boldsymbol{\mathcal{I}}_{K}^{-1} }) + \frac{1}{\breve{b}} \mathbf{v}^{\!\top}\boldsymbol{\mathcal{I}}_{K}^{-1}\boldsymbol{\mathcal{I}}_{K}^{-1}\mathbf{v}\\[0.1cm] \hspace{-0.4cm} {\text{MSE}\hspace{1.0pt}(\hat{\theta})}&\!\!\! \geq&\!\!\! \frac{1}{\breve{b}}\, =\ \mathcal{I}(\theta)^{-1} + \frac{\mathbf{v}^{\!\top}\boldsymbol{\mathcal{I}}_K\mathbf{v}}{\mathcal{I}(\theta)\,(\,\mathcal{I}(\theta)-\mathbf{v}^{\!\top}\boldsymbol{\mathcal{I}}_K\mathbf{v})}.\end{aligned}$$
Probabilistic sparse model {#sec:weak_sparsity_model}
==========================
Regarding the model in (\[eq:basic\_model\]), the data can be explained in different ways. On the one hand, many non-zero components in $\mathbf{x}$ and a large bandwidth, $\theta$, result in many narrow temporal peaks that can yield a good approximation of the observed reflections. On the other hand, it is known that the sensing fiber contains $K$ FBGs, so we expect exactly $K$ reflections. Therefore, a more useful explanation is given by $K$ significant elements in $\mathbf{x}$ with a smaller value of $\theta$, such that $\mathcal{S}$ correctly indicates the reflection delays. Nevertheless, even for a suitable value of $\theta$, the signal $\mathbf{x}$ is usually not exactly sparse but contains many small elements close to zero, e.g. due to measurement noise. In a strongly sparse model, these contributions are not taken into account, which impacts the positions of non-zero elements in $\x$. Hence, it may lead to incorrectly estimated reflection delays. This motivates a weakly sparse model, where the $K$ most significant components indicate the reflection delays. When $\x$ and $\theta$ are both unknown, the reflections delays can only be estimated when prior information of sparsity is incorporated, since $\theta$ depends on $\mathbf{x}$ and vice versa. Severe dictionary coherence aggravates this problem and results in several non-zero components with moderate amplitudes around the true significant elements. The coherence level is even stronger when the dimensionality of the acquired data is further reduced by CS. Thus, an adequate sparse model for $\x$ must compensate for this effect. Classic $\ell_1$-minimization can be interpreted as an MAP estimation problem, where $\mathbf{x}$ has i.i.d. entries with Laplace However, the required performance guarantees for $\ell_1$-minimization, essentially the RIP [@Candes2005; @Candes2008a], are no longer fulfilled in the case of strong dictionary coherence. According to [@Chartrand2007; @Chartrand2008], the RIP conditions can be relaxed for $\ell_p$-minimization, when $0<p<1$. Therefore, we use a prior with stronger selective shrinkage effect, that can be related to constraints on the $\ell_p$-norm in non-convex optimization. Yet, specific characteristics of the signal have to be considered. The measured reflection signal is proportional to the optical power, and the dictionary atoms essentially model the optical power reflected from the individual FBGs. Thus, the prior must also account for the non-negativity of the data. Due to these restrictions, we choose a Weibull prior that resembles a positive version of the horseshoe prior and induces the required selective shrinkage effect: $$\hspace{0.25cm} x_i \sim p(x_i) = {\mathcal{W}(\hspace{0.8pt} x_i \hspace{1.2pt} \boldsymbol{|} \hspace{1.3pt} \lWeibull,\kWeibull \hspace{0.8pt} )}\,, \ x_i > 0, \ i=1,\dots,N, \hspace{-0.3cm}
$$ where $\!\lambda_w,k_w$ are the scale and shape parameters, respectively. Then, the joint prior density of $\mathbf{x}$ is given by $$\label{eq:joint_x_Weibull}
{p(\hspace{0.8pt} \mathbf{x} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \kWeibull,\lWeibull \hspace{0.8pt} )} = \frac{\kWeibull}{\lWeibull^{\kWeibull}} \prod_{i=1}^{N} x_i^{\kWeibull-1}\,{\mathrm{exp}\text{$\left( \!-\lWeibull^{-\kWeibull} \sum_{i=1}^{N}x_i^{\kWeibull}\! \right)$} }\!.\hspace{-0.1cm}$$ Fig. \[fig:kernel\_and\_impact\_on\_pdf\] (top left) shows qualitatively the shape of the considered prior in the bivariate case.\
Based on (\[eq:joint\_x\_Weibull\]) and (\[eq:Gauss\_likelihood\]), we can relate the problem to constrained ML estimation. First, let us consider an interpretation in terms of MAP estimation as in [@Mohammad-Djafari2012], by calculating $\text{arg}\max_{\x}\, \log\,{p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\theta \hspace{0.8pt} )}{p(\hspace{0.8pt} \mathbf{x} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \kWeibull,\lWeibull \hspace{0.8pt} )}$ or, equivalently, $$\begin{aligned}
\nonumber\hspace{-1.2cm} &&\hspace{-0.0cm} \text{arg}\min_{\hspace{-0.3cm}\x}\ -\log\,{p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\theta \hspace{0.8pt} )}{p(\hspace{0.8pt} \mathbf{x} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \kWeibull,\lWeibull \hspace{0.8pt} )}\ = \\
\hspace{-1.2cm} &&\hspace{-0.0cm} \text{arg}\min_{\hspace{-0.3cm}\x}\, \|\y\!-\!\boldsymbol{\Phi}\mathbf{A}(\theta)\x\|_2^2
+ \mu_1\!\sum_{i=1}^N\log(x_i) + \mu_2\!\sum_{i=1}^{N}x_i^{\kWeibull}\!,\end{aligned}$$ where $\mu_1=(1-\kWeibull)$ and $\mu_2=\lWeibull^{-\kWeibull}$ with and $\mu_1,\mu_2 >0$. In order to formulate a related constrained ML problem, let us define two functions, $$\label{eq:constr_fcts}
g_1 = \sum_{i=1}^{N}x_i^{\kWeibull}-\lambda_1^{\kWeibull}\quad\ \ \text{and}\quad\ \ g_2 = \sum_{i=1}^N\log(x_i)-\lambda_2,$$ where $\lambda_1,\lambda_2 \in \mathbb{R}_+$ are related to the coefficients $\mu_1,\mu_2$, respectively. The functions in (\[eq:constr\_fcts\]) can represent inequality constraints of the form $g_1 \leq 0$ and $g_2\leq 0$, that account for the impact of the prior by restricting the search space. constrained version of the ML problem can be formulated by $$\begin{aligned}
\label{eq:optProblem_costFct}
\hspace{0.1cm}\text{arg}\min_{\hspace{-0.3cm}\x \succ \mathbf{0}} &\ \ \, \|\y-\boldsymbol{\Phi}\mathbf{A}(\theta)\x\|_2^2&\\[0.0cm]
\label{eq:optProblem_constr_1}
\text{s.t.}& \ \ \|\mathbf{x}\|_{k_w} &\leq\ \lambda_1 \\[0.1cm] \label{eq:optProblem_constr_2}
\text{and}& \ \ \sum_{i=1}^{N}\log(x_i)& \leq\ \lambda_2 \,.\end{aligned}$$ In this non-convex problem, $\|\mathbf{x}\|_p = (\sum_{i=1}^{N}|x_i|^p)^{1/p}$ denotes the $\ell_p$-norm with . The hyperparameters $\lambda_1, \lambda_2$ control the shrinkage effects. Fig. \[fig:kernel\_and\_impact\_on\_pdf\] (top right) depicts the search space restricted by the constraints (\[eq:optProblem\_constr\_1\])-(\[eq:optProblem\_constr\_2\]). are shown for a fixed value of $\lambda_1$ and $\lambda_2$ in the bivariate case.
Local covariance model for augmented sparsity
---------------------------------------------
In analogy to the concept of block sparsity [@Eldar2010], we can use the specific sparse structure of the signal with respect to the shift-invariant dictionary for CFS to exploit sparsity among groups of variables. The signal contains only $K$ reflections that arrive at temporally separated delays, indicated by the significant components in $\x$. Therefore, we can assume that a significant coefficient is always surrounded by larger groups of non-significant coefficients and any two significant components are always well separated. Also, it is likely that the amplitudes of adjacent non-significant coefficients are similarly close to zero. Borrowing from the ideas of MRFs [@Murphy2012], such local similarity can be modeled by a prior on the differential coefficients, $\Delta\x$, where $\Delta x_i = x_{i+1} - x_{i},\, i=1,\dots,N-1$. It restricts the variation of adjacent amplitudes and establishes a MRF relation between neighboring coefficients in $\x$. Then, non-significant coefficients with larger amplitudes are pulled down to match the majority with almost-zero amplitudes, which promotes additional *collective* shrinkage. However, if a significant coefficient follows a non-significant one (or vice versa), the model should allow for larger changes. Therefore, the differential variation must be locally specified, dependent on the respective amplitudes, in order to avoid undesired shrinkage or equalization.
\
To this end, we define a kernel function for all adjacent pairs of sparse coefficients, i.e. $\forall\ i=1,\dots,N-1$, with hyperparameter $\lambda_{\text{\tiny$\Delta$}}$: $$\label{eq:kernel}
\mathcal{K}(x_{i},x_{i+1}\,\boldsymbol{|}\,\lambda_{\text{\tiny$\Delta$}})\ =\ {\mathrm{exp}\text{$\left( -\lambda_{\text{\tiny$\Delta$}}\frac{|x_{i+1}-x_{i}|}{f_{\text{\tiny$\mathcal{K}$}}(x_i,x_{i+1})} \right)$} }\,.\vspace{0.05cm}$$ The bivariate function $\!f_{\text{\tiny$\mathcal{K}$}}$ controls the similarity level between adjacent coefficients. Within the scope of this work we consider cases, where this function takes the form $i=1,\dots,N-1$, with positive constants $r\leq 1$, $N_x < \infty$. They can be incorporated in ${p(\hspace{0.8pt} \!\x \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} k_w,\!\lambda_w\! \hspace{0.8pt} )}$ to yield a modified joint prior density, $$\hspace{-2.3cm}\tilde{p}(\mathbf{x}\,\boldsymbol{|}\,\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}})\ \ =\ \ \frac{1}{Z_{\text{\tiny$\mathcal{K}$}}} {\mathcal{W}(\hspace{0.8pt} x_N \hspace{1.2pt} \boldsymbol{|} \hspace{1.3pt} \kWeibull,\lWeibull \hspace{0.8pt} )}$$ $$\label{eq:modified_joint_x} \hspace{2.417cm} \times \prod_{i=1}^{N-1}\!\mathcal{K}(x_i,x_{i+1}\,\boldsymbol{|}\,\lambda_{\text{\tiny$\Delta$}})\, {\mathcal{W}(\hspace{0.8pt} x_i \hspace{1.2pt} \boldsymbol{|} \hspace{1.3pt} \kWeibull,\lWeibull \hspace{0.8pt} )},\hspace{-0.6cm}$$ with normalization constant $Z_{\text{\tiny$\mathcal{K}$}}$. For any $\alpha,\beta\in \mathbb{R}_+$, it holds that $0<\mathcal{K}(\alpha,\beta\,\boldsymbol{|}\,\lambda_{\text{\tiny$\Delta$}}) = \mathcal{K}(\beta,\alpha\,\boldsymbol{|}\,\lambda_{\text{\tiny$\Delta$}}) \leq 1$ and $$\begin{aligned}
\hspace{-0.5cm} \left.\tilde{p}(\mathbf{x}\,\boldsymbol{|}\,\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}})\right|_{Z_{\text{\tiny$\mathcal{K}$}}=1}
\!\!&\leq&\! {p(\hspace{0.8pt} \mathbf{x} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \kWeibull,\lWeibull \hspace{0.8pt} )}
$$ is bounded. Hence, there exists a positive constant that normalizes (\[eq:modified\_joint\_x\]) to make $\tilde{p}(\mathbf{x}\,\boldsymbol{|}\,\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}})$ a proper density. Fig. \[fig:kernel\_and\_impact\_on\_pdf\] (bottom) visualizes the function $\mathcal{K}(x_{i},x_{i+1}\,\boldsymbol{|}\,\lambda_{\text{\tiny$\Delta$}})$ and its impact on the original prior in the bivariate case.\
In the view of constraint ML estimation, the modified prior density in (\[eq:modified\_joint\_x\]) can be related to the optimization problem in (\[eq:optProblem\_costFct\])-(\[eq:optProblem\_constr\_2\]) by imposing additional constraints $$\frac{|x_{i+1}-x_{i}|}{f_{\text{\tiny$\mathcal{K}$}}(x_i,x_{i+1})} \leq \mu_i,\quad i=1,\dots,N-1\,.$$ Fig. \[fig:kernel\_and\_impact\_on\_pdf\] (top right) depicts a bivariate example. In order to show the MRF relation between the coefficients, we calculate the conditional densities $\forall\ x_i,\, i=1,\dots,N$. To this end, we conveniently define and get $$\hspace{-0.5cm}\tilde{p}(x_i\,\boldsymbol{|}\,\mathbf{x}_{\setminus i},\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}})\ =\ \,
\tilde{p}(x_i\,\boldsymbol{|}\,x_{i-1},x_{i+1},\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}})$$\
$$\label{eq:px_i_cond}
\hspace{0.4cm}\propto\ {\mathcal{W}(\hspace{0.8pt} x_i \hspace{1.2pt} \boldsymbol{|} \hspace{1.3pt} \kWeibull,\lWeibull \hspace{0.8pt} )}\,\mathcal{K}(x_{i-1},x_{i}\,\boldsymbol{|}\,\lambda_{\text{\tiny$\Delta$}})\,\mathcal{K}
(x_{i},x_{i+1}\,\boldsymbol{|}\,\lambda_{\text{\tiny$\Delta$}}),\hspace{-0.1cm}$$ $$\label{eq:px_1_cond}
\hspace{-0.80cm} \tilde{p}(x_1\boldsymbol{|}\mathbf{x}_{\setminus 1},\!\kWeibull,\!\lWeibull,\!\lambda_{\text{\tiny$\Delta$}}\!)\ \ \ \, \propto\ {\mathcal{W}(\hspace{0.8pt} \!x_1\! \hspace{1.2pt} \boldsymbol{|} \hspace{1.3pt} \kWeibull,\!\lWeibull\! \hspace{0.8pt} )}\,\mathcal{K}(x_{1},\!x_{2}\boldsymbol{|}\lambda_{\text{\tiny$\Delta$}}\!),$$ $$\label{eq:px_N_cond}
\hspace{-0.75cm} \tilde{p}(x_N\boldsymbol{|}\mathbf{x}_{\setminus N}\!,\kWeibull,\!\lWeibull,\!\lambda_{\text{\tiny$\Delta$}}\!)\ \, \propto
{\mathcal{W}(\hspace{0.8pt} \!x_N\! \hspace{1.2pt} \boldsymbol{|} \hspace{1.3pt} \kWeibull,\!\lWeibull\! \hspace{0.8pt} )}\,\mathcal{K}(x_{N\text{-}1},\!x_{N}\boldsymbol{|}\lambda_{\text{\tiny$\Delta$}}\!)\hspace{-0.01cm},
$$ where dependencies appear only between directly adjacent coefficients.\
In order to account for deviations from prior assumptions, we consider randomization of the hyperparameters and assign conjugate inverse Gamma priors to the scale parameters $\lambda_w$ and $\lambda_{\text{\tiny$\Delta$}}$. Finally, given $\lambda_w$ and a normalization constant $Z_{\kWeibull}$, the shape parameter, $\kWeibull > 0$, is assigned the conjugate prior distribution according to [@Fink1997]: $$\label{eq:kw_prior}
\hspace{0.05cm} {p(\hspace{0.8pt} \kWeibull \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} a',b',(d\,')^{k_w}\!\!,\lWeibull \hspace{0.8pt} )} = \frac{\kWeibull^{a'}}{Z_{\kWeibull}} {\mathrm{exp}\text{$\left( \!\!-b'\kWeibull\! - \frac{(d\,')^{\kWeibull}}{\lWeibull}\! \right)$} }, \hspace{-0.15cm}$$ Fig. \[fig:factor\_graph\_localSimilarity\] shows a factor graph for the complete sparsity model with randomized hyperparameters.
Approximate Inference: Hybrid MCMC {#sec:approx_inference}
==================================
In order to accomplish inference in the sparse model, we apply a hybrid MCMC technique, i.e. HMC within Gibbs sampling. The reasons for using HMC are twofold: Firstly, it only requires an analytic expression for the posterior density to be sampled. Secondly, it is efficient in sampling high-dimensional spaces in the presence of correlation. However, as pointed out in [@Neal2011], it can be more efficient to sample the hyperparameters separately, as their posterior distributions are often highly peaked and require a small step size in the HMC algorithm, which limits the general performance. Therefore, we employ an outer Gibbs sampler for approximate inference of the latent variables. In each iteration, $\tilde{p}(\x\,\boldsymbol{|}\,\lWeibull,\kWeibull,\lambda_{\text{\tiny$\Delta$}})$ is sampled using HMC, while all other variables are fixed. Since we are also interested in estimating the noise variance, $\sigma_n^2$, it is assigned an inverse Gamma ($\text{Inv-}\Gamma$) prior and sampled along with the other variables. The resulting model is summarized below: $$\begin{aligned}
\nonumber \mathbf{x}\,\boldsymbol{|}\,\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}}&\sim& \tilde{p}(\x\,\boldsymbol{|}\,\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}})\,\ \ \quad\qquad \; \text{in}\ (\ref{eq:modified_joint_x}),\\
\nonumber \lWeibull &\sim& \text{Inv-}\Gamma(\lWeibull\,\boldsymbol{|}\,a,b),\\
\nonumber \kWeibull\,\boldsymbol{|}\,\lWeibull &\sim& {p(\hspace{0.8pt} \kWeibull \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} a',b',(d\,')^{k_w},\lWeibull \hspace{0.8pt} )}\; \; \; \; \text{in}\ (\ref{eq:kw_prior}),\\
\nonumber \lambda_{\text{\tiny$\Delta$}} &\sim& \text{Inv-}\Gamma(\lambda_{\text{\tiny$\Delta$}}\,\boldsymbol{|}\,a'',b'')\\
\label{eq:sparse_model_summary} \sigma_n^2 &\sim& \text{Inv-}\Gamma(\lambda_{\text{\tiny$\Delta$}}\,\boldsymbol{|}\,a_\sigma,b_\sigma)\,.\end{aligned}$$ We also define $\zeta\in\mathcal{C} = \{\kWeibull,\lWeibull,\lambda_{\text{\tiny$\Delta$}},\sigma_n^2\}$ as a representative variable with corresponding positive, real-valued parameters $a_\zeta\in\{a,a'',a_\sigma\}$ and $b_\zeta\in\{a,a'',a_\sigma\}$, that belong to the respective density functions. Further, the set $\mathcal{C}_{\setminus\zeta}$ denotes the set $\mathcal{C}$ without the respective variable $\zeta$. Fig. \[fig:dependency\_graph\] shows a graphical model that helps to visualize the dependencies in this model. Herein, $\theta$ and $\boldsymbol{\Xi}$ are only valid for strategy $\textbf{\emph{S2}}$, which is discussed in For the particular model in (\[eq:sparse\_model\_summary\]), we assume that the variables $\x, \sigma_n^2$ and $\theta$ are mutually independent. Gibbs sampling requires the full conditional distributions for each parameter of interest. Based on these assumptions, we obtain the relation $$\begin{aligned}
{p(\hspace{0.8pt} \zeta \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \y,\x,\mathcal{C}_{\setminus\zeta} \hspace{0.8pt} )} & \propto & {p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\mathcal{C} \hspace{0.8pt} )}\,{p(\hspace{0.8pt} \zeta \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\mathcal{C}_{\setminus\zeta} \hspace{0.8pt} )}\\
& \propto & {p(\hspace{0.8pt} \y \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \x,\mathcal{C} \hspace{0.8pt} )}\, {p(\hspace{0.8pt} \zeta \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \mathcal{C}_{\setminus\zeta} \hspace{0.8pt} )}\, \tilde{p}(\x\,\boldsymbol{|}\,\mathcal{C}) \,.\end{aligned}$$ Since the prior distributions are all conjugate to the Gaussian likelihood function in (\[eq:Gauss\_likelihood\]), a simple calculation yields the posterior distributions of the parameters involved in the Gibbs sampling procedure. For $\zeta\in\mathcal{C}_{\setminus\kWeibull}$, we obtain $$\label{eq:posterior_zeta_no_kw}
\zeta\,\boldsymbol{|}\,{\y,\x,\mathcal{C}_{\setminus\zeta}} \, \sim\ \text{Inv-}\Gamma(\zeta\,\boldsymbol{|}\,a_\zeta+{\text{\scriptsize$\frac{M}{2}$}},\ b_\zeta +\text{\scriptsize$\frac{1}{2}$}\,)\ \tilde{p}(\x\,\boldsymbol{|}\,\mathcal{C}), $$ and for $\kWeibull$, we obtain $$\label{eq:posterior_kw}
\kWeibull\,\boldsymbol{|}\,{\y,\x,\mathcal{C}_{\setminus\kWeibull}}\ \sim\ \ {p(\hspace{0.8pt} \kWeibull \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \tilde{a}\,',\, \tilde{b}\,',\, \tilde{c}\,' \hspace{0.8pt} )}\
\tilde{p}(\x\,\boldsymbol{|}\,\mathcal{C}),$$ with parameters $\tilde{a}\,' = a\,'+N$, $\ \tilde{b}\,'=b\,'+\sum_{i=1}^{N}\log(x_i)$, and $\tilde{c}\,' = (d\,')^{\kWeibull}+\sum_{i=1}^{N}x_i^{\kWeibull}$. Samples of the posterior variables can be obtained using Metropolis Hastings [@Bishop2006] or HMC.\
![Factor graph of the complete sparse model with local similarity.[]{data-label="fig:factor_graph_localSimilarity"}](./sparse_model_full.pdf){width="0.63\columnwidth"}
The sparse coefficients are sampled using HMC. We briefly describe the idea of this method according adapted to our model for $\x$:\
Within the framework of HMC, the sampling process is described in terms of *Hamilton dynamics*, a concept known from classical physics. It is used to describe the trajectory of a physical system in phase space, based on its potential and kinetic energy. HMC assigns to every sparse coefficient, $x_i$, an associated momentum variable, $\xi_i$, $i=1,\dots,N$, that is responsible for the sampling dynamics. The posterior density to be sampled is related to the potential energy, given by [@Neal2011] $$U(\x\,\boldsymbol{|}\,\y,\mathcal{C})\ =\ -\log \tilde{p}(\x\,\boldsymbol{|}\,\y,\mathcal{C}) - \log(Z_u)\,,$$ where $Z_u$ is a suitable normalization constant. Since $\y$ and $\mathcal{C}$ are fixed, we may drop them and write $U(\x)$ instead. The kinetic energy, $K(\boldsymbol{\xi})$, depends only on the auxiliary variables $\boldsymbol{\xi}\!=\![\xi_1,\dots,\xi_N\!]$. A standard choice for $K(\boldsymbol{\xi})$ corresponds to independent particles in free space with mass $m_i$, i.e. The dynamics of the sampling process are governed by the *Hamiltonian function*, which is given by $\mathcal{H}(\x,\boldsymbol{\xi})\! =\!U(\x)\! +\! K(\boldsymbol{\xi})$ and represents the total system energy. The joint density of $\!(\x,\boldsymbol{\xi})\!$ is defined by [@Neal2011] $$\label{eq:HMC_canonical_density}
p(\x,\boldsymbol{\xi}) = \frac{1}{Z_c} \text{e}^{-\frac{\mathcal{H}(\x,\boldsymbol{\xi})}{T_{\text{\scriptsize sys}}}}\!
= \ \tilde{p}(\x\,\boldsymbol{|}\,\y,\mathcal{C}) \,\prod_{i=1}^{N}\mathcal{N}(\,\xi_i\,\boldsymbol{|}\,0,m_i\,).
$$ Herein, $T_{\text{\scriptsize sys}}$ is called the *system temperature* and $Z_c$ is a normalization constant. The last equation is obtained by setting $T_{\text{\scriptsize sys}}=1$ and $Z_u = Z_c$, while the Gaussian density arises from the special choice of the kinetic energy term. In HMC, a proposal for a new sample is obtained by the final points ($x_i^*,\xi_i^*$) of a trajectory described by Hamilton’s equations of motion. They are calculated $\forall\ (x_i,\xi_i), i\!=\!1,\dots,N$, [@Neal2011]: $$\hspace{0.0cm} \frac{\mathrm{d}x_i}{\mathrm{d}t} = \frac{\xi_i}{m_i}\,, \qquad \frac{\mathrm{d}\xi_i}{\mathrm{d}t} = -\frac{{\frac{\partial }{\partial x_i}}\tilde{p}(\x\,\boldsymbol{|}\,\y,\mathcal{C}) }{\tilde{p}(\x\,\boldsymbol{|}\,\y,\mathcal{C}) }\,.$$ A Metropolis update decides, whether a proposed sample is accepted or rejected, with acceptance probability [@Neal2011] $$\text{P}(\text{accept}) = \min_{}\,(\,1 ,\ {\mathrm{exp}\text{$\left( -\mathcal{H}(x_i^*,\xi_i^*) + \mathcal{H}(x_i,\xi_i) \right)$} }\ )\,.$$
Parametric DL strategies for CFS {#sec:PDL}
================================
In this section, we present two strategies for parametric dictionary learning in CFS. In the first we follow the ideas of hybrid Bayesian inference [@Yuan2009; @Yuan2015] and AM-based DL [@Beck2013], where $\theta$ is a deterministic parameter, that is estimated using the Monte Carlo EM algorithm in [@Bishop2006]. In the second we pursue a full Bayesian approach and consider a probabilistic model for $\theta$. Herein, approximate inference is accomplished by extending the Gibbs sampler in jointly estimate $(\x,\theta,\sigma_n^2)$. depicts the dependency graph for both strategies, where $\theta, \boldsymbol{\Xi}$ belong exclusively\
As pointed out in [@Yuan2009; @Yuan2015], hybrid and full Bayesian strategies have their individual advantages in certain situations. For small sample sizes, Bayesian methods can be superior if good prior knowledge is available [@Yuan2009]. Nonetheless, they are often computationally more complex and insufficient prior information can lead to a small-sample bias, even if a non-informative prior is used [@Yuan2009]. In CFS, the sample size is small and only vague prior knowledge of $\theta$ is available. Therefore, we investigate the performance of both DL strategies based on our probabilistic sparse model. The computational complexity of both strategies is comparable. It is dominated by HMC, i.e. by sampling the high-dimensional vector $\x$ in each iteration of the Gibbs sampler. Regarding $\theta$, the following prior knowledge is assumed: In ***S1***, we roughly restrict the range of values can take, while in ***S2***, we define a non-informative prior over the same range. Recall that $\theta$ effectively describes the filter characteristics of the lowpass To create the dictionary for a certain value of $\theta$ using (\[eq:dict\_atoms\_elements\]), the inverse Fourier transform in (\[eq:sensor\_signa\_IFT\_model\]) has to be evaluated for each atom. Thus, the dictionary is not a simple function of $\theta$ and we restrict ourselves to a discrete set of parameters, with lower and upper bound, $\theta_{\text{\tiny min}}$ and $\theta_{\text{\tiny max}}$, respectively. Since the bandwidth should be positive and bounded, we have $0 < \theta_{\text{\tiny min}}$ and $\theta_{\text{\tiny max}} < \infty$. Then, the set $\Theta$ contains the discrete values $\theta_r, r=1,\dots,\Rtheta$,
Hybrid DL: iterative estimation of $\theta$ and $(\x,\mathcal{C})$ (***S1***)
-----------------------------------------------------------------------------
The dictionary parameters in the CFS problem can be iteratively estimated using a Monte Carlo EM algorithm. First, an initial value, $\theta^{(0)}\!$, has to be chosen. In subsequent iterations with indices $d\!=\!1,\dots,d_{\text{\tiny max}}$, we obtain joint samples $\{\x_l,\mathcal{C}_l\}^{(d)},\,l=1,\dots,L_{\text{\tiny MC}}$, by Gibbs sampling and HMC according to Then, we determine the posterior expectation of $\zeta\in \mathcal{C}$, using the previous estimate $\hat{\theta}^{(d-1)}$: $$\begin{aligned}
\hat{\zeta}^{(d)} &=& \int_{\text{dom}(\zeta)} \zeta \,{p(\hspace{0.8pt} \zeta \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \y,\hat{\theta}^{(d-1)} \hspace{0.8pt} )}\,\mathrm{d}\zeta \\
\label{eq:posterior_mean_zeta} &\approx&\!\! \frac{1}{L_{\text{\tiny MC}}}\sum_{l=1}^{L_{\text{\tiny MC}}}\ \zeta_{\,l}^{(d)}\, {p(\hspace{0.8pt} \zeta_l^{(d)} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \y,\hat{\theta}^{(d-1)} \hspace{0.8pt} )},\end{aligned}$$ where $\text{dom}(\zeta)$ is the domain of $\zeta$. The current estimates of the reflection delays, $\hat{\mathcal{S}}^{(d)}$, are determined by identifying the indices of the $K$ largest elements in the posterior mean of $\x$, denoted by $\hat{\x}^{(d)}$. It is obtained by exchanging $\zeta_{\,l}^{(d-1)}$ with $\x_l^{(d-1)}\!\!$ in (\[eq:posterior\_mean\_zeta\]). Besides, we also estimate the amplitudes of the significant components in $\x$. They can be useful to assess the sparsity level of the solution and to determine the amount of optical power reflected from the FBGs. Since the posterior of $\x$ is multimodal with one narrow peak around zero and another peak at some larger amplitude, the MAP is more suitable for this task. It is given by $$\begin{aligned}
\label{eq:EM_MAP_x}
\hspace{-0.5cm} \{\mathbf{\hat{x}},\hat{\mathcal{C}}\,\}_{\text{\tiny MAP}}^{(d)}\hspace{-0.1cm} &=&\hspace{0.05cm} \text{arg}\max_{\hspace{-0.25cm}\x,\mathcal{C}}\ \log {p(\hspace{0.8pt} \x,\mathcal{C} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \y,\hat{\theta}^{(d-1)} \hspace{0.8pt} )}\\[0.2cm]\label{eq:EM_MAP_x_approx} \hspace{-0.5cm} &\hspace{-2.8cm}\approx&\hspace{-0.6cm} \text{arg}\hspace{-1.08cm}\max_{\hspace{-0.45cm}\{\x_j,\mathcal{C}_j\}\in\{\x_l,\,\mathcal{C}_l\}^{(d)}_{l=1,..,L_{\text{\tiny MC}}}}\hspace{-0.5cm} \log {p(\hspace{0.8pt} \{\x_j,\mathcal{C}_j\}^{(d)} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \y,\hat{\theta}^{(d-1)} \hspace{0.8pt} )}.
$$ However, the estimates of $\mathcal{S}$ obtained from $\hat{\x}_{\text{\tiny MAP}}^{(d)}$ are less accurate than those obtained by the posterior mean. Therefore, the empirical MAP solution is only used to estimate the reflection amplitudes. Next, we calculate the current estimate $\hat{\theta}^{(d)}$ by taking the expected value over $\x,\mathcal{C}$ given $\y,\theta$ (E-step): $$\begin{aligned}
\label{eq:EM_E_step}
\nonumber\hspace{-1.0cm} && \!\!\! {\mathbb{E}_{\text{\scriptsize$\x,\!\mathcal{C}\,$}\boldsymbol{|}\text{\scriptsize$\,\y,\!\theta$}}\hspace{4.0pt} } \log {p(\hspace{0.8pt} \y,\x,\mathcal{C} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \theta \hspace{0.8pt} )}\\ \hspace{-1.0cm} &=&\!\!\!\! \int_{\mathbb{R}_+^N} \int_{\Psi}
\log {p(\hspace{0.8pt} \y,\x,\mathcal{C} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \theta \hspace{0.8pt} )}\, {p(\hspace{0.8pt} \x,\mathcal{C} \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \y,\theta \hspace{0.8pt} )}\,\mathrm{d}\mathcal{C}\,\mathrm{d}\x\\ \label{eq:Qfct}\hspace{-1.0cm} &\approx&\!\!\!\! \frac{1}{L_{\text{\tiny MC}}}\sum_{l=1}^{L_{\text{\tiny MC}}}\ \log {p(\hspace{0.8pt} \y,\{\x_l,\mathcal{C}_l\}^{(d-1)}\! \hspace{1.2pt} \boldsymbol{|} \hspace{1.2pt} \theta \hspace{0.8pt} )}\ \triangleq\ Q(\,\theta\,\boldsymbol{|}\,\hat{\theta}^{(d-1)}).\end{aligned}$$ Herein, $\Psi$ is the product space formed by the individual domains of all variables in $\mathcal{C}$. In the $M$-step, a locally optimal value, $\hat{\theta}^{(d)}$, is obtained by maximizing $\theta$ over the set $\Theta$, i.e. $$\label{eq:EM_M_step}
\hat{\theta}^{(d)}\ =\ \text{arg}\max_{\hspace{-0.3cm}\theta\, \in\, \Theta}\ Q(\,\theta\,\boldsymbol{|}\,\theta^{(d-1)}\,)\,.$$
![Dependency relations for the complete hierarchical model. The variables $\theta$ and $\boldsymbol{\Xi}$ appear exclusively in ***S2***.[]{data-label="fig:dependency_graph"}](./dependency_FULL_BAYES_2.pdf){width="0.63\columnwidth"}
### Initialization of $\theta$ via bisectional search
An adequate initialization, $\theta^{(0)}\!$, can alleviate the problem of local optima in the EM algorithm. In CFS, the desired sparsity level is known to be the number of reflections, $K$. Hence, a good choice for $\theta^{(0)}\!$ yields a solution for $\x$ with $K$ significant non-zero elements. Starting at an arbitrary value $\theta^{(0)}\!\in\Theta$, a bisectional search within $\Theta$ can quickly determine a suitable initial value. After choosing the first value at random, $\Theta$ is subdivided into two parts, containing all larger and all smaller values, respectively. When the number of peaks is too high, the next trial is chosen as the median of the lower division. If it is too low, the next trial is the median of the upper division, and so on. For a properly selected $\theta^{(0)}\!$, ***S1*** converges faster and is more likely to approach (or even attain) the global optimum.
Bayesian DL: joint estimation of ($\x,\mathcal{C},\theta$)$\ $ (***S2***)
-------------------------------------------------------------------------
In strategy ***S2***, we treat $\theta$ as a random variable. Due to its discrete nature, each element is assigned a probability mass, $p_r = p(\theta_r)\,, r=1,\dots,\Rtheta$, where $\sum_{r=1}p(\theta_r)=1$. Then, $\theta$ is distributed over the set of discrete dictionary parameters, $\Theta$, with corresponding probability masses in $\boldsymbol{\Xi} = \{p_1,\dots,p_{\text{\tiny $R_{\small\Theta}$}}\}$. Uncertainty in the *a priori* assigned probability masses is taken into account in terms of a prior on $\boldsymbol{\Xi}$. The Dirichlet (Dir) distribution can be used as the conjugate prior with parameters $\boldsymbol{\nu} = [\nu_1,\dots,\nu_{\text{\tiny$\Rtheta$}}]^{\!\top}$, i.e. $$p(\boldsymbol{\Xi}) =\ \frac{1}{B(\boldsymbol{\nu})}\prod_{r=1}^{\Rtheta}p_r^{\nu_r}\,,$$ where $B(\boldsymbol{\nu})$ denotes the *Beta* function and the variables $\nu_r$, $r=1,\dots,\Rtheta$, describe the number of occurrences of the values in $\Theta$. When a new element, $\theta_q\in\Theta$, is sampled, a posterior count is assigned to that value. After sampling another value in the next iteration, this count is reassigned to the new value. Let $\mathbf{\breve{c}}\in \mathbb{N}^{\Rtheta}$ indicate the current sample, i.e. $c_q=1$ for one index $q\in\{1,\dots,\Rtheta\}$, while all other elements are zero. A non-informative prior is obtained if all values $\theta_q\in\Theta$ are equally likely and each element is assigned a single count. Then, $\nu_r=1\ \forall\ r=1,\dots,\Rtheta$ and a new sample has a strong impact on the posterior distribution. In contrast, for large values, e.g. $\nu_r\! =\! 1000\ \forall\ r\!=\!1,\dots,\Rtheta$, a new count leaves the distribution The complete model is then given by (\[eq:sparse\_model\_summary\]) and, in addition, $$\begin{aligned}
\boldsymbol{\Xi} &\sim & \text{Dir}(\,\boldsymbol{\Xi}\,\boldsymbol{|}\,\boldsymbol{\nu}\,)\\[0.15cm]
\theta\,\boldsymbol{|}\,\boldsymbol{\Xi} &\sim& \text{Cat}(\theta\,\boldsymbol{|}\,\Rtheta,\boldsymbol{\Xi}) \, .\end{aligned}$$ To accomplish approximate inference in this model, the variables $\theta$ and $\boldsymbol{\Xi}$ are included in the Gibbs sampling procedure of Section \[sec:approx\_inference\]. Therefore, the conditional distributions must be determined. Based on the dependencies in Fig. \[fig:dependency\_graph\], and since $\x,\sigma_n^2$ and $\theta$ are assumed to be mutually independent, we find $$\begin{aligned}
\label{eq:theta_Gibbs}
\boldsymbol{\Xi} \,\boldsymbol{|}\,\theta\ =\ \boldsymbol{\widetilde{\Xi}}\ & \sim &\ \text{Dir}(\,\boldsymbol{\Xi}\,\boldsymbol{|}\,\boldsymbol{\nu} + \mathbf{\breve{c}}\,),\\[0.15cm]
\label{eq:Xi_Gibbs}
\theta \,\boldsymbol{|}\,\y,\boldsymbol{\widetilde{\Xi}}\ &\sim&\ \text{Cat}(\,\theta\,\boldsymbol{|}\,\Rtheta,\boldsymbol{\widetilde{\Xi}}\,).
$$
(1,0)[280]{}\
**Algorithm:** Sparse estimation and PDL, strategy ***S1*** & ***S2***
(1,0)[280]{}
----------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Input:** $\mathbf{y},M,\PHI,N,L,T_d,\delta t,r(t,\theta),K,L_{\text{\tiny MC}},d_{\text{max}}$
**Output:** $\hat{\mathcal{S}},\hat{\mathbf{x}},\hat{\theta},\hat{\sigma}_n,d,ee $
**Parameters:** $a,a',a'',a_\sigma,b,b',b'',b_\sigma,d\,',\boldsymbol{\nu}, \Rtheta, \{\theta_r\}_{r=1}^{\Rtheta}$,
internal HMC parameters (c.f. [@Neal2011; @Homan2014]).
**0. Initialize:** $\theta$ at random $\rightarrow \hat{\theta}^{(0)}$ via bisectional search,
$\mathbf{A}(\hat{\theta}^{(0)}),\{\hat{\mathbf{x}}^{(0)}\!\!,\, \hat{\mathcal{C}}^{(0)}\}\!$ as in (\[eq:sparse\_model\_summary\]), (***S2***):$\,d_{\text{max}}\!\!=\!1$
**1. for** $d = 1$ to $d_{\text{max}}$ **do**
**2.** **for** $l=1$ to $L_{\text{\tiny MC}}$ **do**
**3.** Gibbs sampling: (i) $\mathcal{C}_l^{(d)}$ using (\[eq:posterior\_zeta\_no\_kw\]) and (\[eq:posterior\_kw\]),
\(ii) $\x_l^{(d)}$ via HMC.
(***S2***):$\,$ (iii) $\theta_l^{(d)}\!\!,\, \boldsymbol{\Xi}_l^{(d)}$ using (\[eq:theta\_Gibbs\]) and (\[eq:Xi\_Gibbs\])
**4.** **end for**
**5.** Estimate: $\hat{\mathcal{S}}^{(d)}$ from $\hat{\x}^{(d)}$ in (\[eq:posterior\_mean\_zeta\]) with $\zeta_{\,l}^{(d)}\!\rightarrow \x_l^{(d)}\!\!$,
$\hat{\mathcal{C}}^{(d)}$ from (\[eq:posterior\_mean\_zeta\]), $\hat{\x}^{(d)}_{\text{\tiny MAP}}$ from (\[eq:EM\_MAP\_x\_approx\]),
**5.a** (***S1*:**) $\,\hat{\theta}^{(d)} = \text{arg}\max_{\theta\in\Theta}\ Q(\theta\,\boldsymbol{|}\,\hat{\theta}^{(d-1)})$.
**5.b** (***S2*:**) $\,\hat{\theta}^{(d)}$ from (\[eq:posterior\_mean\_zeta\]) with $\zeta_{\,l}^{(d)}\!\rightarrow \theta_l^{(d)}$.
**6.** **if** $\hat{\theta}^{(d)} ==\, \hat{\theta}^{(d-1)}$ **or** $d==d_{\text{max}}$
**7.** **return** $\, \hat{\mathcal{S}}^{(d)}\!,\, \hat{\x}^{(d)}_{\text{\tiny MAP}},\,\hat{\mathcal{C}}^{(d)}\!,\, \hat{\theta}^{(d)}, ee\!=\!\|\y\!-\!\PHI\mathbf{A}(\hat{\theta}^{(d)}\!)\|_2^2$.
**8.** **end if**
**9. end for**
----------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
(1,0)[280]{}
Simulations and experimental data {#sec:performance}
=================================
Let us now evaluate the proposed sparse model and DL strategies. First, we show the qualitative behavior of the algorithms, followed by a quantitative performance analysis in comparison to the method in [@Weiss2016]. To this end, we consider several scenarios of different SNRs, CS sampling matrices and sample sizes. Finally, we apply our algorithms to experimental data taken from a real fiber-optic sensor.
Simulation setup
----------------
We consider $K=3$ uniform FBGs in the sensing fiber, where the observed reflections have a common amplitude, $A_x$, and two reflections are closely spaced. Their delays are indicated by the inidces of the $K$ most significant elements in $\x$, contained in the set $\mathcal{S}$. Subsequently, the dictionary parameter is re-defined relative to its true value, i.e. $\hat{\theta}$ is replaced by $\hat{\theta}/\theta$. Further, we use $\Rtheta = 100$ discrete parameter values, equally spaced between $30\%$ and $150\%$ of the true value. The original signal (prior to CS) contains $L=134$ samples of the measured photocurrent. The dictionary atoms are created using $L$ samples of $r(t\!-\!i\delta t), i=1,\dots,N$, with a delay spacing of We use two types of CS matrices, $\PHI$, with i.i.d. entries drawn from the distributions below:
--------------------------- ----------------------------------------------------------------------------
\(a) Gauss: $\,\mathcal{N}(0,1)$,
\(b) DF [@Achlioptas2003] $\{-1,0,1\}$ with probabilities $\{\frac{1}{6},\frac{2}{3},\frac{1}{6}\}$.
--------------------------- ----------------------------------------------------------------------------
The variables $\{\mathcal{C},\theta,\boldsymbol{\Xi}\}$ are sampled according to For $\x$ we use the ’No-U-Turn’ variant of the HMC which is efficiently implemented in the software package *Stan* [@STAN]. The algorithm ***S1*** is initialized based on a bisectional search and runs at most $d_{\text{max}}=35$ iterations. In ***S2***, we use a non-informative prior for $\theta$, with $p(\theta_r)=1/\Rtheta$ and $\nu_r = 1\ \forall\ r=1,\dots,\Rtheta$.
Visualization and Working Principle
-----------------------------------
The working principle of the algorithms is presented for dB and a Gaussian CS matrix using $M/L=50\%$ of the original samples. Fig. \[fig:visualization\] (top left) depicts the MAP solution for $\x$, obtained by HMC within Gibbs sampling according to Section \[sec:approx\_inference\], where $\theta$ is fixed to the true value. It shows, that collective shrinkage, imposed by the local similarity assumption in the joint prior density of $\x$, yields a highly improved sparsity level in the presence of strong dictionary coherence. Fig. \[fig:visualization\] (top right) shows the posterior density of $\x$ in one dimension. For a non-significant component, it is strongly peaked around zero, and for a significant component, is multimodal with a strong mode around the true amplitude and a smaller mode around zero.
\
\
\[fig:visualization\]
The second row in delineates the evolution of the EM algorithm in ***S1*** over several iterations. Fig. \[fig:visualization\] (center left) shows the current MAP solutions for $\x$, i.e. $\x_{\text{\tiny MAP}}^{(d)}$, zoomed on the two left-sided peaks. Due to a bad initial value for $\theta$, more than $K$ peaks appear in the first iterations. However, as the algorithm proceeds, significant peaks are formed only at the positions of the true significant components (black bullets). Fig. \[fig:visualization\] (center right) shows, that also $\theta$ approaches the true value. Fig. \[fig:visualization\] (bottom left) delineates a typical shape of the function $Q(\theta\,|\,\hat{\theta}^{(d-1)})$ of ***S1*** in (\[eq:Qfct\]), for a properly and badly chosen initial value, $\theta^{(0)}$. A good choice leads to faster convergence, while for a bad choice, the algorithm might get either stuck at a local optimum or requires many EM iterations before the maximum of the Q-function appears close the true value of $\theta$. Finally, Fig. \[fig:visualization\] (bottom right) depicts for ***S2***, the non-informative prior of $\theta$ and a typical posterior density when $\nu_r=1\ \forall\ r=1,\dots,\Rtheta$.
Performance evaluation
----------------------
The performance is evaluated in terms of the root mean-squared error (RMSE). For a vector $\mathbf{v}$ and an estimator $\mathbf{\hat{v}}$, it is given by $\text{RMSE}(\mathbf{v},\mathbf{\hat{v}}) = ({\mathbb{E}_{}\hspace{4.0pt} \!\|\mathbf{v}-\mathbf{\hat{v}}\|_2^2 })^{1/2}$.
\
\
\
\
\
\
\[fig:JOINT\_performance\_50\]
\
\
\
We define $\overline{\text{RMSE}}$ as the approximation, where the expectation is replaced by averaging estimates over 100 Monte Carlo trials. We compare ***S1***, ***S2*** to the PDL-OIAI algorithm in [@Weiss2016], which considers a deterministic sparse model and incorporates a pre-processing routine to handle strong dictionary coherence. To calculate the CRB of Section \[sec:CRB\], the derivative of $r(t,\theta)$ with respect to $\theta$ must be determined for all dictionary elements. Since $r(t,\theta)$ is not a simple function of $\theta$, it can be approximated for a certain value $\theta_0$. For the $(l,\!i)$-th element in $\mathbf{A}'(\theta)$, we obtain $$\hspace{0.01cm} \left.{\frac{\partial }{\partial \theta}}[\mathbf{a}_i(\theta)]_l\right|_{\theta_0}\!\! \approx\ \frac{r(lT_d-\tau_i,\theta_0) - r(lT_d-\tau_i,\theta_0-\Delta\theta)}{\Delta\theta}.\hspace{-0.30cm}
$$ Fig. \[fig:EM\_performance\_50\]-\[fig:JOINT\_performance\_30\] show the results of the proposed and the competetive method. Herein, $\mathbf{s}\in \mathbb{N}^K$ contains all elements in $\mathcal{S}$, $\x_{\mathcal{S}}\in \mathbb{R}_+^K$ contains the coefficients of $\x$ with indices in $\mathcal{S}$. The $\overline{\text{RMSE}}(\x_{\mathcal{S}},\hat{\x}_{\mathcal{S}})$ compares the estimated amplitudes at the positions in $\hat{\mathcal{S}}$ to the true common amplitude, $A_x$, at positions in $\mathcal{S}$. The lower bound of the RMSE for jointly estimating *deterministic* parameters $(\x_\mathcal{S},\theta)$, induced by the CRB derived in Section \[sec:CRB\], is denoted by ’RCRB’.\
Fig. \[fig:EM\_performance\_50\] shows the results for $\{\mathbf{s},\x_{\mathcal{S}},\theta,\sigma_n\}$, obtained by ***S1*** using $50\%$ of the original samples. For fewer samples, the EM algorithm in ***S1*** becomes unstable. depicts the results obtained by ***S2*** using $50\%$ and $30\%$ of the original samples, respectively. It shows, that ***S2*** is more robust against small sample sizes and missing data than ***S1***. Generally, the error is only marginally affected by the type of the CS sampling matrix, i.e. (a) or (b). In all scenarios, ***S1*** and ***S2*** achieve a significantly lower error in estimating $\mathbf{s}$ than PDL-OIAI. At low SNRs, ***S1*** performs better than ***S2***, while ***S2*** becomes better at high SNRs. However, PDL-OIAI estimates $\theta$ with slightly higher accuracy than ***S1*** and ***S2***. When $50\%$ of the original samples are used, the error closely adheres to the RCRB. The amplitudes, $\x_{\mathcal{S}}$, are estimated with similar accuracy by both, ***S1*** and ***S2***, and no improvement is achieved compared to PDL-OIAI. Also, the distance to the RCRB is almost constant at all SNRs. Regarding the noise level, $\sigma_n^2$, ***S1*** yields a slightly smaller estimation error than ***S2***. PDL-OIAI does not provide a simple means for estimating $\sigma_n^2$, which is an advantage of ***S1*** and ***S2***. In the presented results for PDL-OIAI, it is assumed that pure noise samples are available to estimate $\sigma_n^2$. The instability of the RMSE between SNRs of 15 and 17.5 dB in Fig. \[fig:EM\_performance\_50\] might arise from averaging over an insufficient number of samples. It is also possible that the MCMC algorithm took longer to converge to the stationary distribution for SNR$=$20 dB, e.g. due to an unlucky initialization, thus, increasing the error.
Experimental Data
-----------------
To complete our study, we apply **S1** and **S2** to experimental data taken from the real fiber sensor system in [@Nakazaki2009; @Yamashita2009]. It was acquired at the Yamashita laboratory of photonic communication devices at The University of Tokyo, Japan. We consider $L\!=\!134$ original samples of the received sensor signal and use $M/L\!=\!50\%$ of the original samples. The delay spacing between the $N\!\!=\!\!2L$ dictionary atoms is $\delta t\approx 50$ ns. The sensing fiber contains $K=4$ FBGs and the delays of the reflected signals are potentially off-grid. Their positions are approximately at \[7.79, 9.05, 10.27, 12.30\] $\mu$s. We perform 100 Monte Carlo trials to estimate $\{\mathcal{S}, \x_\mathcal{S}, \theta\}$. shows the the original sensor signal and one estimated reflection from FBG$_3$. The shaded area indicates the standard deviation in estimating $\theta$. $\textbf{\emph{S1}}$ estimates a narrower reflection, which also results in slightly different estimates of $\mathcal{S}$. depicts $\hat{\x}_{\hat{\mathcal{S}}}$ at the estimated positions in $\hat{\mathcal{S}}$. The shaded areas represents the standard deviation for $\mathcal{S}$ and the vertical error bars indicate the standard deviation for $\x_{\mathcal{S}}$. Essentially, the results of **S1** and **S2** are comparable, although the variance in the estimates of $\mathcal{S}$ is marginally smaller in the case of $\textbf{\emph{S2}}$. Similar performance was reported for PDL-OIAI in [@Weiss2016].
Discussion {#sec:discussion}
==========
Based on simulations and experimental results, we demonstrate that the proposed sparsity model and our DL strategies, ***S1***, ***S2***, are useful in CFS and can be used for an automated estimation of the reflection delays. In comparison to PDL-OIAI in [@Weiss2016], where the underlying model treats $\x$ and $\theta$ as deterministic parameters, the following general observation can be made: The methods $\textbf{\emph{S1}}$ and $\textbf{\emph{S2}}$, based on a probabilistic sparse model, show comparable performance to PDL-OIAI but do not exceed the performance limit imposed by the non-Bayesian CRB. However, a significant improvement is achieved in estimating $\mathcal{S}$. It should be emphasized is of major importance in WDM-based CFS. It indicates the reflection delays, that are used to infer the quantity or nature of impairments at the FBGs. The amplitudes, $\x_{\mathcal{S}}$, can be used to determine the sparsity level and the amount of optical power reflected from the FBGs. We find that all competing methods estimate $\x_{\mathcal{S}}$ similarly accurate. The real data example shows that ***S1***, ***S2*** are insensitive to signal features that are not explicitly modeled, e.g. the skewness of the reflections or a signal-dependent noise amplitude. This was also reported for\
Our results for $\mathcal{S}$ indicate that the proposed sparsity model is better able to handle strong dictionary coherence than PDL-OIAI, which adopts a dictionary pre-processing routine to reduce the dictionary coherence. We ascribe this ability in part to the favorable selective shrinkage properties of the Weibull prior. Such behavior was previously reported for general heavy-tailed priors in [@Seeger2008; @Polson2010]. Regarding the relation to non-convex optimization, we find that constraints imposed on the $\ell_p$-norm, with $0\!<\!p\!<\!1$, are indeed useful in the presence of strong dictionary coherence. In this context, we support the findings in [@Chartrand2007; @Chartrand2008], that report relaxed RIP conditions when $\ell_1$-minimization is replaced by non-convex optimization methods. Another important factor, that contributes to the ability of handling strong dictionary coherence, is the local similarity model introduced in the joint prior density of $\x$. We observe much sparser solutions due to its collective shrinkage property, without the need for any dictionary pre-processing as in PDL-OIAI. Although this model is designed to deal with the unique features of the CFS dictionary, it can be used for general shift-invariant dictionaries with similar structures and high coherence levels. Therefore, it offers a broader applicability beyond the CFS problem. For parametric DL, all compared methods seem equally suitable for estimating $\theta$, but ***S2*** and PDL-OIAI are more stable for small sample sizes. Since the type of the CS matrix has only marginal impact, DF matrices in [@Achlioptas2003] are favorable. They are easy to implement, require low storage, and reduce the average sampling rate by 66%, since 2/3 of all projections are zero.\
The computational complexity of ***S1*** and ***S2*** is dominated by drawing samples from the posterior of $\x$ using HMC. HMC shows high efficacy in sampling this high-dimensional space in the presence of correlation. It yields samples from the desired posterior, that are weakly sparse with sharp peaks only close to the true positions of the significant components. Compared to optimization methods such as PDL-OIAI, MCMC is slower (c.f. [@Mohamed2012]) but some preliminary efforts are necessary for choosing a proper regularization parameter in the $\ell_1$-minimization problem. The run-time complexity of PDL-OIAI is dominated by a costly but essential data-dependent pre-processing routine to deal with severe dictionary coherence. This can be implemented using parallel processing and might be more efficient in situations, where CFS is used for permanent perturbation monitoring. Nonetheless, it initial estimate of the non-perturbed reference system. For this task, $\mathcal{S}$ can be more accurately estimated using ***S1*** and ***S2***. In contrast to PDL-OIAI, they are also able to estimate the noise level. Combining these methods for calibration and permanent monitoring is a promising perspective for practical systems. A limiting factor in ***S1*** and ***S2*** is the MCMC runtime, i.e. the number of available samples for Monte Carlo integration. Depending on the initial point, sufficient time has to be given for the algorithms to converge to the stationary distribution. Also, ***S1*** may get stuck in local optima but the proposed initialization using a bisectional search can lower this chance and helps to speeds up the convergence of the algorithm. A possible extension of this work can include multiple CS sample vectors to improve the SNR conditions. This might yield more accurate results and stable behavior. A similar technique was proposed in [@Malioutov2005]. Also, as pointed out in [@Weiss2016], additional local dictionary parameters can be considered. Since the reflections in the experimental data are non-uniform, this might improve both robustness and accuracy.
Conclusion {#sec:conclusion}
==========
We present a sparse estimation and parametric dictionary learning framework for Compressed Fiber Sensing (CFS) based on a probabilistic hierarchical sparse model. The significant components in the sparse signal indicate reflection delays, that can be used to infer the quantity and nature of external impairments. In order to handle severe dictionary coherence and to accomodate specific characteristics of the signal, a Weibull prior is employed to promote selective shrinkage. This choice can be related to non-convex optimization based on the To further alleviate the problem of dictionary coherence, we leverage the particular structure of the dictionary and assign a local variance to the differential sparse coefficients. This model can be useful for general shift-invariant dictionaries with similar structure and strong coherence. We propose two parametric dictionary learning strategies, ***S1*** and ***S2***, to estimate the dictionary parameter, $\theta$. In $\textbf{\emph{S1}}$, $\theta$ is treated as a deterministic parameter and estimated using a Monte Carlo EM algorithm. In $\textbf{\emph{S2}}$, a probabilistic hierarchical model for $\theta$ is considered. A hybrid MCMC method based on Gibbs sampling and Hamilton Monte Carlo is used for approximate inference. In simulations and by experimental data, we show the applicability and efficacy of the proposed sparse model, together with the methods ***S1*** and ***S2***, for an automated estimation of the reflection delays and the dictionary parameter in CFS. In a comparative analysis with an existing method, based on a deterministic sparse model, we highlight advantages, disadvantages and limitations, that can serve as a guidance to choose an adequate method for practical systems. To better assess the performance gain of a probabilistic sparse model, the Cram[é]{}r-Rao bound is derived for the joint estimation of deterministic sparse coefficients and the dictionary parameter in CFS. Drawbacks of the proposed methods are the generally high computational costs of MCMC methods, and the lack of simple diagnostic tools for Markov chain convergence and sample independence. Also, ***S1*** suffers from the problem of local optima. As a remedy, we propose a bisectional search to find a proper initialization. In subsequent investigations, multiple CS sample vectors and additional local dictionary parameters can be taken into account. Also, variational Bayes methods can be used to speed up computations.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported by the ’Excellence Initiative’ of the German Federal and State Governments and the Graduate School of Computational Engineering at Technische Universit[ä]{}t Darmstadt.\
The authors would like to thank Professor S. Yamashita and his group at The University of Tokyo, Japan, for kindly providing experimental data of the fiber sensor
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We consider in this paper some class of perturbation for the semilinear wave equation with subcritical (in the conformal transform sense) power nonlinearity. We first derive a Lyapunov functional in similarity variables and then use it to derive the blow-up rate. Though the result is similar to the unperturbed case in its statements, this is not the case of our method, which is new up to our knowledge.'
author:
- |
M. Hamza\
[*Faculté des Sciences de Tunis*]{}\
H. Zaag[^1]\
[*CNRS LAGA Université Paris 13*]{}
title: ' A Lyapunov functional and blow-up results for a class of perturbed semilinear wave equations'
---
[**Keywords:**]{} Wave equation, finite time blow-up, blow-up rate, perturbations.
[**AMS classification :**]{} 35L05, 35L67, 35B20.
**Introduction**
================
This paper is devoted to the study of blow-up solutions for the following semilinear wave equation: $$\label{1} \left\lbrace
\begin{array}{l}
u_{tt}=
\Delta u+ |u|^{p-1} u+f(u)+g(u_t),\qquad (x,t)\in \er^N\times \er_+^* \\
\\
(u(x,0),u_t(x,0))=(u_0(x),u_1(x))\in H^{1}_{loc,u}(\er^N)\times L^{2}_{loc,u}(\er^N),\\
\end{array}
\right.$$ where $p>1$, $f$ and $g$ are locally Lipschitz-continuous satisfying the following condition $$(H_f)\qquad |{f(x)}|\le M(1+|x|^q) \qquad{{\textrm {with}}}\ \ (q<p,\ \ M>0),$$ $$(H_g)\qquad \qquad \qquad |{g(x)}|\le M(1+|x|),\qquad\qquad$$ and $L^{2}_{loc,u}(\er^N)$ and $H^{1}_{loc,u}(\er^N)$ are the spaces defined by $$L^{2}_{loc,u}(\er^N)=\{u:\er^N\rightarrow \er/ \sup_{a\in \er^N}(\int_{|x-a|\le 1}|u(x)|^2dx)<+\infty \},$$ and $$H^{1}_{loc,u}(\er^N)=\{u\in L^{2}_{loc,u}(\er^N),|\grad u|\in L^{2}_{loc,u}(\er^N) \}.$$ We assume in addition that $$1<p<p_c\equiv 1+\frac{4}{N-1}.$$ The Cauchy problem of equation ([\[1\]]{}) is wellposed in $H^{1}_{loc,u}\times L^{2}_{loc,u}$. This follows from the finite speed of propagation and the the wellposdness in $H^{1} \times L^{2}$, valid whenever $ 1< p<1+\frac{4}{N-2}$ . The existence of blow-up solutions for the associated ordinary differential equation of (\[1\]) is a classical result. By using the finite speed of propagation, we conclude that there exists a blow-up solution $u(t)$ of (\[1\]). In this paper, we consider a blow-up solution $u(t)$ of (\[1\]), we define (see for example Alinhac [@A] and [@A1]) $\Gamma$ as the graph of a function $x\rightarrow T(x)$ such that $u$ cannot be extended beyond the set $$D_u=\{(x,t)\ \ \big |t<T(x)\}.$$ The set $D_u$ is called the maximal influence domain of $u$. Moreover, from the finite speed of propagation, $T$ is a $1$-Lipschitz function. Let $\overline{T}$ be the minimum of $T(x)$ for all $x\in \er^N$. The time $\overline{T}$ and the graph $\Gamma$ are called (respectively) the blow-up time and the blow-up graph of $u$.\
Let us first introduce the following non degeneracy condition for $\Gamma$. If we introduce for all $x\in\er^N$, $t\le T(x)$ and $\delta>0$, the cone
$$\label{cone}
C_{x,t,\delta}=\{(\xi,\tau)\neq(x,t)|0\le \tau\le t-\delta|\xi-x|\},$$
then our non degeneracy condition is the following: $x_0$ is a non characteristic point if $$\label{cone1}
\exists\delta_0=\delta_0(x_0)\in (0,1) \ \textrm{ such that}\ u \ \textrm{ is defined on }\ C_{x_0,T(x_0),\delta_0}.$$ We aim at studying the growth estimate of $u(t)$ near the space-time blow-up graph.
In the case $(f, g)\equiv (0,0)$, equation (\[1\]) reduces to the semilinear wave equation: $$\label{mu=0}
u_{tt}=\Delta u+ |u|^{p-1} u,\qquad (x,t)\in \er^N\times [0,\overline{T}).$$
Merle and Zaag in [@MZ3] (see also [@MZ1] and [@MZ2]) have proved, that if $u$ is a solution of (\[mu=0\]) with blow up graph $\Gamma:\{x\rightarrow T(x)\}$, then for all $x_0 \in \er^N$ and $t\in [\frac34T(x_0),T(x_0)]$, the growth estimate near the space-time blow-up graph satisfies $$\begin{aligned}
(T(x_0)-t)^{\frac{2}{p-1}}\frac{\|u(t)\|_{L^2(B(x_0,\frac{T(x_0)-t}{2}))}}{ (T(x_0)-t)^{\frac{N}{2}}}\qquad \qquad\qquad\qquad\qquad\qquad \qquad\qquad\qquad\\%\qquad\\
+ (T(x_0)-t)^{\frac{2}{p-1}+1}\!\Big (\frac{\|u_t(t)\|_{L^2(B(x_0,\frac{T(x_0)-t}{2}))}}{ (T(x_0)-t)^{\frac{N}{2}}}+
\frac{\|\grad u(t)\|_{L^2(B(x_0,\frac{T(x_0)-t}{2}))}}{ (T(x_0)-t)^{\frac{N}{2}}}\Big )\le K,\end{aligned}$$ where the constant $K$ depends only on $N,p,$ and on an upper bound on $T(x_0)$, $\frac1{T(x_0)}$ and the initial data in $ H^{1}_{loc,u}(\er^N)\times L^{2}_{loc,u}(\er^N)$. If in addition $x_0$ is non characteristic (in the sense (\[cone1\])), then for all $t\in [ \frac{3T(x_0)}4, T(x_0)]$,
$$\begin{aligned}
0<\varepsilon_0(N,p)\le (T(x_0)-t)^{\frac{2}{p-1}}\frac{\|u(t)\|_{L^2(B(x_0,{T(x_0)-t}))}}{ (T(x_0)-t)^{\frac{N}{2}}}
\qquad \qquad\qquad\qquad\qquad
\nonumber\\
+ (T(x_0)-t)^{\frac{2}{p-1}+1}\Big (\frac{\|u_t(t)\|_{L^2(B(x_0,{T(x_0)-t}))}}{ (T(x_0)-t)^{\frac{N}{2}}}+
\frac{\|\grad u(t)\|_{L^2(B(x_0,{T(x_0)-t}))}}{ (T(x_0)-t)^{\frac{N}{2}}}\Big )\le K,\end{aligned}$$
where the constant $K$ depends only on $N,p,$ and on an upper bound on $T(x_0)$, $\frac1{T(x_0)}$, $\delta_0(x_0)$ and the initial data in $ H^{1}_{loc,u}(\er^N)\times L^{2}_{loc,u}(\er^N)$.
Following this blow-up rate estimate, Merle and Zaag addressed the question of the asymptotic behavior of $u(x,t)$ near $\Gamma$ in one space dimension.
More precisely, they proved in [@MZ4] and [@MZ5] that the set of non charecteristic points ${\cal R}\subset \er$ is open and that $x\mapsto T(x)$ is of class $C^1$ on ${\cal R}$. They also described the blow-up profile of $u$ near $(x_0,T(x_0))$ when $x_0\in {\cal R}$.
In [@MZarxiv], they proved that $S=\er\backslash {\cal R}$ has an empty interior and that $\Gamma$ is a corner of angle $\frac{\pi}{2}$ near any $x_0\in S$. They also showed that $u(x,t)$ decomposes in a sum of decoupled solitons near $(x_0,T(x_0))$.
Our aim in this work is to generalize the blow-up rate estimate obtained for equation (\[mu=0\]) in [@MZ1] and [@MZ3] in the subcritical case $(p<p_c)$ to equation (\[1\]).
One may think that such a generalization is straightforward and only technical. In fact, that opinion may be valid for all the steps, except for the very first one, that is, the existence of a Lyapunov functional in similarity variables which is far from being trivial. That functional is our main contribution. The existence of the Lyapunov functional is a crucial step towards the derivation of blow-up results for equation (\[1\])
As in [@MZ1] and [@MZ2], we want to write the solution $v$ of the associate ordinary differential equation of (\[1\]). It is clear that $v$ is given by $$v''= v^{p} +f(v)+g(v'),\qquad v({T})=+\infty,$$ and satisfies: $$\label{kappa}
v(t)\sim \frac{\kappa}{ ({T}-t)^{\frac{2}{p-1}}}\ \ \textrm{as}\ t\rightarrow {T},
\ \ \textrm{where}\ \kappa=\Big (\frac{2p+2}{(p-1)^2}\Big )^{\frac{1}{p-1}}.$$ For this reason, we define for all $x_0\in \er^N$, $0< T_0\le T_0(x_0)$, the following similary transformation introduced in Antonini and Merle [@AM] and used in [@MZ1],[@MZ2] and [@MZ3]: $$\label{scaling}
y=\frac{x-x_0}{T_0-t},\qquad s=-\log (T_0-t),\qquad u(x,t)=\frac{1}{(T_0-t)
^{\frac{2}{p-1}}}w_{x_0,T_0}(y,s).$$ The function $w_{x_0,T_0}$ (we write $w$ for simplicity) satisfies the following equation for all $y\in B$ and $s\ge -\log T_0$: $$\begin{aligned}
\label{B}
w_{ss}&=&\frac{1}{\rho}\div(\rho \grad w-\rho(y.\grad w)y)
-\frac{2p+2}{(p-1)^2}w+|w|^{p-1}w-\frac{p+3}{p-1}w_s-2y.\grad w_{s}\nonumber\\
&&+e^{-\frac{2ps}{p-1}}f\Big(e^{\frac{2s}{p-1}}w\Big) +
e^{-\frac{2ps}{p-1}}g\Big(e^{\frac{(p+1)s}{p-1}}(w_s+y.\grad w+\frac{2}{p-1}w)\Big),\qquad\end{aligned}$$ where $\rho=(1-|y|^2)^{\alpha}$ and $\alpha={\frac{2}{p-1}-\frac{N-1}{2}}>0.$
In the new set of variables $(y,s),$ the behavior of $u$ as $t \rightarrow T_0$ is equivalent to the behavior of $w$ as $s \rightarrow +\infty$.
\
We remark that the corresponding terms of the fonctions $f(u)$ and $g(u)$ in the problem (\[B\]) satisfy the following inequalities, for all $s\ge 0$, $$\begin{aligned}
e^{-\frac{2ps}{p-1}}\left|f\left(e^{\frac{2s}{p-1}}w\right)\right| &\le&
CMe^{-\frac{2ps}{p-1}}
+CMe^{-\frac{2(p-q)s}{p-1}}|w|^q\\
&\le& CMe^{-\frac{2(p-q)s}{p-1}}
+CMe^{-\frac{2(p-q)s}{p-1}}|w|^p
,\end{aligned}$$ and $$\begin{aligned}
e^{-\frac{2ps}{p-1}}g\left|\left(e^{\frac{(p+1)s}{p-1}}(w_s+y.\grad w+\frac{2}{p-1}w)\right)\right|\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\\
\le
CMe^{-\frac{2ps}{p-1}}
+CMe^{-s}\left|w_s+y.\grad w+\frac{2}{p-1}w\right|.\end{aligned}$$ For this reason, we can see that in the variable $(y,s)$ the problem (\[B\]) is a perturbation of the particular case where $(f,g)\equiv(0,0)$, when $s\rightarrow +\infty$.
The equation (\[B\]) will be studied in the space $\cal H$ $${\cal H}=\Big \{(w_1,w_2), |
\displaystyle\int_{B}\Big ( w_2^2
+|\grad w_1|^2(1-|y|^2)+w_1^2\Big )\rho {\mathrm{d}}y<+\infty \Big \}.$$ In the whole paper, we denote $\ds{F(u)=\int_0^uf(v)dv}.$\
In the case $(f,g)\equiv(0,0)$, Antonini and Merle [@AM] proved that $$\begin{aligned}
\label{f}
E_0(w)=\displaystyle\int_{B}\Big ( \frac{1}{2}w_s^2
+\frac{1}{2}|\grad w|^2-\frac{1}{2}(y.\grad w)^2+\frac{p+1}{(p-1)^2}w^2
-\frac{1}{p+1}|w|^{p+1}\Big )\rho {\mathrm{d}}y,\end{aligned}$$ is a Lyapunov functional for equation (\[B\]). When $(f,g) \not\equiv (0,0)$, we introduce $$\begin{aligned}
\label{f1}
H(w)&=&E(w)e^{\frac{ p+3}{2\gamma} e^{-\gamma s}}+\theta e^{-2\gamma s},\end{aligned}$$ where $\theta$ is a sufficiently large constant that will be determined later, $$\begin{aligned}
\label{f2}
E(w)&=&E_0(w)+I(w)+J(w),\quad
I(w)=- e^{-\frac{2(p+1)s}{p-1}}\displaystyle\int_{B}F(e^{\frac{2}{p-1}s}w)\rho {\mathrm{d}}y,\nonumber\\
{\textrm {and}} \quad
J(w)&=& -e^{-\gamma s}\displaystyle\int_{B}ww_s\rho {\mathrm{d}}y \qquad {\textrm {with}} \qquad \gamma=\min(\frac12,
\frac{p-q}{p-1}
)>0.%\qquad\qquad\qquad\qquad\end{aligned}$$ We now claim that the functional $H(w)$ is a decreasing function of time for equation $(\ref{B})$, provided that $s$ is large enough.
Here we announce our main result.
(Existence of a Lyapunov functional for equation ([\[B\]]{}))\
\[lyap\] Let $N,p,q$ and $M$ be fixed. There exists $S_0=S_0(N,p,q,M)\in \er$ such that, for all $ s_0\in \er$ and $w$ solution of equation (\[B\]) satisfying $(w,w_s)\in
{\cal
C}([s_0,+\infty),{\cal H})$, it holds that $H$ satisfies the following inequality, for all $s_2>s_1\ge \max(s_0,S_0)$, $$\begin{aligned}
\label{L1}
H(w(s_2))-H(w(s_1))
&\le& -\alpha \int_{s_1}^{s_2}\int_{B}w_s^2(y,s)\frac{\rho}{1-|y|^2}dy ds, \qquad \qquad\end{aligned}$$ where $\alpha=\frac{2}{p-1}-\frac{N-1}2$.
1. One may wander why we take only sublinear perturbations in $u_t$ (see hypothesis ($H_g$)). It happens that any superlinear terms in $u_t$ generates in similarity variables $L^r$ norms of $w_s$ and $\grad w$, where $r>2$, hence, non controllable by the terms in the Lyapunov functional $E_0(w)$ (\[f\]) of the non perturbed equation (\[mu=0\]).
2. Our method breaks down in the critical case $p=p_c$, since in the energy estimates in similarity variables, the perturbations terms are integrated on the whole unit ball, hence, difficult to control with the dissipation of the non perturbed equation (\[mu=0\]), which degenerates to the boundary of the unit ball.
As we said earlier, the existence of this Lyapunov functional (and a blow-up criterion for equation (\[B\]) based in $H$, see Lemma 2.3 below) are a crucial step in the derivation of the blow-up rate for equation (\[1\]). Indeed, with the functional $H$ and some more work, we are able to adapt the analysis performed in [@MZ3] for equation (\[mu=0\]) and get the following result:
(blow-up rate for equation ([\[1\]]{}))\
\[t\] Let $N,p,q$ and $M$ be fixed. Then, there exist $\widehat{S}_0=\widehat{S}_0(N,p,q,M)\in \er$ and $\varepsilon_0=\varepsilon_0(N,p,M)$, such that if $u $ is a solution of ([\[1\]]{}) with blow-up graph $\Gamma:\{x\rightarrow T(x)\}$ and $x_0$ is a non characteristic point, then
i\) For all $s\ge \widehat{s}_0(x_0)=\max(\widehat{S}_0(N,p,q,M),-\log \frac{T(x_0)}4)$, $$0<\varepsilon_0\le \|w_{x_0,T(x_0)}(s)\|_{H^{1}(B)}+
\|\partial_s w_{x_0,T(x_0)}(s)\|_{L^{2}(B)}
\le K,$$ where $w_{x_0,T(x_0)}$ is defined in (\[B\]) and $B$ is the unit ball of $\er^N$.\
ii) For all $t\in [t_0(x_0),T(x_0))$, where $t_0(x_0)=\max (T(x_0)-e^{-\widehat{s}_0(x_0)},\frac{3T(x_0)}4)$, we have
$$\begin{aligned}
&&0<\varepsilon_0\le (T(x_0)-t)^{\frac{2}{p-1}}\frac{\|u(t)\|_{L^2(B(x_0,{T(x_0)-t}))}}{ (T(x_0)-t)^{\frac{N}{2}}}\nonumber\\
&&+ (T(x_0)-t)^{\frac{2}{p-1}+1}\Big (\frac{\|u_t(t)\|_{L^2(B(x_0,{T(x_0)-t}))}}{ (T(x_0)-t)^{\frac{N}{2}}}+
\frac{\|\grad u(t)\|_{L^2(B(x_0,{T(x_0)-t}))}}{ (T(x_0)-t)^{\frac{N}{2}}}\Big )\le K,\end{aligned}$$
where $K=K(N,p,q,\widehat{s}_0(x_0),\|(u(t_0(x_0)),u_t(t_0(x_0)))\|_{
H^{1}\times L^{2}(B(x_0,\frac{e^{-\widehat{s}_0(x_0)}}{\delta_0(x_0)})
)})$ and\
$\delta_0(x_0)\in (0,1)$ is defined in (\[cone1\]).
With this blow-up rate, one can ask whether the results proved by Merle and Zaag for the non perturbed problem in [@MZ4] [@MZ5] [@MZarxiv], hold for equation (\[1\]) (blow-up, profile, regularity of the blow-up graph, existence of characteristic points, etc...). We believe that it is the case, however, the proof will be highly technical, with no interesting ideas (in particular, equation (\[1\]) is not conserved under the Lorentz transform, which is crucial in [@MZ4] [@MZ5] [@MZarxiv], and lots of minor term will appear in the analysis). Once again, we believe that the key point in the analysis of blow-up for equation (\[1\]) is the derivation of a Lyapunov functional in similarity variables, which is the object of our paper.
As in the particular case where $(f,g)\equiv (0,0)$, the proof of Theorem \[t\] relies on four ideas (the existence of a Lyapunov functional, interpolation in Sobolev spaces, some Gagliardo-Nirenberg estimates and a covering technique adapted to the geometric shape of the blow-up surface). It happens that adapting the proof of [@MZ3] given in the non perturbed case (\[mu=0\]) is straightforward, except for a key argument, where we bound the $L^{p+1}$ space-time norm of $w$. Therefore, we only present that argument, and refer to [@MZ1] and [@MZ3] for the rest of the proof.
This paper is divided in two sections, each of them devoted to the proof of a Theorem.
A Lyapunov functional for equation (\[B\])
===========================================
This section is divided in two parts:
- We first prove the existence of a Lyapunov functional for equation (\[B\]).
- Then, we give a blow-up criterion for equation (\[B\]) based on the Lyapunov functional.
Throughout this section, we consider $(w,w_s)\in
{\cal
C}([s_0,+\infty),{\cal H})$ where $w$ is a solution of (\[B\]) and $s_0\in \er$. We aim at proving that the functional $H$ defined in (\[f1\]) is a Lyapunov functional for equation (\[B\]), provided that $s\ge S_0$, for some $S_0=S_0(N,p,q,M)$. We denote by $C$ a constant which depends only on $(p,q,N,M )$. We denote the unit ball of $\er^N$ by $B$.
Existence of a Lyapunov functional
-----------------------------------
\[lyap0\] Let $N,p,q$ and $M$ be fixed. There exists $S_1=S_1(N,p,q,M)\in \er$ such that, for all $ s_0\in \er$ and $w$ solution of equation (\[B\]) satisfying $(w,w_s)\in
{\cal
C}([s_0,+\infty),{\cal H})$, we have the following inequality, for all $s\ge \max (s_0,S_1)$, $$\begin{aligned}
\label{lem21}
\frac{d}{ds}(E_0(w)+I (w))\!\!\!\!&\le&\!\!\!\!-\frac{3\alpha}{2}\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y
+\Sigma_ {0}(s),\end{aligned}$$ where $\Sigma_0$ satisfies $$\begin{aligned}
%\label{sigm}
\Sigma_ {0}(s) &\le&\!\!\!\!C e^{-2s}\int_{B}|\grad w|^2(1-|y|^2)\rho{\mathrm{d}}y+C e^{-2s}\int_{B}w^2\rho{\mathrm{d}}y\\
%\nonumber\\
&&+
Ce^{-\frac{2(p-q)s}{p-1}}\int_{B}\!\!|w|^{p+1}\rho {\mathrm{d}}y
+ C e^{-\frac{2(p-q)s}{p-1}}.\qquad \qquad\end{aligned}$$
Proof: Multipling $(\ref{B})$ by $w_s\rho$, and integrating over the ball $B$, we obtain, for all $s\ge s_0$, (recall from [@AM] that in the case where, $(f,g)\equiv (0,0)$, we have\
$\frac{d}{ds}E_0(w)=-2\alpha \int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y$.) $$\begin{aligned}
\label{E0}
\frac{d}{ds}(E_0(w)+I (w))&=&-2\alpha \int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y+
\underbrace{\frac{2(p+1)}{p-1}e^{-\frac{2(p+1)s}{p-1}}\int_{B}F\Big(e^{\frac{2s}{p-1}}w\Big)\rho {\mathrm{d}}y}_{I_1} \nonumber\\
&&+\underbrace{\frac{2}{p-1}e^{-\frac{2ps}{p-1}}\int_{B}f\Big(e^{\frac{2s}{p-1}}w\Big)w\rho {\mathrm{d}}y }_{I_2}\nonumber\\
&&+\underbrace{ e^{-\frac{2ps}{p-1}}\int_{B}g\Big(e^{\frac{(p+1)s}{p-1}}(w_s+y.\grad w+\frac{2}{p-1}w)\Big)w_s\rho {\mathrm{d}}y}_{I_3}.\end{aligned}$$ By exploiting the fact that $ |{F(x)}|+|x{f(x)}|\le C( 1+|x|^{q+1}),$ we obtain $$\begin{aligned}
\label{I10}
|I_1|+|I_2|&\le& Ce^{-\frac{2(p+1)s}{p-1}}\int_{B}(1+|e^{\frac{2s}{p-1}}w|^{q+1})\rho {\mathrm{d}}y\nonumber\\
&\le& Ce^{-\frac{2(p+1)s}{p-1}}+ Ce^{-\frac{2(p-q)s}{p-1}}\int_{B}|w|^{q+1}\rho {\mathrm{d}}y.\end{aligned}$$ Noticing that $|x|^{q+1}\le C( 1+|x|^{p+1})$, we deduce from (\[I10\]) that for all $s \ge \max (s_0,0)$, $$\begin{aligned}
\label{I100}
|I_1|+|I_2|\le Ce^{-\frac{2(p-q)s}{p-1}}+ Ce^{-\frac{2(p-q)s}{p-1}}\int_{B}|w|^{p+1}\rho {\mathrm{d}}y.\end{aligned}$$ Since $ |g(x)|\le M( 1+|x|)$, we write $$\begin{aligned}
\label{I122}
|I_3|&\le& Ce^{-s}
\int_{B}w_s^2\rho {\mathrm{d}}y +
Ce^{-s}\int_{B}|y.\grad w ||w_s|\rho {\mathrm{d}}y\nonumber\\
&&+ C e^{-s}\int_{B}|ww_s|\rho {\mathrm{d}}y+ C e^{-\frac{2ps}{p-1}}\int_{B}|w_s|\rho {\mathrm{d}}y.\qquad\end{aligned}$$ By exploiting the inequality $ab\le \frac{a^2}{2}+\frac{b^2}{2},$ we obtain $$\begin{aligned}
\label{I13}
C e^{-s}\int_{B}|y.\grad w| | w_s|\rho{\mathrm{d}}y\le
\frac{\alpha}{8}\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y
+C e^{-2s}\int_{B}|\grad w|^2(1-|y|^2)\rho{\mathrm{d}}y.\end{aligned}$$ Similarly, we prove that $$\label{I14}
C e^{-s}\int_{B}|ww_s|\rho {\mathrm{d}}y\le \frac{\alpha}{8}\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y
+C e^{-2s}\int_{B}w^2\rho{\mathrm{d}}y.$$ We infer from the inequality $ |a|\le 1+a^2$ that $$\label{I15}
C e^{-\frac{2ps}{p-1}}\int_{B}|w_s|\rho {\mathrm{d}}y\le C e^{-\frac{2ps}{p-1}}+C e^{-\frac{2ps}{p-1}}\int_{B}w_s^2\frac{\rho}{1-y^2} {\mathrm{d}}y.$$ Combining (\[I122\]), (\[I13\]), (\[I14\]) and (\[I15\]), we conclude that, for all $s \ge \max (s_0,0)$, $$\begin{aligned}
\label{I16}
|I_3|&\le& (Ce^{-s}+
\frac{\alpha}{4})\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y+C e^{-2s}\int_{B}w^2\rho{\mathrm{d}}y\nonumber\\
&&+C e^{-2s}\int_{B}|\grad w|^2(1-|y|^2)\rho{\mathrm{d}}y+ C e^{-\frac{2ps}{p-1}}.\end{aligned}$$ Then, by using (\[E0\]), (\[I100\]) and (\[I16\]), we deduce that, for all $s \ge \max (s_0,0)$, $$\begin{aligned}
\label{E}
\frac{d}{ds}(E(w)+I (w))\!\!\!\!&\le&\!\!\!\!(-\frac{7\alpha}{4}+Ce^{-s}
)\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y
+C e^{-2s}\int_{B}|\grad w|^2(1-|y|^2)\rho{\mathrm{d}}y
\nonumber\\
&&+C e^{-2s}\int_{B}w^2\rho{\mathrm{d}}y+ Ce^{-\frac{2(p-q)s}{p-1}}\int_{B}\!\!|w|^{p+1}\rho {\mathrm{d}}y+C e^{-\frac{2(p-q)s}{p-1}}.%\qquad\qquad\end{aligned}$$ Taking $ S_1=S_1(N,p,q,M)$ large enough, we have the estimate (\[lem21\]). This concludes the proof of Lemma \[lyap0\].
We are now going to prove the following estimate for the functional $J$:
\[lyap1\] Let $N,p,q$ and $M$ be fixed. There exists $S_2=S_2(N,p,q,M)\in \er$ such that, for all $ s_0\in \er$ and $w$ solution of equation (\[B\]) satisfying $(w,w_s)\in
{\cal
C}([s_0,+\infty),{\cal H})$, $J$ satisfies the following inequality, for all $s\ge \max (s_0,S_2)$: $$\begin{aligned}
\label{theta1}
\frac{d}{ds}J(w)&\le &\frac{\alpha}{2}
\!\int_{B}\!w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y+\frac{p+3}2 e^{-\gamma s}E(s)\nonumber\\
&&-\frac{p-1}{4} e^{-\gamma s}\int_{B}(|\grad w|^2(1-|y|^2)\rho{\mathrm{d}}y\\
&&
- \frac{p+1}{2(p-1)}e^{-\gamma s}\int_{B}w^2\rho{\mathrm{d}}y
-\frac{p-1}{2(p+1)} e^{-\gamma s}\int_{B}|w|^{p+1}\rho{\mathrm{d}}y
+\Sigma_ {1}(s),\nonumber\end{aligned}$$ where $\gamma =\min(\frac12,\frac{p-q}{p-1})>0$ and $\Sigma_1(s)$ satisfies $$\begin{aligned}
\label{sigm}
\Sigma_ 1(s) &\le&
C e^{-2\gamma s}\int_{B}w^2\rho
{\mathrm{d}}y+
+C e^{-2\gamma s}\int_{B}\!|\grad w|^2\rho(1-|y|^2)
{\mathrm{d}}y
\nonumber\\
&&+C e^{-2\gamma s}\int_{B}| w|^{p+1}\rho{\mathrm{d}}y+ C e^{-2\gamma s}.\end{aligned}$$
Proof: Note that $J$ is a differentiable function for all $s\ge s_0$ and that $$\begin{aligned}
\frac{d}{ds}J(w)&=&\gamma e^{-\gamma s}\int_{B}ww_{s}\rho{\mathrm{d}}y- e^{-\gamma s}\int_{B}w^2_{s}\rho{\mathrm{d}}y- e^{-\gamma s}\int_{B}
ww_{ss}\rho{\mathrm{d}}y.\end{aligned}$$ By using equation (\[B\]) and integrating by parts, we have $$\begin{aligned}
\label{t1}
\frac{d}{ds}J(w)&=&
- e^{-\gamma s}\int_{B}w^2_{s}\rho{\mathrm{d}}y
+ e^{-\gamma s}\int_{B}(|\grad w|^2-(y.\grad w)^2)\rho{\mathrm{d}}y\nonumber\\
&&+ \frac{2p+2}{(p-1)^2}e^{-\gamma s}\int_{B}w^2\rho{\mathrm{d}}y- e^{-\gamma s}\int_{B}|w|^{p+1}\rho{\mathrm{d}}y\nonumber\\
&&+
(\gamma +\frac{p+3}{p-1}-2N) e^{-\gamma s}\int_{B}ww_s \rho{\mathrm{d}}y-2e^{-\gamma s}\int_{B}ww_s(y.\grad \rho){\mathrm{d}}y\nonumber\\
&&-2e^{-\gamma s}\int_{B}w_s(y.\grad w) \rho{\mathrm{d}}y-e^{-\frac{2ps}{p-1}-\gamma s}\int_{B}wf\Big(e^{\frac{2s}{p-1}}w\Big){\rho}{\mathrm{d}}y\\
&& -
e^{-\frac{2ps}{p-1}-\gamma s}\int_{B}wg\Big(e^{\frac{(p+1)s}{p-1}}(w_s+y.\grad w+\frac{2}{p-1}w)\Big){\rho}{\mathrm{d}}y.\nonumber\end{aligned}$$ By combining (\[f\]), (\[f2\]) and (\[t1\]), we write $$\begin{aligned}
\label{theta}
\frac{d}{ds}J(w)&\le &\frac{p+3}2 e^{-\gamma s}E(s)
-\frac{p-1}{4} e^{-\gamma s}\int_{B}(|\grad w|^2-(y.\grad w)^2)\rho{\mathrm{d}}y\nonumber\\
&&- \frac{p+1}{2(p-1)}e^{-\gamma s}\int_{B}w^2\rho{\mathrm{d}}y
-\frac{p-1}{2(p+1)} e^{-\gamma s}\int_{B}|w|^{p+1}\rho{\mathrm{d}}y\nonumber\\
&&+
\underbrace{(\gamma +\frac{p+3}{p-1}-2N+\frac{p+3}2 e^{-\gamma s}) e^{-\gamma s}\int_{B}ww_s \rho{\mathrm{d}}y}_{J_1}\nonumber\\
&&\underbrace{-2e^{-\gamma s}\int_{B}ww_s(y.\grad \rho){\mathrm{d}}y}_{J_2}
\underbrace{-2e^{-\gamma s}\int_{B}w_s(y.\grad w) \rho{\mathrm{d}}y}_{J_3}\nonumber\\
&&\underbrace{-e^{-\frac{2ps}{p-1}-\gamma s}\int_{B}wf\Big(e^{\frac{2s}{p-1}}w\Big){\rho}{\mathrm{d}}y}_{J_4}\nonumber\\
&&\underbrace{ -
e^{-\frac{2ps}{p-1}-\gamma s}\int_{B}wg\Big(e^{\frac{(p+1)s}{p-1}}(w_s+y.\grad w+\frac{2}{p-1}w)\Big){\rho}{\mathrm{d}}y}_{J_5}\nonumber\\
&&+
\underbrace{\frac{(p+3)}2 e^{-\frac{2(p+1)s}{p-1}-\gamma s}\displaystyle\int_{B}F(e^{\frac{2}{p-1}s}w)\rho {\mathrm{d}}y}_{J_6}.\end{aligned}$$ We now study each of the last five terms. To estimate $J_1$, we use the fact that for all $s\ge \max(s_0,0)$, $$\label{es}
\big |\gamma +\frac{p+3}{p-1}-2N+\frac{p+3}2 e^{-\gamma s}\big |\le C.$$ By using (\[es\]) and the Cauchy-Schwartz inequality we obtain $$\begin{aligned}
\label{J1}
|J_1|& \le& C e^{-\gamma s}
\int_{B}|ww_s| \rho{\mathrm{d}}y
\le
\frac{\alpha}{8}\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y+
C e^{-2\gamma s}\int_{B}w^2\rho
{\mathrm{d}}y.\qquad\qquad\end{aligned}$$ Now we estimate the expression $J_2$. Since we have $\ds{y.\grad \rho=-2\alpha \frac{|y|^2}{(1-|y|^2)}\rho,}$ we can use the Cauchy-Schwartz inequality to write $$\begin{aligned}
\label{J2}
|J_2|& \le& C e^{-\gamma s}
\int_{B}|w_s|(1-|y|^2)^{\frac{\alpha-1}{2}}|w||y|(1-|y|^2)^{\frac
{\alpha-1}{2}}{\mathrm{d}}y\nonumber\\
&\le&\frac{\alpha}{8}\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y+C e^{-2\gamma s}\int_{B}w^2\frac{|y|^2\rho}{1-|y|^2}
{\mathrm{d}}y.\end{aligned}$$ Since we have the following Hardy type inequality for any $w\in
H^{1}_{loc,u}(\er^N)$ (see appendix $B$ in [@MZ1] for details): $$\begin{aligned}
\label{hardyJJ}
\int_{B}w^2\frac{|y|^2\rho}{1-|y|^2}{\mathrm{d}}y
&\le&C\int_{B}|\grad w|^2\rho(1-|y|^2){\mathrm{d}}y+C \int_{B}w^2\rho
{\mathrm{d}}y,\end{aligned}$$ we use (\[J2\]) and (\[hardyJJ\]) to conclude that $$\label{J22}
|J_2|\le\frac{\alpha}{8}\!\int_{B}\!w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y+C
e^{-2\gamma s}\int_{B}\!w^2\rho
{\mathrm{d}}y
+C e^{-2\gamma s}\int_{B}\!|\grad w|^2\rho(1-|y|^2)
{\mathrm{d}}y.\quad$$ By using the Cauchy-Schwartz inequality, we have $$\begin{aligned}
\label{J3}
|J_3|&\le&\frac{\alpha}{8}\int_
{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y
+C e^{-2\gamma s}\int_{B}|\grad w|^2\rho(1-|y|^2){\mathrm{d}}y.\end{aligned}$$ By exploiting the fact that $ |{F(x)}|\le CM( 1+|x|^{q+1})$ and $ |{f(x)}|\le M(1+|x|^q),$ we write $$\begin{aligned}
\label{J4}
|J_4|+|J_6|&\le&
C e^{-\frac{2(p-q)s}{p-1}-\gamma s}\int_{B}(1+| w|^{q+1})\rho{\mathrm{d}}y
\le C e^{-2\gamma s}\int_{B}(1+|w|^{p+1})\rho{\mathrm{d}}y\nonumber\\
&\le&C e^{-2\gamma s}+
C e^{-2\gamma s}\int_{B}|w|^{p+1}\rho{\mathrm{d}}y.\end{aligned}$$ In a similar way, by using the fact that $ |g(x)|\le M(1+|x|)$, we write $$\begin{aligned}
\label{J5}
|J_5|&\le& Ce^{-2\gamma s}
\int_{B}w_s^2\rho {\mathrm{d}}y +
Ce^{-2\gamma s}\int_{B}|y.\grad w ||w|\rho {\mathrm{d}}y\nonumber\\
&&+ C e^{-2\gamma s}\int_{B}w^2\rho {\mathrm{d}}y+ C e^{-2\gamma s}.\qquad\end{aligned}$$ Then, by (\[hardyJJ\]), we have $$\begin{aligned}
\label{J55}
|J_5|&\le& Ce^{-2\gamma s}
\int_{B}w_s^2\frac{\rho}{1-|y|^2} {\mathrm{d}}y +C e^{-2\gamma s}\int_{B}|\grad w|^2\rho(1-|y|^2)
{\mathrm{d}}y\nonumber\\
&& +Ce^{-2\gamma s}\int_{B}w^2\rho {\mathrm{d}}y+ C e^{-2\gamma s}.\end{aligned}$$ Finally, by using (\[theta\]), (\[J1\]), (\[J22\]), (\[J3\]), (\[J4\]) and (\[J5\]) we deduce that $$\begin{aligned}
%\label{theta11}
\frac{d}{ds}J(w)&\le &\frac{p+3}2 e^{-\gamma s}E(s)
-\frac{p-1}{4} e^{-\gamma s}\int_{B}(|\grad w|^2-(y.\grad w)^2)\rho{\mathrm{d}}y\nonumber\\
&&- \frac{p+1}{2(p-1)}e^{-\gamma s}\int_{B}w^2\rho{\mathrm{d}}y
-\frac{p-1}{2(p+1)} e^{-\gamma s}\int_{B}|w|^{p+1}\rho{\mathrm{d}}y\nonumber\\
&&+
C e^{-2\gamma s}\int_{B}w^2\rho
{\mathrm{d}}y+
+C e^{-2\gamma s}\int_{B}\!|\grad w|^2\rho(1-|y|^2)
{\mathrm{d}}y\nonumber\\
&&+(
\frac{3\alpha}{8}+Ce^{-2\gamma s})\!\int_{B}\!w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y+
C e^{-2\gamma s}+C e^{-2\gamma s}\int_{B}| w|^{p+1}\rho{\mathrm{d}}y.\end{aligned}$$ Since $|y.\grad w|\le |y||\grad w|$, it follows that $$\begin{aligned}
\label{last}
\int_{B}|\grad w|^2\rho(1-|y|^2){\mathrm{d}}y&\le &\int_{B}(|\grad w|^2-(y.\grad w)^2) \rho{\mathrm{d}}y.\end{aligned}$$ Taking $S_2=S_2(N,p,q,M)$ large enough, we have easily the estimate (\[theta1\]) and (\[sigm\]). This concludes the proof of Lemma \[lyap1\].
With Lemmas \[lyap0\] and \[lyap1\], we are in a position to prove Theorem \[lyap\].
\
From Lemmas \[lyap0\] and \[lyap1\], we obtain for all $s\ge \max(s_0, S_1, S_2)$, $$\begin{aligned}
\label{E11}
\frac{d}{ds}E(w)&\le &C e^{-2\gamma s}+\frac{p+3}2 e^{-\gamma s}E(w)
-\alpha\!\int_{B}\!w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y\nonumber\\
&&+(C e^{-\gamma s}-\frac{p-1}{4}) e^{-\gamma s}\int_{B}(|\grad w|^2(1-|y|^2)\rho{\mathrm{d}}y\nonumber\\
&&+(C e^{-\gamma s}
- \frac{p+1}{2(p-1)})e^{-\gamma s}\int_{B}w^2\rho{\mathrm{d}}y\nonumber\\
&&+(C e^{-\gamma s}-\frac{p-1}{2(p+1)}) e^{-\gamma s}\int_{B}|w|^{p+1}\rho{\mathrm{d}}y.
%,\quad \forall s\ge s_0.\end{aligned}$$ We now choose $S_0\ge \max(S_1,S_2)$, large enough, so that for all $s\ge S_0$, we have $$\begin{aligned}
\frac{p-1}{4}-C e^{-\gamma s} \ge 0, \qquad
\frac{p+1}{2(p-1)}
-C e^{-\gamma s} \ge 0,\quad \frac{p-1}{2(p+1)}-C e^{-\gamma s}\ge 0.\end{aligned}$$ Then, we deduce that, for all $s\ge \max(S_0,s_0)$, we have $$\begin{aligned}
\label{E111}
\frac{d}{ds}E(w)&\le & C e^{-2\gamma s}+\frac{p+3}2 e^{-\gamma s}E(w)-\alpha \int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y.\end{aligned}$$ Finally, we prove easily that the function $H$ satisfies, for all $s\ge \max(S_0,s_0)$, $$\begin{aligned}
\label{H01}
\frac{d}{ds}H(w)&\le & (Ce^{\frac{ p+3}{2\gamma} e^{-\gamma s}}-2\theta \gamma ) e^{-2\gamma s}-\alpha e^{\frac{ p+3}{2\gamma} e^{-\gamma s}}\int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y\nonumber\\
&\le & (C-2\theta \gamma ) e^{-2\gamma s}-\alpha \int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y.\end{aligned}$$ We now choose $\theta$ large enough, so we have $C-2\theta \gamma \le 0$ and then $$\begin{aligned}
\label{H02}
\frac{d}{ds}H(w)
&\le & -\alpha \int_{B}w_s^2\frac{\rho}{1-|y|^2}{\mathrm{d}}y.\end{aligned}$$ Now (\[L1\]) is a direct consequence of inequality (\[H02\]).\
This concludes the proof of Theorem \[lyap\].
A blow-up criterion in the $w(y,s)$ variable
---------------------------------------------
We now claim the following proposition:
\[L02\] Let $N,p,q,M$ be fixed. There exists $S_3=S_3(N,p,q,M)\ge S_0$ such that, for all $ s_0\in \er$ and $w$ solution of equation (\[B\]) defined to the left of $s_0$, such that $\|w(s)\|_{L^{p+1}(B)}$ is locally bounded, if $H(w(s_3))<0$ for some $s_3\ge \max(S_3,s_0)$, then $w$ blows up in some finite time $S>s_3$.
\
If $w=w_{x_0,T_0}$ defined from a solution of (\[1\]) by (\[scaling\]) and $x_0$ is non characteristic point, then $\|w(s)\|_{H^1(B)}$ is locally bounded and so is $\|w(s)\|_{L^{p+1}(B)}$ by Sobolev’s embedding.
Proof: The argument is the same as in the corresponding part in [@AM]. We write the proof for completeness. Arguing by contradiction, we assume that there exists a solution $w$ on $B$, defined for all time $s\in [s_3,+\infty[$, where $H(w(s_3))<0$. Since the energy $H$ decreases in time, we have $H(w(1+s_3))<0$.
Consider now for $\delta>0$ the function $\widetilde{w}^{\delta}(y,s)$ for $(y,s)\in B\times [1+s_3,+\infty[$ defined by $$\forall s\ge 1+s_3, \forall y\in B,
\widetilde{w}^{\delta}(y,s)=\frac{1}{(1+\delta e^s)^{\frac2{p-1}}} w(
\frac{y}{1+\delta e^s},-\log(\delta+e^{-s})).$$
- \(A) Note that $\widetilde{w}^{\delta}$ is defined in $ B\times [1+s_3,+\infty[$, whenever $\delta>0$ is small enough such that $-\log(\delta+e^{-1-s_3})\ge s_3.$
- \(B) From its construction, $\widetilde{w}^{\delta}$ is also a solution of (\[B\]) (Indeed, let $u$ be such that $w=w_{0,0}$ in defintion (\[scaling\]). Then $u$ is a solution of (\[1\]) and $\widetilde{w}^{\delta}=w_{0,-\delta}$ is defined as in (\[scaling\]); so $\widetilde{w}^{\delta}$ is a solution of (\[B\])).
- \(C) For $\delta$ small enough, we have $H(\widetilde{w}^{\delta}(1+s_3))<0$ by continuity of the function $\delta \mapsto H(\widetilde{w}^{\delta}(1+s_3))$. Then, we write that $H(\widetilde{w}^{\delta}(1+s_3))<0$.
Now, we fix $\delta=\delta_0>0$ such that (A), (B) and (C) hold. Let us note that we have $$\begin{aligned}
\label{c1}
- e^{-\gamma s}\int_{B}w^{\delta_0}w^{\delta_0}_s\rho {\mathrm{d}}y&\ge&-\frac{1}{4}
\int_{B}(w^{\delta_0}_s)^2\rho{\mathrm{d}}y- e^{-2\gamma s}
\int_{B}(w^{\delta_0})^2\rho{\mathrm{d}}y\end{aligned}$$ and from (\[I100\]) $$\begin{aligned}
\label{c2}
- e^{-\frac{2(p+1)s}{p-1}}\displaystyle\int_{B}F(e^{\frac{2}{p-1}s}w^{\delta_0})\rho {\mathrm{d}}y&\ge&
-C e^{-2\gamma s}-C e^{-2\gamma s}\int_{B}|w^{\delta_0}|^{p+1}\rho{\mathrm{d}}y.\end{aligned}$$ By (\[f\]), (\[f2\]), (\[c1\]) and (\[c2\]) we deduce $$\begin{aligned}
E(w^{\delta_0}(s))&\ge& \frac{1}{4}
\int_{B} (w_s^{\delta_0})^2\rho {\mathrm{d}}y
+(\frac{p+1}{(p-1)^2}- e^{-2\gamma s})\int_{B}(w^{\delta_0})^2\rho {\mathrm{d}}y\nonumber\\
&&-(\frac{1}{p+1}+C e^{-2\gamma s})\int_{B}|w^{\delta_0}|^{p+1} \rho {\mathrm{d}}y
-C e^{-2\gamma s}.\end{aligned}$$ We now choose $s_4\ge s_3$ large enough, so that we have $\frac{p+1}{(p-1)^2}- e^{-2\gamma s_4} \ge 0.$ Then, we deduce that we have, for all $s\ge s_4$, $$E(w^{\delta_0}(s))\ge
-(\frac{1}{p+1}+C e^{-2\gamma s})\int_{B}|w^{\delta_0}|^{p+1} \rho {\mathrm{d}}y
-C e^{-2\gamma s}.$$ Since $\rho\le 1$, after a change of variables, we find that $$E(w^{\delta_0}(s))
\ge-\frac{(\frac{1}{p+1}+C e^{-2\gamma s})}
{(1+\delta_0e^s)^{\frac{4}{p-1}+2-N}}\displaystyle\int_{B}|w(z,-\log (\delta_0+e^{-s}))|^{p+1} {\mathrm{d}}z-C e^{-2\gamma s}.$$ Since we have $-\log (\delta_0+e^{-s})\rightarrow -\log (\delta_0)$ as $s\rightarrow +\infty$ and since $\|w(s)\|_{L^{p+1}(B)}$ is locally bounded by hypothesis, by a continuity argument, it follows that the former integral remains bounded and $$\begin{aligned}
%\label{}
E(w^{\delta_0}(s))&\ge&
-\frac{C}{(1+\delta_0e^s)^{\frac{4}{p-1}+2-N}}-C e^{-2\gamma s}\rightarrow 0,\end{aligned}$$ as $s\rightarrow +\infty$ (use the fact that $\frac{4}{p-1}+2-N>0$ which follows from the fact that $p<p_c$). So, from (\[f1\]), it follows that $$\begin{aligned}
\label{d1}
\liminf_{s\rightarrow +\infty}H(w^{\delta_0}(s))\ge 0 .\end{aligned}$$ The inequality (\[d1\]) contradicts the inequality $H(w^{\delta_0}(s_3+1))<0$ and the fact that the energy $H$ decreases in time for $s\ge s_3$. This concludes the proof of Lemma \[L02\].
Boundedness of the solution in similarity variables
===================================================
We prove Theorem 2.2 here. Note that the lower bound follows from the finite speed of propagation and wellposedness in $H^1\times L^2$. For a detailed argument in the similar case of equation (\[mu=0\]), see Lemma 3.1 (page 1136) in [@MZ3].\
We consider $u$ a solution of (\[1\]) which is defined under the graph of $x\mapsto T(x)$, and $x_0$ a non characteristic point. Given some $T_0\in (0,T(x_0)]$, we introduce $w_{x_0,T_0}$ defined in (\[scaling\]), and write $w$ for simplicity, when there is no ambiguity. We aim at bounding $\|(w,\partial_s w)(s)\|_{H^1\times L^2(B)}$ for $s$ large.
As in [@MZ1], by combining Theorem \[lyap\] and Lemma \[L02\] (use in particular the remark after that Lemma) we get the following bounds:
(Bounds on $E$) For all $s\ge \widehat{s}_3=\widehat{s}_3(T_0)=\max(S_3,-\log T_0)$, $s_2\ge s_1\ge \widehat{s}_3$, it holds that \[cor1\] $$\begin{aligned}
\label{cor01}
-C\le E(w(s))\le M_0\nonumber\\
\int_{s_1}^{s_2}\int_{B}w_s^2(y,s)\frac{\rho}{1-|y|^2}dy ds\le M_0,\end{aligned}$$ where $M_0=M_0(N,p,q,M,\widehat{s}_3(T_0),\|(u(t_3),u_t(t_3))\|_{
H^{1}\times L^{2}(B(x_0,\frac{e^{-\widehat{s}_3(T_0)}}{\delta_0(x_0)})
)})$,\
$t_3=t_3(T_0)=T_0-e^{-\widehat{s}_3(T_0)}$, $C=C(N,p,q,M)$ and $\delta_0(x_0)\in (0,1)$ is defined in (\[cone1\]).
Starting from these bounds, the proof of Theorem 1.2 is similar to the proof in [@MZ3] except for the treatment of the perturbation terms. In our opinion, handling these terms is straightforward in all the steps of the proof, except for the first step, where we bound the time averages of the $L^{p+1}_{\rho}(B)$ norm of $w$. For that reason, we only give that step and refer to [@MZ3] for the remaining steps in the proof of Theorem 1.2. This is the step we prove here (In the following $K_1$ denotes a constant that depends only on $p$, $q$, $N$, $M$, $C$, $M_0$, and $\varepsilon$ is an arbitrary positive number in $]0,1[$).
\[pro\](Control of the space-time $L^{p+1}$ norm of $w$)\
For all $s\ge 1+\widehat{s}_3$, $$\label{pro1}
\int_{s}^{s+1}\int_{B}\!\!|w|^{p+1}{\rho}{\mathrm{d}}y{\mathrm{d}}s\le K_1(M_0,C,N,p,q,M ).$$
Proof: For $s\ge 1+\widehat{s}_3$, let us work with time integrals betwen $s_1$ et $s_2$ where $s_1\in [s-1,s]$ and $s_2\in [s+1,s+2]$. By integrating the expression (\[f\]) of $E$ in time between $s_1$ and $s_2$, where $s_2>s_1>\widehat{s}_3$, we obtain: $$\begin{aligned}
\label{et}
\int_{s_1}^{s_2}\!\!E(s) ds&=&\displaystyle\int_{s_1}^{s_2}\int_{B}\!\!\Big ( \frac{1}{2}w_s^2
+\frac{p+1}{(p-1)^2}w^2
-\frac{1}{p+1}|w|^{p+1}\Big )\rho {\mathrm{d}}y{\mathrm{d}}s\\
&&+\frac{1}{2}\displaystyle\int_{s_1}^{s_2}\!\!\int_{B}\!\!\Big (|\grad w|^2-(y.\grad w)^2\Big )\rho {\mathrm{d}}y{\mathrm{d}}s-\int_{s_1}^{s_2}\!\! e^{-\gamma s}\!\!\displaystyle\int_{B}\!\!ww_s\rho {\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&&-\int_{s_1}^{s_2}\!\! e^{-\frac{2(p+1)s}{p-1}}\displaystyle\int_{B}F(e^{\frac{2}{p-1}s}w)\rho {\mathrm{d}}y\mathrm{d}s.\nonumber\end{aligned}$$ By multiplying the equation (\[B\]) by $w\rho$ and integrating both in time and in space over $B\times [s_1,s_2]$, we obtain the following identity, after some integration by parts : $$\begin{aligned}
\label{et1}
&&\Big [\int_{B}\!\!\Big (ww_{s}+(\frac{p+3}{2(p-1)}-N)w^2\Big ) \rho{\mathrm{d}}y\Big ]_{s_1}^{s_2}=
\int_{s_1}^{s_2}\!\!\int_{B}\!\!w^2_{s}\rho{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&&-\int_{s_1}^{s_2}\!\!\int_{B}\!\!(|\grad w|^2-(y.\grad w)^2)\rho{\mathrm{d}}y{\mathrm{d}}s
-\frac{2p+2}{(p-1)^2}\int_{s_1}^{s_2}\!\!\int_{B}\!\!w^2\rho{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&&+
\int_{s_1}^{s_2}\!\!\int_{B}\!\!|w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s+\!2\!\int_{s_1}^{s_2}\!\!\int_{B}\!\!ww_s(y.\grad \rho){\mathrm{d}}y{\mathrm{d}}s
+2\!\!\int_{s_1}^{s_2}\!\!\!\int_{B}\!\!\!w_s(y.\grad w) \rho{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&&+\!\int_{s_1}^{s_2}\!\!\int_{B}\!\!e^{-\frac{2ps}{p-1}}f\Big(e^{\frac{2s}{p-1}}w\Big)w\rho{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&& +\!\int_{s_1}^{s_2}\!\!\int_{B}\!\!
e^{-\frac{2ps}{p-1}}g\Big(e^{\frac{(p+1)s}{p-1}}(w_s+y.\grad w+\frac{2}{p-1}w)\Big)w\rho{\mathrm{d}}y{\mathrm{d}}s.\qquad\end{aligned}$$ By combining the identities (\[et\]) and (\[et1\]), we obtain $$\begin{aligned}
\label{controlp}
&&\frac{(p-1)}{2(p+1)}\int_{s_1}^{s_2}\!\!\int_{B}\!\!|w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s\\
&=&\frac{1}{2}\Big [\int_{B}\!\!\Big (ww_{s}+(\frac{p+3}{2(p-1)}-N)w^2\Big ) \rho{\mathrm{d}}y\Big ]_{s_1}^{s_2}-
\int_{s_1}^{s_2}\!\!\int_{B}\!\!w^2_{s}\rho{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&&+\int_{s_1}^{s_2}\!\!E(s) ds
-\int_{s_1}^{s_2}\!\!\int_{B}\!\!ww_s(y.\grad \rho){\mathrm{d}}y{\mathrm{d}}s
-\int_{s_1}^{s_2}\!\!\int_{B}\!\!w_s(y.\grad w) \rho{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&& -\underbrace{\frac12\!\int_{s_1}^{s_2}\!\!\int_{B}\!\!
e^{-\frac{2ps}{p-1}}g\Big(e^{\frac{(p+1)s}{p-1}}(w_s+y.\grad w+\frac{2}{p-1}w)\Big)w\rho{\mathrm{d}}y{\mathrm{d}}s}_{A_1}\nonumber\\
&&-\underbrace{\int_{s_1}^{s_2}\!\! e^{-\gamma s}\!\!\displaystyle\int_{B}\!\!ww_s\rho {\mathrm{d}}y{\mathrm{d}}s}_{A2}
-\underbrace{\frac12\!\int_{s_1}^{s_2}\!\!\int_{B}\!\!e^{-\frac{2ps}{p-1}}f\Big(e^{\frac{2s}{p-1}}w\Big)w\rho{\mathrm{d}}y{\mathrm{d}}s}_{A_3}
\nonumber\\
&&+\underbrace{\int_{s_1}^{s_2}\!\! e^{-\frac{2(p+1)s}{p-1}}\displaystyle\int_{B}F(e^{\frac{2}{p-1}s}w)\rho {\mathrm{d}}y\mathrm{d}s}_{A4}.\end{aligned}$$ We claim that Proposition \[pro\] follows from the following Lemma where we control all the terms on the right-hand side of the relation (\[controlp\]) in terms of the space-time $L^{p+1}$ norm of $w$:
\[g\] For all $s\ge 1+\widehat{s}_4$, for some $\widehat{s}_4\ge \widehat{s}_3$, for all $\varepsilon>0$, $$\label{control}
\int_{s_1}^{s_2}\int_{B}\!\!|\grad w|^2(1-|y|^2)\rho{\mathrm{d}}y{\mathrm{d}}s\le K_1 +C
\int_{s_1}^{s_2}\!\!\int_{B}\!\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s,$$ $$\label{control1}
\sup_{s\in [s_1,s_2]}\int_{B} \!\!w^2(y,s)\rho{\mathrm{d}}y\le \frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\int_{B}\!\!|w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s.$$ $$\label{control3}
\int_{s_1}^{s_2}\!\!\!\int_{B}\!|w_s y.\grad w| \rho{\mathrm{d}}y{\mathrm{d}}s
\le \frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s,\qquad$$ $$\label{control30}
\int_{s_1}^{s_2}\!\!\!\int_{B}\!|w_sw y.\grad \rho|{\mathrm{d}}y{\mathrm{d}}s
\le \frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s,\qquad$$ $$\label{control4}
\int_{B}|ww_s|\rho{\mathrm{d}}y\le
\int_{B} w_s^2\rho{\mathrm{d}}y+\frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s,$$ $$\label{control5}
\int_{B}( w_s^2(y,s_1)+w_s^2(y,s_2))\rho{\mathrm{d}}y
\le K_1,$$ $$\begin{aligned}
\label{A10}
|A_1|&\le& \frac{K_1}{\varepsilon} +(K_1\varepsilon+
C e^{-s_1})
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s,\end{aligned}$$ $$\begin{aligned}
\label{A20}
|A_2|
&\le &
\frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s,\qquad\end{aligned}$$ $$\begin{aligned}
\label{A30}
|A_3|+|A_4|
&\le& C+ Ce^{-\gamma s_1}\int_{s_1}^{s_2} \!\!\int_{B}|w|^{p+1}\rho {\mathrm{d}}y{\mathrm{d}}s.\end{aligned}$$
Indeed, from (\[controlp\]) and this Lemma, we deduce that $$\begin{aligned}
%\label{}
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s&\le& \frac{K_1}{\varepsilon}+
(K_1\varepsilon+
C e^{-s_1}+C e^{-\gamma s_1})
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s.\qquad\end{aligned}$$ Taking $\widehat{s}_5$ large enough and $\varepsilon$ small enough so that $C e^{-\widehat{s}_5}+C e^{-\gamma \widehat{s}_5}\le \frac14$ and $K_1 \varepsilon \le \frac14$, we obtain (\[pro1\]).
It remains to prove Lemma 3.3.
Proof of Lemma 3.3: For the estimates (\[control\]), (\[control1\]), (\[control3\]), (\[control30\]), (\[control4\]) and (\[control5\]), we can adapt with no difficulty the proof given in the case of the wave equation treated in [@MZ1].
Now, we control the terms $A_1$, $A_2$, $A_3$ and $A_4$. Since $ |g(x)|\le M(1+|x|)$, we write $$\begin{aligned}
\label{I12}
|A_1|&\le& C\int_{s_1}^{s_2}\!\!e^{-s}
\int_{B}w_s^2\rho {\mathrm{d}}y\mathrm{d}s +
C\int_{s_1}^{s_2}\!\!e^{-s}\int_{B}|y.\grad w ||w|\rho {\mathrm{d}}y\mathrm{d}s\nonumber\\
&&+ C \int_{s_1}^{s_2}\!\!e^{-s}\int_{B}w^2\rho {\mathrm{d}}y\mathrm{d}s+ C \int_{s_1}^{s_2}\!\!e^{-\frac{2ps}{p-1}}\int_{B}|w|\rho {\mathrm{d}}y\mathrm{d}s.\qquad\end{aligned}$$ By using (\[cor01\]), we write $$\begin{aligned}
\label{CC0}
C\int_{s_1}^{s_2}\!\!e^{-s}
\int_{B}w_s^2\rho {\mathrm{d}}y\mathrm{d}s \le K_1\end{aligned}$$ Using the fact that $ e^{-s_1}\le1$ and the inequality (\[control1\]), we obtain, $$\begin{aligned}
\label{CC1}
C\!\!\int_{s_1}^{s_2}\!\!e^{-s}\!\!\int_{B}\!\!w^2{\rho}{\mathrm{d}}y{\mathrm{d}}s \le
C\sup_{s\in [s_1,s_2]}\int_{B}w^2{\rho}{\mathrm{d}}y
\le \frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s.
%\qquad\qquad\end{aligned}$$ We infer from (\[CC1\]) and the inequality $ |a|\le 1+a^2$ that $$\begin{aligned}
\label{CC2}
C \int_{s_1}^{s_2}e^{-\frac{2ps}{p-1}}\int_{B}|w|\rho {\mathrm{d}}y\mathrm{d}s&\le& C +C\int_{s_1}^{s_2} e^{-s}\int_{B}w^2\rho {\mathrm{d}}y\mathrm{d}s\nonumber\\
&\le&\frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\int_{B}\!\!|w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s.\end{aligned}$$ Using the Cauchy-Schwarz inequality, we write $$\begin{aligned}
\label{control6}
C\!\! \int_{s_1}^{s_2}\!\!e^{-s}\!\!\int_{B}\!\!|w||y.\grad w|{\rho}{\mathrm{d}}y{\mathrm{d}}s\le
C\!\! \int_{s_1}^{s_2}\!\! e^{-s}\int_{B}\!\!|w||y||\grad w|{\rho}{\mathrm{d}}y{\mathrm{d}}s&&\nonumber\\
\le C \!\!\int_{s_1}^{s_2}\!\! e^{-s}\int_{B}\!\!w^2\frac{|y|^2}{1-|y|^2}{\rho}{\mathrm{d}}y{\mathrm{d}}s
+C e^{-s_1}\!\!\int_{s_1}^{s_2}\!\!\int_{B}\!\!|\grad w|^2(1-|y|^2){\rho}{\mathrm{d}}y{\mathrm{d}}s.&&\end{aligned}$$ By combining (\[CC1\]), (\[control6\]), (\[hardyJJ\]), (\[control\]) and (\[control1\]), we get $$\begin{aligned}
\label{controlp3}
&&C\!\! \int_{s_1}^{s_2}\!\!e^{-s}\!\!\int_{B}\!\!|w||y.\grad w|{\rho}{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&\le& C \int_{s_1}^{s_2}e^{-s}\int_{B}w^2{\rho}{\mathrm{d}}y{\mathrm{d}}s
+C e^{-s_1} \int_{s_1}^{s_2}\int_{B}|\grad w|^2(1-|y|^2){\rho}{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&\le& C e^{-s_1} \int_{s_1}^{s_2}\int_{B}w^2{\rho}{\mathrm{d}}y{\mathrm{d}}s
+K_1 e^{-s_1} +C e^{-s_1}
\int_{s_1}^{s_2}\int_{B}| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s\qquad\nonumber\\
&\le& \frac{K_1}{\varepsilon}
+(K_1 \varepsilon +C e^{-s_1})
\int_{s_1}^{s_2}\int_{B}| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s.\qquad\end{aligned}$$ Using (\[I12\]), (\[CC0\]), (\[CC1\]), (\[CC2\]) and (\[controlp3\]), we obtain $$\begin{aligned}
\label{A1}
|A_1|&\le& \frac{K_1}{\varepsilon} +(K_1\varepsilon+
C e^{-s_1})
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s.\end{aligned}$$ Similarly, we deduce by (\[cor01\]) and (\[control1\]) that $$\begin{aligned}
\label{A2}
|A_2|
&\le&
C\int_{s_1}^{s_2} \!\!\int_{B}\!\!|ww_s|{\rho}{\mathrm{d}}y{\mathrm{d}}s \le\int_{s_1}^{s_2} \!\!\int_{B}\!\!w_s^2{\rho}{\mathrm{d}}y{\mathrm{d}}s+
C\int_{s_1}^{s_2}\!\! \int_{B}\!\!w^2{\rho}{\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&\le &
K_1+
C \sup_{s\in [s_1,s_2]}\int_{B}\!\!w^2{\rho}{\mathrm{d}}y\le
\frac{K_1}{\varepsilon} +K_1\varepsilon
\int_{s_1}^{s_2}\!\!\!\int_{B}\!| w|^{p+1}\rho{\mathrm{d}}y{\mathrm{d}}s.\qquad\end{aligned}$$ Finally, by (\[I100\]), we obtain $$\begin{aligned}
\label{A3}
|A_3|+|A_4|&\le& C\int_{s_1}^{s_2} \!\!e^{-\frac{2(p-q)s}{p-1}}{\mathrm{d}}s+ C\int_{s_1}^{s_2} \!\!e^{-\frac{2(p-q)s}{p-1}}\int_{B}|w|^{p+1}\rho {\mathrm{d}}y{\mathrm{d}}s\nonumber\\
&\le& C+ Ce^{-\gamma s_1}\int_{s_1}^{s_2} \!\!\int_{B}|w|^{p+1}\rho {\mathrm{d}}y{\mathrm{d}}s.\end{aligned}$$ This concludes the proof of Lemma 3.3 and Proposition 3.2 too.
Since the derivation of Theorem 1.2 from Proposition 3.2 is the same as in the non perturbed case treated in [@MZ3] (up to some very minor changes), this concludes the proof of Theorem 1.2.
[AAAA]{}
S. Alinhac. Blow-up for nonlinear hyperbolic equations, volume 17 of Progress in Nonlinear Differential Equations and their Applications. Birkh¨auser Boston Inc., Boston, MA, 1995.
S. Alinhac. A minicourse on global existence and blowup of classical solutions to multidimensional quasilinear wave equations. In Journées “Equations aux Dérivées Partielles” (Forges-les-Eaux, 2002), pages Exp. No. I, 33. Univ. Nantes, Nantes, 2002.
C. Antonini and F. Merle. Optimal bounds on positive blow-up solutions for a semi- linear wave equation. Internat. Math. Res. Notices, (21):1141-1167, 2001. H.A Levine. Instability and non-existence of global solutions of non-linear wave equation of the form $Pu_{tt}
=Au+
F(u)$. Transactions of the American Mathematical Society 192: 1-21, 1974.
H.A Levine. Some additional remarks on the non-existence of global solutions to non-linear wave equation. SIAM Journal on Mathematical Analysis 5 , 138-146, 1974.
F. Merle and H. Zaag. Determination of the blow-up rate for the semilinear wave equation. Amer. J. Math., 125:1147-1164, 2003.
F. Merle and H. Zaag. Determination of the blow-up rate for a critical semilinear wave equation. Math. Annalen, 331(2):395-416, 2005.
F. Merle and H. Zaag. Blow-up rate near the blow-up surface for semilinear wave equations. Internat. Math. Res. Notices, 19:1127-1155, 2005.
F. Merle and H. Zaag. Existence and universality of the blow-up profile for the semilinear wave equation in one space dimension. J. Funct. Anal., 253(1):43-121, 2007.
F. Merle and H. Zaag. Openness of the set of non characteristic points and regularity of the blow-up curve for the 1 d semilinear wave equation. Comm. Math. Phys., 282: 55-86, 2008.
F. Merle and H. Zaag. Existence and characterization of characteristic points for a semilinear wave equation in one space dimension . arxiv 25 novembre 2008.
[**Address**]{}:\
Université de Tunis El-Manar, Faculté des Sciences de Tunis, Département de mathématiques, Campus Universitaire 1060, Tunis, Tunisia.\
e-mail: [email protected]
Université Paris 13, Institut Galilée, Laboratoire Analyse, Géométrie et Applications, CNRS UMR 7539, 99 avenue J.B. Clément, 93430 Villetaneuse, France.\
e-mail: [email protected]
[^1]: This author is supported by a grant from the french Agence Nationale de la Recherche, project ONDENONLIN, reference ANR-06-BLAN-0185.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We propose a quantum spin Talbot effect for an electron beam transmitted through a grating of magnetic nanostructures. Tunable periodic magnetic nanostructures can be used in conjunction with electron-beam illumination to create a spin polarized replica of the transversely periodic exit surface beam a Talbot length away, due to quantum interference. Experiments have been proposed to verify the effect in a two dimensional electron gas. This effect provides a new route to modulate electron spin distributions without a magnetic field. A quantum spin Talbot interferometer and transistor are proposed for spintronics applications.'
author:
- 'W. X. Tang'
- 'D. M. Paganin'
title: Proposal for electron quantum spin Talbot effect
---
The ability to tune scalable semiconductor-based spintronics devices, based on the intrinsic spin of electrons to store and manipulate information, is both important and highly challenging for spin-based electronics since spin injection, spin accumulation and spin modulation of electrons are required [@peter; @albert; @prinz1; @prinz2; @wolf]. Currently, manipulation of the spin during transport between injector and detector via spin precession and spin pumping can be accomplished [@spin], however, those methods have difficulty controlling spin distributions. By contrast, local tunability of spin distributions over nanometer scales is crucial for future solid state quantum computers based on electron spin [@andrea]. Inspired by the progress in fabricating and controlling nanoscale magnetic structures [@parkin], we propose a spin-dependent quantum Talbot effect for electron waves transmitted by a grating composed of magnetic nanostructures, to modulate the spin lattice pattern formed from a spin polarized replica of the structure upon propagation through a Talbot length period and adjustable by controlling the electron wavelength and magnetic nanostructures’ period. This leads to potential applications such as a quantum spin Talbot transistor and a quantum spin Talbot interferometer.
The optical Talbot effect was discovered in 1836 [@Talbot], and later explained by Rayleigh as a natural consequence of Fresnel diffraction. He showed that the Talbot length $\mathcal{Z}$$_{T}$ is given by $\mathcal{Z}_{T}$=$\frac{2a^{2}}{\lambda}$ [@Rayleigh], in the paraxial approximation $a\gg\lambda$, where $a$ is period of the grating and $\lambda$ is the wavelength of the incident light. However, in a non-paraxial regime where $\lambda$$\leqslant$$a$$<$2$\lambda$, the Talbot effect is also operative for nonevanescent components of the scattered beam [@eero]. This effect reveals the wave-nature of both radiation and matter wave fields, examples of the latter including atoms, electrons and plasmons [@chapman; @cronin; @benjamin; @eero; @dennis; @leskova].
![Normalized diffraction intensity and profiles. (a) Maps for spin up $\rho_{+}$ and down $\rho_{-}$ probability density with spin asymmetry $\aleph(x,z; E)$. (b) Two spin asymmetry profiles along $z$ are indicated by symbols $\bigtriangleup$ and $\bigtriangledown$, and fit by Eq. \[Talbot\_fit\], where $a$ = 20 nm, $\lambda$ = 10 nm, $\mathcal{Z}$$_{T}$ = 75 nm from Eq. \[Talbot\_distance\].[]{data-label="rys3"}](fig1.eps){width="40.00000%"}
In this Letter, we calculate a spin polarized non-paraxial Talbot effect for electron matter waves transmitted through a grating composed of magnetic nanostructures. We find that the spin asymmetry of the transmitted field varies with distance from the grating, creating an electron spin replica of the structure a Talbot length away, in a non-paraxial regime where $\lambda$$\leqslant$$a$$<$2$\lambda$. This creates a tunable spin lattice in two-dimensional space, which is a powerful method to manipulate electron spin distributions in solid state systems. We find that the quantum spin interference pattern strongly depends on the wavelength and grating period. Based on our theoretical results, we propose experiments to verify this quantum spin Talbot effect (QSTE) in a two dimensional electron gas (2DEG) system and an atomically flat surface by spin polarized scanning probe microscopy. We propose both a quantum spin Talbot interferometer (QSTI) and quantum spin Talbot transistor (QSTT) devices.
For a grating with period $a$, normally illuminated with a monoenergetic electron plane wave, the two-component spatial electron wave function $\{\psi_+,\psi_-\}^T$ at energy $E$ and at any distance $z \ge 0$ downstream of the exit surface $z=0$ is $$\begin{aligned}
\label{wavefunction}
\psi_{\pm}(x,z;E) = \sum_m c_m^{\pm}(E) \exp\left[ \dot{\imath}(\gamma_{m}x+t_{m}z)\right].\end{aligned}$$ Here, $x$ is the transverse coordinate, $c_m^{\pm}(E)$ denotes the Fourier coefficients of the two independent electron spin projections, $\gamma_{m}$=$\frac{2\pi m}{a}$, and $t$$_{m}$=$\sqrt{(\frac{2\pi}{\lambda})^{2}-\gamma_{m}^{2}}$ [@eero; @dennis; @leskova], $\lambda=h/\sqrt{2m_eE}$ is the de Broglie wavelength, $h$ is Planck’s constant and $m_e$ is the electron mass. $+$ and $-$ represents “spin up”and “spin down” states of electron spin, respectively.
Consider a grating formed by nanoscale magnetic structures, for example, magnetic stripe domains, as shown in Fig. 1. Electron waves have a different complex transmission coefficient depending on the configuration of the incoming electron beam spin state relative to the magnetization direction of magnetic domains (parallel ($\uparrow\uparrow$, $\downarrow\downarrow$) or anti-parallel ($\uparrow\downarrow$, $\downarrow\uparrow$)) [@peter; @albert; @wolf]. Consequently, the spin up (down) electron wave $\psi_{+}$($\psi_{-}$) propagates through the up (down) magnetic domains in the grating, therefore achieving separation of the electron wave depending on the spin state is expected as shown in Fig. 1(a). The electron wave $\psi_+$ passes through the magnetic “up” domains $A$ (green) with 100% transmission while being blocked completely by domains $B$ (red). The color denotes the magnetization direction of a single domain. The corresponding probability density diffracted from the magnetic grating is longitudinally periodic in $z$ with period $\mathcal{Z}_{T}$. Simultaneously, the probability density depends on the spin of the electron wave shown in Fig. 1(a), as given by: $$\begin{aligned}
\label{probability_density}
\rho_{\pm}(x,z; E) = \sum_m \sum_n c_m^{\pm *}(E) c_n^{\pm}(E)
H_{m,n}(x,z;E),\end{aligned}$$ where $$\begin{aligned}
\label{Hmn}
H_{m,n}(x,z;E) = \exp\{\dot{\imath}[(\gamma_{n}-\gamma_{m})x+(t_{n}-t^{\ast}_{m})z]\}.\end{aligned}$$ By definition, the spin asymmetry $\aleph(x,z;E)$ $\equiv$ ($\rho_+$ -$\rho_-$)/($\rho_+$ +$\rho_-$) will have the same longitudinal periodicity as the probability density. Since both the numerator and the denominator have a longitudinal periodicity equal to the Talbot distance, when either $a$$\gg$$\lambda$ or $\lambda$$\leqslant$$a$$<$2$\lambda$, $\aleph(x,z;E)$ implies a [*continuously tunable spin lattice*]{} in two dimensional space as shown in Fig. 1(b). The distribution of two-dimensional spin asymmetry is determined by $\lambda$ and $a$. In Fig. 1, $\lambda$=10 nm and $a$=20 nm; the numerically calculated Talbot distance $\mathcal{Z}$$_{T}$ is 75 nm instead of 80 nm as expected by the conventional formula 2$a$$^{2}$/$\lambda$. The discrepancy is due to the paraxial approximation in conventional Talbot theory. To calculate the non-paraxial $\mathcal{Z}$$_{T}$ by the self-imaging condition $\aleph$($x$,$n$$\mathcal{Z}$$_{T}$;$E$)=$\aleph$($x$,0;$E$), for integer $n$ and $\lambda$$\leqslant$$a$$<$2$\lambda$, we obtain
![Two dimensional spin-distribution $\aleph$($x$,$z$;$E$) and corresponding profiles for $\zeta$ = 0.1, 0.5 and 1.01, respectively.[]{data-label="rys3"}](fig2.eps){width="40.00000%"}
[@eero], $$\begin{aligned}
\label{Talbot_distance}
\mathcal{Z}_{T}=\frac{\lambda}{1-[1-(\lambda/a)^{2}]^{1/2}}.\end{aligned}$$ From Eq. \[Talbot\_distance\], $\mathcal{Z}$$_{T}$= 74.6 nm, consistent with our numerical results based on Eqs. 2–4. For the paraxial approximation $a$$\gg$$\lambda$, Eq. \[Talbot\_distance\] approaches 2$a$$^{2}$/$\lambda$, consistent with the literature [@eero; @dennis; @leskova]. The spin asymmetry distribution (Fig. 1 (b)) in the range 0.5$<$$\zeta$$\leqslant1$, where $\zeta$=$\lambda$/$a$, is $$\begin{aligned}
\label{Talbot_fit}
\aleph(x,z;E)=A_{0}(x)\sin\frac{2\pi z}{\mathcal{Z}_{T}},\end{aligned}$$ where $A_{0}$(x)=$\sin$$\frac{2\pi
x}{a}$$\cdot$$[$$\frac{\pi}{8}$+$\frac{2}{\pi}$$\sin$$^{2}$($\frac{2\pi
x}{a}$)$]$$^{-1}$. Note that evanescent waves have been neglected in calculating the above expression. If $x$= $a$/4, then $A_{0}$$\approx$0.97 as shown in Fig. 1(b). Therefore, spin lattices can be tailored through nm to sub-$\upmu$m depending on $\zeta$ and $a$.
To understand this tunability, we calculated the QSTE for different $\lambda$. We find that the spin asymmetry profile curves show a simple sine relationship when 0.5$<\zeta\leqslant1$. However, if $\zeta<0.5$, the curves have complex structures and small ripples decorate the spin asymmetry distribution (e.g. $\zeta$= 0.1); when $\zeta$ $>$1, evanescent waves imply that the polarization of spin decreases exponentially along $z$ (e.g. $\zeta$= 1.01); Eqs. \[Talbot\_distance\] and \[Talbot\_fit\] are not applicable for these ranges. Movie 1 shows sequential evolution of the $\rho_{\pm}$ and $\aleph$ with $\zeta$ at $a$= 20 nm [@movie1].
In a more realistic model, the electron wave undergoes partial transmission at the antiparallel configuration between spin orientation and magnetization direction of the domain. Considering this, we find that the probability density distributions are blurred but nevertheless distinguishable; even assuming only 1% transmission difference (TD) between two channels. The intensity contrast and spin polarization drop with TD by the same order of magnitude (Fig. 3).
![The spin asymmetry distribution dependent on transmission-rate difference between two channels. $a$=20 nm and $\lambda$=10 nm.[]{data-label="rys3"}](fig3.eps){width="40.00000%"}
For a finite energy spread, assume an incident distribution of electron energies $S_i(E)$. Under this model, we obtain $\overline{\aleph}$($x$,$z$)$\equiv$$\int
S_i(E)$$\aleph$($x$,$z$;$E$)$dE$/$\int S_i(E)$$dE$. The influence of energy spread on the 2D QSTE is calculated by numerical evaluation of $\overline{\aleph}$($x$,$z$), assuming $S_i(E)$ to be uniform from $\lambda$=15 nm to 20 nm, with the results shown in Fig. 4. Surprisingly, a dramatic longitudinal modulation of spin polarization near the grating is observed. Consequently, the spin Talbot distance is also modulated depending on the energy spread $S_i(E)$. In Fig. 4, instead of one peak appearing within each spin Talbot distance, multiple peaks appear. Therefore, to verify the QSTE, a narrow energy spread is desired, or the method should have high energy resolution to distinguish different energy channels.
A 2DEG at interfaces such as in a AlGaAs/GaAs heterostructure is a candidate for testing the effect due to the small energy spread at the Fermi level. In addition, the high mobility of electrons ($>$ 3x10$^{6}$ cm$^{2}$V$^{-1}$s$^{-1}$) and their long spin transportation distance ($>$ hundred $\upmu$m) are suitable properties for spatial imaging of this effect [@holland; @DD; @wolf; @TP; @mark]. The electron de Broglie wavelength at the Fermi energy is unusually long, around 20–100 nm [@TP], making it easy to design a suitable magnetic domain period $a$ and minimize effects caused by nonzero domain wall width [@schmid; @wxtang]. Further concern includes suitable materials for the grating formed by magnetic stripe domains. The wavelength of electrons in metal is normally less than 1 nm, therefore, dilute magnetic semiconductors such as MnGaAs might be suitable to form magnetic domain gratings with similar band structure to AlGaAs/GaAs. Furthermore, by applying a pulsed electron current along the grating, the period $a$ is tunable by domain wall motion in ns [@shinjo; @parkin].
![Polyenergetic spin asymmetry corresponding to $S_{i}(E)$=constant in range $\lambda$= 15–20 nm, using 800 integral steps in the numerical integration; $a$= 20 nm.[]{data-label="rys3"}](fig4.eps){width="40.00000%"}
We have described the properties of a 2DEG system for testing the effect, however, a spin dependent spatial imaging method is needed. A spatial imaging technique has been elegantly applied in spin Hall-effect detection in a 2DEG system by Scanning Magneto-optic Kerr microscopy (SMOKE) [@spinhallmoke]. This is an ideal way to demonstrate the QSTE.
Besides SMOKE, imaging electron flow in a 2DEG at the nanoscale has been achieved based on a scanning probe method [@mark; @spm]. This measure can be applied to verify the QSTE at GaAs/AlGaAs interfaces. In particular, scanning tunnelling microscopy with a spin polarized tip (SP-STM) is an ideal technique for investigating the surface electron wave QSTE [@binnig; @wiesendanger; @oka], as it provides both spin contrast and atomic resolution. Recently, spin-dependent quantum interference within a Co magnetic nanostructure by SP-STM has been reported [@oka]. Inspired by this experiment, we believe SP-STM could be used to see QSTE on an atomically flat surface by a grating formed by an antiferromagnetic atomic chain. One of the advantages of SP-STM is its high energy resolution in $dI$/$dV$ spin-asymmetry spectra to differentiate energy channels [@oka].
![Schematic representing deformation of the quantum spin Talbot pattern caused by the spin-dependent scattering at magnetic dots in front of a grating.[]{data-label="rys3"}](fig5.eps){width="45.00000%"}
We note that a far field non-spin electron Talbot interferometer has been achieved [@benjamin]. Given the quantum spin Talbot pattern obtained in our work, it is straightforward to formulate a quantum spin Talbot interferometer (QSTI)(Fig. 5). In contrast to the non-spin Talbot case, the QSTI is very sensitive to the change in magnetization of magnetic dots located at the front of the grating. This is extremely valuable to read out the magnetic configuration of such dots. The QSTI should also be sensitive to map weak magnetic fields in nanoscale.
![(a) Schematic of quantum spin Talbot transistor. (b) Characteristic quantum spin Talbot resistance depending on grating separation $S$ and $\lambda$, $a$=10 nm and 0.5$<$$\zeta$$\leqslant1$.[]{data-label="rys3"}](fig6.eps){width="45.00000%"}
With electrodes connected to two gratings (G1 and G2), a quantum spin Talbot transistor (QSTT) can be fabricated. By calculating the spin transmission probability $T_{\pm}$$(\lambda,S)$ through the second grating from Eq. \[probability\_density\] as an approximation, we obtain [@wxt] $$\begin{aligned}
\label{resistor}
T(\lambda,S)=T_{+}+T_{-}=\frac{1}{4}+\frac{2}{\pi^{2}}+\frac{4}{\pi^{2}}\cos\frac{2\pi
S}{\mathcal{Z}_{T}(\lambda)}\end{aligned}$$ where $S$ is the separation between G1 and G2. The $T$($\lambda$,$S$) varies from single to multi-peak tuning via $S$ and $\lambda$ in the range 0.5$<$$\zeta$$\leqslant1$, exhibiting behavior quite distinct in comparison to both GMR [@albert] and spin Hall effect transistor [@spinhall]. By further including the effect of electric field, we solve the Schr$\ddot{o}$dinger equation with electrical potential, modifying Eq.\[resistor\] to include an Airy function $Ai(S)$ [@airy]; however, no dramatic change of the QSTT curve shape is expected [@wxt].
In conclusion, we propose an electron quantum spin Talbot effect. Potential applications such as the QSTI and QSTT are present. The success of the experiment will provide a new route to actualize periodic spin state distributions in two-dimensional space, also leading to spintronics applications which will be important for future spin-based technologies.
We thank the referees for their insightful comments. WXT thanks Dr. Zheng Gai at ORNL for illuminating discussions on potential experimental realizations, and encouragement.
[99]{}
P. Grunberg, R. Schreiber, Y. Pang, M. B. Brodsky, and H. Sowers, Phys. Rev. Lett. [**57**]{}, 2442 (1986)
M. N. Baibich, J. M. Broto, A. Fert, F. Nguyen Van Dau, F. Petroff, P. Etienne, G. Creuzet, A. Friederich, and J. Chazelas, Phys. Rev. Lett. [**61**]{}, 2472 (1988)
G. A. Prinz, Phys. Today [**48**]{}(4), 58 (1995).
G. A. Prinz, Science [**282**]{}, 1660 (1998).
S.A. Wolf, D. D. Awschalom, R. A. Buhrman, J. M. Daughton, S. von Molnar, M. L. Roukes, A. Y. Chtchelkanova, and D. M. Tregar, Science [**294**]{}, 1488 (2001)
S. K. Watson, R. M. Potok, C. M. Marcus, and V. Umansky, Phys. Rev. Lett. [**91**]{}, 258301 (2003)
A. Morello $e$$t$ $a$$l$., Nature [**467**]{}, 687 (2010)
L. Thomas, M. Hayashi, X. Jiang, R. Moriya, Ch. Rettner, and S. S. P. Parkin, Nature [**443**]{}, 197 (2006)
H. F. Talbot, Philos. Mag. [**9**]{}, 401 (1836)
Lord Rayleigh, Philos. Mag. [**11**]{}, 196 (1881)
E. Noponen and J. Turunen, Opt. Commun. [**98**]{}, 132 (1993)
M. S. Chapman, et al., Phys. Rev. A [**51**]{}, R14 (1995)
A. D. Cronin and B. McMorran, Phys. Rev. A [**74**]{}, 061602(R) (2006)
B. J. McMorran and A. D. Cronin, New J. Phys. [**11**]{}, 033021 (2009)
M. R. Dennis, N. I. Zheludev, and F. J. G. Abajo, Opt. Express, [**15**]{}, 9692 (2007)
A. A. Maradudin and T. A. Leskova, New J. Phys. [**11**]{}, 033004 (2009)
In movie 1, the wavelength dependent QSTE is presented.
N. Rougemaille and A. K. Schmid, Eur. Phys. J. Appl. Phys. [**50**]{}, 20101 (2010)
Domain wall width $w$$\varpropto$$\sqrt{J/K}$, determined by exchange interaction $J$ and anistropy $K$; $w$ ranges from a few nm to $\mu$m.
M. C. Holland, A. H. Kean, and C. R. Stanley, J. Crys. Growth [**127**]{}, 793 (1993)
D. D. Awschalom, Physica E [**10**]{}, 1 (2001)
T. Chakraborty, P. Pietilainen, eds., The Quantum Hall Effects: Integral and Fractional, 2nd ed., Springer-Verlag, New York (1995)
M. A. Topinka, R. M. Westervelt, and E. J. Heller, Phys. Today [**56**]{}(12), 47 (2003)
A. Dourlat, C. Gourdon, V. Jeudy, K. Khazen, H. J. voon Bardeleben, L. Thevenard, and A. Lemaitre, Physica E [**40**]{}, 1848 (2008)
A. Yamaguchi, T. Ono, S. Nasu, K. Miyake, K. Mibu, and T. Shinjo, Phys. Rev. Lett. [**92**]{}, 077205 (2004)
V. Sih, R. C. Myers, Y. K. Kato, W. H. Lau, A. C. Gossards, and D. D. Awschalom, Nature Phys. [**1**]{}, 31 (2005)
M. A. Topinka, B. J. LeRoy, S. E. J. Shaw, E. J. Heller, R. M. Westervelt, K. D. Maranowski, and A. C. Gossard, Science [**289**]{}, 2323 (2000)
G. Binnig, H. Rohrer, Ch. Gerber, and E. Weibel, Appl. Phys. Lett. [**40**]{}, 178 (1982)
R. Wiesendanger, H.-J. Guntherodt, G. Guntherodt, R. J. Gambino, and R. Ruf, Phys. Rev. Lett. [**65**]{}, 247 (1990)
H. Oka, et al. Science [**327**]{}, 843 (2010)
B. J. McMorran and A. D. Cronin, New J. Phys. [**11**]{}, 033021 (2009)
J. Wunderlich, et al. Science [**330**]{}, 1801 (2010)
D. M. Paganin and W.X. Tang, to be published
S. M. Kennedy, D. M. Paganin and D. E. Jesson, Am. J. Phys. [**76(2)**]{}, 158 (2007)
|
{
"pile_set_name": "ArXiv"
}
|
UUITP-12/02\
hep-th/0210058
[**On the consistency of de Sitter vacua\
**]{}
[Ulf H. Danielsson]{}\
Institutionen för Teoretisk Fysik, Box 803, SE-751 08 Uppsala, Sweden
[[email protected]\
]{}
[**Abstract**]{}
In this paper the consistency of the de Sitter invariant $\alpha $-vacua, which have been introduced as simple tools to study the effects of transplanckian physics, is investigated. In particular possible non renormalization problems are discussed, as well as non standard properties of Greens functions. We also discuss the non thermal properties of the $\alpha $-vacua and the necessity of $\alpha$ to change. The conclusion is that non of these problems necessarily exclude an application of the $\alpha $-vacua to inflation.
October 2002
Introduction
============
The main driving force behind recent work on transplanckian physics in cosmology, is to find out whether physics beyond the Planck scale can give effects on the CMBR spectrum, \[1-30\]. A crucial ingredient in inflationary cosmology is that microscopic quantum fluctuations are magnified by inflation into macroscopic seeds for galaxy formation. In standard inflation the modes of the inflaton field can be carried back in time to eras when they were much smaller than the Planck scale. For small scales the expansion of the universe can be ignored and a unique vacuum can be chosen for the inflaton. This is the Bunch-Davies vacuum. But the construction ignores the Planck scale and the natural expectation that physics beyond the Planck scale is very different from physics at low energies, and not possible to describe using a quantum field theory. The hope would be that transplanckian physics modifies the standard scenario and that the effects are magnified through inflation to leave visible imprints in the CMBR fluctuations.
There are two main approaches to investigate whether the above proposal is viable. One possibility is to use specific models of transplanckian physics, examples include models involving non commutative geometry, and investigate how this will change the predictions of inflation. Another possibility is to leave the details of the transplanckian physics for later work, and impose a cutoff on the theory at the Planck scale, or some other scale like the strings scale, where fundamentally new physics is expected. Our ignorance of the high energy physics is then encoded in the choice of initial conditions for the field modes when they start out at planckian size. Contrary to the standard scenario, the initial conditions are imposed in a situation where the time dependence of the background can not be ignored. Various natural ways of choosing the vacuum, minimal uncertainty, adiabatic to all orders etc., now give different results. The Bunch-Davies vacuum remains as a possibility but there are also other possibilities that also can be argued to be natural. A conservative approach is then to investigate the span of possibilities and see what effects, if any, they give on the CMBR. In [@Danielsson:2002kx] it was argued that the generic size of the effects is expected to be of order $H/\Lambda $ in the spectrum compared with the Bunch-Davies vacuum. To claim smaller effects one would need specific information about the nature of transplanckian physics. It is important to note that the generic effect is not a simple change in normalization of the spectrum but has a definite signature in the form of a modulation.
The simplified approach discussed in [@Danielsson:2002kx][@Danielsson:2002qh], which is not tied to specific models of transplanckian physics, allows for a detailed examination of the consequences and viability of the transplanckian proposal. As observed in [@Danielsson:2002qh] it essentially amounts to an investigation of the physics of the de Sitter invariant vacua introduced in [@chernikov][@Mottola:ar][@Allen:ux][@Floreanini:1986tq], some times called the $\alpha $-vacua. These vacua have also recently been discussed in the context of de Sitter holography, [@Bousso:2001mw][@Spradlin:2001nb].
The purpose of this note is to address recent discussions in the literature, in particular [@Banks:2002nv][@Einhorn:2002nu][@Kaloper:2002cs], where the consistency of the $\alpha $-vacua is questioned. In the first two of the above mentioned works it has been pointed out that quantum field theory with $\alpha $-vacua is not well understood. In particular there are problems with renormalizability and the definition of loop amplitudes, and there are also peculiarities in certain Greens functions that are claimed to suggest that the vacua does not make physical sense.
The outline of the note is as follows. In section two we review the construction of the $\alpha $-vacua, in section three we focus on the problems of non renormalizability, in section four we discuss the large scale structure of the Greens functions, in section five how $\alpha $ might change with time, in section six the non-thermal nature of the $\alpha $-vacua and, finally, we end with some conclusions.
Review of $\protect\alpha $-vacua
=================================
Let us briefly review how the $\alpha $-vacua are constructed. We will focus on a single inflaton field, $\phi \left( x\right) $, and its quantization. The Wightman function $G^{+}\left( x,x^{\prime }\right) $ for the inflaton field is defined as $$G^{+}\left( x,x^{\prime }\right) =\left\langle \Omega \right| \widehat{\phi }\left( x\right) \widehat{\phi }\left( x^{\prime }\right) \left| \Omega
\right\rangle =\int d^{3}k\phi _{\mathbf{k}}\left( x\right) \phi _{\mathbf{k}}^{\ast }\left( x^{\prime }\right) , \label{wightman}$$ where the field is expanded in modes as $$\widehat{\phi }\left( x\right) =\int d^{3}k\left[ \phi _{\mathbf{k}}\left(
x\right) \widehat{a}_{\mathbf{k}}+\phi _{-\mathbf{k}}^{\dagger }\left(
x\right) \widehat{a}_{\mathbf{k}}^{\dagger }\right] . \label{phiexp}$$ Operators are equipped with hats in all expressions. We then assume thatthese modes are obtained through Bogolubov transformations from the Bunch-Davies modes according to $$\phi _{\mathbf{k}}\left( x\right) =A\phi _{\mathbf{k,}BD}\left( x\right)
+B\phi _{-\mathbf{k,}BD}^{\dagger }\left( x\right) ,$$ with $$\left| A\right| ^{2}-\left| B\right| ^{2}=1.$$ A convenient parametrization is to write $$A=\frac{1}{\sqrt{1-e^{\alpha +\alpha ^{\ast }}}}\quad B=\frac{e^{\alpha }}{\sqrt{1-e^{\alpha +\alpha ^{\ast }}}},$$ where $\alpha =-\infty $ is the Bunch-Davies vacuum. For convenience we write $$\phi _{\mathbf{k,}BD}\left( x\right) =\phi _{k,BD}\left( \eta \right) e^{i\mathbf{k}\cdot \mathbf{x}},$$ and demand, for simplicity, that $$\phi _{k,BD}\left( \eta \right) =\phi _{k,BD}^{\dagger }\left( -\eta \right)
.$$
To proceed, let us recall a few elementary aspects of de Sitter geometry. The metric in terms of coordinates useful for inflationary cosmology is given by $$ds^{2}=dt^{2}-a\left( t\right) ^{2}d\mathbf{x}^{2}, \label{metric}$$ where $\ a\left( t\right) =e^{Ht}$ is the scale factor. In terms of the conformal time $\eta =-\frac{1}{aH}$ the metric becomes $$ds^{2}=\frac{1}{H^{2}\eta ^{2}}\left( d\eta ^{2}-d\mathbf{x}^{2}\right) .
\label{confmet}$$ The inflationary universe can be viewed as the upper right triangle of the full de Sitter space, which has a Penrose diagram with the shape of a square, where $\eta \rightarrow -\infty $ corresponds to the Big Bang, while $\eta \rightarrow 0$ is the infinite future. The lower half of de Sitter space can be covered if we also consider positive $\eta $. A crucial ingredient in the study of the $\alpha $-vacua is the antipodal map, which in the above coordinates takes $x=\left( \eta ,\mathbf{x}\right) $ into $\overline{x}=\left( -\eta ,\mathbf{x}\right) $, and acts like a map between the two halves of de Sitter space.
With the help of the antipodal map the Bogolubov transformation can now be written $$\phi _{\mathbf{k}}\left( x\right) =A\phi _{\mathbf{k,}BD}\left( x\right)
+B\phi _{\mathbf{k,}BD}\left( \overline{x}\right) ,$$ and the Wightman function becomes
$$G^{+}\left( x,x^{\prime }\right) =\left| A\right| ^{2}G_{BD}^{+}\left(
x,x^{\prime }\right) +\left| B\right| ^{2}G_{BD}^{+}\left( x^{\prime
},x\right) +AB^{\ast }G_{BD}^{+}\left( x,\overline{x}^{\prime }\right)
+BA^{\ast }G_{BD}^{+}\left( \overline{x},x^{\prime }\right) . \label{planw}$$
After this brief review of the $\alpha $-vacua we will now turn to the main subject of the paper: do the $\alpha $-vacua make physical sense?
Are loop amplitudes ill defined?
================================
In [@Banks:2002nv][@Einhorn:2002nu] problems with the definition of loop amplitudes were pointed out. In both papers one loop amplitudes were investigated and failed to give finite and well defined results. In [@Banks:2002nv] it was shown that non local counter terms in the action, involving insertions at image points, were needed. In [@Einhorn:2002nu] the analytic structure of the Greens functions indicated that pinched singularities in the loop integrations make the results ill defined. These problems are clearly of great interest and it is a challenge to make sense out of the field theory under these circumstances. But none of these problems are necessarily relevant to the issue of transplanckian physics in cosmology.
The reason is simple. The whole point with the transplanckian physics, as explained in the introduction, is to see whether effects beyond quantum field theory can be relevant for the detailed structure of the fluctuation spectrum of the CMBR. Without a planckian cutoff there is no reason to impose initial conditions at any finite scale. The only natural procedure is to go to the infinite past, when the modes are infinitely small, and make the choice there. For these small scales the expansion of the universe is irrelevant and there is a unique natural choice of vacuum. This would be the end of the story in a world without a Planck scale (and dynamical gravity) in which quantum field theory could be trusted to all energies. The issue of $\alpha $-vacua, and transplanckian physics, would never arise. In the real world we do expect quantum field theory to break down at high enough energy to be replaced by something else, presumably string theory. To find out what kind of effects this new physics might have, we can try to modify the transplanckian physics by hand. One way to gain control over the situation is to impose a cutoff and assume quantum field theory to hold for energies below the cutoff. It is now the issue of vacuum choice becomes important. Since the Planck scale is not infinitely smaller than the inflationary Hubble scale, the time dependence of the background has to be taken into account, and there is no unique natural vacuum. In fact, there are certain $\alpha $-vacua that are as natural a product of the unknown transplanckian physics as the Bunch-Davies vacuum. The modest proposal behind [@Danielsson:2002kx] is simply that we should allow for this uncertainty in the possible outcome and investigate the consequences.
To summarize, the issue of $\alpha $-vacua only appears when one takes a planckian cutoff into account. This is perfectly consistent with the results of [@Einhorn:2002nu]. *The* $\alpha $*-vacua not only needs a planckian cutoff to be of physical relevance, they need a planckian cutoff to make physical sense.* In other words, the quantum field theory loop amplitudes needs planckian input to give well defined answers.
In this context one should note that a rough cutoff in the sum over momenta and energies in (\[wightman\]) at the Planck scale, regulates the standard singularities in the Greens functions that occur when the insertions can be joined by a light ray. But it is not really these singularities that make trouble, it is rather those that are due to the image charges. But, as is easily seen, all terms in the Wightman function in (\[planw\]) are regulated with this procedure. With these regulated Greens functions finite loop amplitudes can be constructed, even though their values are sensitive to exactly how the cutoff is implemented, that is, the results depend on transplanckian physics. The main point is that in a quantum gravity theory, where the Planck scale plays an important role, space time points are expected to be effectively smeared to a size of order Planck scale. That is, it is meaningless to claim, for instance, that you sit exactly on a light cone.
Many of the problems with the $\alpha $-vacua are related to the image charges and their apparent non local nature. This is what we turn to next.
Do image charges break causality?
=================================
Another issue that has been brought up concerns the long distance behavior of the theory. As we have seen, the $\alpha $-vacua can easily be constructed by allowing for image charges on the wrong side of de Sitter space. Greens functions might therefore receive contributions directly from a source but also through the image. This has caused concerns that the theory does not make physical sense and one might worry about problems of, e.g., causality.
It is, however, important to bear in mind the physical interpretations of the various Greens functions. The ones which are important for causality are the commutator and the retarded Greens functions, which are independent of the choice of vacuum and always vanish outside of the light cone. When physics depending on these Greens functions is studied, there is no difference between the standard Bunch-Davies vacuum and the more general $\alpha $-vacua. Greens functions that do depend on the choice of vacua, and therefore are different for the $\alpha $-vacua, are, e.g., the Feynman propagator and the Hadamard function. These are the ones that exhibit the extra singularities, outside of the light cone, that have caused some worries. One should note, however, that these Greens functions in general are expected to be non zero outside the light cone; the Bunch-Davies vacuum is no exception. Contrary to the commutator and retarded Greens function they encode information about the vacuum and correlations in the vacuum fluctuations. Unorthodox behavior, like in the case of the $\alpha $-vacua, might seem surprising but does not necessarily imply that the theory is inconsistent.
Of particular interest in this context is the Hadamard function. Note that the Feynman propagator can be written in terms of the vacuum independent retarded and advanced Greens functions and the vacuum dependent Hadamard function. The extra singularities of the Hadamard function occur, if we consider points at equal time in Robertson-Walker coordinates, for points separated by the diameter of de Sitter space. They simply imply that the correlation between fluctuations at such points is enhanced. The vanishing of the commutator guarantees that this can not be used for communications and can not lead to any break of causality. The large separation between the points makes it a little difficult for an observer to actually measure the correlations, although it is suggested in [@Einhorn:2002nu] that interactions might facilitate this. As far as the transplanckian effects are concerned, this is, however, not really the main point. When inflation ends the full structure of the Hadamard function will become visible and available to the CMBR. As suggested in [@Danielsson:2002qh], the end of inflation transforms the meta observables of [@Witten:2001kn] into real observables. Unfortunately, the direct detection of correlations over distances of the order the de Sitter diameter is not possible. At the end of inflation, the modes relevant for the CMBR are much larger than the de Sitter diameter, and the direct as well as the image mediated fluctuations are hidden in the extremely small scales. The only aspect that is claimed to be of relevance for the CMBR is a possible large scale tail.
[@Kaloper:2002cs] reached the same conclusion as in the present work concerning the irrelevance of correlations over spatial distances in the Feynman propagator. The authors of [@Kaloper:2002cs] continued, however, by pointing out that even though the singularities are harmless, a propagator that behaves like in Minkowsky space for small distances does not suffer from them anyway. Small distances and Minkowsky behavior can certainly be reached in the present universe well above the Planck scale. But, as has already been emphasized, the situation is very different during inflation when the Planck scale might be just a few orders of magnitude smaller than the inflationary scale. In such a universe Minkowsky behavior can not be reached before transplanckian physics make quantum field theory irrelevant. This way of arguing for the Bunch-Davies vacuum therefore lacks force.
Finally, it should be noted that it is essential for the consistency of the vacua that we restrict ourselves to half of de Sitter space. As pointed out already in [@Mottola:ar], the direction of time changes on the wrong side of de Sitter. This also guarantees that there is no retarded propagator connecting an image charge with anything on the right half of de Sitter space.[^1]
Can $\protect\alpha $ change with $H$?
======================================
In [@Kaloper:2002cs] it is argued that $\alpha $ can not vary with a changing $H$. The reason, according to [@Kaloper:2002cs], is that local physics can not know about how $H$ changes and consequently $\alpha $ must remain constant. Since present day physics requires $e^{\alpha }$ to be substantially lower than any value that would be of interest to inflation, and $\alpha $ can not change with time, all interesting effects due to transplanckian physics are excluded.
However, one should note that even the Bunch-Davies modes have a dependence on $H$. This is easily seen from the Fourier transform of a Bunch-Davies mode given by $$\phi _{\mathbf{k,}BD}\left( x\right) =-\eta H\frac{1}{\sqrt{2k}}e^{-ik\eta
}\left( 1-\frac{i}{k\eta }\right) e^{i\mathbf{k}\cdot \mathbf{x}},$$ where $k$ is the comoving momentum. The form when the mode is created at the fundamental scale is obtained by using $p=k/a=\Lambda $ and recalling that $\eta =-\frac{1}{aH}$. The result has a dependence on $H$, and the planckian physics preparing the modes in a Bunch-Davies state therefore needs to know about the value of $H$. In a pure quantum field theory world the dependence is naturally generated through the expansion of the universe while the mode is in the transplanckian regime. A further dependence on $H$ through $\alpha
$ is not qualitatively much different and it is easy to see that such a dependence is quite natural and local. The scale factor depends on time through $a\left( t\right) =e^{Ht}$, with the expansion of space as a local effect taking place everywhere. The conjugate time interval associated with a mode of energy $\Lambda $ is $1/\Lambda $, and during this time interval space expands by a factor $e^{H\left( t+1/\Lambda \right) -Ht}=e^{H/\Lambda
}\sim 1+\frac{H}{\Lambda }$. This is true also for spatial distances close to the Planck scale and it is therefore natural to expect effects of the order $\frac{H}{\Lambda }$. This is immediately obvious from looking at the Bunch-Davies modes above, but transplanckian physics might lead to an additional dependence on $\frac{H}{\Lambda }$ through $\alpha $.
The simplest, and most natural dependence, is that we have $$e^{\alpha }\sim \frac{H}{\Lambda },$$ and that this remains true at all times. Not only during inflation. However, in [@Starobinsky:2002rp] it has been pointed out that even this small value of $e^{\alpha }$ could be in conflict with present day measurements of high energy gamma rays. A vacuum of the type above, it is claimed, would lead to a higher rate of particle creation than what is acceptable. The discrepancy is not large, considering the sensitivity of the argument to the detailed astro-physics involved – it is a matter of an order of magnitude or two – but it nevertheless suggests that a more complicated behavior than the one above might be needed. This is precisely the subject of [@Goldstein:2002fc] where the vacuum relaxes towards the Bunch-Davies vacuum once inflation is over. At any rate, high energy gamma rays could, as suggested in [@Starobinsky:2002rp], be another way to probe planckian physics in an expanding universe. One should also investigate how these effects might vary due to local variations in $H$.
Will the $\protect\alpha $-vacua thermalize?
============================================
As discussed in [@Bousso:2001mw][@Einhorn:2002nu][@Kaloper:2002cs] the $\alpha $-vacua are not thermal. One effect of this is that a detector in an $\alpha $-vacuum will not end up in thermal equilibrium. An equilibrium will be reached [@Einhorn:2002nu][@Kaloper:2002cs], but it will not be thermal in character and in general not obey the laws of detailed balance. For a detector with just two levels, equilibrium trivially implies detailed balance between the two levels even though the occupation numbers will be non standard. With three levels, however, the situation is more unusual. Instead of detailed balance between any two levels, there will be a net transition rate from, say, level one to level two, from level two to level three and then back to level one. This is a consequence of the non thermal nature of the background and does not imply any inconsistent physics of the detector. A question one can ask, however, is whether and how such a background can sustain itself. Will there be processes that act to thermalize the background?
In [@Kaloper:2002cs] it is argued based on holography and complementarity that this will indeed be the case. The main argument is that, according to a specific observer in de Sitter space, any perturbation will be redshifted as it approaches the de Sitter horizon and apparently boiled to pieces through the ever higher temperature that the perturbation experiences according to the observer. All perturbations will inevitably be thermalized. One would conclude from this that all traces of physics taking place on smaller scales than the horizon will be erased and the featureless and thermal Bunch-Davies modes are the only thing that remains. However, this is slightly problematic if we take a point of view based on the Robertson-Walker coordinates in (\[metric\]). We can now follow the modes through the de Sitter horizon and see how they freeze when they expand and become larger than the de Sitter radius. The analogue would be to follow an observer who ventures inside a black hole and find that nothing peculiar or dramatic happens near the horizon. And this seems to be the appropriate point of view to take when discussing fluctuations generated through inflation. Holography and complementarity are intriguing concepts that could be important for cosmology, but for this particular exercise they do not seem to be relevant. As observers of the CMBR we are more in the position of an observer inside a black hole rather than one on the outside.
A more conservative approach would be to predict the evolution of the fluctuations without thinking about holography. Any effect that would threaten to thermalize the $\alpha $-vacua must then be due to interactions. To be specific one could imagine a thermal background with a small perturbation that starts out at Planck scale. It is important to note that the temperature is very low with the thermal wave length of the order of the de Sitter radius itself. As a consequence a planckian mode is relatively safe in the beginning. As it redshifts, and its energy decreases towards the de Sitter temperature, the thermalization effects should become more important. This happens over a period of time a few times the inverse temperature. But at the same time the mode will expand past the horizon and freeze. One seems to conclude that the thermalization effects could be important in an interacting theory, but it is not clear that they will suppress any difference with respect to the Bunch-Davies vacuum by orders of magnitude. Much less wipe out any difference completely. Clearly more detailed work is needed to make a precise prediction.
Conclusions
===========
In this short note we have discussed a few of the problems of transplanckian physics that can be addressed in the simplified framework of $\alpha $-vacua. Our conclusion is that the peculiar image singularities of the Feynman propagator, as well as problems with non renormalization, do not necessarily imply the ruling out of transplanckian effects. A true inconsistency at this level would have had dramatic implications. Without observational input we would have been able to exclude all models of transplanckian physics that result in other vacua than the Bunch-Davies, including the non commutative examples discussed in, e.g., [@Easther:2001fi]. We do not believe that such strong conclusions can be drawn from work done so far. We have furthermore considered the way $\alpha $ might change. We do not see any fundamental difficulties in this respect but note that the work of [@Starobinsky:2002rp] shows that there are important constraints coming from present day physics. Issues of thermalization also need further investigation but we do not think that any argument based on holography put forward so far gives any meaningful constraints.
To summarize, we believe that the essence of the arguments of [@Kaloper:2002cs] is the following claim: *it is not consistent to assume a fixed high energy cut off where unknown transplanckian physics delivers states different form the Bunch-Davies vacuum.* An important point of the present paper is that possible problems are not visible in the cutoff theory. All problems must rely on claims about contributions from energy scales above the cutoff, energy densities etc., and speculations about transplanckian physics. Clearly, more work is needed to figure out how unorthodox transplanckian physics really can be.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author would like to thank Lars Bergström, Daniel Domert, Martin Olsson, Hector Rubinstein, Gary Shiu and Konstantin Zarembo for valuable discussions. The author is a Royal Swedish Academy of Sciences Research Fellow supported by a grant from the Knut and Alice Wallenberg Foundation. The work was also supported by the Swedish Research Council (VR).
[99]{} R. H. Brandenberger, “Inflationary cosmology: Progress and problems,” arXiv:hep-ph/9910410. J. Martin and R. H. Brandenberger, “The trans-Planckian problem of inflationary cosmology,” Phys. Rev. D **63**, 123501 (2001) \[arXiv:hep-th/0005209\]. J. C. Niemeyer, “Inflation with a high frequency cutoff,” Phys. Rev. D **63**, 123502 (2001) \[arXiv:astro-ph/0005533\]. R. H. Brandenberger and J. Martin, “The robustness of inflation to changes in super-Planck-scale physics,” Mod.Phys. Lett. A **16**, 999 (2001) \[arXiv:astro-ph/0005432\]. A. Kempf, “Mode generating mechanism in inflation with cutoff,” Phys. Rev. D **63**, 083514 (2001) \[arXiv:astro-ph/0009209\]. C. S. Chu, B. R. Greene and G. Shiu, “Remarks on inflation and noncommutative geometry,” Mod. Phys. Lett. A **16**, 2231 (2001) \[arXiv:hep-th/0011241\]. J. Martin and R. H. Brandenberger, “A cosmological window on trans-Planckian physics,” arXiv:astro-ph/0012031. L. Mersini, M. Bastero-Gil and P. Kanti, “Relic dark energy from trans-Planckian regime,” Phys. Rev. D **64**, 043508 (2001) \[arXiv:hep-ph/0101210\]. J. C. Niemeyer and R. Parentani, “Trans-Planckian dispersion and scale-invariance of inflationary perturbations,” Phys. Rev. D **64**, 101301 (2001) \[arXiv:astro-ph/0101451\]. A. Kempf and J. C. Niemeyer, “Perturbation spectrum in inflation with cutoff,” Phys. Rev. D **64**, 103501 (2001) \[arXiv:astro-ph/0103225\]. A. A. Starobinsky, “Robustness of the inflationary perturbation spectrum to trans-Planckian physics,” Pisma Zh.Eksp. Teor. Fiz. **73**, 415 (2001) \[JETP Lett. **73**, 371 (2001)\] \[arXiv:astro-ph/0104043\].
R. Easther, B. R. Greene, W. H. Kinney and G. Shiu, “Inflation as a probe of short distance physics,” Phys. Rev. D **64**, 103502 (2001) \[arXiv:hep-th/0104102\]. M. Bastero-Gil and L. Mersini, “SN1A data and CMB of Modified Curvature at Short and Large Distances,” Phys. Rev. D **65** (2002) 023502 \[arXiv:astro-ph/0107256\]. L. Hui and W. H. Kinney, “Short distance physics and the consistency relation for scalar and tensor fluctuations in the inflationary universe,” arXiv:astro-ph/0109107. R. Easther, B. R. Greene, W. H. Kinney and G. Shiu, “Imprints of short distance physics on inflationary cosmology,” arXiv:hep-th/0110226. M. Bastero-Gil, P. H. Frampton and L. Mersini, “Modified dispersion relations from closed strings in toroidal cosmology,” arXiv:hep-th/0110167. R. H. Brandenberger, S. E. Joras and J. Martin, “Trans-Planckian physics and the spectrum of fluctuations in a bouncing universe,” arXiv:hep-th/0112122. J. Martin and R. H. Brandenberger, “The Corley-Jacobson dispersion relation and trans-Planckian inflation,” arXiv:hep-th/0201189. J. C. Niemeyer, “Cosmological consequences of short distance physics,” arXiv:astro-ph/0201511. F. Lizzi, G. Mangano, G. Miele and M. Peloso, “Cosmological perturbations and short distance physics from noncommutative geometry,” arXiv:hep-th/0203099. G. Shiu and I. Wasserman, “On the signature of short distance scale in the cosmic microwave background,” arXiv:hep-th/0203113.
R. Brandenberger and P. M. Ho, “Noncommutative spacetime, stringy spacetime uncertainty principle, and density fluctuations,” arXiv:hep-th/0203119. S. Shankaranarayanan, “Is there an imprint of Planck scale physics on inflationary cosmology?,” arXiv:gr-qc/0203060. N. Kaloper, M. Kleban, A. E. Lawrence and S. Shenker, “Signatures of short distance physics in the cosmic microwave background,” arXiv:hep-th/0201158. R. H. Brandenberger and J. Martin, “On signatures of short distance physics in the cosmic microwave background,” arXiv:hep-th/0202142. S. F. Hassan and M. S. Sloth, “Trans-Planckian effects in inflationary cosmology and the modified uncertainty principle,” arXiv:hep-th/0204110.
U. H. Danielsson, “A note on inflation and transplanckian physics,” Phys. Rev. D **66**, 023511 (2002) \[arXiv:hep-th/0203198\]. R. Easther, B. R. Greene, W. H. Kinney and G. Shiu, “A generic estimate of trans-Planckian modifications to the primordial power spectrum in inflation,” arXiv:hep-th/0204129.
U. H. Danielsson, “Inflation, holography and the choice of vacuum in de Sitter space,” JHEP **0207**, 040 (2002) \[arXiv:hep-th/0205227\]. J. C. Niemeyer, R. Parentani and D. Campo, “Minimal modifications of the primordial power spectrum from an adiabatic short distance cutoff,” arXiv:hep-th/0206149. N. A. Chernikov and E. A. Tagirov, “Quantum theory of scalar field in de Sitter space-time,” Ann. Inst. Henri Poincaré, vol. IX, nr 2, (1968) 109.
E. Mottola, “Particle Creation In De Sitter Space,” Phys. Rev. D **31** (1985) 754.
B. Allen, “Vacuum States In De Sitter Space,” Phys.Rev. D **32** (1985) 3136.
R. Floreanini, C. T. Hill and R. Jackiw, “Functional Representation For The Isometries Of De Sitter Space,” Annals Phys. **175** (1987) 345.
R. Bousso, A. Maloney and A. Strominger, “Conformal vacua and entropy in de Sitter space,” arXiv:hep-th/0112218.
M. Spradlin and A. Volovich, “Vacuum states and the S-matrix in dS/CFT,” arXiv:hep-th/0112223.
T. Banks and L. Mannelli, “De Sitter vacua, renormalization and locality,” arXiv:hep-th/0209113. M. B. Einhorn and F. Larsen, “Interacting Quantum Field Theory in de Sitter Vacua,” arXiv:hep-th/0209159. N. Kaloper, M. Kleban, A. Lawrence, S. Shenker and L. Susskind, arXiv:hep-th/0209231.
M. K. Parikh, I. Savonije and E. Verlinde, “Elliptic de Sitter space: dS/Z(2),” arXiv:hep-th/0209120. E. Witten, “Quantum gravity in de Sitter space,” arXiv:hep-th/0106109. A. A. Starobinsky and I. I. Tkachev, “Trans-Planckian particle creation in cosmology and ultra-high energy cosmic rays,” arXiv:astro-ph/0207572. K. Goldstein and D. A. Lowe, arXiv:hep-th/0208167.
[^1]: Note however recent work, [@Parikh:2002py], where other possibilities to handle this problem are discussed.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Pauli spin blockade (PSB) is a significant physical effect in double quantum dot (DQD) systems. In this paper, we start from the fundamental quantum model of the DQD with the electron-electron interaction being considered, and then systematically study the PSB effect in DQD by using a recently developed non-perturbative method, the hierarchical equations of motion (HEOM) approach. The physical picture of the PSB is elucidated explicitly and the gate voltage manipulation is described minutely, which are both qualitatively consistent with the experimental measurements. When dot-dot exchange interaction is involved, the PSB effect may be lifted by the strong antiferromagnetic exchange coupling.'
author:
- WenJie Hou
- Dong Hou
- YuanDong Wang
- JianHua Wei
- YiJing Yan
title: 'Non-perturbative Theory of Pauli Spin Blockade'
---
\[Intro\]INTRODUCTION
=====================
Pauli spin blockade (PSB) is an important physical effect appears in double quantum dot (DQD) systems, which was discovered experimentally in vertically coupled GaAs/AlGaAs DQD as early as 2002 [@ono2002vertical]. The basic picture is that the hopping of electrons between two dots will be influenced by their spin configuration if the total excess electron number of the system is 2 with occupation state $(N_1, N_2)=(2,0)$, (1,1) or (0,2), as a consequence, the current - voltage($I-V$) curve will show a rectification behaviour. Obviously, the PSB is caused by the universal Pauli’s exclusion principle. It receives extensive studies in various quantum dot systems with different structures from vertical to lateral dots [@johnson2005lateral] and from double to three dots [@busl2013bipolar], as well as in dots with different semiconductor materials from GaAs/AlGaAs to Si [@liu2008silicon]. Recently, the PSB has been used to fabricate and readout the singlet-triplet spin qubit, which will promote the development of the quantum information [@hao2014electron].
Some important characters in the PSB regime have been investigated by virous theoretical groups. Those include: 1) the correlation between the PSB effect and occupation of the two-electron triplet state [@fransson2006pauli]; 2) the dynamical nuclear spin polarization by hyperfine interaction [@deng2005nuspin]; 3) the nonthermal broadening effect of tunneling current [@Kuo2011raeffect]; 4) the leakage-current line shapes from inelastic cotunneling [@Coish2011cotunneling]; 5) the spin-flip phonon-mediated charge relaxation in double quantum dots [@Danon2013crelax]; and 6) the PSB and the ultrasmall magnetic field effect in organic magnetoresistance [@Danon2013UMag]. The Pauli master equation (PME) with second order Fermi¡¯s golden rule is the main approach in above works to archive the transition rates. Other approaches (such as nonequilibrium Green¡¯s functions) are not yet so popular in literatures [@Stepanenko2012stsplit; @Hsieh2012eprop; @Amaha2014trielec].
We would like to comment that the PME is not accurate enough for the PSB theory. Firstly, the DQD is a typical quantum open system with infinity degree of freedoms of the total density matrix, while the PME only concerns the diagonal terms of the reduced density matrix and treats the dot-electrode couplings by low-order perturbation schemes; and secondly, the DQD is also a typical strongly correlated system with infinity degree of freedoms of the electron-electron ([*e-e*]{}) interactions, while the PME either neglects this important interaction or treats it in the single electron level.
Obviously, for the theoretical study on such fundamental physics processes as the PSB, a non-perturbative approach is highly required to deal with the basic quantum model involving the [*e-e*]{} interactions. The hierarchical equations of motion (HEOM) approach we newly developed can meet this requirement, which nonperturbatively resolves the combined effects of dot-electrode dissipation, [*e-e*]{} interactions, and non-Markovian memory [@Jin08234703; @Li2012dresponse; @2008jcp184112; @2009jcp124508; @2008njp093016; @2009jcp164708; @Zhe121129; @2015Cheng033009]. In this paper, we start from the Anderson multiple impurity model to describe the DQD, fully considering the [*e-e*]{} interaction and the dot-electrode couplings. By using the HEOM approach, we deal with this quantum model non-perturbatively to accurately obtain some observations, such as the spectral function, occupancy of electron spin and current, etc. Our theory not only can reveal the physical picture of the PSB clearly but also can elucidate its dependence on various parameters such as the gate voltages, dot-dot coupling, and exchange correlation between spins in different dots. Besides, the external field manipulation is also convenient to be involved in our theory.
The paper is organized as follows. In Sec.II we briefly review our model and the non-perturbative HEOM approach. In Sec.III, we present our accurate solutions relating to the PSB effect, those results include: III.A. the physical picture of the PSB; III.B. the gate voltage manipulation of the PSB; and III.C. the lift of the PSB by the dot-dot exchange interaction. In Sec.IV we give the summary of our work.
\[Theo\]THEORY AND FORMULAS
===========================
The HEOM is a general formula for the quantum open systems composed of three parts: the system (quantum dots here), the bath (two electrodes here) and the system-bath couplings. Let us introduce the total Hamiltonian (Anderson impurity model) for the DQD as follows, $$\begin{aligned}
\label{ha}
H_{T}=H_{S}+H_{B}+H_{SB}\end{aligned}$$ where $H_{S}$ is the Hamiltonian for the two coupled dots $$\begin{aligned}
\label{hs}
H_{S}=\sum_{i=1,2\sigma} \epsilon_{i\sigma}\hat{a}^\dag_{i\sigma}\hat{a}_{i\sigma} + \frac{U}{2}\sum_{i=1,2\sigma} n_{i\sigma}n_{i\bar{\sigma}}
{\nonumber \\}+t\sum_{\sigma}(\hat{a}^\dag_{1\sigma}\hat{a}_{2\sigma}+\text{H.c.})
\end{aligned}$$ here $\epsilon_{i\sigma}$ indicates the on-site energy of the electron with spin $\sigma$ ($\sigma=\uparrow,\downarrow$) on dot $i (i=1,2)$, $\hat{a}_{i\sigma}^\dag$ and $\hat{a}_{i\sigma}$ correspond the creation and annihilation operators for an electron with spin $\sigma$. $n_{i\sigma}=\hat{a}^\dag_{i\sigma}\hat{a}_{i\sigma}$ is the electron number operator of dot $i$, and $U$ is the Coulomb interaction between electrons with spin $\sigma$ and $\bar{\sigma}$ (opposite spin of $\sigma$) within one dot. $t$ is the inter-dot coupling, determined by the overlapping integral of electron wave functions. $\text{H.c.}$ stands for the Hermitian conjugate.
For brevity, in what follows, we use the symbol $\mu$ to denote the electron orbital (including spin, space, *etc.*) in the system , i.e., $\mu=\{{\sigma},i...\}$. The Hamiltonian of the electrodes is described as a noninteracting Fermi bath, $$\begin{aligned}
\label{hb}
H_{B}=\sum_{k\mu\alpha=L,R}\epsilon_{k\alpha}\hat{d}^\dag_{k\mu\alpha}\hat{d}_{k\mu\alpha}
\end{aligned}$$ with $\epsilon_{k\alpha}$ being the energy of an electron with wave vector $k$ in the $\alpha$ lead, and the $\hat{d}^\dag_{k\mu\alpha}$($\hat{d}_{k\mu\alpha}$) corresponding creation (annihilation) operator for an electron with the $\alpha$-reservoir state $|k\rangle$ of energy $\epsilon_{k\alpha}$. The dot-electrode coupling Hamiltonian is $$\begin{aligned}
\label{hd}
H_{SB}=\sum_{\mu}[f^\dag_{\mu}(t)\hat{a}_{\mu} +\hat{a}^\dag_{\mu}f_{\mu}(t)]
\end{aligned}$$ in the bath interaction picture. Here, $f^\dag_{\mu}=e^{iH_{B}t}[\sum_{k\alpha}t^{*}_{\alpha k
\mu}\hat{d}^\dag_{k\mu\alpha}]e^{-iH_{B}t}$ is stochastic interactional operator and satisfies the Gauss statistics with $t_{\alpha k\mu}$ denoting the transfer coupling matrix element. The influence of electrodes on the dots is considered through the hybridization functions with a Lorentzian form, $\Delta_{\alpha}({\omega})\equiv\pi\sum_{k} t_{\alpha k\mu}t^\ast_{\alpha k\mu} \delta({\omega}-{\epsilon}_{k\alpha})=\Delta
W^{2}/[2({\omega}-\mu_{\alpha})^{2}+W^{2}]$, with $\Delta$ being the effective impurity-lead coupling strength, $W$ being the band width, and $\mu_{\alpha}$ being the chemical potentials of the $\alpha$ lead.
Obviously, [Eq.]{} is a strongly correlated Hamiltonian with infinite degree of freedoms, which is hard to exactly solve by the the Schrödinger equation directly. Fortunately, we can derive the accurate HEOM for the reduced density matrix (together with the auxiliary ones) from the basic path integral equations (influence functional theory) without take any approximations [@Jin08234703]. The HEOM that governs the dynamics of the DQD takes the form of $$\begin{aligned}
\label{HEOM}
\dot\rho^{(n)}_{j_1\cdots j_n} =& -\Big(i{\cal L} + \sum_{r=1}^n \gamma_{j_r}\Big)\rho^{(n)}_{j_1\cdots j_n}
-i \sum_{j}\! {\cal A}_{\bar j}\, \rho^{(n+1)}_{j_1\cdots j_nj}
{\nonumber \\}&
-i \sum_{r=1}^{n}(-)^{n-r}\, {\cal C}_{j_r}\,
\rho^{(n-1)}_{j_1\cdots j_{r-1}j_{r+1}\cdots j_n}\end{aligned}$$ where the $n$th-order auxiliary density operator $\rho^{(n)}$ can be defined via auxiliary influence functional $\mathcal{F}^{(n)}_{\textbf{j}}$ as $$\begin{aligned}
\label{HEOM}
\rho^{(n)}_{\textbf{j}}(t)\equiv \mathcal{U}^{(n)}_{\textbf{j}}(t,t_{0})\rho(t_{0})\end{aligned}$$ with the reduced Liouville-space propagator, $$\begin{aligned}
\label{HEOM}
\mathcal{U}^{(n)}_{\textbf{j}}(\psi,t;\psi_{0},t_{0}) \equiv \int^{\psi[t]}_{\psi_{0}[t_{0}]}\mathcal{D}\psi e^{i\mathcal{S}[\psi]} \mathcal{F}^{(n)}_{\textbf{j}}[\psi] e^{-i\mathcal{S}[\psi^{'}]}\end{aligned}$$ $\mathcal{S}[\psi]$ is the classical action functional of the reduced system. The definition of the auxiliary influence functional $\mathcal{F}^{(n)}_{\textbf{j}}$ together with its equations is referred to in [@Jin08234703].
We denote $\textbf{j}=\{j_1\cdots j_n\}$ and $\textbf{j}_{r}=\{j_1\cdots j_{r-1}j_{r+1}\cdots j_n\}$, the action of superoperators respectively is $$\begin{aligned}
\label{HEOM}
{\cal A}_{\bar j}\, \rho^{(n+1)}_{\textbf{j}j} = a^{\bar{o}}_{\mu} \rho^{(n+1)}_{\textbf{j}j}
+(-)^{n+1} \rho^{(n+1)}_{\textbf{j}j}a^{\bar{o}}_{\mu}\end{aligned}$$ $$\begin{aligned}
\label{HEOM}
{\cal C}_{j_r}\,\rho^{(n-1)}_{\textbf{j}_{r}} = \sum_{\nu} \{{\cal C}^{o}_{\alpha\mu\nu} a^{o}_{\nu}\rho^{(n-1)}_{\textbf{j}_{r}}
&
-(-)^{n-1}\, {\cal C}^{\bar{o}}_{\alpha\nu\mu}\,\rho^{(n-1)}_{\textbf{j}_{r}}a^{o}_{\nu}\}\end{aligned}$$ In this formalism, $a_{\mu}^{o}$ ($a_{\mu}^{\bar{o}}$) corresponds the creation (annihilation) operator for an electron with the $\mu$ electron orbital. The reduced system density operator $\rho^{(0)}(t) \equiv {\rm tr}_{B}[\rho_{{{\rm total}}}(t)]$ and auxiliary density operators $\{\rho^{(n)}_{j_1\cdots j_n}(t); n=1,\cdots,L\}$ are the basic variables, here $L$ denotes the terminal or truncated tier level. The Liouvillian of dots, $\mathcal{L}\,\cdot \equiv \hbar^{-1}[H_{{\text{\scriptsize sys}}}, \cdot\,]$, contains the [*e-e*]{} interactions. The index $j \equiv (o\mu m)$ corresponds to the transfer of an electron to/from ($o=+/-$) the impurity state $|\sigma\rangle$, associated with the characteristic memory time $\gamma_m^{-1}$. The correlation function ${\cal
C}^{o}_{\alpha\mu\nu}(t-\tau)=\langle
f^{o}_{\alpha\mu}(t)f^{\bar{o}}_{\alpha\nu}(\tau)\rangle_{B}$ follows immediately the time-reversal symmetry and detailed-balance relations.
We set the initial total system at equilibrium where $\mu_{\alpha}=\mu^{eq}=0$. The system will leave equilibrium after applying a voltage to the left (L) and right (R) leads, and there will be a current flowing into the $\alpha$-lead $I_{\alpha}(t)$ $$\begin{aligned}
\label{hd}
I_{\alpha}(t)=i\sum_{\mu}\mathrm{tr}_{s}[{\rho^\dag_{\alpha \mu}(t)\hat a_{\mu} -\hat a^\dag_{\mu}\rho^-_{\alpha \mu}(t)}]
\end{aligned}$$Here, $\rho^\dag_{\alpha \mu}=(\rho^-_{\alpha \mu})^\dag$ is the first-tier auxiliary density operator obtained by solving Eq.(5). Through the extended Meier-Tannor parametrization method and multiple-frequency-dispersed hierarchy construction, we can achieved the closed HEOM formalism [@Jin08234703]. As a result, the current from $L$ to $R$ lead can be denoted $I(t)=I_{L}(t)=-I_{R}(t)$.
For evaluation of dynamical variables of the DQD system, we focus on correlation function between two arbitrary dynamical operators, $\widetilde{C}_{AB}(t) \equiv {\langle}\hat A(t)\hat B(0){\rangle}=
{\rm tr}_{{{\rm total}}}[\hat A(t)\hat B(0) \rho^{\rm eq}_{{{\rm total}}}(T)]$. Here, the Heisenberg operators and thermal equilibrium density operator $\rho^{\rm eq}_{{{\rm total}}}(T)$ are all defined in the total space. A linear response theory for quantum open systems [@Wei2011arxiv; @Li2012dresponse] has been established, based on which $\widetilde{C}_{AB}(t)$ is retrieved exactly within the HEOM framework. Let $C_{AB}({\omega}) \equiv \frac{1}{2}\int dt\, e^{i{\omega}t}
\widetilde{C}_{AB}(t)$, which satisfies the detailed balance relation of $C_{BA}(-\omega)=e^{-\hbar\omega/k_B T}C_{AB}(\omega)$. The system spectral function is obtained as $J_{AB}({\omega}) \equiv \frac{1}{2\pi} \int dt\, e^{i {\omega}t}
\langle \{ \hat A(t), \hat B(0) \} \rangle
= \frac{1}{\pi} \left(1 + e^{-\hbar\omega/k_B T}\right) C_{AB}({\omega})$. With $\hat A=\hat a_{\mu}$ and $\hat B=\hat a^{{\dagger}}_{\mu}$, it recovers the spectral function of the impurity state $\mu$, *i.e.*, $A_{\mu}(\omega) \equiv J_{\hat a_{\mu}\hat
a^{{\dagger}}_{\mu}}(\omega) = -\frac{1}{\pi}{\rm Im}\,
G_{\mu\mu}(\omega)$. Here, $G_{\mu\mu}(\omega)$ is the retarded Green’s function.
In our calculations, we treat the results as converging if the errors in numerical results of each element of the density matrix or the matrix of spectral function between the truncation $L=N$ and $L=N+1$ are less than $5\%$, then sufficiently accurate current will be output. In the follow calculations we adopt $L=4$ .
The main advantages of the HEOM approach applying to the DQD systems are as follows: 1) the HEOM theory is established based on the Feynman-Vernon path-integral formalism, in which all the system-bath correlations are taken into consideration; 2) the HEOM method is nonperturbative. In principle, the HEOM formalism is formally exact for noninteracting electron reservoirs. It also resolve nonperturbatively the combined effects of *e-e* interactions; 3) the HEOM is a high-accuracy numerical approach. It has the ability to achieve the same level of accuracy as the latest high-level NRG method [@Li2012dresponse]. Its main disadvantage lies in the increasing computational cost as the system temperature decreases.
\[Rest\]RESULTS AND DISCUSSION
==============================
\[Pict\]Physical picture of Pauli spin blockade
-----------------------------------------------
The spin degeneracy in [Eq.]{} is not convenient for the detailed analysis of the PSB, thus we theoretically lift the spin degeneracy in dot 1 by means of a local magnetic field $B_{1}$ applied onto it, with its direction paralleling to the down spins. $B_{1}$ is chosen to be strong enough to make the energy level of up-spin electrons $\epsilon_{1\uparrow}$ much higher than $\epsilon_{1\downarrow}$. In experiments, a local-like inhomogeneous Zeeman field can be archived by a novel split micromagnet [@Brun2011dqbit]. By adjusting gate voltage $V_{1}$, we then set $\epsilon_{1\downarrow}$ at the equilibrium Fermi level $E_{F}$ ($E_{F}=\mu_{L}=\mu_{R}$ at zero-bias). Under such condition, $E_{F}$ coincides with the center of the peak of the transition for down-spin electrons in dot 1 to jump from the zero occupied level to single occupied one, as the spectral function $A_{1\downarrow} (\omega)$ shown in [Fig.\[fig1\]]{}(a). This kind of single-occupied transition peak of $A_{1\uparrow} (\omega)$ is pushed far higher than $E_{F}$ by the large $B_{1}$, which makes the up-spin current negligibly small, thus $A_{1\uparrow} (\omega)$ is not shown in figures \[see [Fig.\[fig1\]]{}(a)-(c)\] but its little contribution to the total current is still counted \[see [Fig.\[fig1\]]{}(d)\].
Without the coupling $t$ between dot 1 and 2, the local $B_{1}$ should take no direct effects on the spins in dot 2. By adjusting gate voltage $V_{2}$, we set the double-occupied transition (from the single occupied level to double occupied one) peaks of $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$ to coincide with $E_{F}$ at $t\sim0$, as shown in [Fig.\[fig1\]]{}(a). Proceeding from this set-up, we gradually adjust the inter-dot coupling strength and other parameters to elaborate the physical picture of the PSB in details.
![(Color online). In the case of the spin non-degeneracy in dot 1, (a) equilibrium($V=0$) spectral function $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$; (b) nonequilibrium spectral function $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$ at positive bias($V=+2$ mV);(c) nonequilibrium spectral function $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$ at negative bias ($V=-2$ mV); and (d) the dependence of the occupation numbers $N_{1\downarrow}$, $N_{2\uparrow}$, $N_{2\downarrow}$ as well as the total current $I$ on the bias voltage $V$. The parameters adopted are $\varepsilon_{1\uparrow}=\varepsilon_{1\downarrow}=\varepsilon_{2\uparrow}
=\varepsilon_{2\downarrow}=-1.0$, $U_{1}=U_{2}=2.0$, $V_{g1}=-2.5$, $V_{g2}=1.0$, $W_{L}=W_{R}=4.0$, $\Gamma_{L}=\Gamma_{R}=2.0$, $\Delta E_{B1}=1.5$, $T=0.1$ and $t=0.001$ (in unit of meV).[]{data-label="fig1"}](fig1){width="3.5in"}
We start our study on the PSB from the following parameters: the potential energy $\varepsilon_{1\uparrow}=\varepsilon_{1\downarrow}=\varepsilon_{2\uparrow}
=\varepsilon_{2\downarrow}=-1.0$ meV; on-site [*e-e*]{} interaction strength $U_{1}=U_{2}=2.0$ meV; gate voltage $V_{g1}=-2.5$ meV and $V_{g2}=1.0$ meV. The left and right electrodes are chosen to be symmetric, with their DOS being Lorentzian-type; bandwidth being $W_{L}=W_{R}=4.0$ meV and the electrode-dot coupling being $\Gamma_{L}=\Gamma_{R}=2.0$ meV. The Zeeman splitting energy caused by the local magnetic field $B_{1}$ is $\Delta E_{B1}=1.5$ meV; and the temperature $T=0.1$ meV. The dot-dot coupling in [Fig.\[fig1\]]{} is near zero, $t=0.001$ meV.
The spectral functions $A(\omega)$ shown in [Fig.\[fig1\]]{}(a) is calculated by HEOM at the equilibrium state (bias voltage $V=0$), where only $A_{1\downarrow}(\omega)$ is shown in dot 1. [Fig.\[fig1\]]{}(b) and (c) show nonequilibrium spectral functions $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$ at positive ($V=+2$ mV) and negative ($V=-2$ mV) bias, respectively. [Fig.\[fig1\]]{}(d) depicts the dependence of the occupation numbers $N_{1\downarrow}$, $N_{2\uparrow}$, $N_{2\downarrow}$ as well as the total current $I$ on the bias voltage $V$ ($N-V/I-V$ curves), where $N$ is obtained from the diagonal elements of the reduced density matrix, which is also equal to the weighted integral of corresponding $A(\omega)$ under the Fermi level at the steady state, $\rho=\frac{1}{2\pi}\int A(\omega)f(\omega)d\omega$.
As shown in [Fig.\[fig1\]]{}, without the bias being applied ($V=0$), $N_{1\downarrow}\approx0.5$, $N_{2\uparrow}=N_{2\downarrow}\approx0.7$. The fractional charges results from the following two reasons: 1) the value of the Fermi-Dirac function $f(\omega)$ in the vicinity of $E_F$ continuously changes from 0 to 1 rather than a integer; and 2) the spectral function $A(\omega)$ near $E_F$ shows a finite-width peak structure broadened by the dot-electrode interaction, thus the weighted integral value below the Fermi surface is less than 1. The total electron number in the two-dot system is $Q_{T}=N_{1\uparrow}+N_{1\downarrow}+N_{2\uparrow}+N_{2\downarrow}\approx2.0$ $(N_{1\uparrow}\sim0)$, which remains almost constant even in the non-equilibrium (steady-state) transport process if the external (electric and/or magnetic) field is not too large.
Then, the symmetrical positive ($\mu_{L}\rightarrow \mu_{L}+eV/2;\mu_{R}\rightarrow \mu_{R}-eV/2$) and negative ($\mu_{L}\rightarrow \mu_{L}-eV/2;\mu_{R}\rightarrow \mu_{R}+eV/2$) bias are applied to the configuration shown in [Fig.\[fig1\]]{}(a). From the changes of $A(\omega)$, $N$ and $I$ with $V$ shown in [Fig.\[fig1\]]{}(b)-(d), we can see that the spectral function and occupation number of up-spin and down-spin electrons keep degenerate in dot 2, and neither positive nor negative bias can lift the degeneracy. The reason lies in the very small coupling between two dots, $t\sim 0$.
The $I-V$ curve \[see [Fig.\[fig1\]]{}(d)\] shows distinct asymmetric behavior that the steady value of the current at $V<0$ is much larger than that at $V>0$, actually the former is almost twice of the latter. However, such asymmetry is not the same as the rectification since the positive steady current is not small enough (comparing to the negative one ) to define a blockage effect. That result can be explained as follows:
In principle, the dot-to-dot electron transfer can induce an antiferromagnetic exchange between them with the strength $J_{AF}\sim 4t^2/U$. Enough strong $J_{AF}$ will lock the ground state of the isolate double dot system into a spin singlet state $S(1,1)$ whose energy is lower than the triplet state $T(1,1)$ in the order of $J_{AF}$. In [Fig.\[fig1\]]{}, $t\sim 0$ means $J_{AF}\sim 0$, thus states of $S(1,1)$ and $T(1,1)$ are nearly degenerate at $V=0$.
When positive bias applied, $\mu_{R}<0$, as shown in [Fig.\[fig1\]]{}(d), $N_{2\uparrow}$ and $N_{2\downarrow}$ decreases gradually with $\mu_{R}$ decreasing . The relation of $N_{2\uparrow}\approx N_{2\downarrow}$ at $V=0$ will keep unchanged at $V>0$, and the number will reach a steady value after $V>1$ mV, $N_{2\uparrow}\approx N_{2\downarrow}\approx 0.5$. At the same time, $N_{1\downarrow}$ increases gradually with $\mu_{L}(\mu_{L}>0)$ increasing and tends to $1.0$ after $V>1$ mV. The nonequilibrium spectral function at $V=+2$ mV in [Fig.\[fig1\]]{}(b) further confirms the above process. It means that the states $S(1,1)$ and $T(1,1)$ remain degenerate at $V>0$, and each of them has $50\%$ probability after $V>1$ mV. It is believed that electron spin is conserved through direct hopping process,thus the initial state $T(1,1)$ can only transfer to $T(0,2)$, if one electron is driven from dot 1 to dot 2 by the positive bias. However, the state $T(0,2)$ is not permitted to exist due to the Pauli’s exclusion principle (the exception from the excess freedom such as the orbital or valley not considered in the present work). Therefore, only half of the total initial states \[$S(1,1)$\] can contribute to the transport current at $V>0$ via the transition $S(1,1)\rightarrow S(0,2)$.
As shown in [Fig.\[fig1\]]{}(d), the negative bias makes $\mu_{R}>0$, then $N_{2\uparrow}$ and $N_{2\downarrow}$ increases gradually with $\mu_{R}$ increasing but remains $N_{2\uparrow}\approx N_{2\downarrow}$ which is close to $1.0$ after $V<-1$ mV. Meanwhile, $N_{1\downarrow}$ decreases gradually with $\mu_{L}(\mu_{L}<0)$ decreasing and approaches to $0$ after $V<-1$ mV. The nonequilibrium spectral function at $V=-2$ mV in [Fig.\[fig1\]]{}(c) further confirms the above process. It suggests that the negative bias will stabilize the $S(0,2)$ state which $100\%$ contributes to the current, since the Pauli’s exclusion principle takes no effect on the transition from $S(0,2)$ to $S(1,1)$. That argument is verified by the $I-V$ curve in [Fig.\[fig1\]]{}(d), which shows that the steady current at $V<0$ is almost twice larger than that at $V>0$.
![(Color online). In the case of the spin non-degeneracy in dot 1, (a) equilibrium($V=0$) spectral function $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$; (b) nonequilibrium spectral function $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$ at positive bias($V=+2$ mV);(c) nonequilibrium spectral function $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$ at negative bias ($V=-2$ mV); and (d) the dependence of the occupation numbers $N_{1\downarrow}$, $N_{2\uparrow}$, $N_{2\downarrow}$ as well as the total current $I$ on the bias voltage $V$. The parameters adopted are the same as those in [Fig.\[fig1\]]{}, except $t=0.08$ meV here.[]{data-label="fig2"}](fig2){width="3.5in"}
We then increase the inter-dot coupling strength $t$ from $0.001$ to $0.08$ meV and keep other parameters unchanged for the purpose of comparison. The calculated $A(\omega, V=0)$, $A(\omega, V=+2$ meV$)$, $A(\omega, V=-2$ meV$)$ and the $N-V/I-V$ curve are depicted in [Fig.\[fig2\]]{}(a)-(d) respectively. From [Fig.\[fig2\]]{}(a) for the case of $V=0$, we can see that the degeneracy of $S (1,1)$ and $T(1,1)$ has been lifted by the $t$-induced antiferromagnetic interaction $J_{AF} (\sim 4t^{2}/U)$. Since the energy of $S(1,1)$ is lower than that of $T(1,1)$, $A_{2\uparrow}(\omega)$ moves downward and $N_{2\uparrow}$ increases from $\sim 0.7$ at $t\sim 0$ to $\sim 0.72$. Meanwhile, $N_{2\downarrow}$ decreases from $\sim 0.7$ to $\sim 0.68$, as shown in [Fig.\[fig2\]]{}(d).
The PSB will take place when the positive bias is applied to the set-up shown in [Fig.\[fig2\]]{}(a). At $V>0$, $N_{2\uparrow}$ and $N_{2\downarrow}$ changes to different directions \[see [Fig.\[fig2\]]{}(d)\] instead of synchronous varying at $t\sim 0$ \[cf. [Fig.\[fig1\]]{}(d)\]. As shown in [Fig.\[fig2\]]{}(d), $N_{2\uparrow}$ decreases rapidly as $V$ positively increasing and tends to $0$ after $V>1.0$ mV; whereas $N_{2\downarrow}$ increases and approaches $1.0$ after $V>1.0V$ mV. As for $N_{1\downarrow}$, it gradually increases from $0.5$ to $1.0$ with increasing $V$, much like the change of $N_{2\uparrow}$. As a result of above changes, only state $T(1,1)$ is retained under positive bias and $S(1,1)$ will no longer exist after $V>1.0$ mV. As already explained, $T(1,1)$ can not transfer to $T(0,2)$ to create any current due to the Pauli’s exclusion principle, thus PSB occurs naturally. The forbidden electron transition from dot 1 to 2 outputs near zero current, as the $I-V$ curve shows in [Fig.\[fig2\]]{}(d). That is exactly the PSB effect observed in experiments at a moderate coupling strength $t$, and the finite current in the interval of $V\in[0,1.0$ mV$]$ is the so-called leakage current corresponding to the process of $S(1,1)$ being depleted gradually. After $V>1.0$ mV, the current enters its total-blocked zone with a near zero value.
When negative bias applied as shown in [Fig.\[fig2\]]{}(d), with the increase of $\mu_{R}$, $N_{2\uparrow}$ increases gradually and becomes saturated at $N_{2\uparrow}\sim 1.0$ after the bias $V<-1.0$ mV, while $N_{2\downarrow}$ decreases and keeps at about $ 0.5+\delta$ ($\delta\approx0.13$) after $V<-1.0$ mV. At the same time, $N_{1\downarrow}$ decreases with $\mu_{L}$ decreasing and maintains $N_{1\downarrow}\approx 0.5-\delta$ after $V<-1.0$ mV. Although $N_{1\downarrow}>0$ in this case, it won’t cause any PSB effect of the down-spin electrons, for the reason that the $\mu_{L}$ below the Fermi surface can provide enough space to accept the electrons transferring from dot 2 to 1, as the nonequilibrium spectral function shown in [Fig.\[fig2\]]{}(c). As a consequence, the considerable current is output in the $I-V$ curve. It should be noted that the unit of current in [Fig.\[fig2\]]{}(d) is pA instead of nA in [Fig.\[fig1\]]{}(d).
Summarizing [Fig.\[fig1\]]{} and \[fig2\], our theory appropriately describes the physical mechanism and picture of the PSB effect in DQD systems, by means of a strong local magnetic field applied onto dot 1 to lift its spin degeneracy. We are now on the position to elucidate the PSB effect under more general conditions. In [Fig.\[fig3\]]{}, we depict the $N-V/I-V$ curves at $t=0.001$ meV \[[Fig.\[fig3\]]{}(a)\] and $0.08$ meV \[[Fig.\[fig3\]]{}(b)\], where the other parameters are the same as those in [Fig.\[fig1\]]{} and [Fig.\[fig2\]]{} accordingly, except that the local magnetic field is absent in both cases, i.e. $B_1=0$. By comparing the $I-V$ curve in [Fig.\[fig3\]]{}(a) to that in [Fig.\[fig1\]]{}(d), we can see that the additional up-spin channel barely changes the current at $V>0$ but will increase that at $V<0$ if the dot-dot coupling is very weak. The reason for the former lies in the equally dividing of $N_{1\downarrow}$ in [Fig.\[fig1\]]{}(d) by $N_{1\uparrow}$ and $N_{1\downarrow}$ in [Fig.\[fig3\]]{}(a) and the current changing little. The reason for the latter is that the up- and down-spin channels contribute to the current independently in the limit of $t\approx 0$, thus the current will be enhanced by additional channels. If distinct PSB effect occurs, the situation will become much different. As indicated by the comparison of the $I-V$ curve in [Fig.\[fig3\]]{}(b) to that in [Fig.\[fig2\]]{}(d), the additional up-spin channel hardly changes the current either in the PSB region ($V>0$) or in the conductive region ($V<0$). By analyzing the corresponding $N-V$ curves and nonequilibrium spectral functions (the figures not shown), we find in the PSB region, the probability of $T(\downarrow, \downarrow)$ in [Fig.\[fig2\]]{}(d) is equally divided into $T(\uparrow,\uparrow)$ and $T(\downarrow,\downarrow)$ in [Fig.\[fig3\]]{}(b), and the current keeps its value. In the conductive region, the single-spin transport channel $S(0,\uparrow \downarrow)\rightarrow S(\downarrow,\uparrow)$ is equally divided by the degenerate double-spin ones, $S(0,\uparrow \downarrow)\rightarrow \frac{1}{2}S(\downarrow,\uparrow)+\frac{1}{2}S(\uparrow,\downarrow)$, and the total current remain unchanged.
![(Color online). Without the local magnetic field applied ($B_1=0$), the dependence of the occupation numbers$N_{1\uparrow}$, $N_{1\downarrow}$, $N_{2\uparrow}$, and $N_{2\downarrow}$ as well as the total current $I$ on the bias voltage $V$ at $t=0.001$ meV (a) and $0.08$ meV (b). The other parameters in (a) and (b) are the same as those in [Fig.\[fig1\]]{} and [Fig.\[fig2\]]{} accordingly.[]{data-label="fig3"}](fig3){width="2.3in"}
\[Gmodu\]Gate voltage modulation on PSB
---------------------------------------
In experiments on PSB, the gate voltages $V_{1}$ and $V_{2}$ are two important parameters which respectively manipulate the on-site energy of dot 1 and dot 2. We thus theoretically investigate the variation of the PSB with them and summarize the results in [Fig.\[fig4\]]{}, where the dot-dot coupling strength is chosen as $t=0.05$ meV, a small value but large enough to induce the PSB effect, with the purpose of making the boundaries shown in [Fig.\[fig4\]]{} clear and distinguishable. [Fig.\[fig4\]]{} (a) depicts the positive current at $V=0.4$ mV changing with respect to the parameters in the $V_{1}-V_{2}$ plane, in the form of the 3D colormap surface image together with the 2D bottom contour projection. The other parameters are the same as those in [Fig.\[fig3\]]{}, except both $V_{1}$ and $V_{2}$ are variables now within the range of $[-U,U]$($U=2$ meV). In the 2D projection image, we schematically mark off the boundary of the stability diagrams by dotted lines. Actually, the quadrangles shown in the figure should changes to hexagon at finite $t$, however, the boundary line is hard to accurately determine in theory. The schematic stability diagrams shown in [Fig.\[fig4\]]{} is just for reference purposes.
![(Color online). 3D colormap surface image with the 2D bottom contour projection of the current (absolute value) as a function of the gate voltage $V_{1}$ and $V_{2}$. The bias voltages used are (a) $V=0.4$ mV, and (b) $V=-0.4$ mV. In the 2D projection image, the dotted lines schematically mark off the boundary of the stability diagrams. []{data-label="fig4"}](fig4){width="2.5in"}
[Fig.\[fig4\]]{} clearly shows how ($V_{1},V_{2}$) modulate ($N_1,N_2$) and the PSB effect. The DQD system remains the stability charge-occupied state (1,1) within $-U/2\leq V_{1}\leq U/2$ and $-U/2\leq V_{2}\leq U/2$, and no current occurs in the center of this area due to the Coulomb blockade. However, finite current may be output at the four top corners of (1,1) state, resulting from the charge transferring in and out, which can be listed as $(0,0)\rightleftharpoons (1,1), (2,0)\rightleftharpoons (1,1),(1,1)\rightleftharpoons (2,2)$, and $(1,1)\rightleftharpoons (0,2)$ (clockwise from lower left). By referring the figure, one can see that the current at the corners of $(0,0)\rightleftharpoons (1,1)$ and $(1,1)\rightleftharpoons (2,2)$ is symmetric about the bias voltage with no PSB effect occurring. In addition, the current at above two corners also shows symmetric behaviors along the diagonal line ($V_{1}=V_{2}$), which comes from the electron-hole symmetry satisfied by our Hamiltonian, [Eq.]{}. As shown in [Fig.\[fig4\]]{}, current at the corners of $(2,0)\rightleftharpoons (1,1)$ and $(1,1)\rightleftharpoons (0,2)$ exhibits rectifying characters about positive and negative bias. The former case has been elaborated in [Fig.\[fig1\]]{} to \[fig3\], and the small current ($|I|<1$ pA) shown in [Fig.\[fig4\]]{}(b) corresponds to the leakage current at low bias shown in [Fig.\[fig3\]]{}. The latter case of $(1,1)\rightleftharpoons (0,2)$ is very similar except that the PSB effect takes place at the positive bias.
By referring [Fig.\[fig4\]]{}, one can see that large current more than 4 pA can occur at the top corners of $(1,1)$ state, namely the center points of transition between $(1,1)$ and other stability states. In [Fig.\[fig4\]]{}(a), those points correspond to $(-U/2,-U/2), (U/2,-U/2)$ and $(U/2,U/2)$ in the $(V_{1},V_{2})$ parameter plane, versus $(-U/2,-U/2), (-U/2,U/2)$ and $(U/2,U/2)$ in [Fig.\[fig4\]]{}(b). What special about those points is that one quantum transition peak in dot 1 will resonate with another one in dot 2 at the Fermi surface, as the nonequilibrium spectral functions shown in [Fig.\[fig5\]]{}, where [Fig.\[fig5\]]{}(a) to (d) corresponds to the point $(-U/2,-U/2)$ to $(-U/2,U/2)$ in the clockwise order. Taking [Fig.\[fig5\]]{}(a) as an example, we can see that the resonance on point $(-U/2,-U/2)$ takes place between the transition of $(0\rightarrow 1)$ in dot 1 and that of $(0\rightarrow 1)$ in dot 2 at the Fermi surface, which induces a very large current, $I\sim4.2$ pA. If ($V_{1},V_{2}$) deviates those points parallel to the $V_{1}=V_{2}$ diagonal line, the resonance between the transition peaks will still exist, but no longer coincide with the Fermi surface. As a consequence, the current will gradually decrease into the Coulomb blockade region, after some peak-like structures, as shown in the 3D colormap surface image in [Fig.\[fig4\]]{}. If ($V_{1},V_{2}$) deviates the four top corners of (1,1) state along any direction of $V_{1}\neq V_{2}$, none of the resonance will survive, and then the current will decays to very small value ($I<0.5$ pA) quickly.
![(Color online). Non-equilibrium spectral functions at four top corners of (1,1) state in [Fig.\[fig4\]]{}. (a), (b), (c) and (d) respectively represent the point $(-U/2,-U/2), (U/2,-U/2)$, $(U/2,U/2)$ and $(-U/2,U/2)$ in the $(V_{1},V_{2})$ parameter plane.[]{data-label="fig5"}](fig5){width="3.8in"}
Although both peaks of the transition of $(0\rightarrow 1)$ in dot 1 and that of $(1\rightarrow 2)$ in dot 2 are coincide with the Fermi surface on the point $(-U/2,U/2)$, the PSB prohibits the resonance between them as shown in [Fig.\[fig5\]]{}(d), and only very small leakage current ($I<1$ pA) occurs in the vicinity of $(-U/2,U/2)$ \[see [Fig.\[fig4\]]{}(a)\]. When the direction of the bias is reversed from $V>0$ to $V<0$, $(-U/2,U/2)$ will change from a PSB point to a resonance one, and then very large current occurs at this point. Accordingly, the PSB point moves to $(U/2,-U/2)$, as shown in [Fig.\[fig4\]]{}(b).
As indicated in [Fig.\[fig4\]]{}, our theoretical results of the gate voltage modulation on PSB is qualitatively consistent with the experimental measurements. The shape of conduct regions is also similar to the triangles observed in experiments, but not exactly the same. The difference may come from the reason that the Anderson multi-impurity model can not describe all the details in experiments. For instance, when the gate voltage $V_{1}$ on dot 1 changes, it has been confirmed by experiments that the effects on dot 2 are induced not only through the direct coupling $t$ but also through a capacitive coupling. We believe the former has been well described in our theory, but the latter has not yet.
\[Exag\] Lift of PSB by exchange interaction
--------------------------------------------
Now we extend the Anderson two-impurity model to adding the term of dot-dot exchange interaction, which describes the coupling between the local spins of two dots, with the Hamiltonian as follows: $$\label{EXH}
H_{E}=J\hat{S_{1}}\cdot\hat{S_{2}}$$ where $\hat{S_{i}}$ is the spin operator of dot $i(i=1,2)$. $J$ is coupling strength between spins in different dots, which could be positive(antiferromagnetic) or negative(ferromagnetic).
Fundamentally, $J$ originates from the exchange term of the [*e-e*]{} interaction (potential Energy) between two dots, thus plays an equal important role as the kinetic energy $t$. The manipulation of $J$ is a fascinating issue closely relevant to quantum information [@Hanson2007siqd]. Despite the practical difficulties, experiments may achieve this goal indirectly, for example, N. J. Craig [*et al.*]{} have demonstrated the control of the strength and the sign of $J$ between two dots coupled through an open conducting region [@Craig2004RKKY]. In the present work, we investigate the effect of $J$ on PSB. When $t\neq 0$, the total exchange interaction $J_{T}$ principally equals to the sum of $J$ and the antiferromagnetic one $J_{AF}$ induced by $t$, that is $J_{T}\approx J+4t^2/U$. In order to highlight the role of $J$, we thus choose a relatively small $t$ ($t=0.05$ meV) at fixed $U=2.0$ meV to produce a very small $J_{AF}$ ($J_{AF}\sim 0.005$ meV).
[Fig.\[fig6\]]{} shows our results of gate voltage modulation (bias $V=0.4$ mV) on PSB at various $J$ with different signs and values, where [Fig.\[fig6\]]{}(a), (b), (c) and (d) correspond to $J=-0.32$, $-0.08$, $0.08$ and $0.32$ meV, respectively. By referring [Fig.\[fig6\]]{}, one can see that $J$ substantially affects the current near the points $(V_1, V_2)=(-U/2,U/2)$ and $(U/2,-U/2)$ where the PSB effects take place, while it slightly dose to the current near $(-U/2,-U/2)$ and $(U/2,U/2)$ where no PSB effects. It indicates that the exchange interaction can directly change the characters of the PSB.
![(Color online). 2D contour plot of the current (bias V=$0.4$ mV) as a function of the gate voltage $V_{1}$ and $V_{2}$ at various $J$. (a), (b), (c) and (d) correspond to $J=-0.32$, $-0.08$, $0.08$ and $0.32$ meV, respectively. []{data-label="fig6"}](fig6){width="3.5in"}
Let us focus on the change of the leakage current of the PSB around $(-U/2,U/2)$. It corresponds to a very small value at $J=-0.32$ meV \[[Fig.\[fig6\]]{}(a)\], which indicates that the ferromagnetic exchange interaction tends to enhance the PSB effect. Generally speaking, the leakage current will increase with the increasing of the algebra value of $J$, as shown in [Fig.\[fig6\]]{}. For example, the leak current increases from near zero to $\sim 0.5$ nA \[[Fig.\[fig6\]]{}(b)\] as $J$ increasing from $-0.32$ to $-0.08$ meV. Changing the sign of $J$ still maintains this kind of tendency, and the antiferromagnetic exchange interaction seems to suppress the PSB. When $J$ continually increases to $0.08$ meV \[[Fig.\[fig6\]]{}(c)\], the leakage current is clearly visible and in the range of $1\sim 2$ nA. If $J$ positively increases to a enough large value, e.g. $J=0.32$ meV \[[Fig.\[fig6\]]{}(d)\], the leakage current will increase distinctly, even to the same order of magnitude ($\sim 4$ nA) as the conductive current. In this case, we argue that the PSB has been lifted by strong antiferromagnetic exchange interaction.
The lift of the PSB is a significant feature of [Fig.\[fig6\]]{}, which deserves careful study to reveal its mechanism. We thus apply the local magnetic field $B_{1}$ again onto dot 1 to lift is spin degeneracy and investigate the transport of down-spin electrons in the PSB region. [Fig.\[fig7\]]{} shows the nonequilibrium spectral functions at the point of $(V_1, V_2)=(-U/2, U/2)$, with the Zeeman energy caused by $B_{1}$ being $\Delta E_{B1}=1.5$ meV and other parameters being the same as those in [Fig.\[fig6\]]{}. [Fig.\[fig7\]]{}(a)-(d) corresponds to [Fig.\[fig6\]]{}(a)-(d), respectively.
At $J=-0.32$ meV, the dot-dot exchange interaction is ferromagnetic, which makes the energy level of $T(1,1)$ lower than $S(1,1)$. As a consequence, the single occupied transition peak of down-spin in dot 2 is lower than that of up-spin, since that peak in dot 1 has been locked as down-spin one at the Fermi surface. It means that the down-spin electron in dot 2 has almost fully occupied the single level under the Fermi surface, which will prevent the hopping of electrons with the same spin and thus enhance the PSB effects. Continuously increasing the algebra value of $J$ to $-0.08$ meV will not essentially change above process, and only slightly increase the weight of the down-spin holes in the double occupied transition peak at the Fermi surface, which makes the leakage current increase to a small nonzero value, as shown in [Fig.\[fig6\]]{}(b).
![(Color online). In the case of the spin non-degeneracy in dot 1, nonequilibrium (bias V=$0.4$ mV) spectral function $A_{1\downarrow}(\omega)$, $A_{2\uparrow}(\omega)$ and $A_{2\downarrow}(\omega)$ at $(V_1, V_2)=(-U/2, U/2)$. (a), (b), (c) and (d) correspond to $J=-0.32$, $-0.08$, $0.08$ and $0.32$ meV, respectively.[]{data-label="fig7"}](fig7){width="3.5in"}
When $J>0$, the spin exchange interaction is anti-ferromagnetic, which makes the energy level of $S(1,1)$ is lower than that of $T(1,1)$. This case should help electrons in dot 1 transfer to dot 2 by means of the transition $S(1,1)\rightarrow S(0,2)$, so the leak current will increase, which has been confirmed by the change of the spectral functions shown in [Fig.\[fig7\]]{}(c) and [Fig.\[fig7\]]{}(d). We can see that after the sign change of $J$, the weight of the down-spin holes in the double occupied transition peak increases further, offering more space for down-spin electrons to transfer from dot 1, and consequently increases the leak current value. Since $J$ is small in [Fig.\[fig7\]]{}(c), the leak current can only increase to a relatively small value, as shown in [Fig.\[fig6\]]{}(c). In [Fig.\[fig7\]]{}(d), large enough $J$ ($J=0.32$ meV) makes the weight of down-spin hole increase remarkably, thus induces the large leak current shown in [Fig.\[fig6\]]{}(d) which lifts the PSB effect. By referring [Fig.\[fig6\]]{} and \[fig7\], we can see that the lift of PSB by $J$ is a continuous process instead of a sudden change.
![(Color online). In the case of the spin non-degeneracy in dot 1, the dependence of the occupation numbers $N_{1\downarrow}$, $N_{2\uparrow}$, and $N_{2\downarrow}$ as well as the total current $I$ on the bias voltage $V$ at $J=-0.32$ meV (a) and $0.32$ meV (b).[]{data-label="fig8"}](fig8){width="2.5in"}
To further verify the mechanism of the lift of PSB by the exchange interaction, we calculate the $N-V/I-V$ curves at different $J$ and summarize the results in [Fig.\[fig8\]]{}, where [Fig.\[fig8\]]{}(a) corresponds to $J=-0.32$ meV and [Fig.\[fig8\]]{}(b) to $J=0.32$ meV. By comparing those two figures in [Fig.\[fig8\]]{}, one can see that $N_{1\downarrow}$ ($J=-0.32$ meV)$\approx N_{1\downarrow}$ ($J=0.32$ meV) at bias $V=0$, but both $N_{2\downarrow}$ and $N_{2\uparrow}$ are distinctly different at $J=\pm0.32$ meV. Specifically, when $J=-0.32$ meV and $V=0$, $N_{2\downarrow}>N_{2\uparrow}$, which means that the single occupation of down-spin electron in dot 2 is sufficient and the weight of hole with the same spin in double occupation is small, thus the transfer of down-spin electrons from dot 1 to 2 will be blocked. With the bias positively increasing, as shown in [Fig.\[fig8\]]{}(a), $N_{2\downarrow}$ gradually increases and approaches to 1.0 after $V>1.0$ mV, as a consequence, the PSB effect is enhanced and the current decreases to a near zero value. Negatively increasing bias will produce a steady current at about $-4.0$ nA after $V<-1$ mV as shown in the same figure.
On the other hand, when $J=0.32$ meV and $V=0$, $N_{2\downarrow}<N_{2\uparrow}$, which means that the single occupation of down-spin electron in dot 2 is small and the weight of hole with the same spin in double occupation is large, thus the PSB of down-spin electrons from dot 1 to 2 will be lifted. With the bias positively increasing, as shown in [Fig.\[fig8\]]{}(b), $N_{2\downarrow}$ slowly increases and then stabilises at 0.9 after $V>1.0$ mV. It suggests that the weight of hole in double occupation keeps finite, which will induce considerable leakage current even at large positive $V$. Negatively increasing bias will produce a steady current at about $-6.0$ nA after $V<-1$ mV, much larger than that at $J=-0.32$ meV \[cf. [Fig.\[fig8\]]{} (a) and (b)\]. The reason may lie in the fact that antiferromagnetic interaction tends to form $S(1,1)$ which is in favor of the transition of $S(0,2)\rightarrow S(1,1)$.
We are now on the position to compare the different roles of $J$ and $t$ on PSB. For this purpose, we define a physical quantity $I_C:I_L$ as the ratio between the conductive current $I_C$ at $V=-2.0$ mV and the leakage one $I_L$ at $V=2.0$ mV at the point of $(V_1, V_2)=(-U/2, U/2)$, and then summarize the dependence of $I_C:I_L$ on $J$ in [Fig.\[fig9\]]{} together with $I_C:I_L$ on $t$ in its insert. In principle, the larger the ratio $I_C:I_L$ is, the smaller the leakage current, and the more distinct the PSB effect, in case that $I_C$ does not change significantly. As shown in the figure, the ratio $I_C:I_L$ can reach 21 at $J=-0.32$ meV, indicating a well-defined PSB effect, as shown in [Fig.\[fig6\]]{}(a) and [Fig.\[fig8\]]{}(a). Positively increasing $J$ will induce a continuously decreasing of $I_C:I_L$, smoothly passing through the zero point, $I_C:I_L\sim15$ at $J=0$, and finally approaching a small value $I_C:I_L\sim 3$ at $J=0.32$ meV, indicating a total lift of the PSB effect. At $J<-0.2$ meV and $J>0.2$ meV, one can see that the decrease of $I_C:I_L$ with $J$ exhibits a near-linear behavior, but a nonlinear swell around the zero point is clearly shown in the figure. It suggests that $J=0$ is a crossover point from the appearance to the lift of the PSB effect (at a fixed small $t$).
![(Color online). The dependence of the ratio $I_C(V=-2.0{\rm mV}):I_L(V=2.0{\rm mV})$ on $J$ at the point of $(V_1, V_2)=(-U/2, U/2)$. The insert shows the dependence of the ratio $I_C:I_L$ on $t$.[]{data-label="fig9"}](fig9){width="2.5in"}
By referring the insert of [Fig.\[fig9\]]{}, one can see that the dependence of $I_C:I_L$ on $t$ exhibits a peak structure. More specifically, starting from the non-PSB point at $t\sim0$ \[see [Fig.\[fig1\]]{} and [Fig.\[fig3\]]{}(a)\], increasing $t$ will firstly result in a rapid increase of $I_C:I_L$ to a value about 16 at $t=0.8$, denoting a well-defined PSB effect, as shown in [Fig.\[fig2\]]{} and [Fig.\[fig3\]]{}(b). By further increasing $t$ ($t>0.8$), the $I_C:I_L$ will not increase any more, but slowly decrease to a small value about 5 at $t=0.3$. It represents that large $t$ does not mean distinct PSB effects which should only takes place at moderate $t$. That result is consistent with the experimental measurements in literatures[@Hanson2007siqd; @Zwanenburg2013siliconqd].
\[sec:level3\]SUMMARY
=====================
In summarize, we systematically investigate the Pauli spin blockade in double quantum dot systems. We start from the Anderson multiple impurity model to describe the system, fully considering the electron-electron interaction and the dot-electrode couplings. By using the hierarchical equations of motion approach, we deal with this quantum model non-perturbatively to accurately obtain the equilibrium and nonequilibrium spectral functions, occupation numbers and current, etc.
By means of a strong local magnetic field applied onto dot 1 to lift its spin degeneracy, our theory appropriately describes the physical mechanism and picture of the Pauli spin blockade effect in double quantum dot systems, followed by a general discussion without the local field applied. Then, the gate voltage manipulation of the spin blockade is elaborated in detail by our theory. Our results are proved to be qualitatively consistent with the experimental measurements.
We further extend the Anderson multiple impurity model to involve the dot-dot exchange coupling, and carefully study its effect on the spin blockade by changing the strength and the sign of the coupling. It is found that the ferromagnetic exchange interaction tends to enhance the spin blockade, while the antiferromagnetic one to suppress it. What is more, the Pauli spin blockade effect may be lifted by the strong antiferromagnetic exchange coupling.
\[sec:level3\]ACKNOWLEDGEMENT
=============================
The support from the NSF of China (No.11374363) and the Research Funds of Renmin University of China (Grant No. 11XNJ026) is gratefully appreciated.
[1]{}
K Ono, DG Austing, Y Tokura, and S Tarucha, Science. [**297**]{}, 1313 (2002).
A. C. Johnson, J. R. Petta, and C. M. Marcus, Phys. Rev. B. [**72**]{}, 165308 (2005).
M Busl, G Granger, L Gaudreau, R S[á]{}nchez, A Kam, M Pioro-Ladriere, SA Studenikin, P Zawadzki, ZR Wasilewski, AS Sachrajda, et al, Nat. Nano. [**8**]{}, 261 (2013).
H. W. Liu, T. Fujisawa, Y. Ono, H. Inokawa, A. Fujiwara, K.Takashina, and Y. Hirayama, Phys. Rev. B. [**77**]{}, 073310 (2008).
Xiaojie Hao, Rusko Ruskov, Ming Xiao, Charles Tahan, and HongWen Jiang, Nat. Comm. [**5**]{}, 3860 (2014).
Jonas Fransson and M R[å]{}sander, Phys. Rev. B. [**73**]{}, 205333 (2006).
Changxue Deng and Xuedong Hu, Phys. Rev. B. [**71**]{}, 033307 (2005).
David M.-T. Kuo, Shiue-Yuan Shiau, and Yia-chung Chang, Phys. Rev. B. [**84**]{}, 245303 (2011).
W. A. Coish and F. Qassemi, Phys. Rev. B. [**84**]{}, 245407 (2011).
J. Danon, Phys. Rev. B. [**88**]{}, 075306 (2013).
Jeroen Danon, Xuhui Wang, and Aure¡älien Manchon, Phys. Rev. Lett. [**111**]{}, 066802 (2013).
Dimitrije Stepanenko, Mark Rudner, Bertrand I. Halperin, and Daniel Loss, Phys. Rev. B. [**85**]{}, 075416 (2012).
Chang-Yu Hsieh,Yun-Pil Shim, and Pawel Hawrylak, Phys. Rev. B [**85**]{}, 085309 (2012).
S. Amaha, W. Izumida, T. Hatano, S. Tarucha, K. Kono, and K. Ono, Phys. Rev. B. [**89**]{}, 085302 (2014).
J. S. Jin, X. Zheng, and Y. J. Yan, J. Chem. Phys. [**128**]{}, 234703 (2008).
ZhenHua Li, NingHua Tong, Xiao Zheng, Dong Hou, JianHua Wei, Jie Hu, and YiJing Yan, Phys. Rev. Lett.[**109**]{}, 266403 (2012).
X. Zheng, J. S. Jin, and Y. J. Yan, J. Chem. Phys. [**129**]{}, 184112 (2008).
X. Zheng, J. Y. Luo, J. S. Jin, and Y. J. Yan, J. Chem. Phys. [**130**]{}, 124508 (2009).
X. Zheng, J. S. Jin, and Y. J. Yan, New J. Phys. [**10**]{}, 093016 (2008).
X. Zheng, J. S. Jin, S. Welack, M. Luo, and Y. J. Yan, J. Chem. Phys. [**130**]{}, 164708 (2009).
X. Zheng, R. X. Xu, J. Xu, J. S. Jin, J. Hu, and Y. J. Yan, Prog. Chem. [**24**]{}, 1129 (2012).
YongXi Cheng, WenJie Hou, YuanDong Wang, ZhenHua Li, JianHua Wei and YiJing Yan, New J. Phys. [**17**]{}, 033009 (2015).
J. H. Wei and Y. J. Yan, arXiv:1108.5955 (2011).
R. Brunner,Y. S. Shin, T. Obata, M. Pioro-Ladriere, T. Kubo, K. Yoshida, T. Taniyama, Y. Tokura, and S. Tarucha, Phys. Rev. Lett. [**107**]{},146801 (2011).
R. Hanson, L. P. Kouwenhoven, J. R. Petta, S. Tarucha, L. M. K. Vandersypen, Rev. Mod. Phys. [**79**]{}, 1217 (2007).
N. J. Craig, J. M. Taylor, E. A. Lester, C. M. Marcus,M. P. Hanson, A. C. Gossard, Science. [**304**]{}, 565 (2004).
F.A. Zwanenburg, A. S. Dzurak, A. Morello, and M.Y.Simmons, L. C. L. Hollenberg, G. Klimeck, S. Rogge, S. N. Coppersmith and M. A. Eriksson, Rev. Mod. Phys. [**85**]{}, 961 (2013).
|
{
"pile_set_name": "ArXiv"
}
|
[**On the Geometry and Homology of\
Certain Simple Stratified Varieties**]{}\
[**T. Hübsch**]{}[^1] and [**A. Ramān**]{}[^2]\
[*Department of Mathematics and\
Department of Physics and Astronomy\
Howard University\
Washington, DC 20059*]{}\
[**ABSTRACT**]{}\
Introduction, Results and Summary
=================================
In string theory, rather than being an assumed arena, the spacetime is identified with the dynamically determined ‘ground state variety’ of a (supersymmetric) $\s$-model [@gsw:string; @wit:phases; @pol:string]. In the simplest physically interesting and nontrivial case [@chsw:strcys; @wit:phases], the spacetime is of the form $M^{3,1}{\times}K$, where $K$ is a compact Calabi-Yau 3-fold modeled from the (bosonic subset of the) ‘field space’ of the $\s$-model[^3], $\cF=\{p,s_0,{\cdots},s_4\}\simeq\IC^6$, which admits a $\IC^*$ action: $$\hat{\lm}:~\{p,s_0,{\cdots},s_4\} \mapsto
\{\lm^{-5}p,\lm s_0,{\cdots},\lm s_5\}~,\qquad \lm\in\IC^*~.
\Label{e:ProjMap}$$ The ‘ground state variety’ is defined to be [@wit:phases; @agm:chngtp] $$\cV~\define~[\,(\vd W)^{-1}(0)-0\,]/\hat{\lm}~,
\Label{e:HolQ}$$ with the $\hat{\lm}$-invariant holomorphic ‘superpotential’ $$W~\define~p{\cdot}G(s)~. \Label{e:W=pG}$$ Alternatively, we denote by $\hat{\lbar}$ the $|\lm|=1$ restriction of the map (\[e:ProjMap\]), and define the ‘potential’ $$U_r~\define~\|\vd W\|^2+D_r^2~, \Label{e:U=dW+D}$$ where $$D_r~\define~\|s\|^2-5|p|^2-r~,\qquad r\in\IR~. \Label{e:gaugeD}$$ Then $$\cV~\simeq~[\,U_r^{-1}(0)-0\,]/\hat{\lbar}~.
\Label{e:SimQ}$$
Owing to the positive definiteness of $U_r$, $$U_r^{-1}(0) ~=~ (\vd W)^{-1}(0) ~\cap~ D_r^{-1}(0)~. \Label{e:RMfld}$$ Furthermore, the $\hat{\lm}$-invariance of $W\,{=}\,pG$ implies that $G(s)$ is a degree-5 homogeneous complex polynomial $$G(\lm s_0,{\cdots},\lm s_4) ~=~ \lm^5 G(s_0,{\cdots},s_4)~,
\Label{e:HomHoloG}$$ whereupon the zero locus of $\vd W$ is the intersection of the cones $$(\vd W)^{-1}(0)~=~G^{-1}(0)~\cap~(p{\cdot}\vd_sG)^{-1}(0)~.$$ The above definition may then be rephrased as follows:
Given the polynomials $G(s)$ and $D_r$ as defined in Eqs. (\[e:HomHoloG\]) and (\[e:gaugeD\]), respectively, the ‘ground state variety’ is $$\begin{array}{rcl}
\cV&=&\Big\{G^{-1}(0)\cap
(p{\cdot}\vd_sG)^{-1}(0)-0\,\Big\}\Big/\hat{\lm}~,\\[2mm]
&=&\Big\{G^{-1}(0)\cap
(p{\cdot}\vd_sG)^{-1}(0)\cap
D_r^{-1}(0)-0\,\Big\}\Big/\hat{\lbar}~,
\Label{e:GndStVar}
\end{array}$$ where the $S^1$-action, $\hat{\lbar}$, in the latter (symplectic) quotient is the $|\lm|=1$ restriction of the $\IC^*$-action (\[e:ProjMap\]) in the former (holomorphic) quotient.
$\cV^+$ ($\cV^-$) shall denote the restriction of $\cV$ to positive (negative) values of $r$ in .
As we show in more detail in Section \[s:GSV\], the dependence $U_r$, in , turns the ‘ground state variety’ into a 1-parameter family of (stratified) varieties[^4], and the subtraction of zero in Eqs. (\[e:GndStVar\]) separates the two branches, $\cV^\pm$, defined with $r{>}0$ and $r{<}0$, respectively. Moreover, the dependence on $r$, as defined originally in the gauged linear $\s$-model [@wit:phases], is complicated near $r{=}0$ by quantum corrections and we restrict to $r{\neq}0$.
Generalizations involve ([*a*]{}) additional $p$ and $s$ variables, ([*b*]{}) additional corresponding terms in the superpotential (\[e:W=pG\]), and ([*c*]{}) additional maps (\[e:ProjMap\]) and their modifications where the exponents of $\lm$ are different integers: negative for the $p$’s, positive for the $s$’s. The generalization ([*a*]{}) turns the $\IC^*$ action (\[e:ProjMap\]) into a more general toric action, while ([*a*]{}) enlarges the field space and ([*b*]{}) modifies the “moment map” (\[e:W=pG\]) accordingly. The resulting ‘ground state varieties’ will thus include intersections of hypersurfaces in products of toric varieties [@wit:phases; @agm:chngtp; @ful:torvar; @gun:torvar]. All of these are of the form given in the definition \[d:GndStVar\], with the ‘ingredients’ $\{(p,s),\hat\lm,D_r,W\}$ duly modified.
Our main result is:
Let $\cV$, the ‘ground state variety’ of the gauged linear $\s$-model [@wit:phases], be defined in Definition \[d:GndStVar\]. Then $\cV^+$ is a stratified variety [@gmp:strata] when the polynomial $G(s)$ is non-transversal at $n$ isolated rays, $\sh$. Then, $\cV^+=\shM\cup\bigcup_iA_i$ with $\shM=G^{-1}(0)$ smooth except at $n$ isolated nodes, where the $n$ noncompact ‘antennae’ $A_i\simeq\IC^1$ attach. For $\dim_\IC\cV^+{=}3$, the minimal holomorphic compactification $\ba{\cV}^+=\shM\cup\bigcup_i\bar{A}_i$ satisfies: $\ba{\cV}^+$ is an exoflop of the small resolution(s) of $\shM$ in the sense of Ref. [@agm:chngtp], and $\oplus_qH^{2q}(\ba{\cV}^+)$ satisfies the “Kähler package” of requirements [@gmp:strata], and is compatible with mirror symmetry [@yau:mirror] and string theory [@str:nullBH; @hub:strcoh].
This article is organized as follows: Section \[s:GSV\] shows that the ‘ground state variety’ becomes stratified as $G(s)$ becomes non-transversal, and we explore the induced (exo-)strata[^5] and their union. Section \[s:Hom\] explores the contribution of the induced (exo-)strata to the homology of the ‘ground state variety.’ Section \[s:DefRes\] re-examines the ‘conifold transition’ of Refs. [@cgh:confld] in view of Theorem \[t:main\].
The Ground State Variety
========================
We now turn to analyze the geometry of the ground state variety, as determined by the choice of the homogeneous holomorphic polynomial $G(s)$. Such polynomials typically depend on a multitude of parameters; when properly accounted for redundancies, these span (a subspace of) the moduli space of the ground state variety. Thus, we automatically have a family of ground state varieties, fibered over this (partial) moduli spaces. Works in the literature, Ref. [@wit:phases] and the subsequent studies, all assumed $G(s)$ to be transversal and so have explored the generic fibre of this family. We begin by analyzing this case in some detail, and then turn to the less generic mild degenerations of the fibre.
The transversal case
--------------------
$G(s)$ being transversal[^6], $G= \rd G=0$ only at $s=0$. In this case, the zero locus of $\vd W=(G,p{\cdot}\vd_s G)$ is a union of two branches[^7]: $$(G)^{-1}(0)\cap(p{\cdot}\vd_sG)^{-1}(0)
~=~\Big\{p{=}0,~s:G(s)=0\Big\}\cup\Big\{p,~s{=}0\Big\}~.
\Label{e:union}$$
So, following the first (holomorphic quotient) part of definition \[d:GndStVar\], we have that $$\cV~=~\Big\{p{=}0,~s:G(s)=0\Big\}\Big/\hat{\lm}
~\cup~\Big\{p,~s{=}0\Big\}\Big/\hat{\lm}~,
\Label{e:qunion}$$ where the quotients are taken after the fixed point of the $\hat{\lm}$-action, $\{s,p=0\}$, is excised. Now, since $\{s{\neq}0\}/\hat{\lm}$ is $\IP^4$, then $$\Big\{p{=}0,~s\neq0:G(s)=0\Big\}\Big/\hat{\lm}~=~\cM \Label{e:CYM}$$ is the Calabi-Yau quintic hypersurface in $\IP^4$.
The second term in the union (\[e:qunion\]) is $$\Big\{p{\neq}0,~s{=}0\Big\}\Big/\hat{\lm}~\simeq~\IC^*/\IC^*
~\simeq~\{\hbox{pt.}\}/\ZZ_5~. \Label{e:LGO}$$ This is the ‘fuzzy point’ [@agm:chngtp] of the Landau-Ginzburg orbifold. $\ZZ_5$ is the subroup of $\hat\lm$ which leaves $G(s)$ invariant and so acts trivially on both $W{=}p{\cdot}G(s)$ and on $p$.
(170,125)(-20,0) (0,0) (-6,19)[$s_0$]{} (17,5)[$s_1$]{} (48,2)[$s_2$]{} (113,15)[$s_3$]{} (130,29)[$s_4$]{} (73,8)[The cone $G^{-1}(0)$]{} (19,22)[$\IP^4$]{} (10,15)[$(r{>}0)$]{} (26,27)[$\sqrt{r}$]{} (59,17)[$\cM$]{} (35,95)[$p$]{} (-9,80)[The Landau-Ginzburg]{} (14,75)[‘fuzzy point’]{} (20,65)[$\sqrt{-r/5}$]{} (22,58)[$(r{<}0)$]{} (88,60)[$s_0$]{} (111,50)[$s_1$]{} (103,103)[$s_2$]{} (157,97)[$s_3$]{} (160,61)[$s_4$]{} (113,112)[The cone $G^{-1}(0)$]{} (100,80)[$\IP^4$]{} (97,75)[$(r{>}0)$]{} (114,85)[$\cM$]{}
The two quotients in the union (\[e:qunion\]) are thus manifestly disconnected: the former, (\[e:CYM\]), lies entirely in the $\{p{=}0,~s{\neq}0\}$-subspace of the field space $\cF$, whereas the latter, (\[e:LGO\]), lies well in the complementary $\{p{\neq}0,~s{=}0\}$-subspace. The above is illustrated in Fig. \[f:smooth\]. Even in the transversal case, $\cV$ may be regarded as a stratified variety, consisting of two disconnected objects: a (complex) 3-dimensional one and a (complex) 0-dimensional one, each of which containing a single variety: $\cM$ and $\{\hbox{pt.}\}/\ZZ_5$, repsectively.
In fact, the second component, $\{\hbox{pt.}\}/\ZZ_5$, actually lies in the ‘second sheet’ of the field space $\cF$. To see this, it will be useful to also present $\cV$ using the alternate (symplectic quotient) definition \[d:GndStVar\]:
1. When $r\gg0$, $D_r^{-1}(0)\neq0$ implies that $\|s\|^2\neq0$, and so $\vd_sG\neq0$ as $G$ is transversal. Then, $(\vd W)^{-1}(0)$ lies entirely in the $(p{=}0)$ $s$-hyperplane, and $\cV$ is the $\hat\lm$-quotient, , the complex base of the cone $G^{-1}(0)$. Then, $\cV=\cM\id[\{p=0\}\cap G^{-1}(0)]/\hat{\lm}$. Since $\{s{\neq}0\}/\hat{\lm}=\IP^4$, the [*projective*]{} Calabi-Yau quintic hypersurface (\[e:CYM\]) is $G^{-1}(0)/\hat{\lm}=\cM\into\IP^4$.
2. When $r\ll0$, $D_r^{-1}(0)\neq0$ implies that $|p|^2\neq0$, and so $\|s\|^2{=}0$ since $G$ is transversal. Then, $(\vd W)^{-1}(0)$ lies entirely in complex the $p$-plane, and is the ‘fuzzy point’ [@agm:chngtp], $\Big\{|p|=\sqrt{|r|/5}\Big\}/\ZZ_5$, of the Landau-Ginzburg orbifold (\[e:LGO\]).
At the critical point $r=0$, these two branches formally collapse to the the highly degenerate point $p,s=0$, which is the branching point of the two ‘sheets’ of the field space $\cF$. This point is, by definition, excised before taking the quotients (\[e:HolQ\]). Indeed, for applications to string theory, the preceding analysis is not to be trusted in the region near $p,s{=}0$ since quantum corrections modify the map (\[e:gaugeD\]) and so also the structure of the quotients in definition \[d:GndStVar\]; see Ref. [@wit:phases]. For this reason, we will mostly concern ourselves with the $r\gg0$ ‘sheet’, and comment on occasion on the $r\ll0$ ‘sheet’, but leave any ‘connection’ between the two ‘sheets’ unexplored for now.
The conifold with exocurves
---------------------------
Unlike Ref. [@wit:phases] and subsequent work, we will be concerned with ground state varieties using homogeneous holomorphic polynomials $G(s)$ which are non-transversal along $n$ isolated (complex) directions: $$\vd G(s)~=~0 \qquad\To\qquad s~=~\sh_j~,\quad j=1,{\cdots},n~.
\Label{e:nodalG}$$ Clearly, $\{\sh_j\}\simeq\IC^1$, and we denote by $\cB^n\define\sqcup_{j=1}^n\{\sh_j\}$ the ‘bouquet’ of $n$ $\IC^1$’s all meeting at the origin. Since $G(s)$ is holomorphic and homogeneous, $\vd G(s)=0$ implies $s{\cdot}\vd_sG(s)=5G(s)=0$ and the $G(s)=0$ condition is automatically satisfied on $\cB^n$. Thus, we find that $$(G)^{-1}(0)\cap(p{\cdot}\vd_sG)^{-1}(0)
~=~\Big\{p{=}0,~s\neq\sh_j:G(s)=0\Big\}\cup\Big\{\{p\}{\times}\cB^n\Big\}~.
\Label{e:Union}$$
So, following the first (holomorphic quotient) part of definition \[d:GndStVar\], we now have that $$\cV~=~\Big\{p{=}0,~s\neq\sh_j:G(s)=0\Big\}\Big/\hat{\lm}
~\cup~\Big\{\{p\}{\times}\cB^n\Big\}\Big/\hat{\lm}~;
\Label{e:QUnion}$$ again, the quotients are taken after the fixed point of the $\hat{\lm}$-action, $\{s,p=0\}$, is excised. Now, since $\{s{\neq}0,\sh_j\}/\hat{\lm}$ equals $\IP^4$ without its points where $G(s)|_{\IP^4}$ is non-transversal, then $$\Big\{p{=}0,~s\neq0,\sh_j:G(s)=0\Big\}\Big/\hat{\lm}
~=~\shM-\hbox{Sing}(\shM) \Label{e:nonConifold}$$ is the non-singular (and non-compact) part of the conifold[^8] $\shM\into\IP^4$. Note that $$\begin{array}{rcl}
\hbox{Sing}(\shM)
&\define& \Big\{p{=}0,~s{=}\sh_j:G(s)=0\Big\}\Big/\hat{\lm}\\[2mm]
&=& \bigcup_{j=1}^n\Big\{p{=}0,~s{=}\sh_j\Big\}\Big/\hat{\lm}
~=~ \bigcup_{j=1}^n \shx_j \subset \shM~, \Label{e:SingM}
\end{array}$$ since $G(\sh_j){=}0$; $\shx_j$ are the singular points of $$\shM~\define~[\{p=0\}\cap G^{-1}(0)]/\hat{\lm}~. \Label{e:shM}$$
The second quotient in the union (\[e:QUnion\]) is quite more intricate. Setting $p{=}0$ in , we see that $\{\{p\}{\times}\cB^n\}/\hat{\lm}$ is non-empty in the $r{>}0$ ‘sheet’ of the field space $\cF$, and also that it includes the points $\{p{=}0,\sh_j{\neq}0\}/\hat{\lm}=\hbox{Sing}(\shM)$. On the other hand, $$\Big(\Big\{\{p\}{\times}\cB^n\Big\}\Big/\hat{\lm}\Big)
~=~\sqcup_{j=1}^n A_j~,\qquad
A_j~\define~\{p,\sh_j\}\Big/\hat{\lm}~. \Label{e:Antennae}$$ Each of the $A_j$’s contains precisely one of the singular points of $\shM$, as given in $$\shx_j~=~\{p{=}0,~\sh_j\}\Big/\hat{\lm}~=~A_j\cap\shM~.$$ Thus, ground state variety (\[e:QUnion\]), which is the [*connected*]{} union of (\[e:Antennae\]) and of (\[e:nonConifold\]), is then $$\cV~=~\shM ~\cup~ \sqcup_{j=1}^n A_j~. \Label{e:AntennaM}$$ That is, the ($r{>}0$ ‘sheet’ of the) ground state variety is the conifold $\shM$, with an exocurve, $A_j$, attached at each singular point.
In the other, $r{<}0$ ‘sheet’ of the field space $\cF$, the first term in the union (\[e:QUnion\]) turns out to be empty. On the other hand, the second one is not since $\{p,~s=\sh_j\}$ does include the complex $p$-plane in which shows that $r<0$. In this case, the second term in the union (\[e:QUnion\]) again turns out to be of the form (\[e:Antennae\]), except this time the $A_j$’s have a single common point, the Landau-Ginzburg orbifold (\[e:LGO\]).
Alternatively, consider the symplectic quotient: impose the vanishing of $D_r$, , intersect with $D_r^{-1}(0)$, and pass to the $S^1$-quotient. To this end, consider each term in the union (\[e:Union\]) separately. Now $D_r=0$ implies that $$\|s\|^2-5|p|^2~=~r~>~0~,\quad\To\quad \|s\|^2~\geq~|r|~>~0~.
\Label{e:snot0}$$ The ground state variety now is the $S^1$-quotient of the union: $$\Big\{p{=}0,s\neq\sh_j:G(s)=0,\|s\|^2=r\Big\}
~\cup~
\Big\{p,~s{=}\sh_j:\|s\|^2=r_+\Big\}~, \Label{e:UnionQ}$$ where $$r_+~=~5|p|^2+|r|~. \Label{e:rplus}$$ Note that the $p=0$ points of the second component, where $\|s\|^2=r_+{=}r$, the $$\Big\{(p,s){=}(0,\sh_j):\|s\|^2=r\Big\}\Label{e:Sings}$$ points are the $s\to\sh_j$ limiting points of the first component, since $G(\sh_j){=}0$. The $S^1$ quotient of these are the (nodal) singular points of the conifold (\[e:shM\]), and they connect the two terms in the union (\[e:UnionQ\]). This then becomes $\shM\cup\sqcup_jA_j$, just as obtained using the holomorphic quotient (\[e:AntennaM\]).
Now $D_r=0$ implies that $$\|s\|^2-5|p|^2~=~r~<~0~,\quad\To\quad |p|~\geq~\sqrt{|r|/5}~>~0~.
\Label{e:pnot0}$$ This renders the first term in the union (\[e:qunion\]) empty, and the ground state variety now is: $$\Big\{p,~s{=}\sh_j:\|s\|^2=r_-\Big\}\Big/S^1~, \Label{e:LGOQ}$$ where now $$r_-~=~\|\sh_j\|^2+|r|~. \Label{e:rminus}$$ Note that at $s=0=r_-$, where $|p|=\sqrt{|r|/5}$, the $$\Big\{p,~s{=}0:5|p|^2=r\Big\}\Big/S^1$$ point is common to all components of the second component, and is the ‘fuzzy point’ of the Landau-Ginzburg orbifold (\[e:LGO\]).
The foregoing proves the following
With the ‘ingredients,’ $\{(p,s),\hat\lm,D_r,W\}$, defined as in , (\[e:gaugeD\]) and (\[e:W=pG\]), the $r{>}0$ ‘sheet’ of the ground state variety (Definition \[d:GndStVar\]), $\cV^+$, becomes a stratified variety, $\shM\,\cup\,\sqcup_jA_j$ when $G(s)$ is non-transversal as specified in .\
The ‘main’ stratum (\[e:nonConifold\]) has complex dimension 3, while the ‘exocurves’ (\[e:Antennae\]) minus the singular points $\shx_j$ form the complex dimension 1 stratum; the singular points, $\sqcup_j\shx_j=\hbox{Sing}(\shM)$, form the complex dimension 0 stratum.
Under the same conditions as in Lemma \[l:GndStVar\], the $r{<}0$ ‘sheet’ of the ground state variety, $\cV^-$, is the stratified variety: the union of the exocurves (\[e:Antennae\]), $\cup_jA_j$, connected at the ‘fuzzy point’ of the Landau-Ginzburg orbifold (\[e:LGO\]).
The $r{<}0$ stratified variety consists of the exocurves $A_j$ minus the ‘fuzzy point’ which form the complex dimension 1 stratum, and the ‘fuzzy point’ (\[e:LGO\]) which forms the complex dimension 0 stratum.
The resulting non-transversal ground state variety is illustrated in Fig. \[f:singular\]
(170,110)(-20,0) (0,0) (39,17)[$s_0$]{} (62,3)[$s_1$]{} (85,1)[$s_2$]{} (156,18)[$s_3$]{} (166,31)[$s_4$]{} (105,1)[$\sh_1$]{} (142,6)[$\sh_2$]{} (91,9)[$\shx_1$]{} (119,11)[$\shx_2$]{} (53,17)[$\IP^4$]{} (52,12)[$(r{>}0)$]{} (69,26)[$\sqrt{r}$]{} (105,18)[$\shM$]{} (81,95)[$p$]{} (68,66)[$\sqrt{-r/5}$]{} (69,60)[$(r{>}0)$]{} (114,63)[$\sqrt{r_-}$]{} (100,92)[$A^-_1$]{} (129,94)[$A^-_2$]{} (121,53)[$\sqrt{r_+}$]{} (99,60)[$A^+_1$]{} (153,71)[$A^+_2$]{} (-3,45)[$s_0$]{} (15,40)[$s_1$]{} (8,88)[$s_2$]{} (56,82)[$s_3$]{} (50,52)[$s_4$]{} (31,93)[$\sh_1$]{} (46,80.5)[$\sh_2$]{} (24,46)[$\IP^4$]{} (21,41)[$(r{>}0)$]{} (16,72)[$\shM$]{}
Since the ’fuzzy point’ of the Landau-Ginzburg orbifold (\[e:LGO\]) may be, formally, considered as the (negative size) collapse (or, perhaps more properly, analytic continuation) of the 3-dimensional Calabi-Yau variety $\cM$, the same relation remains between $\cV^+$ and $\cV^-$, regardless of the (non)transversality of $G(s)$.
The exocurves
-------------
We now turn to study the exocurves, $A_j$, in some detail. In particular, we prove:
In the $r>0$ ‘sheet’ of the field space, $\cF$, the exocurves (\[e:Antennae\]) are $$A^+_j~\simeq~\IC\IP^1_{[-5,1]}~\simeq~\IC^1~.$$
In the $r>0$ ‘sheet,’ the definition (\[e:Antennae\]) of the exocurve: $$A^+_j~\define~\{p,\sh_j\}\Big/\hat{\lm}~, \Label{e:Antenna}$$ includes implicitly that $\|s\|^2\geq|r|>0$ owing to , and the superscript ‘$+$’ reminds that $r{>}0$. That is, $$(p,\sh_j)~\cong~(\lm^{-5}p,\lm\sh_j)~,\qquad \lm\in\IC^*~,
\Label{e:proj}$$ which defines $A^+_j$ as the weighted projective space $A^+_j=\IP^1_{[-5,1]}$, proving the first part of (\[e:Antenna\]). This case, however, differs from the usual consideration of weighted projective spaces [@dim:weight] in that the weights, $-5$ and $1$, are of opposite sign. Still, we proceed by considering the two candidate charts: $$U_p=(p,\sh_j)_p~\cong~(1,u_p)~,\quad p\neq0~,\qquad
u_p\define\sh_j\,p^{1/5}~, \Label{e:Up}$$ and $$U_s=(p,\sh_j)_s~\cong~(u_s,1)~,\quad \sh_j\neq0~,\qquad
u_s\define p\,(\sh_j)^5~. \Label{e:Us}$$ In both cases, the equivalences are obtained using the map (\[e:proj\]), however with $\lm=p^{1/5}$ in the first case, and $\lm=(\sh_j)^{-1}$ in the second. Now, in the second candidate chart, $U_s$, the limit point $p,u_s\to0$ is included, and so $$U_s=(p,\sh_j)_s~\cong~(u_s,1)~\simeq~\IC^1$$ is a proper chart. On the other hand, in the first candidate chart, $U_p$, the limit point $\sh_j,u_p\to0$ is [*excluded*]{} by the inequality (\[e:snot0\]), so that $$U_p=(p,\sh_j)_p~\cong~(1,u_p)~\simeq~\IC^*$$ is [*not*]{} a proper chart. In its place, we should introduce two $\IC^1$-like charts which cover $U_p$. However, this will not really be necessary since and (\[e:Us\]) imply that $$u_s=p(\sh_j)^5~\mapsto~u_p^5~, \Label{e:glue}$$ which is a 1-to-5 holomorphic map outside $u_s{=}0$. That is, $U_p=(1,u_p)\simeq\IC^*$ is a 5-fold cover of $(U_s{-}0)=(u_s,1)_{u_s{\neq}0}\simeq\IC^*$; $u_s{=}0$ is of course the branching point of this holomorphic covering. Therefore, $A^+_j$ may be parametrized by $u_s$ and so $A^+_j\simeq U_s$. With (\[e:glue\]) as the ‘glueing map,’ we then have that the $j^{th}$ exocurve is: $$A^+_j~\simeq~\IP^1_{[-5,1]}~=~U_p \cup U_s ~=~ U_s ~\simeq~ \IC^1.
\Label{e:lemma2a}$$
In the $r<0$ ‘sheet’ of the field space, $\cF$, the exocurves (\[e:Antennae\]) are $$A^-_j~\simeq~\IC\IP^{1,-}_{[-5,1]}~\simeq~\IC^1/\ZZ_5~.$$
In the $r<0$ ‘sheet,’ the definition (\[e:Antenna\]–\[e:proj\]) still guarantees that $A^-_j\simeq\IP^{1,-}_{[-5,1]}$, but now enforces $|p|^2\geq|r|>0$, as indicated by the superscript ‘$-$’. We again proceed by considering the two candidate charts (\[e:Up\]) and (\[e:Us\]). This time, it is in the first candidate chart, $U_p$, that the limit point $\sh_j,u_p\to0$ is included, and so $$U_p=(p,\sh_j)_p~\cong~(1,u_p)~\simeq~\IC^1$$ is a proper chart. Similarly, in is now the second candidate chart, $U_s$, from which the limit point $p,u_s\to0$ is [*excluded*]{} by the inequality (\[e:pnot0\]), so that $$U_s=(p,\sh_j)_s~\cong~(u_s,1)~\simeq~\IC^*$$ is [*not*]{} a proper chart. Again, it is not necessary to introduce two $\IC^1$-like charts to cover $U_s$, since and (\[e:Us\]) again imply the 1-to-5 holomorphic map (\[e:glue\]) now outside $u_p{=}0$. Now $(U_p{-}0)=(1,u_p{\neq}0)\simeq\IC^*$ is a 5-fold cover of $U_s=(u_s,1)\simeq\IC^*$, and $u_p{=}0$ is of course the branching point of this holomorphic covering. Therefore, $A^-_j$ now must be parametrized by $u_p$ which is 5-fold redundant except at $u_p{=}0$. Therefore, we now have that the $j^{th}$ exocurve is: $$A^-_j~\simeq~\IP^{1,-}_{[-5,1]}~=~(U_p \cup U_s)/\ZZ_5~=~ U_p/\ZZ_5
~\simeq~ \IC^1/\ZZ_5. \Label{e:lemma2b}$$
Note that the $\ZZ_5$ quotient in Lemma \[l:Antennae-\] precisely corresponds to the $\ZZ_5$ quotient in . Indeed, this says that the “fuzzy point” of the Landau-Ginzburg orbifold (\[e:LGO\]) becomes $$(\sqcup_{j=1}^n A^-_j\Big) ~\simeq~ (\IC^1/\ZZ_5)^{\vee5}~,$$ , the ‘plum product’ of five copies of the $\IC^1/\ZZ_5$ cone, all connected at the vertex—the “fuzzy point” (\[e:LGO\]).
The exocurves are illustrated in Fig. \[f:antenna\].
(170,135)(-20,0) (0,0) (160,5)[$\sh_j$]{} (80,2)[$\shx_j$]{} (88,18)[$\shM$]{} (5,33)[$\sqrt{-r/5}$]{} (5,26)[$(r{<}0)$]{} (-5,125)[$p$]{} (23,15)[$\sqrt{r}$, $(r{>}0)$]{} (10,53)[$\sqrt{r_+}=\sqrt{r+5|p|^2}$]{} (132,115)[$A^+_j\simeq\IC^1$]{} (126,109)[the $r{>0}$ ‘sheet’]{} (121,103)[of the exocurve]{} (58,66)[$=\sqrt{r_-}$]{} (58,73)[$\sqrt{|r|/5+\|\sh_j\|^2}$]{} (83,122)[$A^-_j\simeq\IC^1/\ZZ_5$]{} (73,116)[the $r{<0}$ ‘sheet’]{} (64,110)[of the exocurve]{}
A comparison
------------
For comparison, we include a similar analysis of $\IP^1_{[5,1]}$. In contrast to the non-compact $\IP^1_{[-5,1]}$, the weighted projective space $\IP^1_{[5,1]}$ will prove to be compact.
Again, it is possible to view $\IP^1_{[5,1]}$ both as a holomorphic quotient, $$\IP^1_{[5,1]} ~\simeq~ \left\{q,s\right\}/\hat{\m}$$ where $$\hat\m\colon (q,s) \mapsto (\m^5q,\m s)~,\qquad \m\in\IC^*~,
\Label{e:projmap}$$ and also as a symplectic quotient, $$\IP^1_{[5,1]} ~\simeq~
\left\{\left\{q,s\right\}\cap \D_r^{-1}(0)\right\}/S^1$$ where the $S^1$-action is the restriction of (\[e:projmap\]) to $|\m|=1$ and now becomes $$\D_r~=~|s|^2+5|q|^2-r~,\qquad r\in\IR~. \Label{e:gauged}$$ Following the proofs of Lemmas \[l:Antennae+\] and \[l:Antennae-\], we consider the latter.
The vanishing of $\D_r$ now simply states that $$|s|^2+5|q|^2 ~=~r~\geq~0~, \Label{e:rPos}$$ where $r=0$ would force $s=0=q$, the trivial solution. Restricting then to $r\neq0$, implies the positive definiteness of $r$ on $\D_r^{-1}(0)$, so that there is only the $r>0$ sheet. Indeed, this is precisely why the $r<0$ sheet of $\IP^1_{[-5,1]}$ appears to be rather unfamiliar an object. Owing to the inequality (\[e:rPos\]), $s,q$ must not vanish [*simultaneously*]{}; either one of them however may very well vanish while the other one is nonzero. Thus, unlike in the case of $\IP^1_{[-5,1]}$, we now have two perfectly proper coordinate charts: $$U_q=(q,s)_q~\cong~(1,u_q)~\simeq~\IC^1~,\quad\hbox{using}\quad
\m=q^{-1/5}~,~ q\neq0~,\quad u_q=s\,q^{-1/5}~,$$ and $$U_s=(q,s)_s~\cong~(u_s,1)~\simeq~\IC^1\quad\hbox{using}\quad
\m=s^{-1}~,~s\neq0~,\quad u_s=q\,s^{-5}~.$$ The two chart coordinates, $u_q$ and $u_s$, respectively, [*can*]{} attain the value of 0, since $s,u_q\to0$ is permitted in $U_q$ where $q\neq0$, and $q,u_s\to0$ is permitted in $U_s$ where $s\neq0$. Finally, the two charts are glued through the relation $$u_s~=~u_q^{-5}~,\qquad\hbox{where}\quad u_s,u_q\neq0~,
\Label{e:glue1}$$ which provides a 1-5 map: $$\left\{U_s-0\right\} ~\tooo{~1-5~}~ \left\{U_q-0\right\}~.$$ To render the map (\[e:glue1\]) single-valued, we may glue together $U_s$ and $U_q/\ZZ_5$: $0\in U_s$ becomes ‘$\infty$’ added to $U_q/\ZZ_5$, and $0\in
U_q/\ZZ_5$ becomes ‘$\infty$’ added to $U_s$. The resulting space, $\IP^1_{[5,1]}=U_s\cup(U_q/\ZZ_5)$, then is compact and smooth except at $0\in U_q/\ZZ_5$, where $\IP^1_{[5,1]}$ has a $\ZZ_5$ quotient singularity, , a conical singularity with $(1-\inv5)2\p=8\p/5$ deficit angle.
Note, however, that by Delorme’s Lemma [@dol:weight], $\IP^1_{[k,1]}\approx\IP^1_{[1,1]}\equiv\IP^1\simeq S^2$. The relationship ‘$\approx$’ here denotes a $k$-to-$1$ map of the coordinates as used here, but an isomorphism of the corresponding coordinate rings, which then extends to an isomorphism of the respective spaces [@dol:privco].
A one-point compactification of exocurves
-----------------------------------------
For later convenience and use, we describe here a one-point compactifications of the exocurves.
It is straightforward from the proof of Lemmas (\[l:Antennae+\]) that the limiting point $s,u_p\to\infty$ may be added to $U_p$. Upon the inversion of its variables, this now becomes a proper coordinate chart: $$\Tw{U}_p = (1,w_p)~\simeq~\IC^1~,\qquad
w_p\define u_p^{-1}={\sh}_j^{-1}p^{-1/5}~.$$ Clearly, the glueing map now becomes $$u_s~=~w_p^{-5}~\colon~
\left\{U_s-0\right\}~\tooo{~1-5~}~\left\{\Tw{U}_p-0\right\}~.
\Label{e:glue2}$$ To render the glueing map (\[e:glue2\]) single-valued, we form $$\bar{A}{}_j^+~\define~U_s\cup(\Tw{U}_p/\ZZ_5)~\simeq~\IP^1_{[5,1]}~.
\Label{e:CompAnt}$$ As shown above, this is compact and isomorphic to $\IP^1\simeq S^2$.
Cohomology and Homology of $\bar\cV$
====================================
As presented in Theorem 1, the stratified variety can be written as $$\bar \cV~=~\shM ~\cup~ \bigcup_{j=1}^n \bar {A}_j~, \qquad
\shM \cap \bar{A}_j = \shx_j~. \Label{e:AntennaM2}$$
The Mayer-Vietoris principle then induces the long exact cohomology sequence $$\3\to H^q(\bar\cV) \to H^q(\shM)\oplus H^q(\cup_j\bar{A}_j)\to
H^q(\shM\cap\cup_j\bar{A}_j)\to H^{q+1}(\bar\cV)\to\3
\Label{e:LongHomSeq}$$ Since $$\shM\cap\cup_j\bar{A}_j=\cup_j(\shM\cap\bar{A}_j)=\sqcup_j\shx_j~,$$ owing to our assumption that $\shx_j$ are isolated (non overlapping) nodes, and $$H^q(\cup_j\shx_j)=\bigoplus_{j=1}^n H^q(\shx_j)\simeq\d_{q,0}\IC^{\oplus n}~,$$ the long exact sequence (\[e:LongHomSeq\]) breaks into five isomorphisms: $$H^q(\bar\cV) = H^q(\shM)\oplus H^q(\cup_j\bar{A}_j)~,
\qquad\hbox{for~}q=2,{\cdots},6~, \Label{e:IsoMV}$$ and $$0\to H^0(\bar\cV)\tooo{~\a~} H^0(\shM)\oplus H^0(\cup_j\bar{A}_j)\tooo{~\b~}
H^0(\sqcup_j\shx_j)\to H^1(\bar\cV)\to0~.$$ The above map $\b$ is induced from the injective inclusion $\sqcup_j\shx_j=\shM\cap(\sup_j\bar{A}_j)\to\shM\sqcup(\cup_j\bar{A}_j)$, and so is surjective. Then: $$H^1(\bar\cV)=\emptyset~, \quad\hbox{and}\quad
H^0(\bar\cV) \simeq \IC~. \Label{e:BegMV}$$
Contributions from the antennae
-------------------------------
Recall from a previous section that $\bar {A_j} \simeq\IP^1\simeq S^2 $. Then, $$H^q(\bar{A}_j)~~\left\{
\begin{array}{ll}
\simeq\IC^1 &\mbox{for~} q=0,2~,\\[1mm]
=\emptyset &\mbox{otherwise.}
\end{array}\right.$$ With this, for $\shM$ with $n$ simple nodes, and (\[e:BegMV\]) would seem to imply that $H^q(\cV)$ should equal to $H^q(\shM)$, except for $q=2$, where it ought to be augmented by $H^2(\cup_j\bar{A}_j)\simeq\oplus\IC^{\oplus n}$.
This, however, is not correct: the (area) 2-forms of the $n$ antennae are not independent cohomology elements. As described in detail in Ref. [@hub:cymani], $N$ mutually exclusive subsets of the $n$ $\shx_j$’s lie on corresponding 4-cycles $C^{(4)}_k\subset\shM$, $k=1,{\cdots},N$. Let $J_k$ denote the multiindex containing the indices, $j$, of all $\shx_j$’s that lie on $C^{(4)}_k$. Clearly then, $$\bar{A}_j \cap C^{(4)}_k ~=~ \left\{
\begin{array}{ll}
\shx_j & \mbox{if~}j\in J_k\mbox{~{\it i.e.}~}\shx_j\in C^{(4)}_k~,\\[1mm]
\emptyset & \mbox{otherwise.}
\end{array}\right. \Label{e:ClasSing}$$ Considering then the [*homology*]{} elements in $H_q(\bar\cV)$, dual to the cohomology group obtained in and (\[e:BegMV\]), and denoting them by square brackets, we have: $$[\bar{A}_j] \cap [C^{(4)}_k] ~=~ \left\{
\begin{array}{ll}
1 & \mbox{if~}j\in J_k~,\\[1mm]
0 & \mbox{otherwise.}
\end{array}\right. \Label{e:DualHom}$$ Owing to this result, it follows that $$\left.\begin{array}{lll}
[\bar{A}_j]=[\bar{A}_{j'}] &\mbox{if~~}j,j'\in J_k~,\\[1mm]
[\bar{A}_j]\neq[\bar{A}_{j'}] &\mbox{otherwise.}
\end{array}\right\} \Label{e:EquAnt}$$ That is, the $n$ antennae, $\{\,\bar{A}_j\,\}$ contribute only $N$ inequivalent 2-cycles, so $$H_2(\cup_j\bar{A}_j) ~\simeq~ \IC^{\oplus N} ~\simeq~ H^2(\cup_j\bar{A}_j)~.
\Label{e:AntCoh}$$
The combined result
-------------------
Combining , (\[e:BegMV\]) and (\[e:AntCoh\]) proves
Let $\bar\cV$ as defined in , where $\shM$ is a conifold with only $n$ isolated nodes ($\shx_j$) lying on $N$ distinct 4-cycles $C^{(4)}_k$, and $\bar{A}_j$ as defined in . Then, $$H^q(\bar\cV) = \left\{
\begin{array}{l}
H^q(\shM)\quad \mbox{for~} q\neq2~,\\[1mm]
H^2(\shM)\oplus H^2(\cup_j\bar{A}_j) \simeq H^2(\shM)\oplus\IC^{\oplus N}~.
\end{array}\right.$$
As it stands, with $H^3(\bar\cV)=H^3(\shM)$, the [*complete*]{} $H^*(\bar\cV)$ can have neither Poincaré duality nor a Hodge decomposition. Both are obstructed by the fact that the 3-cycle(s) which pass through the $\shx_j$’s remain without dual 3-cycle(s) [@hub:cymani]. In fact, the subgroup of $H^3(\bar{\cV})$ generated by the 3-cycles passing through the $\shx_j$’s may well be odd-dimensional, making this obstruction manifest.
However, $\oplus_qH^{2q}(\bar\cV)$ subgroup does exhibit both Poincaré duality and a Hodge decomposition. As usual, $H^2(\bar{A}_j)\simeq\IC^1$ is generated by the volume $(1,1)$-form on $\bar{A}_j\simeq\IP^1$. Moreover, dually to the homology result (\[e:DualHom\]), the volume $(1,1)$-forms, $\w_{(1,1)}^j$, of all $\bar{A}_j$’s which intersect $C^{(4)}_k$ are dual to the $(2,2)$-form $\w_{(2,2)}^k$, itself dual to $C^{(4)}_k$. In fact, the double dualities[^9] $$\w_{(1,1)}^j\dual[\bar{A}_j]\Dual[C^{(4)}_k]~, \quad\mbox{and}\quad
[\bar{A}_j]\Dual[C^{(4)}_k]\dual\w_{(2,2)}^k~, \qquad\mbox{for~}j\in J_k~,
\Label{e:DDual}$$ establishes the isomorphisms $$\w_{(1,1)}^j\simeq[C^{(4)}_k]~, \quad\mbox{and}\quad
[\bar{A}_j]\simeq\w_{(2,2)}^k~, \qquad\mbox{for~}j\in J_k~,
\Label{e:Isom}$$ whereupon implies that also $$\left.\begin{array}{lll}
[\w_{(1,1)}^j]=[\w_{(1,1)}^{j'}] &\mbox{if~~}j,j'\in J_k~,\\[1mm]
[\w_{(1,1)}^j]\neq[\w_{(1,1)}^{j'}] &\mbox{otherwise.}
\end{array}\right\} \Label{e:EqVAnt}$$ On a more fomal level, implies that also $$[\w_{(1,1)}^j] \cup [\w_{(2,2)}^k] ~=~ \left\{
\begin{array}{ll}
1 & \mbox{if~}j\in J_k~,\\[1mm]
0 & \mbox{otherwise.}
\end{array}\right. \Label{e:DualCoh}$$ Clearly, the evaluation map of the cup product here cannot be the integration (of the wedge product of the indicated forms) over the stratified variety $\bar\cV$ in any conventional sense. Instead, it may be taken to reduce to the evaluation over the point of common support, $\shx_j=\bar{A}_j\cap C^{(4)}_k$ if $j\in J_k$, and is vacuous otherwise.
Owing to the isomorphisms (\[e:Isom\]), the $[\bar{A}_j]\Dual[C^{(4)}_k]$ duality (when $j\in J_k$) implies the desired Poincaré duality of $\w_{(1,1)}^j\Dual\w_{(2,2)}^k$, for all $j\in J_k$. Let $\vev{\w_{(2,2)}^k}$ denote the subgroup of $H^{(2,2)}(\shM)$ generated by the $\w_{(2,2)}^k$’s. The quotient $H^{(2,2)}(\shM)/\w_{(2,2)}^k$ is then generated by the $(2,2)$-forms dual to 4-cycles which do not pass through $\shx_j$; this quotient is easily seen to form an additive group, exhibiting both poincaré duality and Hodge decomposition.
The foregoing then proves:
Let $\bar\cV$ as defined in , where $\shM$ is a conifold with only $n$ isolated nodes ($\shx_j$) lying on $N$ distinct 4-cycles $C^{(4)}_k$, and $\bar{A}_j$ as defined in . Then, $$\oplus_q H^{2q}(\bar\cV) = \left\{
\begin{array}{l}
H^{2q}(\shM)\quad \mbox{for~} q\neq1~,\\[1mm]
H^2(\shM)\oplus H^2(\cup_j\bar{A}_j) \simeq H^2(\shM)\oplus\IC^{\oplus N}~.
\end{array}\right.$$ has both an induced Hodge decomposition and Poincaré duality, as induced by the double dualities (\[e:DDual\]).
Deformations, Resolutions and the Mirror Map
============================================
We have originally restricted $\shM$ to conifolds with only nodes (, double points, or $A_1$ hypersurface singularities), $\shx_j$. Their local neighborhood is isomorphic to the cone $\IC^4/Q$, where $Q$ is a non-degenerate quadratic polynomial over $\IC^4$. In a small resolution, this neighborhood is replaced with a copy of the total space of an ${\cal O}(-1,-1)\define{\cal O}(-1){\oplus}{\cal O}(-1)$ bundle over $\IP^1\simeq S^2$. In short, a small resolution replaces each node $\shx_j$ with a $(-1,-1)$-curve, $\IP^1_{\natural,j}\simeq S^2$. Since there are two topologically distinct ways to do this at each node $\shx_j$, a conifold $\shM$ with $n$ would appear to have $2^n$ small resolutions, $\nM_I$. However, all nodes $\shx_j$ which lie on a single 4-cycle $C^{(4)}_k\subset\shM$ must be resolved “compatibly”: all the corresponding 2-spheres $\IP^1_{\natural,j}\in\nM_I$ intersect $C^{(4)}_{\natural,k}\in\nM_I$ (the proper transform of $C^{(4)}_k\subset\shM$) in a single point and so must all represent the same element of $H_2(\nM_I)$, the one that is dual to $C^{(4)}_{\natural,k}$. With the use of , this implies that $$\left.\begin{array}{ll}
[\w_{(1,1)}^{\natural,j}]=[\w_{(1,1)}^{\natural,j'}] &\mbox{if~~}j,j'\in J_k~,\\[1mm]
[\w_{(1,1)}^{\natural,j}]\neq[\w_{(1,1)}^{\natural,j'}] &\mbox{otherwise,}
\end{array}\right\} \Label{e:EqVAntSmR}$$ and $$[\w_{(1,1)}^{\natural,j}] \cup [\w_{(2,2)}^{\natural,k}] ~=~ \left\{
\begin{array}{ll}
1 & \mbox{if~}j\in J_k~,\\[1mm]
0 & \mbox{otherwise.}
\end{array}\right. \Label{e:DualCohSmR}$$ for $$\begin{aligned}
H_2(\nM_I)\ni\IP^1_{\natural,j}
&\dual& \w_{(1,1)}^{\natural,j}\in H^{(1,1)}(\nM_I)~, \\
H_4(\nM_I)\ni C^{(4)}_{\natural,k}
&\dual& \w_{(2,2)}^{\natural,k}\in H^{(2,2)}(\nM_I)~.\end{aligned}$$ Of course, in , the cup product is indeed obtained as the ordinary wedge product, integrated over the (smooth) manifold $\nM_I$ Consequently, the multiplicity of small resolutions to $I=1,{\cdots},2^N$, where $N$ is the number of $H_2(\nM_I)$ elements which the small resolution exceptional sets, $\IP^1_{\natural,j}$ represent, , the number of $H_4(\nM_I)$ elements, $C^{(4)}_{\natural,k}$, which are the proper transforms of the 4-cycles that pass through the nodes $\shx_j\in\shM$. [@hub:cymani].
The formal identity of the –(\[e:DualCoh\]) with the –(\[e:DualCohSmR\]) then proves:
Let $\shM$ be a Calabi-Yau complex 3-dimensional algebraic variety with only a finite number of isolated nodes, $\shx_j$. Let $\nM_1$ and $\nM_2$ denote two small resolutions of $\shM$, related by a flop: $\nM_1\buildrel{{\rm f}}\over{\longleftrightarrow}\nM_2$. Finally, let $\bar{\cV}$ be the compactification of the stratified variety (\[e:AntennaM2\]). Then the flop involution $\nM_1\buildrel{{\rm f}}\over{\longleftrightarrow}\nM_2$ generalizes to a triple of (exo)flops: $$\mbox{\begin{picture}(100,45)(0,0)
\put(-5,0){\epsfxsize=100mm\epsfbox{TriFlop.eps}}
\put(-12,25){$\cM^\flat$}
\put(16,24){defo}
\put(38,30){$\bar{\cV}$}
\put(58,31){exoflop}
\put(52,10){exoflop}
\put(95,33){$\nM_1$}
\put(95,7){$\nM_2$}
\put(91,19){flop}
\end{picture}}$$
The map “[*defo*]{}” to the left is realized as follows: Deformations smooth $\shM$ by replacing the local cones $\IC^4/Q$ centered at each node, $\shx_j$, with a real 3-bundle over a copy of $S^3$. It is easy to see that a deformation of $G(s)$ from the non-transversal choice to a transversal choice in section \[s:TransV\] precisely induces the smoothing of the ground state variety from a (compactified) stratified variety of the type described in section \[s:SingV\] to a smooth Calabi-Yau 3-fold of the type described in section \[s:TransV\]. This provides the map $\cM^\flat\buildrel{{\rm defo}}\over{\longleftrightarrow}\bar{\cV}$ in the diagram in Lemma \[l:TriFlop\].
Finally, we note that the above described homology of $\bar{\cV}$ excluding the middle dimension, which we have not discussed herein, satisfies the requirements given in Ref. [@hub:strcoh], and so is compatible with the ‘mirror map.’ The extension of this result to include the (co)homology groups in the middle dimension remains an open question for now and we hope to return to it in a future effort.
#### Acknowledgments:
The authors would like to thank M. Goresky (Institute for Advanced Study, Princeton, NJ) and N. Ramachandran (University of Maryland, College Park, MD) for their helpful comments.
[text]{} plus1fill
P. Aspinwall, B. Greene and D. Morrison: (1994)414–480 .
P. Candelas, G. Horowitz, A. Strominger and E. Witten: (1985)46.
P. Candelas, P.S. Green and T. Hübsch: (1990)49–102.
A. Dimca: [*Singularities and Coverings of Weighted Complete Intersections*]{}, [*R.Ang.J.Math.*]{} [**366**]{}(1986)184–193.
I. Dolgachev: [*Weighted Projective Varieties*]{} in Group Actions and Vector Fields, [*Lect.N.Math.*]{}[**956**]{}(1982) 34–71.
I. Dolgachev: private communication.
W. Fulton: [*Toric Varieties*]{} (Princeton University Press, Princeton, 1990).
J. Cheeger, M. Goresky and R. MacPherson, [*Ann. Math. Studies*]{} [**102**]{} (1982) 303-340.
M.B. Green, J.H. Schwarz and E. Witten: [*String Theory*]{}, (Cambridge University Press, Cambridge, 1987).
E. Gunter: [*Convex Bodies and Toric Varieties*]{} (Springer-Verlag, New York, 1990)
T. Hubsch: [*Calabi-Yau Manifolds: A Bestiary for Physicists*]{}, (World-Scientific, Singapore, 1992).
T. Hubsch; (1997)521.
J. Polchinski: [*String Theory*]{}, (Cambridge University Press, Cambridge, 1997).
A. Strominger: (1995)96-108.
E. Witten: (1993)159–222.
S.-T. Yau, ed.: [*Mirror Manifolds*]{} (International Press, Hong-Kong, 1990);\
B. Greene and S.-T. Yau, eds.: [*Mirror Manifolds II*]{} (International Press, Hong-Kong, 1996).
[^1]: [email protected]; On leave from the “Rudjer Bošković” Institute, Zagreb, Croatia.
[^2]: [email protected]
[^3]: To avoid obscuringly complicated notation, we focus on a simple example and discuss generalizations later.
[^4]: The general category of ‘stratified varieties’ is specified for example in the works [@gmp:strata]. Our situation is far simpler: we will encounter unions of several (complex, algebraic) varieties of complex dimension $0,\cdots,3$, possibly connected at codimension$\geq1$ subspaces.
[^5]: We will use the prefix ‘exo’ to denote (components of) strata that are ‘external’ to the ‘main’ stratum.
[^6]: Transversality ensures that the projective hypersurface defined by $G=0$ is smooth.
[^7]: The subsequent analysis for non-transversal $G$, the case of our real interest, is more detailed and shown below. The Reader can then recover the presently omitted details as a special case; see also Ref. [@wit:phases].
[^8]: Following Ref. [@cgh:confld], a [*conifold*]{} is a variety which is smooth except for a finite number of isolated conical singularities. Furthermore, herein we will consider only varieties with nodes (double points).
[^9]: By $\dual$ we denote the standard homology–cohomology duality, and use $\Dual$ for Poincaré duality.
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'M. Kunert'
- 'A. Marecki'
- 'R. E. Spencer'
- 'A. J. Kus'
- 'J. Niezgoda'
date: 'Received 12 November 2001/ Accepted 5 April 2002'
subtitle: 'I. MERLIN images of arc-second scale objects'
title: 'FIRST-based survey of Compact Steep Spectrum sources'
---
Introduction
============
Compact Steep Spectrum (CSS) sources [@kapahi; @pw82] form a well defined class of radio sources; they are powerful, compact (projected linear sizes $\leq$ 20 kpc and hence angular sizes of the order of a few arcseconds) and possess steep ($\alpha\geq 0.5$, $S\propto\nu^{-\alpha}$) spectra. CSSs are identified with quasars, radio galaxies and Seyferts. @wsrps84 found a morphological separation between CSS quasars and galaxies [see also @spencer89; @fanti90] which is similar to that observed for larger sources: radio galaxies generally have a simple double radio structure (sometimes with weak radio jets and weak radio cores) whilst quasars show either a triple structure (with a strong central component consisting of a bright jet) or complex structure.
An astrophysical interpretation of the CSS phenomenon has been given in @fanti90. They show that in principle small apparent sizes of CSSs could result from the projection of “normal” Large Symmetric Objects (LSO), however only less than 25% of these objects are expected to be larger sources seen close to the line of sight. Most of them are supposed to be intrinsically small objects randomly oriented on the sky and their sub-galactic apparent linear dimensions can be explained by two main hypotheses. According to the first one CSS sources are confined by the interaction of the jet with an inhomogeneous, dense and possibly turbulent medium in the host galaxy which inhibits a normal development [@bmh84]. In this scenario CSSs are so called [*frustrated*]{} objects. The second hypothesis [@pm82; @c85; @mp88] — and this one has gained more observational support recently [see e.g. @fanti2000] — suggests that CSS sources may be the [*young*]{} stages of future LSOs and so the compactness of these objects is just an evolutionary effect: they are small because they have not had enough time to expand to supergalactic scales. If that hypothesis is correct, CSSs can be regarded as an intermediate class between even smaller Compact Symmetric Objects (CSOs) [@r96] and LSOs and, together, these three groups of radio sources make up an evolutionary sequence. One of the main argument in favour of the evolution is that CSOs and some CSS sources, namely Medium-sized Symmetric Objects (MSOs) [@aug98] which are unbeamed CSSs, have similar morphologies to LSOs. On the other hand, the lobe speeds in CSO sources are high: $\sim0.2c$ [@oc98; @ocp98], so CSOs quickly evolve into larger objects and MSOs seem to be perfect candidates to become post-CSOs. Most of the CSS sources known so far have sufficiently high radio luminosity that even assuming a strong decrease in luminosity as they evolve, they remain good candidates for future large scale Fanaroff–Riley class II (FRII) [@fr74] sources with high radio luminosity. An additional support of this view comes from the morphological similarity of many CSSs to FRIIs as expected if the evolution is self-similar.
All studies of CSS sources made by 1995 i.e. until publication of papers by @dffss95 and @sang95, were based on the so called 3CRPW sample consisting of 54 sources [@spencer89]. The next step in such investigations can be made through extension of the available sample of CSSs toward weaker sources. For example, @saikia01 selected a sample of 42 candidates from the S4 survey [@pt78] which is complete to 0.5 Jy at 5 GHz and @fanti01 — hereafter F2001 — derived a new sample of 87 CSSs with flux densities $\geq 0.8\mathrm{Jy}$ at 408 MHz from B3-VLA survey [@vig89]. Here we present a genuine method of finding weak CSSs which makes use of [*Faint Images of Radio Sky at Twenty*]{} (FIRST) [@wbhg97]. Its fine resolution ($5\farcs4$) is the crucial feature for that purpose. A strong motivation to conduct the research in this direction came from the fact that a number of CSS sources weaker than those in the 3CRPW sample have already been mapped with the VLA at 8.4 GHz by @pbww92[^1]. It appears that those sources have angular sizes of the same range as all the CSSs known so far, yet they are significantly weaker. There are two plausible explanations: a) these sources are those CSSs we see almost exactly face-on, so their Doppler boosting is minimal, or b) these sources form a new class of “weak” CSSs.
In 1996 we proposed observations based on newly released early results of FIRST aimed to discriminate between cases a) and b). If case a) is true then those ones with counter-jets should dominate among these new CSS sources; sources of this kind were rare in the 3CRPW sample. If case b) is true then there is a chance to find an elegant analogy: strong CSSs are “miniature FRIIs” whereas weak CSSs are “miniature FRIs”. Taking into account the evolutionary scenario and assuming that radio sources evolve in a self-similar manner we might also say that strong CSSs from the 3CRPW sample evolve towards FRIIs whereas weak CSSs evolve towards FRIs. Testing such a possibility is among the goals of a series of papers resulting from interferometric (MERLIN, EVN, VLBA) observations of our FIRST-based sample of CSS candidates.
Source 4C RA (J2000) DEC (J2000) $m_R$ $m_v$ $z$ $F_{1.4\mathrm{GHz}}$ $F_{4.85\mathrm{GHz}}$ $\alpha_{1.4\mathrm{GHz}}^{4.85\mathrm{GHz}}$
-------- -------- -------------- ------------- --------- ------- ------- ----------------------- ------------------------ -----------------------------------------------
+30.13 08 04 42.148 30 12 37.91 18.1 19.2 1.446 1189 404 0.87
+40.19 08 09 03.158 40 32 56.72 $>20.8$ 21.1 —– 437 179 0.72
+33.22 08 53 21.100 32 55 00.60 —– —– —– 465 208 0.65
12 04 06.859 39 12 18.17 19.6 21.4 0.445 468 162 0.85
12 35 35.706 41 37 07.40 17.9 20.8 0.25 651 276 0.69
\[table1\]
Sample selection
================
To select weak CSS sources from the FIRST catalogue we made the following steps:
1. From the source list based on Green Bank (GB) surveys at 21 and 6 cm @wb92 we selected those sources lying within the then current limits of the FIRST survey when FIRST covered the area of declination $28\degr$ — $42\degr$, having steep spectra ($\alpha >0.5$) and being stronger than 150 mJy at 6 cm. (This flux density limit was chosen in order to produce a sample of manageable size.) The above declination limits indicate that the overlap between our sample and the B3-VLA survey based sample of F2001 (their limits are $37\degr 15\arcmin$ — $47\degr 37\arcmin$) is not large.
2. We identified FIRST sources with those GB survey sources. We found, quite expectedly, that thanks to a dramatic difference in the resolution, the majority of sources appearing as single in the GB survey turn out to be double (or multiple) on FIRST maps and so they are represented either as compact pairs or clusters of pointlike sources in the FIRST catalogue.
3. We rejected all such cases i.e. we selected only those sources that are single entities in the FIRST catalogue i.e. more compact than the FIRST beam (54) and surrounded by an empty field. We adopted 1 arcmin as a radius of that field. Such a procedure allows us to make sure that we deal with isolated objects and not parts of larger objects.
4. We, again, checked whether our targets fulfill the spectrum steepness criterion: instead of GB-survey flux densities at 21 cm we used more accurate values from FIRST. We rejected candidates with flat spectra ($\alpha \leq 0.5$).
5. We found that all already known CSS sources lying within our R.A. and declination limits have been correctly selected so far. Obviously we rejected them.
6. We rejected the Gigahertz Peaked Spectrum (GPS) sources because — in our opinion — they constitute a separate class. The main reason for this is that GPSs are an order of magnitude more compact than CSSs and their spectra have a different shape. Our research was focused on “true” CSSs and not GPSs[^2] To this end we identified our preliminary candidates with objects listed in 365 MHz Texas catalogue [@douglas]. We passed only those objects which have non-inverted spectra between 365 and 1400 MHz. In other words the turnover frequencies of our sources lie below 365 MHz.
Finally we selected 60 candidates for CSS sources. Radio selected samples normally suffer from redshift information scarcity and it was the case here. Therefore, for the majority of our candidates it was not possible to calculate their distances and to judge which of them fulfill the linear size criterion for CSSs which, obviously, is of primary importance for the physics and evolution issues. Instead we used the angular size criterion which is still helpful for making sure we reject objects with excessively large linear sizes. Since the resolution of FIRST is 54, we realised from the beginning that a number of those candidates may *not* fulfill the angular size criterion and so a rejection of sources with angular sizes larger than certain limit yet pointlike according to FIRST was planned as the first step after completion of the initial survey of all those 60 targets. Assuming that the linear sizes of a CSS source should remain below 20 kpc for currently adopted cosmological parameters, in particular for $H_0=72$ km $\mathrm s^{-1} \mathrm {Mpc}^{-1}$ [@hstkp01], we adopted $3\arcsec$ for such a criterion.
Observations and data reduction
===============================
The initial survey was performed with MERLIN at 5 GHz. At that frequency MERLIN attains a resolution of $0\farcs04$ which is sufficient to make a final selection of actual CSSs from the list of our candidates. We made snapshot observations of the sample of CSS sources defined above in 1997. Our targets were observed 6 times in 10 min. scans spread evenly over a 12-hour track. Six MERLIN telescopes were used. A typical $u-v$ coverage accomplished in those observations is shown on Fig. \[uvc\].
{width="8cm"} {width="8cm"}
{width="8cm"} {width="8cm"}
Phase calibrator sources chosen from the MERLIN Calibrator List [@pbww92] were observed twice per target scan for 1–2 min. Poor weather conditions allowed us to observe only a part of our sample in 1997, however we successfully observed and mapped about 3 dozen sources. At this point we rejected the sources which were too large to be regarded as CSS sources using a $3\arcsec$ limit for the angular size. The remaining 21 objects were indeed new CSS sources. We divided them into 3 groups:
{width="8cm"} {width="8cm"}
- relatively large ones with sizes ranging from $1\arcsec$ to $3\arcsec$ (6 sources);
- relatively compact ones with typical sizes of $0\farcs5$ and double structure (9 sources);
- relatively compact ones with typical sizes of $0\farcs5$ and complex structure (6 sources).
Objects from groups 2 and 3 as well as the full list of objects in our sample will be described in subsequent papers. In this paper we focus on the first group i.e. on the objects possessing similar sizes to classical CSSs, yet less luminous[^3]. Further observations have been made using MERLIN at 1.6 GHz in ‘snapshot’ mode of six sources from the first group to make calculation of components’ spectral indices possible. Again six MERLIN telescopes were used. A typical $u-v$ coverage accomplished in those observations is shown on Fig. \[uvl\]. Here we present the observations of five sources (see Table \[table1\]) from the first group because the data for the sixth source (1236+327) were corrupt. Initial amplitude calibration was derived from daily observations of the unresolved source OQ208 giving a calibration error of $<5\%$ in flux density. The preliminary data reduction including phase-referencing was made using the AIPS-based PIPELINE procedure developed at JBO. The phase-calibrated images created with PIPELINE were refined in AIPS using several cycles of self-calibration and — in case of 1.6 MHz observations — amplitude self-calibration was applied at the end. The corrected data were mapped with IMAGR and the final maps are shown on Figs. \[0801+303\_maps\] to \[1233+418\_maps\]. The lowest contour represents roughly a $3\sigma$ level.
We measured the flux densities of the components labelled on those maps and arrayed the results in Table \[table2\]. For the 3 targets also observed by F2001 (, , ) we quoted their 4.86 GHz VLA measurements of respective components in the third column of that table. The differences in the 4.86 GHz flux densities measured with the VLA and MERLIN are attributed to the sparse $u-v$ coverage attained by MERLIN during the observations in snapshot mode. As a result, some flux pertinent to extended structures is missing on our 5 GHz maps. For calculation of the spectral indices (see Table \[table2\]) we used therefore VLA (A-conf.) 4.86 GHz fluxes taken from F2001 (when available) instead of those derived from our MERLIN maps.
-------------------- --------- ------------- ------------- -------
Source/ Sp.
component 1.6 GHz [5.0 GHz]{} [4.9 GHz]{} index
VLA
C 455.5 145.0
B1 118.3 40.9
B2 24.1
C 46.4 28.9 34.9 0.26
A 77.6 26.6 47.3 0.46
B 136.1 39.4 63.5 0.71
C 23.0 47.2
A1 19.3 5.1
A2 194.3 14.3
B1 224.7 48.4 104.6 0.71
B2 120.4 9.6 59.4 0.66
C 281.5 98.5 141.5 0.64
A 254.7 28.7 97.5 0.89
-------------------- --------- ------------- ------------- -------
: Flux densities of sources’ principal components
\[table2\]
has clearly unbeamed lobes and its redshift is known; therefore we estimated the angular sizes $\theta_{x}$, $\theta_{y}$ from the 1.6-GHz maps using the JMFIT program from the AIPS package and based on these values the physical parameters of the lobes were found using formul[æ]{} from @miley80. The results are shown in Table \[table3\]. The calculations of physical parameters were made for deceleration parameter $q_0=0.5$ which is used throughout this paper.
{width="8cm"} {width="8cm"}
{width="8cm"} {width="8cm"}
Notes on individual sources
===========================
[****]{}. This radio source is identified with a QSO of redshift $z=1.446$ [@hb89]. The 1.6-GHz observation reveals a triple structure. The central, brightest component C is a core and the two structures B1 and B2 straddling it are lobes. The B1 lobe is much brighter than B2 (Fig. \[0801+303\_maps\], left panel). The 5-GHz map shows a double structure of the source: the B1 lobe and more complex structure of the central component. The element C seen on the 1.6-GHz map is split here into two components: C1, which is probably the actual core, and C2, which is a part of the jet (Fig. \[0801+303\_maps\], right panel). The spectrum of the component C is the flattest in the part identified with C1 and becoming steeper towards C2. The spectrum of component B1 is steep and the component B2 has presumably even steeper spectrum; it does not appear on the 5-GHz map at all.
[****]{}. This source is an unconfirmed quasar (a “blue object”) of unmeasured redshift [@greg98; @vig99]. The map resulting from the 1.6-GHz observation shows a triple structure of the source. The brightest component C is a core, the A component is a part of the jet structure directed towards us and the B component is a lobe (Fig. \[0805+406\_maps\], left panel). A triple structure of the source appears also on the 5-GHz map. Components C and A are compact and have higher luminosity than more extended component B (Fig. \[0805+406\_maps\], right panel). The spectrum of component C is flat and its spectral index amounts to $\alpha=0.26$. Also the component A has a rather flat spectrum ($\alpha=0.46$) and only the component B is featured by a truly steep spectrum ($\alpha=0.71$). That means the source is clearly a core-jet-lobe structure. Another confirmation of this comes from F2001; their maps additionally show a diffuse western component which may be related to the (hidden) counter-jet.
[****]{}. There is no optical identification of this source [@mwhb01] so we do not know either its magnitude or redshift. The 1.6-GHz map shows the very complex structure of the source. The brightest component C is a core and structures A1 and A2 are fragments of a jet but the A1 component has lower radio emission than component A2. The brighter elements of A2 structure presumably represent the part of the jet seen at a smaller angle to the line of sight and so more Doppler-boosted (Fig. \[0850+331\_maps\], left panel). The 5-GHz map also shows a complex structure of the source. The component C seen on the 1.6-GHz map, here falls apart into small elements which, again, seem to be fragments of the wiggling jet, except component C1 which is probably the actual core. We can hardly see the component A1 and the structure A2 also consists of a few small elements (Fig. \[0850+331\_maps\], right panel).
------ -------------- -------------- ----------------------- ------------------------------------------------ ------------------ ---------- -----------
Lobe $\theta_{x}$ $\theta_{y}$ $L_{1.6\mathrm{GHz}}$ $L_{5\mathrm{GHz}}$ $B_{me}$ $u_{me}$ $u_{tot}$
[\[$10^{-3}$G\]]{} [\[$10^{-9}\mathrm{erg}~\mathrm{cm}^{-3}$\]]{} \[$10^{58}$erg\]
B1 449 318 5.70 2.66 0.143 1.89 0.216
B2 705 249 3.06 1.51 0.103 0.977 0.171
------ -------------- -------------- ----------------------- ------------------------------------------------ ------------------ ---------- -----------
\[table3\]
[****]{}. This source is identified with a radio galaxy of redshift $z=0.445$ [@wandaly96]. The 1.6-GHz map shows a symmetric, double structure of the source. The two extended components B1 and B2 are lobes but the lobe B1 has stronger radio emission (Fig. \[1201+394\_maps\], left panel). The 5-GHz map shows the same symmetric, double structure with two lobes B1 (also brighter) and B2 (Fig. \[1201+394\_maps\], right panel). The spectra of the lobes are steep and their spectral indices amount to $\alpha=0.71$ for B1 and $\alpha=0.66$ for B2. Physical parameters for both lobes are shown in Table \[table3\]. 1.6 GHz luminosities were calculated based on fluxes derived from our MERLIN maps; 5 GHz luminosities were calculated based on fluxes taken from F2001.
[****]{}. This source is identified with a radio galaxy redshifted to $z=0.25$ [@murgia99]. The 1.6-GHz map shows a core-jet structure; the source consists of two components: the core (component C) and an elongated jet structure (component A). Emission from the jet fragment lying closer to the core is high and it fades along the jet structure (Fig. \[1233+418\_maps\], left panel). The 5-GHz observation shows a triple structure of the source. The brightest component C is probably a core. The jet component A is split here into two elements A1 and A2 (Fig. \[1233+418\_maps\], right panel). All the components have steep or very steep spectra: the spectral index of the component C is $\alpha=0.64$ and the spectral index of the component A amounts to $\alpha=0.89$.
Discussion
==========
New MERLIN 1.6 and 5-GHz maps show many details of the structures of weak CSS sources. None of these have both jets visible — most of them consist of a core, one-sided jet and sometimes a lobe. Jets of and have quite complex structures. There are no hotspots in the lobes of these sources although the lobes in are edge-brightened as in FRIIs. For all principal components of the 3 sources belonging both to our and F2001 samples we calculated the spectral indices. Cores have been found for four sources: , , and . Three objects — two quasars and one radio galaxy — follow a division in radio morphology similar to that for LSOs and CSSs from the 3CRPW sample: galaxies are simple doubles whilst quasars show triple or complex structures. The fourth object, a radio galaxy , is an exception — it shows an asymmetric structure with a core and one-sided jet. (The remaining 5th object has no optical identification.) With an exception of all sources are moderately beamed.
The symmetric, double structure of and absence of a visible core indicate that this source lies almost in the sky plane so the calculated projected linear size — $l=10.1 h^{-1}$ kpc — is probably close to the physical size. is therefore MSO-type.
Because of the apparent lack of beaming in this source plus the fact we know its redshift and consequently the luminosity, we checked how this object would fit into the Fanaroff–Riley classification scheme [@fr74]. To this end we calculated the spectral index between 365 and 1400 MHz using the 365-MHz flux density from the Texas catalogue (1343 mJy). We found that $\alpha_{0.365\mathrm{GHz}}^{1.4\mathrm{GHz}}= 0.78$ is quite consistent with $\alpha_{1.4\mathrm{GHz}}^{4.85\mathrm{GHz}}=0.85$ so using both of them and assuming that the indices remain valid down to 178 MHz we estimated the 178-MHz flux to be 2359 mJy or 2723 mJy respectively. The mean value derived from the above figures yields $L_{178\mathrm{MHz}}=6.45\times10^{26}h^{-2}\mathrm{W~Hz^{-1}}$. According to Fanaroff & Riley, the boundary value of $L_{178\mathrm
{MHz}}\approx 3\times10^{25} h^{-2}\mathrm{W~Hz^{-1}}$ indicates the division of large-scale objects into two types: the FRII sources, which are beyond the luminosity boundary, and FRI sources which lie below. Our estimate is more than an order of magnitude higher than that dividing luminosity and indicates that indeed belongs to the FRII class.
@lo96 found that the radio luminosity of the FRI/FRII divide varied with optical luminosity. For a mixed sample of sources the dividing luminosity at 1.4 GHz is around $1.4\times10^{24} h^{-2}\mathrm{W~Hz^{-1}}$ for $M_R=-21.2$ which is the case of the galaxy identified with . The rest frame radio luminosity of at 1.4 GHz is $1.2\times10^{26} h^{-2}\mathrm{W~Hz^{-1}}$ and so the source is again expected to be an FRII.
@fanti95 and @r96 using self-similar models predict that the luminosity is expected to decrease rapidly with size as the source evolves. @ob97 also show that a strong decrease in luminosity with size is expected as CSS sources evolve into LSOs and so might become an FRI-type object in the future. On the other hand this conjecture seems to be unlikely taking into account the clear FRII-like morphology of (an edge brightened double lobed structure). Self-similar evolution would result in the source maintaining the same morphology as it evolved. It is only for a flat external density profile that a source might be expected to maintain or increase in luminosity as predicted for GPS sources [@snellen00] and so rather special conditions are required for the source to stay as an FRII. To add to the confusion, @zirb97 found that FRI sources tend to be in richer groups than FRIIs and so may be in a flatter external density profile. Clearly the evolution of these sources is uncertain and further work on the lower luminosity CSSs is required.
Conclusions
===========
MERLIN has been used to survey a sample of 60 weak Compact Steep Spectrum sources. This paper deals with five relatively large (arcsecond scale) sources. According to the evolutionary scheme compact doubles or, broadly speaking, CSOs are the progenitors of the extended doubles [@pm82; @c85] and the CSS sources form an evolutionary link between those most compact/youngest objects and the classical double FRIIs. All CSS sources known so far have high radio luminosities and the radio structures of those which are unbeamed have FRII structures. It seemed reasonable to suspect that the lower radio luminosity CSSs could be the progenitors of less luminous FRI objects. Our investigations of a new sample of weak CSS sources were motivated by the above-mentioned view. The triple or double structures of the five CSS sources and the presence of one-sided core–jet structures indicate they are more similar to FRII objects than FRIs. The radio structure of is also similar to the structure of FRII object because of edge-brightening of the lobes. The remaining four sources seem to be moderately beamed. Their structures consist of a core and one-sided jet. None of these sources have counter-jets. This means that the low luminosities of these sources are not a consequence of a lack of significant Doppler boosting. We claim, therefore, they constitute a new class of “weak” CSS sources.
MERLIN is a UK National Facility operated by the University of Manchester on behalf of PPARC.
This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
Part of this research was made when MK stayed at JBO and received a scholarship provided by the EU under the Marie Curie Training Site scheme.
Augusto, P., Browne, I. W. A. & Wilkinson, P. N. 1998, Ap&SS, 261, 261
van Breugel, W. J. M., Miley, G. K. & Heckman, T. A. 1984, AJ 89, 5
Carvalho, J. C. 1985, MNRAS, 215, 463
Dallacasa, D., Fanti, C., Fanti, R., Schilizzi, R. T. & Spencer, R. E. 1995, A&A, 295, 27
Douglas, J. N., Bash, F. N., Arakel Bozyan, F. & Torrence, G. W. 1996, AJ, 111, 1945
Fanaroff, B. L. & Riley, J. M. 1974, MNRAS, 167, 31P
Fanti, R., Fanti, C., Schilizzi, et al. 1990, A&A, 231, 333
Fanti, C., Fanti, R., Dallacasa, D., et al. 1995, A&A, 302, 317
Fanti, C., Pozzi, F., Fanti, R., et al. 2000, A&A, 358, 499
Fanti, C., Pozzi, F., Dallacasa, D., et al. 2001, A&A, 369, 380 (F2001)
Freedman, W. L., Madore, B. F., Gibson, B. K., et al. 2001, ApJ 553, 47
Gregorini, L., Vigotti, M., Mack, K.-H., Zonnchen, J. & Klein, U. 1998, A&AS, 133, 129
Hewitt, A. & Burbidge, G. 1989, [*magnetic tape*]{}
Kapahi, V. K. 1981, A&AS, 43, 381
Ledlow, M. J., & Owen, F. N. 1996, AJ, 112, 9
McMahon, R. G., White, R. L., Helfand, D. J. & Becker, R. H. 2001, ApJS (in press) \[astro-ph/0110437\] [http://sundog.stsci.edu/first/APM/]{}
Miley, G. 1980, ARA&A, 18, 165
Murgia, M., Fanti, C., Fanti, R., et al. 1999, A&A, 345, 769
Mutel, R. L., Phillips, R. B. 1988, in The impact of VLBI on Astrophysics and Geophysics, ed. M. J. Reid, J. M. Moran, IAU Symp. 129 (Kluwer), 73
O’Dea, C. P. & Baum, S. A. 1997, AJ, 113, 148
O’Dea, C. P. 1998, PASP, 110, 493
Owsianik, I., Conway, J. E. 1998, A&A, 337, 69
Owsianik, I., Conway, J. E. & Polatidis, A. G. 1998, A&A, 336, L37
Patnaik, A. R., Browne, I. W. A., Wilkinson, P. N. & Wrobel, J. M. 1992, MNRAS, 254, 655
Pauliny-Toth I. I. K., Witzel A., Preuss E., et al. 1978, AJ, 83, 451
Peacock, J. A. & Wall, J. V. 1982, MNRAS, 198, 843
Phillips, R. B. & Mutel, R. L. 1982, A&A, 106, 21
Readhead, A. C. S., Taylor, G. B., Xu, W., et al. 1996, ApJ, 460, 612
Saikia, D. J., Jeyakumar, S., Salter, C. J., et al. 2001, MNRAS, 321, 37
Sanghera, H. S., Saikia, D. J., Lüdke, E., et al. 1995, A&A, 295, 629
Snellen, I. A. G., Schilizzi, R. T., Miley, G. K., et al. 2000, MNRAS, 319, 445
Spencer, R. E., McDowell, J. C., Charlesworth, M., et al. 1989, MNRAS, 240, 657
Vigotti, M., Grueff, G. & Perley, R. 1989, AJ, 98, 419
Vigotti, M., Gregorini, L., Klein, U. & Mack, K.-H. 1999, A&AS, 139, 359
White, R. L., Becker, R. H., Helfand, D. J. & Gregg, M. D. 1997, ApJ, 475, 479
Wan, L. & Daly, R. A. 1996, ApJ, 467, 145
White, R. L. & Becker, R. H. 1992, ApJS, 79, 331
Wilkinson, P. N., Spencer, R. E., Readhead, A. C. S., Pearson, T. J. & Simon, R. S. 1984, in VLBI and Compact Radio Sources, ed. R. Fanti, K. Kellerman, G. Setti, IAU Symp. 110 (Kluwer), 25
Zirbel, E. L. 1997, ApJ, 476, 489
[^1]: In principle @pbww92 sought candidates for pointlike phase calibrators among flat-spectrum objects in Green Bank surveys but the procedure they used was different from that used by e.g. @wb92. When we supplemented their list with flux densities from @wb92 it turned out that some of their candidates actually had steep-spectra. Consequently the VLA observations revealed resolved structure and so these objects could not be used as calibrators. In this way @pbww92 made a serendipitous discovery of several weak CSSs.
[^2]: A review of both GPS and CSS classes has been given by @odea98.
[^3]: Three of them also belong to F2001 sample.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The simultaneous photometric and spectroscopic observations of the RR Lyrae variables in the globular cluster, M3, published in @data [Paper I] made it possible to perform Baade-Wesselink (BW) analysis of a large sample of Blazhko stars for the first time. The BW distances of Blazhko stars turned out to be unreliable, as significantly different distances were obtained for the stars of the Blazhko sample and also for the same star in different modulation phases. Even the results of small modulation-amplitude Blazhko stars may be doubtful. This result warns that the application of the BW method to Blazhko stars is not trustworthy.
Keeping the distance fixed for each Blazhko star in each modulation phase, a significant difference between the spectroscopic and the photometric radius ($R_{\textrm{sp}}$, $R_{\textrm{ph}}$) variations is detected. The phase and amplitude variations of $R_{\textrm{sp}}$ follow the changes of the light curve during the Blazhko cycle but the $R_{\textrm{ph}}$ curve seems to be not (or only marginally) affected by the modulation. The asynchronous behaviour of $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ supports the interpretation of the Blazhko effect as a depth-dependent phenomenon, as the spectroscopic radius variation reflects the radial displacement of the line-forming region high in the atmosphere, while the photospheric radius variation is derived from the information of the observed visual-band light emitted mostly by the lower photosphere. The stability of $R_{\textrm{ph}}$ may be interpreted as a strong argument against the non-radial-mode explanation of the Blazhko phenomenon.
author:
- |
J. Jurcsik$^{1}$[^1], G. Hajdu$^{2,3}$\
$^{1}$Konkoly Observatory of the Hungarian Academy of Sciences, H–1525 Budapest PO Box 67, Hungary\
$^{2}$Instituto de Astrofísica, Pontificia Universidad Católica de Chile, Av. Vicuña Mackenna 4860, 782-0436 Macul, Santiago, Chile\
$^{3}$Instituto Milenio de Astrofísica, Santiago, Chile\
date: 'Accepted 2017 ..... Received 2017 ..., in original form '
title: 'The first Baade-Wesselink analysis of Blazhko RR Lyrae stars: discrepancies between photometrically and spectroscopically determined radius variations'
---
\[firstpage\]
stars: horizontal branch – stars: oscillations (including pulsations) – stars: variables: RR Lyrae – Galaxy: globular clusters: individual: M3 – techniques: photometric – techniques: radial velocities
Introduction
============
The main goal of the Baade-Wesselink (BW) method is to determine the distance and the radius of radial pulsators using photometric and radial-velocity (RV) data. The basic concept is to measure the line-of-sight radial displacement of the stellar surface via spectroscopic tools, and the simultaneous angular-diameter variation of the stellar disk using photometric (recently also interferometric) methods. The distance and the mean radius value of the star are then determined by fitting the amplitudes of the two radii variations via appropriate scaling. The BW method, and its variants were successfully applied to several radial-mode pulsators, from the pre-/post- and main sequence $\delta$Sct stars (e.g. @bm, @hae) to the horizontal-branch giant RR Lyrae (RRL) variables (without completeness, e.g. @jcsl [@f94; @ccf; @lj90] and @sk93) and to the super-giant Cepheids (as a summary of the BW distances of Cepheids read e.g. @dib and @gron).
The BW analysis relies on some basic assumptions, which may hold some uncertainties. The exact value of the projection factor, which converts radial velocity to pulsation-velocity variation is still under active investigation [@ker; @nar], and any inaccuracy of the colour-temperature and bolometric-correction transformations – based on synthetic atmosphere-model results – can modify the BW results. As the radius variations measured spectroscopically and photometrically reflects the radial motions at different depths, and differences between the motions of these layers may occur in dynamic atmospheres, depth-dependent phenomena (shocks, velocity gradients, etc.) may have a strong bias on the BW distances. The applicability of the method also requires that the pulsation has to be fully radial.
Though some of the conditions of the applicability of the BW method listed above are not fully satisfied during the pulsation of RRL stars (e.g. the shocks propagating in the atmosphere are strongly depth dependent), and any inaccuracy of the $p$-factor and the synthetic model-atmosphere data biases the BW results, the error of the BW distances of RRL stars has been estimated to be $3-10$ per cent for galactic field and globular-cluster RRL stars in previous studies [e.g. @jcl; @ccp; @storm2].
The light curves of about half of the fundamental-mode RRL stars are not stable [@kbs; @k10]; amplitude and phase variations, i.e. the Blazhko effect is detected. Recently, Blazhko modulation of the light curves of Cepheids in the Galactic field [@blcep] and the Magellanic Clouds [@so15; @s17] has also been discovered. Moreover, [@ander14; @ander16] showed that the small-amplitude modulations of the RV curves observed in some Cepheids may affect their BW distances, too.
The applicability of the BW method on Blazhko RRL stars may be questioned until the physical mechanism inducing the phenomenon is explained. Besides this, the requirement of simultaneous spectroscopic and photometric observations is the other reason why no BW analysis of any Blazhko star has been published yet. Although extensive spectroscopic observations of some Blazhko RRL stars have been published [@cg96; @cp13], they lack simultaneous multi-band photometry needed for the application of the BW method.
To perform the first BW analysis of a large set of RRL star in a globular cluster and of Blazhko stars in different phases of the modulation, simultaneous photometric and spectroscopic observations of the variables of M3 were secured in 2012. The results of the photometric observations of overtone/double-mode variables were published in [@overtone]. The photometric data of RRab stars, together with the RV measurements and the results of the BW analysis of single-mode variables were presented in Paper I.
BW distances of 26 stable-light-curve RRL stars of M3 were determined in Paper I. 10.5 kpc mean distance of the cluster with 0.2 kpc formal error of the mean value was derived. The random, uncorrelated errors connected to the uncertainties of the magnitudes of the individual stars and the fitting process of the $R_{\textrm{sp}}$ and the $R_{\textrm{ph}}$ curves are significantly reduced if the sample comprises stars of the same distance, as for globular clusters. Indeed, the 0.2 kpc error does not exceed the real dispersion of the distances of the stars, as the tidal radius of M3 is about 0.1 kpc. As the global properties, which influence the systematic errors of the BW distances (reddening, \[Fe/H\], zero point of the photometry) are thought to be known with relatively high precision for M3, the major sources of the systematic biases are connected to the uncertainties of the $p$-factor and the synthetic model-atmosphere data. [@jcl] estimated the effect of these factors to be 0.13 mag on the distance modulus, i.e. 6 per cent of the distance. Using the more recent model-atmosphere data [@kurucz] and an updated value for the $p$-factor [@n04], the systematic errors of the derived distances might be somewhat smaller, however, only the Gaia parallaxes will determine the accuracy of the BW distances correctly. The data published in Paper I provide the first observations suitable to perform direct BW analysis of Blazhko stars in different phases of the modulation. The obtained results are very intriguing, and rise suspicion concerning the applicability of the BW analysis for Blazhko stars. The aim of this paper is to document these issues, and not to determine the actual distance and radius values of the studied stars.
Data, method and BW distances of Blazhko stars {#2}
==============================================
![The quasi-simultaneity of the photometric and spectroscopic observations is documented on the example of V035, a large modulation-amplitude Blazhko star with a 57-d modulation period. Nine-days-long segments of the $V$ light curve and the RV data are plotted; the single-mode Fourier solutions of the data, derived in two different ways for the light curve, are drawn by dotted/dashed lines. The fits shown are used as the simultaneous $V$ and RV data in the analysis. \[v35\]](v35.eps){width="9.6cm"}
{width="18.1cm"}
The photometric and RV data of RRL stars in M3 were published in [@overtone] and in Paper I. The full description of the application of the BW method was also given in Paper I.
Exclusively, the RV data derived from the Hectoechelle@MMT observations are used in the analysis of Blazhko stars, as they are condensed enough to provide accurate RV curves for separate Blazhko phases. The two runs of the Hectoechelle@MMT observations cover the pulsation phases more or less completely at two epochs separated by $\sim30$ days for 14 stars showing light-curve modulation. The data are also suitable for the analysis of eight more Blazhko stars in one Blazhko phase. Most of these Blazhko stars are fundamental-mode RRab type variables, but RV data have been obtained for one of the overtone-mode RRc stars (V140) exhibiting Blazhko modulation, too. Altogether, we can study the results of the BW method in 36 discrete phases of the modulation using the data of 22 Blazhko stars.
The procedure of determining the photometric data parallel with the RV observations for Blazhko stars is twofold: single-mode Fourier fits of $5-10$ d segments in the vicinity of the spectroscopic observations of either the direct photometric observations, or of the synthetic data generated according to the full light-curve solution are used. The differences between these synthetic data are, however, marginal, and the choice of using any of these solutions has no significant effect on the results. The photometric and spectroscopic data-coverage in the vicinity of the two Hectoechelle@MMT runs is shown in Fig. \[v35\], for a Blazhko star (V035) with a 57-d modulation period and with a strongly variable light curve, as an example. The dotted/dashed lines indicate the synthetic data used in the analysis. Two solutions for the synthetic $V$ light curve, derived as described above, are displayed.
The complete BW analysis carried out for Blazhko stars has led to a contradictory result. The derived distances cover an unrealistically large, $7.7-14.5$ kpc range, which is significantly larger than the range of the BW distances obtained for stable-light-curve RRL stars. As a result of the BW analysis of 26 stable-light-curve RRL stars, 10.5 kpc mean value for the distance of M3 was derived in Paper I, with the individual distances falling in a narrower, $9.7 -11.4$ kpc distance range. Although both the light and RV curves of Blazhko stars utilised in the analysis are less accurate than the mean light and RV curves of stable-light-curve variables, especially for variables with strongly and rapidly varying light curves, it cannot account for the large spread of the BW distances derived for Blazhko stars.
The random errors of the individual distances of stable-light-curve RRL stars were dominated by the uncertainties arising from the selection of the phase interval omitted from the fitting process of the $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ curves. Depending on the choice of the part of the $R$ curves to match, up to 0.5 kpc differences were obtained. This is the case for Blazhko stars, too, but the effect is as large as $1.0-1.5$ kpc for some stars. The error connected to the incomplete phase coverage of the RV curve and the uncertainity of the light cuvre at the dates of the RV observations is tested using different possible solutions for the actual RV and light curves. The distances derived this way remain within $0.5-1.0$ kpc range, typically. The effect of the random errors of the photometric zero point of the individual stars on the distance estimates is supposed to be less than 0.3 kpc for both Blazhko and non-Blazhko RRL stars, as the accuracy of the zero points of the ISIS flux photometry depends on the accuracy of the magnitudes of the variables determined on the reference frames, which is $0.01-0.05$ mag depending on the crowding. These errors, added in quadrature, give an estimate of $1.1-1.8$ kpc for the accuracy of the individual distances of Blazhko stars. Consequently, the $\sim7$ kpc full range of the distance values obtained for Blazhko stars are as large as $4-6\,\sigma$ of the estimated random errors.
Not only the scatter of the individual distances derived for Blazhko stars are larger than the errors would indicate, but as large as $3-5$ kpc ($2-4\,\sigma$) differences between the two distances obtained for Blazhko stars with RV data suitable for the analysis at two epochs are obtained. Such a large discrepancy between the derived distances of the same star using different sets of the observations cannot have any physically plausible interpretation, if the basic assumptions of the BW analysis are valid in both modulation phases. Consequently, we have to conclude that the BW method fails to give reliable distances for Blazhko stars.
![Results of the analysis of four large-modulation Blazhko RRab stars in two different phases of the modulation are plotted. The full $V$ light curves, the light curves matching the two RV observation runs are highlighted, the RV observations and their fits, and the derived $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ variations are shown from top to bottom. The distance is fixed to $d=10.5$ kpc. The continuous and dotted lines denote the radius variations in the large- and the small-amplitude phases of the modulation. Some of the RV curves are scanty; different solutions to fill these gaps are shown for both the small- and the large-amplitude phases of V035, and for the small-amplitude phase of V045 and for the large-amplitude phase of V130. Two different representations of the photometric data (as described in Sect. \[2\]) are considered for the small-amplitude phase of V045 and for both the small- and the large-amplitude phases of V130, in order to indicate the effect on the resultant $R_{\textrm{ph}}$ curves. []{data-label="bw4"}](bw4.eps){width="8.8cm"}
![Phase differences and amplitude ratios between different combinations of the $V$, $V_{\textrm{mLC}}$, $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ variations of Blazhko stars versus the phase difference (left-hand panel) and amplitude ratio (right-hand panel) of $V$ and $V_{\textrm{mLC}}$. $V$ and $V_{\textrm{mLC}}$ denote the actual and the mean $V$ light curve, respectively. First-order Fourier amplitudes and phases are utilised. The first three plots from top (the relations between $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$, between $R_{\textrm{sp}}$ and $V_{\textrm{mLC}}$ and between $V$ and $R_{\textrm{ph}}$) follow the amplitude and phase relations between $V$ and $V_{\textrm{mLC}}$. The slope of the best-fit lines to the data (continuous lines) equals with 1.0 (dashed lines) within the uncertainty limits in these plots. The two bottom plots document the relations between $V_{\textrm{mLC}}$ and $R_{\textrm{ph}}$ and between $V$ and $R_{\textrm{sp}}$. These parameters are independent from the phase and amplitude relations between $V$ and $V_{\textrm{mLC}}$, as the steepness of the linear fits to these data are close to zero. The exact, one-to-one (three plots from the top) and zero-steepness relations (two plots from the bottom) are shown by dotted lines for comparison. The black circles denote the 0:0 and 1:1 positions in the topmost relations of the left- and right-hand panels, respectively. Different solutions for the RV fits of scanty data, and for the light curves parallel with the RV observations are used in the analysis for some of the stars. These multiple results are connected by lines. The differences between these data hardly exceed the symbols’ size for most of the cases. The filled blue triangles denote the relations between $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ of V140, the only RRc star in the sample. The other phase and amplitude relations of V140 are not plotted, because they are significantly different for overtone- and fundamental-mode variables.[]{data-label="bwbl"}](bwbl.ereda.eps){width="8.7"}
Discrepancies between $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ in Blazhko stars {#radii}
==============================================================================
In order to find the reason of the failure of the BW method for Blazhko stars, we have reanalysed the data using a fixed, $d=10.5$ kpc distance, as determined for stable-light-curve variables of M3 (Paper I), for each modulated star in each modulation phase, with RV observations enough for the analysis.
The results are shown in Fig. \[bw3x\] for some Blazhko stars, whose RV observations cover the maximum or the minimum phase of the modulation. The $V$ light curves, with the synthetic fit matching the dates of the RV observations set out, the RV data and the Fourier fit, and the radius variations determined from the spectroscopic ($R_{\textrm{sp}}$: dotted lines) and the photometric ($R_{\textrm{ph}}$: continuous lines) data as described in Paper I, phased with the pulsation period are shown for 14 Blazhko stars according to the fixed-distance BW solution. The top- and bottom-line panels correspond to Blazhko-maximum and Blazhko-minimum data, respectively.
What stands out most, is that the amplitude of the $R_{\textrm{ph}}$ variation ($A_{R_{\textrm{ph}}}$) seems to be smaller than the amplitude of the $R_{\textrm{sp}}$ variation ($A_{R_{\textrm{sp}}}$) at the large-amplitude phase of the modulation and, on the contrary, $A_{R_{\textrm{ph}}}$ is larger than $A_{R_{\textrm{sp}}}$ at around Blazhko minimum. Although the minimum of the $R_{\textrm{ph}}$ variation is less deep than the minimum of the ${R_{\textrm{sp}}}$ curve for many stable RRL stars (this is why this part is usually omitted from the BW analysis of large-amplitude variables exhibiting strongly violent atmospheric dynamics in these pulsation phases), the opposite sign difference that is detected in the small-amplitude phase of the modulation, is not observed in any of the stable RRab stars as documented in the right-hand panel of figure 8. in Paper I.
Phase shift between the $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ curves is also evident in some Blazhko stars (V034, V045, V066, V117 and V140); besides the amplitude modulation of the light curve the phase modulation also has a large amplitude in these stars.
The behaviour of the $R_{\textrm{sp}}$ and the $R_{\textrm{ph}}$ curves shown in Fig. \[bw3x\] suggests that, while $R_{\textrm{sp}}$ follows the amplitude and phase changes observed in the light and RV variations, the $R_{\textrm{ph}}$ curve remains relatively stable in both amplitude and phase.
This possibility is checked first by using Blazhko stars that have suitable data for the analysis at two significantly different phases of the modulation. Fig. \[bw4\] illustrates the results for four large modulation-amplitude Blazhko stars in different phases of the modulation. Although the RV curve is somewhat ambiguous because of the scanty data in one or both Blazhko phases for these stars, it has only a marginal effect on the results, as documented in the figure. Different solutions to fill in the gaps of the RV curves are shown for V034, V045 and V130, but the differences between the resultant $R_{\textrm{sp}}$ curves are negligible. The uncertainties of the $R_{\textrm{ph}}$ curves arise from the uncertainty of the synthetic light and colour curves simultaneous with the RV data. As described in Sect. \[2\] it was derived either according to the full-light-curve solution, or from the photometric observations obtained close in time to the RV measurements. $R_{\textrm{ph}}$ curves derived using both of these methods are shown for V045 and V130 in Fig. \[bw4\]. Again, the difference between these results is much smaller than the phase and/or amplitude difference between the simultaneous $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ curves. Consequently, the uncertainties involved in the method cannot account for the discrepant behaviour of the $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ curves.
Each of the four stars shown in Fig. \[bw4\] exhibits large-amplitude phase and amplitude modulations. As supposed to be normal, both the RV and the $R_{\textrm{sp}}$ variations follow the amplitude- and the phase-changes of the light-curve. In contrast, the $R_{\textrm{ph}}$ curves seem, indeed, to remain unaffected by the modulation in all of the four stars shown as examples. Therefore, these results support the assumption that the shape of the $R_{\textrm{ph}}$ curve is not affected by the modulation.
Secondly, the amplitude and phase stability of the $R_{\textrm{ph}}$ curves of Blazhko stars has also been checked using the data of all the Blazhko stars observed spectroscopically. Phase differences and amplitude ratios of the different combinations of the actual ($V$) and the mean light curves ($V_{\textrm{mLC}}$) and the $R_{\textrm{ph}}$ and $R_{\textrm{sp}}$ radius variations are examined. The amplitudes and phases of the first order of appropriate-order Fourier-series solutions are considered for all the parameters investigated. These are less sensitive to any defect of the $R_{\textrm{ph}}$ appearing in the $0.8-1.0$ pulsation-phase interval in the large-amplitude phase of the modulation than the values of the higher-order Fourier components or the total amplitude of the variations and the phase of the absolute minimum values. Moreover, the first-order components can be determined accurately even for the very sinusoidal shape $R$ curves, which higher-order Fourier components are uncertain. The amplitude and phase of the $V_{\textrm{mLC}}$ are derived from the fit of the full $V$ light curve.
The phase differences and amplitude ratios between the parameter pairs, ($R_{\textrm{sp}};R_{\textrm{ph}}$), ($R_{\textrm{sp}};V_{\textrm{mLC}}$), ($V;R_{\textrm{ph}}$), ($V_{\textrm{mLC}};R_{\textrm{ph}}$) and ($V;R_{\textrm{sp}}$), versus the phase and amplitude relations of the actual and the mean light curves for 36 different Blazhko phases of 22 Blazhko stars are shown in Fig. \[bwbl\]. The linear fits to the data are drawn in the plots by continuous lines. For comparison, dotted lines represent the corresponding linear fits with 1.0 and 0.0 steepness.
If the hypothesis that the changes of $R_{\textrm{sp}}$ follow the modulation of the light curve but the Blazhko effect does not influence the variations of $R_{\textrm{ph}}$ is true, then the phase differences and amplitude ratios between $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ should have to vary according to the phase differences and amplitude ratios between the actual and the mean light curves. Fig. \[bwbl\] documents that this is indeed what happens. The steepness of the fitted lines to the $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ phase differences and amplitude ratios versus $\Phi_V - \Phi_{V_{\textrm{mLC}}}$ and $A_V / A_{V_{\textrm{mLC}}}$ (the top row in the figure) equals to 1.0 within the limits of the uncertainty, and the same is true for the relations between $R_{\textrm{sp}}$ and $V_{\textrm{mLC}}$ and between the actual light curve and $R_{\textrm{ph}}$, as well.
The two bottom plots in the panels of Fig. \[bwbl\] document the same result from the opposite aspect; if the amplitudes and phases of $R_{\textrm{ph}}$ and $R_{\textrm{sp}}$ are locked to the amplitudes and phases of the mean and the actual light curves, respectively, then the relations between $R_{\textrm{ph}}$ and the mean light curve and between $R_{\textrm{sp}}$ and the actual light curve has to remain constant, i.e. they are independent from the amplitude and phase relations between the actual and the mean light curves. As the steepness of the fitted lines equals to zero within the error ranges, these relations also verify our statement.
The 0:0 and 1:1 coordinates, marked by black circles in the first row of the left- and right-hand panels of Fig. \[bwbl\], correspond to the positions where the phases and the amplitudes of both the actual and the mean light curves, and the $R_{\textrm{ph}}$ and $R_{\textrm{sp}}$ curves are equal. These positions are very close to the fitted lines; according to the best linear fits, the phase difference and amplitude ratio between the actual and the mean light curves are $-0.05$ rad and 0.94 where the amplitudes and the phases of the $R_{\textrm{ph}}$ and $R_{\textrm{sp}}$ curves are the same. Correspondingly, the phases of $R_{\textrm{ph}}$ and $R_{\textrm{sp}}$ coincide the best at the Blazhko phase when the first-order Fourier phase of the actual light curve is the same or it is slightly smaller than the phase of the mean light curve, i.e. the actual light curve is at the same phase or is shifted to slightly larger phase values than the mean light curve on a phased light-curve plot. Similarly, the amplitudes of the $R_{\textrm{ph}}$ and $R_{\textrm{sp}}$ radius curves are matching the best when the first-order Fourier amplitude of the actual light curve is close to or is a bit smaller than the amplitude of the mean light curve.
However, this result does not necessarily imply that the pulsation of Blazhko stars would be normal when the actual light curve is close in amplitude and phase to the mean light-curve, as the amplitudes of the mean light curves of most of the Blazhko stars are significantly smaller than the amplitudes of similar-period non-Blazhko stars (see figure 1. in Paper I).
Although there is no doubt that the amplitude ratios and phase differences shown is Fig. \[bwbl\] confirm that the variations of $R_{\textrm{sp}}$ reflects the variations of the light curve, while the amplitude and phase of $R_{\textrm{ph}}$ remain relatively stable during the Blazhko cycle, the scatter of the plots shown in Fig. \[bwbl\] is notable; it is substantially larger than the uncertainties would indicate. Therefore, the outlined regularity has to be taken as a tendency rather than a rule.
BW results for small modulation-amplitude Blazhko stars {#small}
=======================================================
{width="17.9cm"}
[ll@l@rrr]{} Var.&P$_{\mathrm{puls}}$ \[d\]& P$_{\mathrm{Bl}}$ \[d\]&\
& & & & &\
V048 OoI? & 0.62783& 150 & 8.4 0.2 & & 9.5 0.2\
V062 OoI & 0.65240& 280 & & & 10.2 0.4\
V071 OoI & 0.54905& 46.7& 10.6 1.7 & 13.1 1.0 & 11.3 1.4\
V073 OoI & 0.67350& 52.5& 13.7 0.3 & 8.6 0.3 & 12.4 0.2\
V104 OoII & 0.56993& 108 & 11.0 0.4 & 9.5 0.4 & 10.8 0.4\
V114 OoI & 0.59773& 54.4& & & 10.6 0.6\
\
\
![Results of the analysis of three small modulation-amplitude Blazhko stars in two different phases of the modulation; $a$ denotes the first, $b$ the second run of the observations. The distance is fixed to 10.5 kpc. Note that the amplitude of $R_{\mathrm {ph}}$ is smaller than the amplitude of $R_{\mathrm {sp}}$ in the largest amplitude phase of the modulation (V071b and V073a), and $R_{\mathrm {ph}}$ has somewhat larger amplitude than $R_{\mathrm {sp}}$ in the smallest amplitude phase (V073b and V104b) even in these marginally modulated stars, similarly to the detected behaviour of large modulation-amplitude stars.[]{data-label="bwsa"}](bwsa.eps){width="9.cm"}
We have investigated the small modulation-amplitude variables separately, in order to decide how reliable the BW distances are if the modulation affects the light-curve shape only marginally.
RV data suitable for the BW analysis of six small modulation-amplitude Blazhko stars were published in Paper I. The same results as shown for the Blazhko maximum and minimum phases of large modulation-amplitudes RRL stars in Fig. \[bw4\], are shown for the six small modulation-amplitude Blazhko stars in Fig. \[bw6\]. The distance is fixed to 10.5 kpc here, too. The BW results shown in Fig. \[bw6\] were obtained by utilising the combined set of the RV data of the two observation runs, and the corresponding synthetic light curves were determined form the segments of the photometric time-series close in time to the spectroscopic measurements. Although the matching of the $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ curves seems to be satisfactory for most of these stars, some discrepancy between the amplitudes seems to be present, e.g. for V071 and V073.
The results of the BW analysis of the small modulation-amplitude stars, with the distance as a free parameter, are summarised in Table \[satab\]. The pulsation and Blazhko periods and the BW distances derived from the data of the first and the second run of the spectroscopic observations (if any or both can be determined separately), and from the complete RV data set are given in the columns. As the photometric zero-point errors of the stars are the same in each phases of the modulation, the errors of the distances are estimated by fitting different parts of the $R$ curves. Table \[satab\] lists the mean values of the distances and their rms values derived for the data sets from 16 different solutions omitting $0.1-0.4$ phase intervals at around the minimum-to-maximum phases from the fitting process.
Comparing the results given in Table \[satab\] with the $10.5\pm0.2$ kpc mean value of the distances derived for stable RRL stars, we find that about half of the BW distances of the small modulation-amplitude Blazhko stars determined for a given phase of the modulation differ by $2-10\,\sigma$, $2-3$ kpc, from the mean, and for V048 and V073 the distances derived from the mean light curves and the complete RV data differ by $5-10\,\sigma$ from $10.5$ kpc, as well.
The mean value of the distances derived using all the RV data of the small modulation-amplitude stars is 10.8 kpc with 0.4 kpc formal error. Omitting V048, the mean is $11.1\pm 0.4$ kpc (see the remark on V048 at the end of this section). This is $0.6$ kpc larger than the mean value of the distances derived for stable RRL stars, and its error is also larger than obtained for the non-Blazhko stars. However, the $1.5\,\sigma$ difference between the mean values of the small-modulation-amplitude Blazhko and the non-Blazhko RRL stars is not statistically significant.
The analysis of the small modulation-amplitude Blazhko stars with good phase-coverage RV curves obtained for two epochs indicate the same systematic bias of the results as obtained for the large modulation-amplitude stars (see Fig. \[bwsa\]). Namely, the amplitude of $R_{\textrm{sp}}$ is smaller than the amplitude of $R_{\textrm{ph}}$ at the small-amplitude phase of the modulation (V073b and V104b), and the opposite is true at the large-amplitude phase (V071b and V073a).
Consequently, the BW analysis may result in erroneous distance estimate even for very small modulation-amplitude Blazhko stars both if data of a given Blazhko phase and if the mean RV and light-curve are considered in the analysis.
[*Remark on V048:*]{} The photometric data of V048 is contradictory, according to the mean $BVI$ magnitudes the star is a luminous, OoII variable, however, its amplitude is hardly larger than the amplitudes of similar-period OoI RRab stars in M3, even at the largest-amplitude phase of the modulation. The amplitudes and mean magnitudes of the star according to previous photometric studies [@ca05; @be06] indicated similar ambiguities. One possibility to resolve this inconsistency is to assume that the star is an OoI variable but, to explain the 0.1 mag too bright magnitudes, it is supposed to be closer by $\sim0.5$ kpc to us than the cluster itself. However, the $-141.4$kms$^{-1}$ mean RV of V048 is only 5.5kms$^{-1}$ larger than the cluster’s average RV, i.e. the difference is significantly smaller than the escape velocity. Therefore, it is unlikely that this star would indeed be an ‘escaper’. Most probably, an unresolved companion contaminates the photometry, making the star brighter and of smaller amplitude than normal. If this is the case, the BW analysis of V048 is unreliable independently from the Blazhko effect of the star.
Discussion and conclusions
==========================
We have shown in Sects. \[radii\] and \[small\] that the discrepant behaviour of the $R_{\textrm{ph}}$ and the $R_{\textrm{sp}}$ radius variations leads to the failure of the application of the BW method for Blazhko stars. As the BW method determines the distance via appropriate fitting of the amplitude of the $R_{\textrm{ph}}$ radius curve (angular radius variation) to the amplitude of the $R_{\textrm{sp}}$ curve (line-of-sight radius variation), the divergent behaviour of their amplitudes and phases is the reason of the inadequacy of the BW method for Blazhko stars. Applying the BW method, the distance will be overestimated if $R_{\textrm{ph}}$ has a smaller amplitude than $R_{\textrm{sp}}$ (around the large-amplitude phase of the modulation) and it will be underestimated if $R_{\textrm{ph}}$ has a larger amplitude than $R_{\textrm{sp}}$ (around the small-amplitude phase of the modulation).
What is the origin of the discrepant behaviour of the $R_{\textrm{sp}}$ and the $R_{\textrm{ph}}$ curves of Blazhko stars? A natural explanation would be that the pulsation of Blazhko stars is not fully radial; non-radial components also act as proposed e.g. in [@dm]. However, the relative stability of the amplitudes and phases of the $R_{\textrm{ph}}$ curves compared to the variations of $R_{\textrm{sp}}$ should imply that the orientation of the non-radial oscillation is the same for each star and at each Blazhko phase, that is obviously far from being realistic.
Another possibility is that we see a depth dependent phenomenon. @sch was the first who proposed to explain the observed line doubling and H$\alpha$ emission of pulsating variables by atmospheric shocks. Following the pioneering hydrodynamical study [@h72] of modelling the shock propagation, @fo92 showed that two shocks are successively generated during one period in RR Lyrae, and that the main shock reaches its maximum amplitude very high in the atmosphere. Based on observations and model predictions, more details on the complexity of the dynamics of the atmosphere of RR Lyrae stars, and on the role of the different shocks in explaining the Blazhko effect were discussed already in several papers [@mat; @cg98; @cvg08; @gi13; @cp13; @gf14].
The detected strong correlation between the strength of the amplitude and phase modulations and the amplitude and phase differences between the radial motions of the different depth layers reflected by the changes of $R_{\textrm{sp}}$ and $R_{\textrm{ph}}$ as shown in Fig. \[bwbl\], indicates also that the atmospheric shocks responsible for the desynchronisation of the different depth layers have indeed an important role in the explanation of the Blazhko phenomenon.
Looking at the details, [@c14] postulated that the Blazhko effect is generated by the interaction of a multi-shock structure with an out-flowing wind in the corona. The detection of HeII emission at the large-amplitude phase of the modulation [@p11; @gfl13] showed that the dynamics of the higher atmosphere of Blazhko stars can be even more violent than the dynamics of stable RRab stars. @gi13 proposed that the intensity of the main shock, producing the bump preceding light minima, varies during the Blazhko cycle. According to his description, after reaching a critical strength of the main shock, the photospheric motions become desynchronised at around Blazhko maximum. A variant of this idea was suggested in @cp13. Although the phenomena delineated in these papers differ in details, both description predict that the atmospheric motions become synchronised at Blazhko minimum and are desynchronised at Blazhko maximum in order to explain the observed Van Hoof effect and the doubling/broadening of metallic lines at Blazhko maximum [@cp13].
The observed variations of the $R_{\textrm{sp}}$ and the $R_{\textrm{ph}}$ curves also indicate that the motions of the atmosphere are strongly desynchronised. However, it seems that the radial displacement of line-forming regions of the upper part of the atmosphere is synchronised with the radial displacement of the lower, photospheric layers, the measured light and colour information originate from, during the modulation phase when the amplitude and the phase of the light variation is close to (or is the same as) the amplitude and phase of the mean light curve. Significant desynchronisation between the different atmospheric layers occur both at the large- and the small-amplitude phases of the modulation.
Despite of the large variety of the amplitude and phase modulations of the visual-band light curves of Blazhko stars, there is an indication that in the $K$ band, only marginal if any modulation is detected [@n15]. Hence, the pulsation seems to be stable at the depth of the $K$ band radiation.
Summarising all these results, we have information on the modulation of the pulsation from three different depths. First, the RV of the metallic lines reflects the motion of the topmost, line-forming region. Then, we can analyse the variations of the integrated visual-band light, radiated by the photosphere, and finally the K band variations hold information on the deepest regions of the photosphere, as the K-ban radiation comes from a larger depth than the visual-band light because of the near-IR dip in the opacities.
The RV variation of the metallic lines shows that the Blazhko effect manifests in this region as strongly varying radius changes of these layers during the modulation cycle. The observed modulation of the visual-band light and colour curves corresponds to changes in the pulsation temperature and luminosity variations of the photosphere during the Blazhko period. At the same time, only marginal if any modulation of the $R_{\textrm{ph}}$ curves – derived from the same visual-band photometric information – is detected. It means that at the depth of the optical-band radiation no or marginal modulation characterises the pulsation radius variation. Finally, at an even larger depth, the modulation seems to completely disappear, as the K-band data tends to indicate.
What is the most surprising, is that not only the amplitude modulation disappears at a larger depth, but the phase modulation, too. Although the phase shift between the radial-velocity curves of V130 is as large as 1.3 radian ($0.2P_{\mathrm{puls}}$), no phase shift between the $R_{\textrm{ph}}$ curves is evident as shown in Fig. \[bw4\]. The phase modulation of the Blazhko phenomenon was interpreted as changes in the pulsation period e.g. by [@st; @mw] and [@c14]. However, the recent result seems to contradict this possibility, as at the engine of the pulsation, at the partial ionisation zones, the period looks to be stable as indicated by the stability of the $R_{\textrm{ph}}$ and the $K$-band light curves. The modulation affects (both in phase and in amplitude) the temperature and luminosity variations during the pulsation of the inner part of the photosphere first, then modify the radial displacement of the outer part of the envelope as well.
As a summary, the first attempt to perform BW analysis of Blazhko stars in different phases of the modulation led to the conclusion that the modulation is a strongly depth-dependent phenomenon, that influences the pulsation of the up-most regions of the atmosphere only. This result supports the interpretation of the phenomenon by the interaction of different shocks instead of the non-radial or the resonant-mode [@kb] explanations, as the detected depth dependency of the modulation properties seems to contradict the latter ideas. It is hard to reconcile non-linear mode interactions with the strictly repeating nature of expansion and contraction of the deeper layers of the atmospheres of Blazhko RRL stars, as inferred from the present study, and from the reduced/missing nature of the modulation in the $K$-band.
However, to prove that, indeed, shock interaction is responsible for the modulation of the pulsation, 3-dimension modelling of the propagation of the different shocks in the atmosphere of RRL stars would be needed. It has to be noted as well that there are some properties of the Blazhko phenomenon, e.g. the relative stability and regularity of the modulation in some cases, and the lack of any connection between the physical properties of the stars and the occurrence of the modulation, which are hard to explain if the modulation is induced by shock interaction. Many details need to be investigated further in order to understand how the shock model of the modulation works. It would be desirable to observe at least one Blazhko RRL variable simultaneously with optical and near-infrared photometry, as well as spectroscopically through a complete modulation cycle, to definitely establish the change of the behaviour of the different layers of RRL star atmospheres during the Blazhko cycle.
Acknowledgments {#acknowledgments .unnumbered}
===============
The comments of the anonymous referee helped to improve the lucidity and readability of the paper significantly. We are grateful to J. Nuspl and L.G. Balázs for fruitful discussions and A. Sollima for commenting the possible status of V48. GH acknowledges support by the Chilean Ministry for Economy, Development, and Tourism’s Programa Iniciativa Científica Milenio through grant IC 120009, awarded to the Millennium Institute of Astrophysics (MAS); by Proyecto Basal PFB-06/2007; by FONDECYT Regular 1141141, and by CONICYT-PCHA/Doctorado Nacional grant 2014-63140099.
Anderson R. I. et al., 2016, ApJS, 226, 18 Anderson R. I., 2014, A&A, 566, L10 Benkő J. M., Bakos G. Á. Nuspl J., 2006, MNRAS, 372, 1657 Burki G., Meylan G., 1986, A&A, 159, 261 Buchler R., Kolláth Z., 2011, ApJ, 731, 24 Cacciari C., Clementini G., Fernley J. A., 1992, ApJ, 396, 219 Cacciari C., Clementini G., Prevot, L., Buser, R., 1989, A&A, 209, 141 Cacciari C., Corwin T. M., Carney B. W., 2005, AJ, 129, 267 Castelli F., Kurucz R. L., 2003, in Piskunov N., Weiss W.-W., Gray D.-F., eds, Proc. IAU Symp. 210, Modelling of Stellar Atmospheres. Cambridge Univ. Press, Cambridge, p. 20 Chadid M., Gillet D., 1996, A&A, 319, 154 Chadid M., Gillet D., 1998, A&A, 335, 255 Chadid M., Preston G. W., 2013, MNRAS, 434, 552 Chadid M., Vernin J., Gillet D., 2008, A&A, 491, 537 Chadid M. et al., 2014, AJ, 148, 88 Di Benedetto G. P., 2013, MNRAS, 430, 546 Dziembowski W., Mizerski T., 2004, Acta Astron., 54, 363 Fernley J., 1994, A&A, 284, 16 Fokin A. B., 1992, MNRAS, 256, 26 Gillet D., 2013, A&A, 554, 46 Gillet D., Fokin A. B., 2014, A&A, 565, 73 Gillet D., Fabas N., Lèbre A., 2013, A&A, 553, 59 Groenewegen M. A. T., 2013, A&A, 550, 70 Hill S. J., 1972, ApJ, 178, 793 Jones R. V., Carney B. W., Storm J., Latham D. W., 1992, ApJ, 386, 646 Jones R. V., Carney B. W., Latham D. W., 1988, ApJ, 326, 312 Jurcsik J. et al., 2009a, MNRAS, 393, 1553 Jurcsik J. et al., 2009b, MNRAS, 400, 1006 Jurcsik J. et al., 2015, ApJS, 219, 25 Jurcsik J. et al., 2017, MNRAS, 468, 1317 (Paper I) Kervella P. et al., A&A, 600, 127 Kolenberg K. et al., 2010, ApJ, 713, L198 Liu T., Janes K. A., 1990, ApJ, 354, 273 Mathias P., Gillet D., Fokin A.B., Chadid M., 1995, A&A, 298, 843 Molnár L. et al., 2017, MNRAS, 466, 4009 Nardetto N., Fokin A., Mourard D., Mathias Ph., Kervella P., Bersier D., 2004, A&A, 428, 131 Nardetto N. et al., 2017, A&A, 597, 73 Navarrete C. et al., 2015, A&A, 577, 99 Preston G. W., 2011, AJ, 141, 6 Ripepi V. et al., 2014, MNRAS, 437, 906 Schwarzschild M., 1952, in Trans. IAU, ed. P.T. Oosterhoff (Cambridge: Cambridge Univ. Press), 8, 881 Skillen I., Fernley J. A., Stobie R. S., Jameson R. F., 1993, MNRAS, 265, 301 Smolec R., 2017, MNRAS, 467, 2349 Soszyński I. et al., 2015, Acta Astron., 65, 329 Storm J., Carney B. W., Latham D. W., 1994, A&A, 290, 443 Stothers R., 2006, ApJ, 652, 643
[^1]: E-mail: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study a set of crossed 1D systems, which are coupled with each other via tunnelling at the crossings. We begin with the simplest case with no electron-electron interactions and find that besides the expected level splitting, bound states can emerge. Next, we include an external potential and electron-electron interactions, which are treated within the Hartree approximation. Then, we write down a formal general solution to the problem, giving additional details for the case of a symmetric external potential. Concentrating on the case of a single crossing, we were able to explain recent experinents on crossed metallic and semiconducting nanotubes \[J. W. Janssen, S. G. Lemay, L. P. Kouwenhoven, and C. Dekker, Phys. Rev. B **65**, 115423 (2002)\], which showed the presence of localized states in the region of crossing.'
author:
- 'D. Makogon, N. de Jeu, and C. Morais Smith'
title: Coupled quantum wires
---
Introduction
============
Physics in 1D systems manifests a number of peculiar phenomena, such as spin-charge separation, conductance quantization,[@BLandauer] and anomalous low-temperature behavior in the presence of backscattering impurity.[@Kane1] It is reasonable to expect that the more complex structures composed of crossed 1D systems, such as crossings and arrays, should exhibit some particular features as well. Although the transport properties of crossed 1D systems and their arrays have been thoroughly studied both theoretically[@Komnik] and experimentally[@Gao; @Fuhrer; @Postma], the electronic structure of these systems is much less understood and the interpretation of existing experimental results is challenging. Recent scanning tunnelling microscopy (STM) experiments on a metallic carbon nanotube crossed with a semiconducting one[@Janssen] have shown the existence of localized states at the crossing which are not due to disorder. However, these localized states do not appear systematically in all experiments, i.e. the effect is highly dependent on the nature of the carbon nanotubes (metallic or semiconducting), of the barrier formed at the crossing, etc. Aiming at clarifying this problem, we present in this paper a detailed study of tunnelling effects between crossed 1D systems in the presence of potential barriers for massive quasiparticle excitations. Because effects of electron-electron interactions can be reasonably incorporated in a random phase approximation (RPA),[@DzLarkin; @DasSarma] we study a simpler model, accounting for electron-electron interactions only within Hartree approximation. The outline of this paper is the following: in section II we introduce the model that we are going to use to describe the array of crossed nanowires. In section III we consider a particular case of free electrons and write down explicit solutions for the case of one and four crossings. Section IV contains formal general solution with additional details given for the case of a symmetric external potential. We demonstrate the effect of tunnelling on the electronic structure of single crossings in Section V and qualitatively discuss different possibilities depending on the external potential. Section VI contains quantitative analysis and comparison with available experimental data of the electronic structure of single crossing for different values of parameters. Our conclusions and open questions are presented in Section VII.
The Model
=========
We consider a system composed of two layers of crossed quantum wires with interlayer coupling. The upper layer has a set of parallel horizontal wires described by fermionic fields $\psi_{j}(x)$, whereas the lower layer contains only vertical parallel wires described by the fields $\varphi_{i}(y)$. The wires cross at the points $(x_i,y_j)$, with $i,j \in Z$ and the distance between layers is $d$, with $min(|x_i-x_{i+1}|,|y_j-y_{j+1}|)\gg d$, see Fig.1.\
{width="6cm"}
\
The partition function of the system reads $$Z=\int d[\psi_j] d[\psi^*_j] d[\varphi_{i}]
d[\varphi_{i}^*]e^{-S/\hbar},$$ with the total action given by $$S=S_0+S_{\rm sct}+S_{\rm int}.\label{Taction}$$ The first term accounts for the kinetic energy and external potential $V^{\rm ext}_{j}(x)$, which can be different in each wire and may arise, e.g., due to a lattice deformation, when one wire is built on top of another, $$\begin{aligned}
\nonumber
S_{0}&=&\sum_j\int_0^{\hbar \beta}d\tau\int dx\psi_{j}^*
(x,\tau)G^{-1}_{jx}\psi_{j}(x,\tau)\\
&+& \sum_i\int_0^{\hbar\beta}d\tau\int dy\varphi_{i}^*
(y,\tau)G^{-1}_{iy}\varphi_{i}(y,\tau),\label{action0}\end{aligned}$$ where $$\begin{aligned}
\nonumber
G^{-1}_{jx}&=&\hbar\frac{\partial}{\partial\tau}-\frac{\hbar^2}{2
m}\frac{d^2}{d x^2}+V^{\rm ext}_{j}(x)-\mu_x,\\
G^{-1}_{iy}&=&\hbar\frac{\partial}{\partial\tau}-\frac{\hbar^2}{2
m}\frac{d^2}{d y^2}+V^{\rm ext}_{i}(y)-\mu_y. \label{Greenf}\end{aligned}$$ Here, $\mu_{x,y}$ denotes the chemical potential in the upper ($\mu_{x}$) or lower ($\mu_{y}$) layer.
The second term of Eq. (\[Taction\]) describes scattering at the crossings $(x_i,y_j)$, $$S_{\rm sct}=\sum_{ij}\int_0^{\hbar\beta}d\tau H_{ij},$$ where $$\nonumber
H_{ij}=\left[ \psi_{j}^*
(x_i,\tau) \quad \varphi^*_{i}(y_j,\tau) \right]\left( \begin{array}{cc} U_{ij} & T_{ij}\\
T^*_{ij} & \tilde{U}_{ij}
\end{array}
\right)
\left[ \begin{array}{c} \psi_{j}(x_i,\tau)\\ \varphi_{i}(y_j,\tau)
\end{array}
\right].$$ Notice that the matrix element $U_{ij}$ describing intra-layer contact scattering can, in principle, be different from $\tilde{U}_{ij}$, but both must be real. On the other hand, the contact tunnelling (inter-layer) coefficient between the two crossed wires $T_{ij}$ can be a complex number, since the only constraint is that the matrix above must be Hermitian.
The third term in Eq. (\[Taction\]) accounts for electron-electron interactions, $$\begin{aligned}
\nonumber
S_{\rm int}&=&\frac{1}{2}\sum_j\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\int dx\int dx'\psi_{j}^*
(x,\tau)\psi_{j}^* (x',\tau')V^{\rm
e-e}(x-x')\psi_{j}(x,\tau)\psi_{j}(x',\tau')\\
&+&
\frac{1}{2}\sum_i\int_0^{\hbar\beta}d\tau\int_0^{\hbar\beta}d\tau'\int
dy\int dy'\varphi_{i}^* (y,\tau)\varphi_{i}^* (y',\tau')V^{\rm
e-e}(y-y')\varphi_{i}(y,\tau)\varphi_{i}(y',\tau').\label{actionint}\end{aligned}$$
Free electrons case
===================
We start by considering a very simplified case, namely, free electrons (no electron-electron interaction, $V^{\rm e-e}(x)=0$ and no external potential, $V^{\rm ext}_{j}(x)=0$). Moreover, we assume $\tilde{U}_{ji}=U_{ji}=0$ and put $\mu_x=\mu_y=\mu$. The interlayer tunnelling is assumed to be equal at each crossing point $T_{ij}=T$ and to have a real and positive value. In such a case, the partition function consists of only Gaussian integrals. We can then integrate out the quantum fluctuations, which reduces the problem to just solving the equations of motion. Considering a real time evolution and performing a Fourier transformation in the time variable, we are left with the following equations of motion for the fields: $$\begin{aligned}
\nonumber
\left(-\frac{\hbar^2}{2 m}\frac{d^2}{d x^2}-E\right)\psi_{j}(x)+T\sum_{l} \delta(x-x_l)\varphi_{l}(y_j)&=&
0,\\
\left(-\frac{\hbar^2}{2 m}\frac{d^2}{d
y^2}-E\right)\varphi_{i}(x)+T\sum_{l}
\delta(y-y_l)\psi_{l}(x_i)&=&
0,\label{EOM1}\end{aligned}$$ where $m$ denotes the electron mass and $E$ is the energy of an electron state. Firstly, we evaluate the solutions for the case of free electrons without tunnelling and then we investigate how the addition of tunnelling changes the results. The solution for the free electron case consists of symmetric and antisymmetric normalized modes, $$\psi_{s}(x)=\frac{1}{\sqrt{L}}\cos(k_s x), \qquad
\psi_{a}(x)=\frac{1}{\sqrt{L}}\sin(k_a x), \label{freesol}$$ respectively. The corresponding momenta $k_s$ and $k_a$ depend on the boundary conditions: with open boundary conditions $k_s=\pi
(2n+1)/2L$, $k_a=\pi n/L$ and with periodic boundary conditions $k_s=k_a=\pi n/L$ for a wire of length $2L$ and $n$ integer. To find the solution for the case with tunnelling $T\neq0$, we have to solve Eqs. (\[EOM1\]). These equations are linear, therefore, the solution consists of a homogeneous and an inhomogeneous parts, $$\psi_{j}(x)=\psi_{j}^{\rm hom}(x)+\psi_{j}^{\rm inh}(x),$$ which are $$\psi_{j}^{\rm hom}(x)=A_j e^{i k x}+B_j e^{-i k x},$$ $$\psi_{j}^{\rm inh}(x)=\frac{T m}{\hbar^2 k
}\sum_{l}\varphi_{l}(y_j)\sin(k|x-x_l|). \label{Inh1}$$ Imposing open boundary conditions, $\psi_{j}(\pm
L)=0$, we find $$\begin{aligned}
\nonumber
A_j e^{i k L}+B_j e^{-i k L}+\psi_{j}^{\rm inh}(L)&=&
0,\\
A_j e^{-i k L}+B_j e^{i k L}+\psi_{j}^{\rm inh}(-L)&=&
0.\label{EOM2}\end{aligned}$$ Writing the above equations in a matrix notation and inverting yields $$\nonumber
\left( \begin{array}{c} A_j\\ B_j
\end{array}
\right)=\frac{-1}{2 i \sin(2 k L)}\left( \begin{array}{cc} e^{i k L} & -e^{-i k L}\\
-e^{-i k L} & e^{i k L}
\end{array}
\right)
\left( \begin{array}{c} \psi_{j}^{\rm inh}(L)\\ \psi_{j}^{\rm inh}(-L)
\end{array}
\right).$$ Substituting explicitly the expression for $\psi_{j}^{\rm inh}(\pm
L)$ given by Eq. (\[Inh1\]) and using the mathematical identity $$\begin{aligned}
\nonumber
\left( e^{i k x} \quad e^{-i k x} \right)\left( \begin{array}{cc} e^{i k L} & -e^{-i k L}\\
-e^{-i k L} & e^{i k L}
\end{array}
\right)
\left( \begin{array}{c} \sin(kL-kx_l)\\ \sin(kL+kx_l)
\end{array}
\right)\\ \nonumber
=\cos\left(2 k L\right)\cos(kx-kx_l)-\cos(kx+kx_l),\end{aligned}$$ leads, after simplifications, to the solution $$\begin{aligned}
\nonumber
\psi_{j}(x)&=&-T\sum_{l} G(x,x_l)\varphi_{l}(y_j),\\
\varphi_{i}(y)&=&-T\sum_{l} G(y,y_l)\psi_{l}(x_i),\label{FSol1}\end{aligned}$$ where, for open boundary conditions, $$\begin{aligned}
\nonumber
G_o(x_i,x_j,E)&\equiv&\frac{m}{\hbar^2 k\sin(2 k L)}
[\cos(kx_i+kx_j)\\ &-&\cos(2kL-k|x_i-x_j|)],\label{SimpleRep}\end{aligned}$$ and the energy $E$ is related to $k$ as $E=\hbar^2 k^2/2m$. Similar calculations can be performed for the case of periodic boundary conditions, yielding Eq. (\[FSol1\]) with $$G_p(x_i,x_j,E)\equiv\frac{m}{\hbar^2 k\sin(k
L)}\cos(kL-k|x_i-x_j|).\label{PeriodRep}$$
Two crossed wires
-----------------
In particular, for the simplest case of a single horizontal and a single vertical wires, with just one crossing at $(x_0,y_0)$, the solution is: $$\begin{aligned}
\nonumber
\psi(x)&=&-T G(x,x_0,E)\varphi(y_0)\\
\varphi(y)&=&-T G(y,y_0,E)\psi(x_0).\label{FSol2}\end{aligned}$$
{width="6cm"}
\
By substituting $(x,y)=(x_0,y_0)$, we find that at the crossing point $$\begin{aligned}
\nonumber
\psi(x_0)&=&-T G(x_0,x_0,E)\varphi(y_0)\\
\varphi(y_0)&=&-T G(y_0,y_0,E)\psi(x_0).\label{FSol3}\end{aligned}$$ The consistency condition requires that $$\left| \begin{array}{cc} 1 & T G(x_0,x_0,E)\\
T G(y_0,y_0,E) & 1
\end{array}
\right|=0,$$ or $$T^2 G(x_0,x_0,E) G(y_0,y_0,E)=1.\label{FreeSpectr}$$ The solution is even simpler if $(x_0,y_0)=(0,0)$. Then, for open boundary conditions, the symmetric modes are $$\begin{aligned}
\nonumber
\psi(x)&=&\frac{\varphi(0)T m}{\hbar^2 k \cos(k
L)}\sin(k L-k|x|), \\ \nonumber
\varphi(y)&=&\frac{\psi(0)T m}{\hbar^2 k \cos(k
L)}\sin(k L-k|y|),$$ and the antisymmetric modes are left unchanged in comparison with Eqs. (\[freesol\]). Also, $$G(0,0,E)=\frac{m \tan(k L)}{\hbar^2k}, \label{nonintfs}$$ and the secular equation (\[FreeSpectr\]) becomes $$\left[\frac{T m\tan(k L)}{\hbar^2 k}\right]^2=1,$$ which splits into two transcendental equations $$\begin{aligned}
\nonumber
k^+&=&-\frac{T m}{\hbar^2}\tan(k^+ L), \\ \nonumber
k^-&=&\frac{T m}{\hbar^2}\tan(k^- L).$$ The first one describes the shifted values of scattering states energies, whereas the second equation has an additional bound state solution with $E<0$, if $T>T_0=\hbar^2/mL$. The appearance of the bound state is exclusively due to the presence of tunnelling. For an electron in a wire of length $2L=10^3$ nm the corresponding value is $T_0=7.62\times 10^{-5}$ eV$\cdot$nm and for quasiparticles the value of $T_0$ is typically larger, inversely proportional to their effective mass. Defining then $\kappa\equiv-i k^-$ and taking the thermodynamic limit $L\rightarrow \infty$, we find $|\kappa|=T m/\hbar^2$ with the corresponding bound state energy $$E=-\frac{T^2 m}{2\hbar^2},\label{BEnergy}$$ and the wave function given by $$\psi(x)=\frac{\sqrt{|\kappa|}}{2}e^{-|\kappa x|}.\label{Expsol}$$ The factor $1/2$ instead of $1/\sqrt{2}$ comes from the fact that now an electron can tunnel into the other wire, where its wavefunction $\varphi(0)=-\psi(0)$. Eqs. (\[BEnergy\]) and (\[Expsol\]) hold for both open and periodic boundary conditions. Since the threshold value $T_0$ is quite small, the bound state should exist for a typical crossing with relatively good contact. However, the energy of the state is extremely small, $E\sim 10^{-8}$ eV if $T\sim T_0$. Qualitatively similar results were found by numerical computation[@Schult; @Carini] of the ground-state energy of an electron trapped at the intersection of a cross formed by two quantum wires of finite width.
Four crossed wires
------------------
For the case of two wires in the upper and two in the lower layers, there are four crossings. In this case, the self consistent equations read $$\left[ \begin{array}{c} \psi_{1}(x_1)\\
\psi_{1}(x_2)\\ \psi_{2}(x_1)\\
\psi_{2}(x_2)
\end{array}
\right]=M(x_1,x_2,E)
\left[ \begin{array}{c} \varphi_{1}(y_1)\\
\varphi_{1}(y_2)\\ \varphi_{2}(y_1)\\
\varphi_{2}(y_2)
\end{array}
\right]$$ and $$\left[ \begin{array}{c} \varphi_{1}(y_1)\\
\varphi_{1}(y_2)\\ \varphi_{2}(y_1)\\
\varphi_{2}(y_2)
\end{array}
\right]=M(y_1,y_2,E)\left[ \begin{array}{c} \psi_{1}(x_1)\\
\psi_{1}(x_2)\\ \psi_{2}(x_1)\\
\psi_{2}(x_2)
\end{array}
\right],$$ where $$M(x_1,x_2,E)=-T\left( \begin{array}{cccc} G(x_1,x_1,E) & 0 & G(x_1,x_2,E) & 0\\
G(x_1,x_2,E) & 0 & G(x_2,x_2,E) & 0\\
0 & G(x_1,x_1,E) & 0 & G(x_1,x_2,E)\\
0 & G(x_1,x_2,E) & 0 & G(x_2,x_2,E)
\end{array}
\right).$$ The secular equation then has the form $$\det[M(x_1,x_2,E)M(y_1,y_2,E)-I]=0,\label{SpectralEq}$$ which yields a rather complicated transcendental equation ($I$ is the identity matrix). The spectral equation for bound states $E<0$ can be significantly simplified in the thermodynamic limit $L\rightarrow \infty$. Then, with $k=i \kappa$, for both open and periodic boundary conditions, the matrix elements become $$G(x_i,x_j,E)=\frac{m}{\hbar^2 |\kappa|}
e^{-|\kappa(x_i-x_j)|}\label{fundbound}$$ and the secular equation in Eq. (\[SpectralEq\]) has 4 solutions with negative energy described by $$\begin{aligned}
\nonumber
E&=&-\frac{T^2 m}{2\hbar^2}(1-a_1-a_2+a_1 a_2),\\ \nonumber
E&=&-\frac{T^2 m}{2\hbar^2}(1+a_1-a_2-a_1 a_2),\\ \nonumber
E&=&-\frac{T^2 m}{2\hbar^2}(1-a_1+a_2-a_1 a_2),\\ \nonumber
E&=&-\frac{T^2 m}{2\hbar^2}(1+a_1+a_2+a_1 a_2). \nonumber
$$ Here, $a_1\equiv e^{-|\kappa(x_2-x_1)|}$, $a_2\equiv
e^{-|\kappa(y_2-y_1)|}$, and $E=-\hbar^2 \kappa^2/2m$ (notice the implicit dependence of $a_1$ and $a_2$ on $E$). The value of $a_i$ depends exponentially on the distance between the crossing points. In the limit $|x_2-x_1|,|y_2-y_1|\rightarrow \infty$ the value of $a_1,a_2\rightarrow 0$, which correspond to four independent crossings with the bound state energy $E=-T^2 m/{2\hbar^2}$, the same value as we found in the previous case (see Eq.(\[BEnergy\])).
A regular lattice of crossed wires
----------------------------------
Consider now a regular square lattice, with lattice constant $a$. Then, one has $x_l=al$ and $y_j=aj$. From symmetry arguments, the wave functions should be $\psi_{j}(x)=\psi_{0}(x)e^{iK_y aj}$ and $\varphi_{l}(y)=\varphi_{0}(y)e^{iK_x al}$. After substituting them into Eq. (\[FSol1\]) and using Eq. (\[fundbound\]) we find $$\begin{aligned}
\nonumber \psi_j(x)=&-&T\varphi_0(y_j)\frac{me^{i K_x l_x
a}}{\hbar^2 \kappa}\left[\frac{\sinh(\kappa x -
\kappa a l_x)e^{i K_x a }}{\cosh(\kappa a )-\cos(K_x a )}\right.\\
\nonumber &-&\left.\frac{\sinh(\kappa x - \kappa
(l_x+1)a)}{\cosh(\kappa a )-\cos(K_x a )}\right],\\ \nonumber
\varphi_l(y)=&-&T\psi_0(x_l)\frac{me^{i K_y l_y a}}{\hbar^2
\kappa}\left[\frac{\sinh(\kappa x -
\kappa a l_y)e^{i K_y a }}{\cosh(\kappa a )-\cos(K_y a )}\right.\\
\nonumber &-&\left.\frac{\sinh(\kappa y - \kappa
(l_y+1)a)}{\cosh(\kappa a )-\cos(K_y a )}\right],\end{aligned}$$ where $l_x,l_y \in Z$, such that $al_x\leq x<a(l_x+1)$ and $al_y\leq y<a(l_y+1)$. Therefore, $\psi_{j}(x_l)=\psi_{0}(0)e^{i(K_x al+K_y aj)}$ and $\varphi_{l}(y_j)=\varphi_{0}(0)e^{i(K_x al+K_y aj)}$, with $\psi_{0}(0)$ and $\varphi_{0}(0)$ related by $$\begin{aligned}
\nonumber
\psi_0(0)&=&-T\frac{m}{\hbar^2 \kappa}\frac{\sinh(\kappa
a )}{\cosh(\kappa a )-\cos(K_x a )}\;\varphi_0(0),\\
\varphi_0(0)&=&-T\frac{m}{\hbar^2 \kappa}\frac{\sinh(\kappa
a )}{\cosh(\kappa a )-\cos(K_y a )}\;\psi_0(0).\label{BandSol}\end{aligned}$$ Thus, the spectral equation reads $$\nonumber
1=\frac{(mT)^2}{(\hbar^2 \kappa)^2}\frac{\sinh^2(\kappa
a )}{[\cosh(\kappa a )-\cos(K_x a )][\cosh(\kappa a )-\cos(K_y a
)]}.$$ By performing an analytic continuation $k=i \kappa$ in Eq.(\[BandSol\]), we find an equations similar to the one obtained previously by Kazymyrenko and Douçot[@Kazymyrenko] when studying scattering states in a lattice. The spectral equation describes a band formed by bound states with energies $-T/a<E<0$. The momenta $K_x$ and $K_y$ run in the interval $-\pi<K_xa,K_ya<\pi$ if $T\geq T_f=2\hbar^2/ma$ or inside the region $|\sin(K_xa/2)\sin(K_ya/2)|\leq T/T_f$ if $T<T_f$. Similar results were calculated,[@Dickinson] estimated,[@CastroNeto] and measured[@Zhou] in the context of hybridization between vertical and horizontal stripe modes in high-Tc superconductors.
A more general case
===================
Now we consider a more general model, which takes into account the presence of an inhomogeneous potential $V^{\rm ext}_{j}(x)$ arising from possible lattice deformations, and includes electron-electron interactions $V^{\rm e-e}(x)$, which will be treated at a mean field level, within the Hartree approximation $V^{\rm e-e}_{{\rm H} j}(x)$. Each crossing $(x_i,y_j)$ is considered as a scattering point with tunnelling $T_{ij}$ and scattering potential $U_{ij}$. The corresponding equations of motion then read $$\begin{aligned}
\nonumber
D_{jx}\psi_{j}(x)+\sum_{l}[U_{lj}\psi_{j}(x_{l})+T_{lj}\varphi_{l}(y_j)]\delta(x-x_l)&=&
0,\\\nonumber
D_{iy}\varphi_{i}(x)+\sum_{l}[\tilde{U}_{il}\;\varphi_{i}(y_l)+T^*_{il}\psi_{l}(x_{i})
]\delta(y-y_l)&=&
0,$$ where $$\begin{aligned}
\nonumber
D_{jx}&=&-\frac{\hbar^2}{2 m}\frac{d^2}{d x^2}+V_{j}(x)-E,\\ \nonumber
D_{iy}&=&-\frac{\hbar^2}{2 m}\frac{d^2}{d
y^2}+V_{i}(y)-E,$$ with $V_{j}(x)=V^{\rm ext}_{j}(x)+V^{\rm e-e}_{{\rm H} j}(x)$. This model is solved most easily through the Green’s function satisfying $$\nonumber
D_{jx_1} G_j(x_1,x_2,E)=\delta (x_1-x_2)$$ with $$\nonumber
G_j(x_1,x_2,E)= G_j^*(x_2,x_1,E),$$ and the corresponding open boundary conditions, $$\nonumber
G_j(x_1,L,E)=0,\quad G_j(x_1,-L,E)=0,$$ or the periodic ones $$\begin{aligned}
\nonumber
G_j(x_1,L,E)= G_j(x_1,-L,E),\\ \nonumber
{ G_j}'(x_1,L,E)={ G_j}'(x_1,-L,E),\end{aligned}$$ where the prime denotes the derivative with respect to $x_1$. Note that we consider real time Green’s function for a particular wire (not the whole system). The solution to the model is $$\begin{aligned}
\nonumber
\psi_{j}(x)&=&-\sum_{l}[U_{lj}\psi_{j}(x_{l})+T_{lj}\varphi_{l}(y_j)] G_j(x,x_l,E),\\
\varphi_{i}(y)&=&-\sum_{l}[\tilde{U}_{il}\;\varphi_{i}(y_l)+T^*_{il}\psi_{l}(x_{i})
] G_i(y,y_l,E),\label{GenSol}\end{aligned}$$ which we require to be normalized $$\sum_{l}\left(\int |\psi_{l}(x)|^2 dx+\int
|\varphi_{l}(y)|^2dy\right)=1.$$ The self consistency condition for the value of the functions at crossing points $(x_i,y_j)$ yields the equations $$\begin{aligned}
\nonumber
\sum_{l}[(U_{lj}
G_j(x_i,x_l,E)+\delta_{il})\psi_{j}(x_{l})\\\nonumber +T_{lj}
G_j(x_i,x_l,E)\varphi_{l}(y_j)]&=&0,\\\nonumber
\sum_{l}[(\tilde{U}_{il}\; G_s(y_j,y_l,E)+\delta_{jl})\varphi_{i}(y_l)\\
+T^*_{il} G_i(y_j,y_l,E)\psi_{j}(x_{i}) ]&=&0.\label{Cons}\end{aligned}$$ To find nontrivial solutions for the fields $\psi_{j}(x)$ and $\varphi_{i}(y)$, the system of homogeneous equations in Eq.(\[Cons\]) has to be linearly dependent and hence the solution is represented by the null space of the system. This means that after writing the equations in a matrix form, the determinant of the matrix should be zero, thus leading to a spectral equation for $E$. Moreover, bound state solutions in the thermodynamic limit $L\rightarrow \infty$ satisfy both open and periodic boundary conditions, since $\psi(\pm L)\rightarrow 0$ and $\psi'(\pm
L)\rightarrow 0$.
To understand better the dependence of the Green’s function $
G_j(x_i,x_l,E)$ on $E$, we represent the function through the solutions of the homogenous equations, $$D_{jx}\psi_j(x)=0.$$ We omit the index $j$ in what follows for simplicity. The most general and common representation, which holds for any static potential, reads as follows: $$G(x_1,x_2,E)=\sum_{n}\frac{\psi_{\varepsilon_{n}}^*(x_1)\psi_{\varepsilon_{n}}(x_2)}{\varepsilon_{n}-E}.
\label{fsrep0}$$ Here, the function $\psi_{\varepsilon}(x)$ is the solution of the homogenous equation $$\left(-\frac{\hbar^2}{2 m}\frac{d^2}{d
x^2}+V(x)-\varepsilon\right)\psi_\varepsilon(x)=0,$$ and the spectrum $\{\varepsilon_{n}\}$ is obtained by imposing the corresponding boundary conditions. Notice that in the present representation of $ G(x_1,x_2,E)$ the functions $\psi_{\varepsilon_{n}}(x)$ have to be orthonormal. By writing $ G(x_1,x_2,E)$ in the form given in Eq. (\[fsrep0\]), the following identity arises $$\int dx'
G(x_1,x',E) G(x',x_2,E)=\frac{\partial G(x_1,x_2,E)}{\partial
E}.$$ The case $x_1=x_2=0$ for free electrons is illustrated in Fig.\[fig3\], where Eq. (\[nonintfs\]) is plotted. If some external potential is present, $ G(x_0,x_0,E)$ has the same form but the positions of the poles are shifted and the corresponding values are different.
![\[fig3\]$ G(0,0,E)$ in units of $m/\hbar^2$ versus $E$ in units of $\hbar^2/2mL^2$. ](Spectrum.eps){width="6cm"}
If no regularization is used, the calculations for $E>0$ must be performed in the finite size limit, otherwise with $L\rightarrow\infty$ the energy distance between different modes vanishes and the poles situated on the real positive half axis merge to form a branch cut singularity. This behavior can be readily seen on the example of Eq. (\[nonintfs\]), where can perform an analytic continuation, considering $k\rightarrow k+
ik'$. Then, in the limit $L\rightarrow\infty$, $\tan(kL + ik'L)=i
{\rm sgn}(k')$, and the function $ G(x_0,x_0)$ changes sign as one goes from the upper to the lower complex half plane for $k\neq0$.
Now we represent the Green’s function through the solutions of the homogenous equation $$\left(-\frac{\hbar^2}{2 m}\frac{d^2}{d
x^2}+V(x)-E\right)\psi(x)=0. \label{Srdng}$$ This is a second order differential equation, therefore, it should have two linearly independent solutions, which we call $\psi_1(x)$ and $\psi_2(x)$. Then the Green’s function is $$G(x_1,x_2,E)=\left\{\begin{array}{c} A_{-}\psi_1(x_1)+B_{-}\psi_2(x_1), x_1\leq x_2\\
A_{+}\psi_1(x_1)+B_{+}\psi_2(x_1), x_1 >x_2
\end{array} \right.
\label{fsrepf},$$ where the expressions for the coefficients $A_{-},B_{-},A_{+},B_{+}$ (functions of $x_2$), are derived in the Appendix A. In particular, for a symmetric potential $V(x)$, we can choose a symmetric $\psi_s(x)$ and an antisymmetric $\psi_a(x)$ solutions as linearly independent, i.e., $\psi_1(x)=\psi_s(x)$ and $\psi_2(x)=\psi_a(x)$. Thus we find $$G(x,0,E)=\frac{m\psi_{a}(L)}{\hbar^2{\psi_{a}}'(0)}\left[\frac{\psi_{s}(x)}{\psi_{s}(L)}-\frac{\psi_{a}(|x|)}{\psi_{a}(L)}\right]$$ and $$G(0,0,E)=\frac{m\psi_{s}(0)}{\hbar^2{\psi_{a}}'(0)}\frac{\psi_{a}(L)}{\psi_{s}(L)}.
\label{fs00n}$$ To obtain the results in the thermodynamic limit $L\rightarrow
\infty$, it is useful to rewrite $ G(x_1,x_2)$ using quantities which do not depend on $L$ explicitly. For example, $$G(x,0,E)= G(0,0,E)\frac{\psi_{s}(x)}{\psi_{s}(0)}-\frac{m}{\hbar^2}\frac{\psi_{a}(|x|)}{{\psi_{a}}'(0)}.$$ After substitution of Eqs. (\[freesol\]) into Eq.(\[fsrepf\]) and simplification, for the case of noninteracting electrons we find $$\begin{aligned}
\nonumber
G(x_1,x_2,E)&=&\frac{m}{\hbar^2k\sin(2 k L)}
[\cos(kx_1+kx_2)\\
\nonumber &-&\cos(2kL-k|x_1-x_2|)],\nonumber\end{aligned}$$ which is the same expression as in the previous section (see Eq.(\[SimpleRep\])). This is a posteriori justification of the use of the same letter $ G(x_1,x_2,E)$ in the first section. The case of a harmonic potential is considered in Appendix B.
a single crossing
=================
Now we apply our results including tunnelling and external potential to the simpler case of only two crossed wires, aiming to compare our findings with experiments. Using the general solution given by Eq. (\[GenSol\]), and considering $T=T^*$, we can write $$\begin{aligned}
\nonumber
\psi(x)&=&-[U\psi(x_{0})+T\varphi(y_0)] G_1(x,x_0,E),\\
\nonumber \varphi(y)&=&-[\tilde{U}\;\varphi(y_0)+T\psi(x_{0}) ]
G_2(y,y_0,E).\label{Solsingle}\end{aligned}$$ By substituting $(x,y)=(x_0,y_0)$, we find that at the crossing point $$\begin{aligned}
\nonumber
[1+U G_1(x_0,x_0,E)]\psi(x_{0})+T G_1(x_0,x_0,E)\varphi(y_0)&=&0,\\
\nonumber [1+\tilde{U} G_2(y_0,y_0,E)]\varphi(y_0)+T
G_2(y_0,y_0,E)\psi(x_{0})&=&0.\label{Solsinglecross}\end{aligned}$$ The consistency condition requires that $$\left| \begin{array}{cc} 1+U G_1(x_0,x_0,E) & T G_1(x_0,x_0,E)\\
T G_2(y_0,y_0,E) & 1+\tilde{U} G_2(y_0,y_0,E)
\end{array}
\right|=0,$$ or $$\begin{aligned}
\nonumber
0&=&[1+U G_1(x_0,x_0,E)][1+\tilde{U} G_2(y_0,y_0,E)] \\
\nonumber &-&T^2 G_1(x_0,x_0,E) G_2(y_0,y_0,E).\end{aligned}$$ The meaning of this equation becomes clearer in the symmetric case, when $U=\tilde{U}$ and $ G_1(x_0,x_0,E)= G_2(y_0,y_0,E)= G$. In this case, it reduces to a quadratic equation, which bears two solutions, $$\nonumber
G_+=\frac{-1}{U+T},\qquad G_-=\frac{-1}{U-T}.$$ Notice that they differ by the sign in front of the tunnelling amplitude $T$, which is shifting the potential $U$. Such symmetry effectively reduces the problem to 1D with effective potential $U_{\rm eff}\delta(x_0)$. Hence, we have $$\begin{aligned}
\nonumber
\psi(x_{0})&=&\varphi(y_0),\quad \; \; U_{\rm eff}^+=U+T,\\
\psi(x_{0})&=&-\varphi(y_0),\quad U_{\rm
eff}^-=U-T.\label{SymSolsingle}\end{aligned}$$ The shift of the energy levels in a wire due to the presence of the $\delta$ potential can be visualized with the help of the Green’s function expansion, where one has $$G(x_0,x_0,E)=\sum_{n}\frac{|\psi_{\varepsilon_{n}}(x_0)|^2}{\varepsilon_{n}-E}=\frac{-1}{U_{\rm
eff}}.$$ In the case with $U_{\rm eff}=0$, the energies are exactly those of the poles and, therefore, remain unshifted. However, since $
G(x_0,x_0,E)=-1/U_{\rm eff}$, the curve actually describes how the energies of the modes change as we keep increasing $-1/U_{\rm
eff}$ from $-\infty$ if $U_{\rm eff}>0$ or decreasing $-1/U_{\rm
eff}$ from $+\infty$ if $U_{\rm eff}<0$. In the latter case, we can run into the region with $E<0$, which would correspond to the appearance of a bound state. Nevertheless, to obtain an exact solution, it is more convenient to work with the expression for $
G(x_0,x_0,E)$ in terms of the wave functions, $$G(0,0,E)=\frac{m\psi_{s}(0)}{\hbar^2{\psi_{a}}'(0)}\frac{\psi_{a}(L)}{\psi_{s}(L)}=\frac{-1}{U_{\rm
eff}},$$ where we assumed $x_0=0$ for simplicity.
comparison with experiments
===========================
![\[figexp\] Voltage versus length diagram, which shows the experimentally observed density of states. Notice the existence of two localized states in black. (Extracted from Ref. ).](boundstatesexp.ps){width="6cm"}
Now, we will compare our theoretical findings with experimental results. We concentrate mostly on the analysis of a system consisting of two crossed single wall carbon nanotubes (SWNTs): a metallic on top of a semiconducting (MS) one.[@Janssen] In its unperturbed state, the band structure of a SWNT can be understood by considering the electronic structure of graphene. Due to its cylindrical shape, the transverse momentum of one particle excitations in a SWNT has to be quantized, whereas the longitudinal momentum may vary continuously. Combining this condition with the assumption that the electronic structure is not very different from that of graphene, one finds two different situations, depending on the topology of the SWNT: there are no gapless modes and the nanotube is semiconducting, or two gapless modes are present and the nanotube is called metallic. Analyzing the spectroscopic measurements performed along the metallic nanotube (see Fig. \[figexp\]) and comparing with the unperturbed electronic structure, one notices two main changes. First, a small quasi gap opens around the Fermi energy level $\varepsilon_F$ between $\varepsilon_F-0.2$ eV and $\varepsilon_F+0.3$ eV in the spectrum of the massless modes (corresponding to zero transverse momentum). Second, two peaks are visible at $\varepsilon_0=\varepsilon_F-0.3$ eV and $\varepsilon_1=\varepsilon_F-0.6$ eV in the region around the crossing, corresponding to localized states between the Fermi energy and the van Hove singularity at $\varepsilon_{\rm vH
}=\varepsilon_F-0.8$ eV. Such states are not visible above the Fermi energy, thus suggesting that the electron-hole symmetry is broken by the presence of some external potential. The latter may appear due to lattice distortions and the formation of a Schottky barrier at the contact between the nanotubes. [@Odintsov; @OdintsovYo] In the following, we show that if the potential is strong enough, localized states can form in the spectrum of the massive mode corresponding to the van Hove singularity with energy $\varepsilon=\varepsilon_{\rm vH }-E$. Therefore, the observed localized states should have $E_0=-0.5$ eV and $E_1=-0.2$ eV.
To incorporate in a more complete way the effects of the Schottky barrier and lattice deformation, we assume $V^{\rm ext}(x)$ to have a Lorentzian shape, $$V^{\rm ext}(x)=-\frac{\tilde{V}}{1+x^2/b^2}.\label{Potential}$$
Firstly, we study the influence of this potential alone on the electronic structure, i.e. we assume that there is no tunnelling $T=0$, and no electron-electron interactions. Exact numerical solution of the Schrodinger equation shows that an approximation of the potential in Eq. (\[Potential\]) by the harmonic one does not change the solution qualitatively. Therefore, we consider $V^{\rm ext}(x)\approx-\tilde{V}(1-x^2/b^2)$, which describes a harmonic oscillator with frequency $\omega=\sqrt{2\tilde{V}/mb^{2}}$ and corresponding spectra $E_{n}=-\tilde{V}+(n+1/2)\sqrt{2\hbar^{2}\tilde{V}/mb^{2}}$ for $E_{n}<0$. Moreover, it is reasonable to assume that the strength of the barrier $\tilde{V}$ is of the same order as the energy of the bound states and that the potential is localized on the same length scale as the localized states. Hence, we take $\tilde{V}=0.7$ eV and $b=4$ nm. It follows then from our calculations that the difference between neighboring energy levels is quite small and there are many bound states present in the case when $m$ is the actual electron mass. However, assuming $m$ to be an effective electron mass, with $m=0.025 \; m_e$, which is of the same order as the experimentally estimated values $m=0.037 \;
m_e$[@JarilloHerrero] and $m=0.06 \; m_e$,[@Radosavljevic] we find exactly two pronounced bound states: the first one has $E=-0.5$ eV and is described by the symmetric wavefunction $\psi_s(x)$ as shown in Fig. \[fig4\],
![\[fig4\]$\psi_s(x)$(nm$^{-1/2}$) versus $x$ (nm) ](psis.eps){width="6cm"}
whereas the other has $E=-0.2$ eV and is described by the antisymmetric wavefunction $\psi_a(x)$, see Fig. \[fig5\].
![\[fig5\]$\psi_a(x)$(nm$^{-1/2}$) versus $x$ (nm). ](psia.eps){width="6cm"}
Considering Fig. \[fig4\], we observe that the localization size of the state is around $10$ nm, which agrees well with the experimental data. On the other hand, the state shown in Fig.\[fig5\] has a zero value exactly at the crossing and is rather spread, a behavior which is not observed experimentally. Besides these two, a number of other states are also present in the vicinity of the van Hove singularity with $E>-0.1$ eV.
Secondly, we take into account electron-electron interactions to consider other possibilities to obtain two pronounced bound states. Unfortunately, our approach only allows us to incorporate electron-electron interactions at the mean-field level by using the Hartree selfconsistent approximation $$V^{\rm e-e}_H(x)=\int dx' V^{\rm e-e}(x-x')n(x'),$$ where $n(x)$ is the electron density, given by $$n(x)=\sum_k |\psi_k(x)|^2 n_F(\varepsilon_k-\mu).$$ Here the summation $k$ goes over energy levels and $n_F(\varepsilon)$ is the Fermi distribution. Although it is known that in 1D systems quantum fluctuations play an extremely important role, we nevertheless start with the mean-field approximation as a first step to incorporate them in RPA. Moreover, we believe that their presence does not qualitatively change the obtained results. To render the numerical calculation simpler, we consider a delta-like interaction potential, which leads to $$V^{\rm e-e}_H(x)=V_0 n(x),$$ By estimating the effective interaction strength $V_0\sim 2\pi
\hbar v_F$ from the Luttinger liquid theory, we obtain that $V_0\sim 3.4$ eV$\cdot$nm for $v_F=8.2\times 10^7$ cm/s.[@Lemay] Suppose that the lowest energy state with $E=-0.5$ eV is occupied by an electron with a certain spin. Then, there is a possibility to add to the same state an electron with an opposite spin. However, due to the repulsive Coulomb interaction the energy of the two-electron state becomes $E=-0.2$ eV for $V_0=3.15$ eV$\cdot$nm. The corresponding self consistent solution is presented in Fig. \[fig4int\].
![\[fig4int\]$\psi_s(x)$(nm$^{-1/2}$) versus $x$ (nm) ](psisint.eps){width="6cm"}
The state has the same shape as in Fig. \[fig4\], but is a bit more spread. By comparing the density of states (DOS) distribution with scanning tunnelling spectroscopy (STS) data for the crossing,[@Janssen] we observe that the inclusion of electron-electron interactions (Fig. \[fig4int\]) provides a much better agreement between theory and experiment for the $E=-0.2$ eV bound state than in the previous case (Fig.\[fig5\]).
Thirdly, we take into account tunnelling between the wires. Qualitatively, this leads to the splitting of energy levels and redistribution of charge density in the wires, thus effectively reducing the strength of electron-electron interactions. Since we have no information about the electronic structure of the semiconducting nanotube, to make a quantitative estimation we assume that the effective mass is equal in both wires and that the potential is also the same. In such a case, from symmetry arguments the electron density should be evenly distributed in both wires even for a very weak tunnelling. Therefore, the electron-electron interactions should be twice stronger than in the case without tunnelling, namely, $V_0=6.3$ eV$\cdot$nm to achieve the same energy value. Moreover, if the tunnelling coefficient is large enough, the splitting of the energy levels becomes significant and detectable. We can estimate the coefficient $T$, if we assume that it has the same order for SM, metallic-metallic (MM), and semiconducting-semiconducting (SS) nanotube junctions. The SS and MM junctions have Ohmic voltage-current dependance, characterized by the conductance $G$. Moreover, we can estimate the transmission coefficient of the junction as $G/G_0 \sim (T/2\pi \hbar
v_F)^{2}$, for $G/G_0\ll 1$. For MM junctions experimental measurements[@Dickinson] typically yield $G/G_0\sim 10^{-2}$, thus corresponding to $T\sim 0.34$ eV$\cdot$nm. For example, for $T=0.28$ eV$\cdot$nm and $\tilde{V}=0.44$ eV in Eq.(\[Potential\]), without electron-electron interactions we find that the system has [*two bound states*]{}. The lowest energy bound state with $E=-0.5$ eV is shown in Fig.\[fig6\].
![\[fig6\]$\psi_-(x)$(nm$^{-1/2}$) versus $x$ (nm). ](modeminus.eps){width="6cm"}
Compared with Fig. \[fig4\], the state has a peak exactly at the crossing, corresponding to a local increase of the DOS. The other bound state with $E=-0.2$ eV is shown in Fig. \[fig7\].
![\[fig7\]$\psi_+(x)$(nm$^{-1/2}$) versus $x$ (nm). ](modeplus.eps){width="6cm"}
Contrary to the previous case, the state has a deep at the crossing, corresponding to a local decrease of the DOS. However, these local change in DOS is too small to be observable in the present experimental data. If we now include electron-electron interactions with $V_0=3.15$ eV$\cdot$nm and add a second electron with different spin to the system, we find that the new state has $E=-0.267$ eV and acquires the shape shown in Fig. \[fig8\].
![\[fig8\]$\psi_-(x)$(nm$^{-1/2}$) versus $x$ (nm). ](modeminusint.eps){width="6cm"}
The last result suggests that there are yet other possible interpretations of the experimental results. Firstly, if the potential in the metallic SWNT is significantly decreased due to screening effects but a Schottky barrier in the semiconducting SWNT can reach considerable values, sufficient for the formation of the bound states, then the latter are also going to be present in the metallic SWNT due to tunnelling between SWNTs. Secondly, there is still a possibility to find a bound state existing purely due to tunnelling, i.e., without external potential, as was shown in Eq. (\[Expsol\]), and a second bound state may arise with different energy due to Coulumb repulsion between electrons with different spins. However, this is most probably not the case we have in the experiments, because due to electron-hole symmetry such states would exist also above the Fermi energy, a result which is not observed experimentally.
Conclusions
===========
We presented several possibilities to explain the observed localized states at the crossing of metallic and semiconducting nanotubes.[@Janssen] All of them require the existence of an external potential in the metallic and/or semiconducting SWNT to break the electron-hole symmetry, since the localized states were seen only below the Fermi energy. Most probably, such a potential comes from a Schottky barrier and the effect of lattice distortions is minimal, since such localized states were, up to now, observed only for MS crossings and not for MM or SS ones. Moreover, the effective mass of quasiparticle excitations should be of order $m=0.025 \; m_e$, where $m_e$ is the actual electron mass, to generate only a few bound states localized on a region of approximately $10$ nm with energy of order of $0.5$ eV. The best agreement with the experimental data is obtained by assuming that the second bound state has a different energy due to the Coulumb repulsion between electrons with different spins. The role of tunnelling in the observed electronic structure is not clear and allows for many interpretations. To avoid such ambiguity, the electronic structure of the semiconducting nanotube should be measured as well. Moreover, to be sure that the available STS measurements indeed represent the electronic structure of the nanotube and are free of artifacts introduced by the STM tip [@LeRoy] several measurements with different tip height should be performed.
ACKNOWLEDGMENTS
===============
We are very grateful to S. G. Lemay for useful discussions.
Here we consider the Green’s function as a function of one variable $x_1$ and fix $x_2$ for a moment. Since $ G(x_1,x_2,E)$ is the Green’s function, we require it to satisfy proper boundary conditions $ G(\pm L,x_2,E)=0$, be continuous $ G(x_2-0,x_2,E)=
G(x_2+0,x_2,E)$, and also $ G'(x_2-0,x_2,E)-
G'(x_2+0,x_2,E)=2m/\hbar^2$. Substituting Eq. (\[fsrepf\]) into the above requirements one finds $$P \left( \begin{array}{c} A_+\\
A_-\\ B_+\\
B_-
\end{array}
\right)= \frac{2 m}{\hbar^2}\left( \begin{array}{c} 0\\
0\\ 0\\
1
\end{array}
\right),\label{Acoef}$$ where $$P\equiv\left( \begin{array}{cccc} \psi_{1}(L) & 0 & \psi_{2}(L) & 0\\
0 & \psi_{1}(-L) & 0 & \psi_{2}(-L)\\
\psi_{1}(x_2) & -\psi_{1}(x_2) & \psi_{2}(x_2) & -\psi_{2}(x_2)\\
-\psi_{1}'(x_2) & \psi_{1}'(x_2) & -\psi_{2}'(x_2) & \psi_{2}'(x_2)
\end{array}
\right).$$ Multiplying the Eq. (\[Acoef\]) by the matrix $P^{-1}$ we find $$\nonumber
\left( \begin{array}{c} A_+\\
A_-\\ B_+\\
B_-
\end{array}
\right)= C \left( \begin{array}{c} \psi_{2}(L)[\psi_{2}(-L)\psi_{1}(x_2)-\psi_{1}(-L)\psi_{2}(x_2)]\\
\psi_{2}(-L)[\psi_{2}(L)\psi_{1}(x_2)-\psi_{1}(L)\psi_{2}(x_2)]\\
-\psi_{1}(L)[\psi_{2}(-L)\psi_{1}(x_2)-\psi_{1}(-L)\psi_{2}(x_2)]\\
-\psi_{1}(-L)[\psi_{2}(L)\psi_{1}(x_2)-\psi_{1}(L)\psi_{2}(x_2)]
\end{array}
\right),$$ where $$\nonumber
C\equiv\frac{2m}{\hbar^2
W_r}[\psi_{1}(L)\psi_{2}(-L)-\psi_{1}(-L)\psi_{2}(L)]^{-1}.$$ The Wronskian $$\nonumber
W_r\equiv \psi_{1}(x_2)\psi_{2}'(x_2)-\psi_{2}(x_2)\psi_{1}'(x_2),$$ is nonzero for linearly independent functions and its value does not depends on the point $x_2$.
Suppose that Eq. (\[Srdng\]) has a solution $\psi(x)$ which is neither symmetric nor antisymmetric. Thus, for symmetric potentials $\psi(-x)$ is also a solution and both of them are linearly independent. Furthermore, we can then compose a symmetric $\psi_s(x)=(\psi(x)+\psi(-x))/2$ and an antisymmetric $\psi_a(x)=(\psi(x)-\psi(-x))/2$ solutions. In particular, for a harmonic potential $V(x)=m \omega^2 x^2/2$, one can find such a solution $$\psi(x)=e^{-\frac{m \omega x^2}{2\hbar}}H\left(\frac{E}{\hbar
\omega}-\frac{1}{2},\sqrt{\frac{m \omega}{\hbar}}x \right),$$ where $H(\nu,x)$ is the Hermite polynomial for integer $\nu$. It follows then that $$\psi_s(0)=2^{\frac{E}{\hbar
\omega}-\frac{1}{2}}\frac{\sqrt{\pi}}{\Gamma(\frac{3}{4}-\frac{E}{\hbar
\omega})}$$ and $$\psi_a'(0)=-2^{\frac{E}{\hbar \omega}}\sqrt{\frac{2\pi\omega m
}{\hbar}}\frac{1}{\Gamma(\frac{1}{4}-\frac{E}{\hbar \omega})}.$$ Moreover, in the thermodynamic limit $L\rightarrow\infty$, $$G(0,0,E)=\frac{1}{2\hbar}\sqrt{\frac{m }{\omega
\hbar}}\frac{\Gamma(\frac{1}{4}-\frac{E}{\hbar
\omega})}{\Gamma(\frac{3}{4}-\frac{E}{\hbar
\omega})}.\label{HarmFSol}$$ The Eq. (\[HarmFSol\]) approaches asymptotically the expression for free fermions, as $\omega \rightarrow 0$ for $E<0$, $$G(0,0,E)\rightarrow \frac{1}{\hbar}\sqrt{\frac{-m }{2
E}}.\label{FreeFSol}$$
[10]{} M. Büttiker, Y. Imry, R. Landauer, and S. Pinhas, Phys. Rev. B **31**, 6207 (1985). C. L. Kane and M. P. A. Fisher, Phys.Rev. Lett. **68**, 1220 (1992). A. Komnik and R. Egger, Phys. Rev. Lett.**80**, 2881 (1997). B. Gao, A. Komnik, R. Egger, D. C. Glattli, and A. Bachtold, Phys. Rev. Lett.**92**, 216804 (2004). M. S. Fuhrer, J. Nygard, L. Shih, M. Forero, Young-Gui Yoon, M. S. C. Mazzoni, Hyoung Joon Choi, Jisoon Ihm, Steven G. Louie, A. Zettl, and Paul L. McEuen, Science **288**, 494 (2000). H. W. Ch. Postma, M. de Jonge, Zhen Yao, and C.Dekker, Phys. Rev. B **62**, 10653 (2000). J. W. Janssen, S. G. Lemay, L. P. Kouwenhoven, and C. Dekker, Phys. Rev. B **65**, 115423 (2002). I. E. Dzyaloshinskii and A. I. Larkin, Zh. Eksp. Toer. Fiz.**65**, 411 (1973) \[So. Phys. JETP **38**, 202 (1974)\]. S. Das Sarma and E. H. Hwang, Phys. Rev. B **54**, 1936 (1996). R. L. Schult, D. G. Ravenhall, and H. W. Wyld, Phys. Rev. B **39**, 5476 (1989). J. P. Carini, J. T. Londergan, K. Mullen, and D. P. Murdock, Phys. Rev. B **46**, 15538 (1992). K. Kazymyrenko, B. Douçot, Phys. Rev. B **71**, 075110 (2005). P. H. Dickinson and S. Doniach, Phys. Rev. B **47**, 11447 (1993). A. H. Castro Neto and F. Guinea, Phys. Rev. Lett.**80**, 4040 (1998). X. J. Zhou, P. Bogdanov, S. A. Kellar, T. Noda, H. Eisaki, S. Uchida, Z.Hussain, and Z.X. Shen, Science **286**, 268 (1999). A. Odintsov, Phys. Rev. Lett.**85**, 150 (2000). A. Odintsov and Hideo Yoshioka, Phys. Rev. B **59**, 10457 (1999). P. Jarillo-Herrero, S. Sapmaz, C. Dekker, L. P. Kouwenhoven, and H. S. J. van der Zant, Nature **427**, 389 (2004). M. Radosavljevic, J. Appenzeller, Ph. Avouris, and J. Knoch, Appl. Phys. Lett. **84**, 3693 (2004). S. G. Lemay, J. W. Janssen, M. van den Hout, M. Mooij, M. J. Bronikowski, P. A. Willis, R. E. Smalley, L. P. Kouwenhoven, and C.Dekker, Nature (London) **412**, 617 (2001). B. J. LeRoy, I. Heller, V. K. Pahilwani, C.Dekker, and S. G. Lemay, Nano Lett. **7**, 2937 (2007).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We report the full complex dielectric function of high-purity $\textrm{Ba}_{0.68}\textrm{K}_{0.32}\textrm{Fe}_2\textrm{As}_2$ single crystals with $T_{\mathrm{c}}=38.5\ \textrm{K}$ determined by wide-band spectroscopic ellipsometry at temperatures $10\leq T\leq300\ \textrm{K}$. We discuss the microscopic origin of superconductivity-induced infrared optical anomalies in the framework of a multiband Eliashberg theory with two distinct superconducting gap energies $2\Delta_{\mathrm{A}}\approx6\ k_{\mathrm{B}}T_{\mathrm{c}}$ and $2\Delta_{\mathrm{B}}\approx2.2\ k_{\mathrm{B}}T_{\mathrm{c}}$. The observed unusual suppression of the optical conductivity in the superconducting state at energies up to $14\ k_{\mathrm{B}}T_{\mathrm{c}}$ can be ascribed to spin-fluctuation–assisted processes in the clean limit of the strong-coupling regime.'
author:
- 'A. Charnukha'
- 'O. V. Dolgov'
- 'A. A. Golubov'
- 'Y. Matiks'
- 'D. L. Sun'
- 'C. T. Lin'
- 'B. Keimer'
- 'A. V. Boris'
title: 'Eliashberg approach to superconductivity-induced infrared anomalies in $\bf\textrm{Ba}_{0.68}\textrm{K}_{0.32}\textrm{Fe}_2\textrm{As}_2$'
---
The discovery of iron-based superconductors [@kamihara] has generated significant experimental and theoretical effort to unravel the mechanism of high-temperature superconductivity in these compounds. This effort has yielded a comprehensive experimental description of the electronic structure at the Fermi level, which includes multiple Fermi surface sheets in a good agreement with density functional calculations [@Paglione_review_2010; @Johnston_Review_2010]. Partial nesting between at least two of these sheets leads to a spin-density-wave instability that renders the metallic parent compounds antiferromagnetic. In the superconducting compounds spin fluctuations become the source of strong repulsive interband interactions and might give rise to superconductivity with different signs on these sheets [@Mazin_NatureInsights_2010].
The most incisive experimental data have been obtained on high-quality single-crystals of iron pnictides with the so-called 122 structure, for instance $\textrm{BaFe}_2\textrm{As}_2$, with K substituted for Ba or Co for Fe, resulting in hole and electron doping, respectively. In all of these materials five Fermi surface sheets have been identified in calculations and confirmed by numerous independent experimental studies [@Paglione_review_2010; @Johnston_Review_2010]: in the reduced Brillouin-zone scheme these are three hole pockets at the $\Gamma$ point and two almost degenerate electron pockets at the $X$ point with nesting between hole and electron sheets. Among all of these $122$ materials, the optimally hole-doped compound $\textrm{Ba}_{0.68}\textrm{K}_{0.32}\textrm{Fe}_2\textrm{As}_2$ (BKFA) has the highest transition temperature of $38.5\ \textrm{K}$. Due to their exceptional quality, crystals of this compound are well suited as a testbed for theoretical models. A four-band Eliashberg theory with strong interband couping has already proven successful in accounting for the transition temperature, as well as the temperature dependence of the free energy and superconducting gaps of this compound [@PhysRevLett.105.027003]. This analysis has made clear that a satisfactory description of the bulk thermodynamical properties in the superconducitng state can only be obtained via strong coupling to spin fluctuations or other bosons with spectral weight below $50\ \textrm{meV}$.
![\[fig:firexperiment\] Real part of the (a) optical conductivity and (b) dielectric function in the far-infrared spectral region. Two characteristic superconductivity energy scales are present: $6\ k_{\mathrm{B}}T_{\mathrm{c}}$ and $14\ k_{\mathrm{B}}T_{\mathrm{c}}$.](Fig1.pdf){width="3.4in"}
In this Letter we extend the approach of Ref. to describe the far-infrared properties of the same BKFA single crystal. We show that major characteristic features of superconductivity can be explained within a strong-coupling Eliashberg approach with two distinct values of the superconducting energy gap $2\Delta_{\mathrm{A}}\approx6\ k_{\mathrm{B}}T_{\mathrm{c}}$ and $2\Delta_{\mathrm{B}}\approx2.2\ k_{\mathrm{B}}T_{\mathrm{c}}$, in quantitative agreement with angle-resolved photoemission [@PhysRevB.79.054517; @PhysRevLett.105.117003; @PhysRevB.83.020501], scanning-tunneling microscopy [@Shan-Wen_STM_2011] and specific-heat measurements [@PhysRevLett.105.027003]. We also demonstrate that within this approach the qualitative differences in the infrared spectra of electron- and hole-doped 122 compounds are reproduced by strong-coupling calculations in the clean and dirty limits (weak and strong impurity scattering), respectively.
The optimally-doped BKFA single crystals were grown in zirconia crucibles sealed in quartz ampoules under argon atmosphere [@Lin_BKFA_growth_2010]. From DC resistivity, magnetization and specific-heat measurements we obtained $T_{\mathrm{c}}=38.5\pm0.2\ \textrm{K}$. The sample surface was cleaved prior to every optical measurement. The full complex dielectric function $\varepsilon(\omega)$ was obtained in the range $0.01-6.5\ \textrm{eV}$ using broadband ellipsometry, as described in Ref. [@boris:027001]. In this work we focus on the itinerant charge carrier response contained within the far-infrared spectral range measured at the infrared beamline of the ANKA synchrotron light source at Karlsruhe Institute of Technology, Germany. The contribution of the interband transitions has been eliminated based on a dispersion analysis in the entire spectral range [@supplementary.material.prl].
![\[fig:scattering\] Optical scattering rate obtained from the experimental data at $41$ and $10\ \textrm{K}$ within the extended-Drude model, with the contribution of the interband transitions subtracted (blue and black lines, respectively) and at $41\ \textrm{K}$ without subtraction (green line). Dash-dotted line indicates the saturation level of the high-energy optical scattering rate.](Fig2.pdf){width="3.4in"}
The optical response of BKFA in the far-infrared spectral range is shown in Fig. \[fig:firexperiment\](a, b) respectively for the real parts of optical conductivity $\sigma(\omega)=\sigma_1(\omega)+i\sigma_2(\omega)$ and dielectric function $\varepsilon(\omega)=1+4\pi i\sigma(\omega)/\omega$. It is dominated by the contribution of the itinerant charge carriers manifested in negative values of $\varepsilon_1(\omega)$. Figure \[fig:firexperiment\](a) also reveals a peak around $15-20\ \textrm{meV}$ in $\sigma_1(\omega)$ with a concomitant upturn in $\varepsilon_1(\omega)$ at higher temperatures indicating the presence of a collective excitation. The superconducting transition is accompanied by the suppression of the optical conductivity up to $50\ \textrm{meV}$ ($14k_{\mathrm{B}}T_{\mathrm{c}}$). The corresponding missing area in $\sigma_1(\omega)$ between $41$ and $10\ \textrm{K}$, $\int_{0^+}^{6\Delta_{\mathrm{A}}}\Delta\sigma_1(\omega)d\omega=(8\lambda_{\mathrm{L}}^2)^{-1}$, manifests itself as a characteristic $-1/(\lambda_{\mathrm{L}}\omega)^2$ contribution to $\varepsilon_1(\omega)$ in Fig. \[fig:firexperiment\](b) in the superconducting state. The London penetration depth $\lambda_{\mathrm{L}}=2200$ Å extracted from these data is consistent with other measurements [@PhysRevLett.101.107004]. At energies close to the optical superconducting gap $2\Delta_{\mathrm{A}}\approx20\ \textrm{meV}$ one of the directly measured ellipsometric angles $\Psi(\omega)$ approaches its critical value of $45^\circ$ at the superconducting transition, which implies that the reflectivity of the sample approaches unity and its optical conductivity $\sigma_1(\omega)$ is close to zero. Remarkably, Fig. \[fig:firexperiment\](a) shows a quasilinear dependence of $\sigma_1(\omega)$ in the superconducting state from $2\Delta_{\mathrm{A}}$ to as high as $14k_{\mathrm{B}}T_{\mathrm{c}}$, in a stark contrast to the electron-doped 122 compounds [@PhysRevB.82.174509; @PhysRevB.81.214508; @PhysRevB.82.100506]. In the latter the optical conductivity at $2\Delta_{\mathrm{A}}$ decreases abruptly upon cooling below $T_{\mathrm{c}}$, but only a weak superconductivity-induced modification is observed at higher energies. The quasilinear behavior in BKFA cannot be reconciled with the widely used for pnictides Mattis-Bardeen theory [@Zimmermann199199], a weak-coupling extension of the BCS theory to finite impurity scattering. As all optimally-doped 122 pnictide superconductors appear to be in the strong-coupling regime, the Eliashberg theory [@PhysRev.156.470] has to be used in order to obtain an adequate description of the optical properties.
![\[fig:zerocrossing\] (a) Real part of the optical conductivity at 41K (blue line). The contribution of itinerant charge carriers (blue area) is obtained by subtracting all interband transitions $\sigma^{\mathrm{inter}}_1(\omega)$ (gray area) from the optical response. (b) Real part of the dielectric function at 41K (blue line). The free-charge-carrier response $\varepsilon^{\mathrm{it}}_1(\omega)$ (open circles) is obtained by eliminating all interband transitions $\varepsilon^{\mathrm{inter}}_1(\omega)$ (black solid line). The blue dashed line indicates the screened plasma frequency at 41K.](Fig3.pdf){width="3.4in"}
Signatures of a boson pairing mediator of the Eliashberg theory come from a qualitative analysis of the optical conductivity within the extended Drude model. It implies that the optical scattering rate is related to the far-infrared optical response as $\gamma(\omega)=\textrm{Re}[\omega_{\mathrm{pl}}^2/4\pi\sigma^{\mathrm{it}}(\omega)]$, where the superscript ’it‘ stands for ’itinerant‘ and implies that the contribution of all interband transitions [*must*]{} be subtracted from the experimentally obtained optical conductivity. The optical scattering rate shows clear evidence of an intermediate boson irrespective of complications due to the multiband character of the compound. Figure \[fig:scattering\] plots $\gamma(\omega)$ of BKFA at $41\ \textrm{K}$ (blue line) and $10\ \textrm{K}$ (black line) for $\omega_{\mathrm{pl}}=[8\textrm{SW}_{\mathrm{it}}]^{1/2}=[8\int_{0}^{\infty}\sigma_1^{\mathrm{it}}(\omega)d\omega]^{1/2}=1.6\ \textrm{eV}$ with all interband transitions subtracted in both cases ($\textrm{SW}_{\mathrm{it}}$ corresponds to the blue shaded area in Fig. \[fig:zerocrossing\](a)). In the superconducting state, no scattering is expected up to photon energies exceeding the binding energy of the Cooper pairs. Thus the onset of the optical scattering rate marks the optical energy gap $2\Delta_{\mathrm{A}}=20\ \textrm{meV}$. Saturation of $\gamma(\omega>50\ \textrm{meV})$ at $1100\ \textrm{cm}^{-1}$ indicates that the boson spectral function is contained well below $50\ \textrm{meV}$ [@Shulga1991266]. It is important to emphasize that, due to the multiband character of the iron pnictides, an analysis of the optical scattering rate in the framework of a single-band Eliashberg theory is potentially misleading. Moreover, also shown in Fig. \[fig:scattering\] is a spectrum that directly results from the experimental data, without accounting for the interband transitions. It becomes clear that an increase in the scattering rate at higher energies that might be ascribed to strong electron correlations can result from an unsubtracted contribution of the interband transitions to the complex optical conductivity. This is especially important in iron pnictides since the lowest lying interband transition at about $0.5\ \textrm{eV}$ contributes to an anomalously large value of the low-energy dielectric permittivity $\varepsilon_\infty$ [@footinbib_fir2011] due to the high polarizability of the Fe-As bonds [@2010arXiv1009.5915C]. In order to reconcile the bare plasma frequency of $1.6\ \textrm{eV}$ (see Fig. \[fig:zerocrossing\](a)) with the zero-crossing in $\varepsilon_1(\omega)$ at $0.2\ \textrm{eV}$ (blue line in Fig. \[fig:zerocrossing\](b)) $\varepsilon_\infty$ has to be as large as 60, consistent with the contribution of the interband transitions $\varepsilon^{\mathrm{inter}}_1(\omega)$ determined by means of the dispersion analysis, as shown in Fig. \[fig:zerocrossing\](b). Such $\varepsilon_\infty$ is thus an order of magnitude larger than in any other high-temperature superconductor (e. g. $\approx5$ in cuprates [@A.V.Boris04302004]). Recently, a similarly high value in a conventional superconductor was inferred from reflectivity measurements on elementary bismuth [@PhysRevLett.104.237401].
To determine the microscopic origin of the high-energy anomaly $2\Delta_{\mathrm{A}}<\hbar\omega<14k_{\mathrm{B}}T_{\mathrm{c}}$ in the real part of the optical conductivity in Fig. \[fig:firexperiment\](a) we use a four-band Eliashberg theory that proved successful in explaining thermodynamical data obtained on the same compound [@PhysRevLett.105.027003]. The $4\times4$ matrix of coupling constants and 4 densities of states characterizing this model are highly constrained by thermodynamic, transport and photoemission data [@PhysRevLett.105.027003; @PhysRevB.79.054517; @PhysRevLett.105.117003; @PhysRevB.83.020501; @Shan-Wen_STM_2011], and the same set of parameters is used here. In principle, an additional set of 4 plasma frequencies and a $4\times4$ matrix of intraband/interband impurity scattering rates has to be taken into account to describe the optical response. However, this parameter set can be strongly reduced based on the following considerations.
A substantial simplification is made possible by a projection of the four-band model onto an effective two band model motivated by the observation of two distinct groups of superconducting energy gaps in a variety of experiments [@PhysRevLett.105.027003; @PhysRevB.79.054517; @PhysRevLett.105.117003; @PhysRevB.83.020501; @Shan-Wen_STM_2011]. These gaps can be identified as a single gap $\Delta_B$ on the outer hole-like Fermi surface and a group of three gaps of magnitude $\sim\Delta_A$ on the inner hole-like and the two electron-like Fermi surfaces. Minimizing the ground-state energy subject to this grouping constraint yields an effective two-band model [@supplementary.material.prl]. Furthermore, as the superconducting transition temperature of BKFA appears to be only weakly correlated with the residual resistivity (which is a measure of the impurity scattering), off-diagonal elements of the impurity scattering matrix can be neglected (see Table S1 in Ref. ).
Given the boson spectrum centered at $13\ \textrm{meV}$ (see supplementary online material in Ref. ) consistent with the energy of the spin resonance excitation in this compound [@Osborn_INS_BKFA_2008; @footinbib_SF2011] one obtains the following two-band model coupling matrix: $\lambda_{\mathrm{AA}}=4.36,\ \lambda_{\mathrm{BB}}=0.2,\ \lambda_{\mathrm{AB}}=-0.35,\ \lambda_{\mathrm{BA}}=-0.5$, with the fractional density of states being $N_{\mathrm{A}}/(N_{\mathrm{A}}+N_{\mathrm{B}})=0.59$ [@supplementary.material.prl]. The first effective intraband coupling constant is an order of magnitude larger than predicted for the intraband electron-phonon coupling [@boeri:026403]. It does not, however, bear any physical meaning by itself but rather incorporates contributions from [*three*]{} different bands. We reiterate that the coupling matrix has been inferred from prior measurements. In this way, only two intraband impurity scattering rates enter as free parameters of the theory in addition to the plasma frequencies of the bands.
![\[fig:eliashberg\_sigma\] (a) Real part of the far-infrared conductivity obtained within the two-band Eliashberg theory (see text) at $40\ \textrm{K}$ (blue lines) and $10\ \textrm{K}$ (black lines) in the clean limit $\gamma_{\mathrm{A}}=\gamma_{\mathrm{B}}=1\ \textrm{cm}^{-1}$ (solid lines) and dirty limit $\gamma_{\mathrm{A}}=\gamma_{\mathrm{B}}=200\ \textrm{cm}^{-1}$ (dashed lines). (b) Optical scattering rate in the clean limit from the same model. The gray area shows the normalized boson spectral function $B(\omega)$ used in the calculation, displaced from zero by $2\Delta_{\mathrm{A}}$ to assist interpretation in the superconducting state [@footinbib_SF2011].](Fig4.pdf){width="3.4in"}
In our calculation we consider two clean bands with $\gamma_{\mathrm{A}}=\gamma_{\mathrm{B}}=1\ \textrm{cm}^{-1}$. As the bare plasma frequencies of all bands are similar it follows that the spectral weight of band $\mathrm{A}$ has to be much larger than that of band $\mathrm{B}$. Assigning $80\%$ of the spectral weight to the effective band we obtain the results presented as solid lines in Fig. \[fig:eliashberg\_sigma\](a). The high-energy anomaly at $14k_{\mathrm{B}}T_{\mathrm{c}}$ is naturally captured by the model without resorting to additional gaps. Its energy is given by $2\Delta_{\mathrm{A}}+\Omega$, where $\Omega$ is the characteristic frequency of the boson spectrum, as shown in Fig. \[fig:eliashberg\_sigma\](b) (gray shaded area, displaced from zero by $2\Delta_{\mathrm{A}}$). This calculation also accounts for the fact that only the biggest superconducting gap is visible in the optical responce of BKFA due to a small contribution of band B ($20\%$ of the spectral weight) to the overall optical conductivity. This leads to two possible levels of the impurity scattering rate of band B, which has to be either very small $\gamma_{\mathrm{B}}\approx1\ \textrm{cm}^{-1}$ or very large at about $1000\ \textrm{cm}^{-1}$. The latter value provides a better description of the optical scattering rate (see interactive simulation in [@supplementary.material.prl]) and DC transport [@2010arXiv1011.1900G]. However, such a large disparity between the charge carriers is hard to reconcile with the Hall and de Haas–van Alphen experiments, which imply that the impurity scattering rate of the holes is no more than one order of magnitude higher than that of the electrons [@PhysRevB.80.140508; @PhysRevLett.101.216402]. This residual uncertainty notwithstanding, our results show that the impurity scattering rate of band A must be very small, because the energy $2\Delta_{\mathrm{A}}+\Omega$ is no longer discernible in the simulated spectra when $\gamma_{\mathrm{A}}$ increases (interactive simulation in [@supplementary.material.prl]). The region of linear increase of $\sigma_1(\omega)$ is related to the linear segment in the boson spectrum and can only be observed in a very clean material.
The same reduced two-band model can be applied to the case of $\textrm{BaFe}_{1.85}\textrm{Co}_{0.15}\textrm{As}_2$ (BFCA). In this compound the spin resonance excitation occurs at a very similar energy of $10\ \textrm{meV}$ [@Inosov_BFCA_2009]. A boson spectrum centered at this energy is also consistent with Andreev-reflection measurements [@PhysRevLett.105.237002]. Recently, a comprehensive specific-heat study of this compound at different Co-doping levels has been carried out [@0295-5075-91-4-47008]. The analysis of the experimental data in the framework of the two-band $\alpha$-model indicates that the largest gap develops in the band with the largest electronic density of states, providing further evidence that several bands contribute to the strongly-coupled band in the reduced two-band model. Figure \[fig:eliashberg\_sigma\](a) (dashed lines) shows that a calculation within the same reduced two-band model qualitatively reproduces the far-infrared optical conductivity of BFCA [@PhysRevB.82.174509; @PhysRevB.81.214508; @PhysRevB.82.100506] when both bands are assumed to be dirty with $\gamma_{\mathrm{A}}=\gamma_{\mathrm{B}}=200\ \textrm{cm}^{-1}$ and a redistribution of the spectral weight between the bands is taken into account as $\omega^2_{\mathrm{pl,A}}\approx\omega^2_{\mathrm{pl,B}}$. The model captures the two prominent superconductivity-induced anomalies clearly observed in experiments: the steep onset of absorption at the value of the small gap $2\Delta_{\mathrm{B}}$ and the weaker superconductivity-induced changes of the optical conductivity extending up to $18k_{\mathrm{B}}T_{\mathrm{c}}$. The redistribution of the spectral weight between the bands in BFCA compared to BKFA implied by our analysis is justified by doping with different carriers in the two compounds, while the large difference in the their impurity scattering rates is a natural consequence of the difference in doping mechanisms by chemical substitution, which directly affects the FeAs layers in BFCA, but not in BKFA.
In summary, a qualitative description of superconductivity-induced optical anomalies in the far-infrared optical conductivity of $\textrm{Ba}_{0.68}\textrm{K}_{0.32}\textrm{Fe}_2\textrm{As}_2$ is obtained in the framework of an effective two-band Eliashberg theory with a strong coupling to spin fluctuations reduced from its four-band counterpart. The linear increase of absorption above the larger superconducting gap can only be observed when the effective band is extremely clean. The same model in the dirty limit provides a good qualitative explanation of the optical conductivity of the optimally electron-doped BFCA consistently in the strong-coupling regime.
This project was supported by the German Science Foundation under grant BO 3537/1-1 within SPP 1458. We gratefully acknowledge Y.-L. Mathis for support at the infrared beamline of the synchrotron facility ANKA at the Karlsruhe Institute of Technology and P. Popovich for taking part in some of the measurements.
[10]{}
Y. Kamihara [*et al.*]{}, , 3296 (2008).
J. Paglione and R. L. Greene, , 645 (2010).
D. C. Johnston, , 803 (2010).
I. I. Mazin, , 183 (2010).
P. Popovich [*et al.*]{}, , 027003 (2010).
D. V. Evtushinsky [*et al.*]{}, , 054517 (2009).
Y. Zhang [*et al.*]{}, , 117003 (2010).
K. Nakayama [*et al.*]{}, , 020501 (2011).
L. Shan [*et al.*]{}, , 060510 (2011).
G. L. Sun [*et al.*]{}, , 1, doi:10.1007/s10948-010-1123-z (2011).
A. V. Boris [*et al.*]{}, , 027001 (2009).
.
G. Li [*et al.*]{}, , 107004 (2008).
J. J. Tu [*et al.*]{}, , 174509 (2010).
K. W. Kim [*et al.*]{}, , 214508 (2010).
R. P. S. M. Lobo [*et al.*]{}, , 100506 (2010).
W. Zimmermann [*et al.*]{}, , 99 (1991).
S. B. Nam, , 470 (1967).
S. V. Shulga, O. V. Dolgov, and E. G. Maksimov, , 266 (1991).
A. Charnukha [*et al.*]{}, , 219, doi:10.1038/ncomms1223 (2011).
A. V. Boris [*et al.*]{}, , 708 (2004).
N. P. Armitage [*et al.*]{}, , 237401 (2010).
A. D. Christianson [*et al.*]{}, , 930 (2008).
L. Boeri, O. V. Dolgov, and A. A. Golubov, , 026403 (2008).
A. A. Golubov [*et al.*]{}, .
L. Fang [*et al.*]{}, , 140508 (2009).
A. I. Coldea [*et al.*]{}, , 216402 (2008).
D. S. Inosov [*et al.*]{}, , 178 (2009).
M. Tortello [*et al.*]{}, , 237002 (2010).
F. Hardy [*et al.*]{}, , 47008 (2010).
[**Supplementary online material for the article\
Eliashberg approach to superconductivity-induced infrared anomalies in $\bf\textrm{Ba}_{0.68}\textrm{K}_{0.32}\textrm{Fe}_2\textrm{As}_2$**]{} 0.17in A. Charnukha$^1$, O. V. Dolgov$^1$, A. A. Golubov$^2$, Y. Matiks$^1$, D. L. Sun$^1$, C. T. Lin$^1$, B. Keimer$^1$, and A. V. Boris$^1$ 0.05in [*$^\mathit{1}$Max-Planck-Institut für Festkörperforschung, Heisenbergstrasse 1, D-70569 Stuttgart, Germany*]{}\
[*$^\mathit{2}$Faculty of Science and Technology and MESA+ Institute of Nanotechnology, 7500 AE Enschede, The Netherlands*]{}
Dispersion analysis
===================
The full complex dielectric function $\varepsilon(\omega)=\varepsilon_1(\omega)+i\varepsilon_2(\omega)$ obtained experimentally in the range from $12\ \textrm{meV}$ to $6.7\ \textrm{eV}$ was analyzed in the Drude-Lorentz model:$$\varepsilon(\omega)=1-\frac{\omega_{\mathrm{pl}}^2}{\omega^2+i\gamma\omega}+\sum_{j=1}^{n}\frac{\Delta\varepsilon_j\omega_{0j}^2}{(\omega_{0j}^2-\omega^2)-i\Gamma_j\omega}\nonumber,$$where $(\omega_{\mathrm{pl}},\ \gamma)$ are the plasma frequency renormalized by interaction with the mediating boson and optical scattering rate, and $(\Delta\varepsilon_j,\omega_{0j},\ \Gamma_j)$ are the DC permittivity contribution, center frequency and the width of the Lorentzian oscillators used to model the interband transitions, respectively. The results of this analysis are presented in Fig. \[fig:dispanalysis\] for $10\ \textrm{K}$ (blue line - experimental data, black lines - separate Lorentz contributions). To display the scale of the temperature-induced variation of the interband transitions the experimental spectrum at $300\ \textrm{K}$ is also shown (red line). The lowest interband transition in this material lies around $0.5\ \textrm{eV}$ and significantly contributes to the AC polarizability of the system, as is evident from Fig. 3(b) of the main text. The residual optical response (open circles and inset in Fig. 3(b)) was studied after the subtraction of all thus determined interband transitions down to $0.5\ \textrm{eV}$. An unscreened bare plasma frequency of $1.6\ \textrm{eV}$ at $41\ \textrm{K}$ was consistently obtained from the spectral weight of the residual response $\textrm{SW}=\int_{0}^{\infty}\sigma_1^{\mathrm{it}}(\omega)d\omega$ as $\omega_{\mathrm{pl}}=\sqrt{8\textrm{SW}}$ and a simultaneous fit of the real and imaginary parts of the dielectric function at high energies.
![\[fig:dispanalysis\]Real part of the (a) optical conductivity and (b) dielectric function (b) at 300K (red line) and 10K (blue line). Interband transitions inferred from the dispersion analysis (gray lines). The shaded region is enlarged in Fig. 3(b) of the main text.](FigS1.pdf){width="\columnwidth"}
Effective two-band model for pnictides
======================================
Multiband Eliashberg formalism
------------------------------
The Eliashberg theory of superconductivity \[S,S\] extended to the multiband case [@PhysRevB.72.024504] has already proven successful in describing the thermodynamical properties of iron pnictides [@PhysRevLett.105.027003s]. Here we employed the same formalism to account for the far-infrared optical response of these compounds. The main ingredient of the theory is the total spectral function of the electron-boson interaction $B(\omega)$ (Eliashberg function; analogous to that of the electron-phonon interaction $\alpha ^{2}F(\omega)$). In a four band system it can be decomposed into 16 functions $B(\omega) _{ij}$, where $i$ and $j$ label the four Fermi surface sheets ($i,j=1,2,3,4$). The *standard* Eliashberg functions determine the superconducting and thermodynamical properties such as the superconducting transition temperature and gaps, electronic specifc heat, de Haas-van Alphen mass renormalizations etc. and are defined as $$B(\omega)_{ij}=\frac{1}{N_{i}}\sum_{{\bf k,k}^{\prime },\nu }\left| g_{{\bf k,k}^{\prime }}^{ij,\nu }\right| ^{2}\delta
(\varepsilon _{{\bf k}}^{i})\delta (\varepsilon _{{\bf k^{\prime }}}^{j})\delta (\omega -\omega _{{\bf k-k^{\prime }}}^{\nu }),\nonumber$$ where $N_{i}$ is the partial density of states per spin on the $i$’th sheet of the Fermi surface, and $g_{{\bf k,k}^{\prime
}}^{ij}$ is the matrix element of electron-boson interactions.
Transport and electrodynamical properties are defined by 16 [*transport*]{} Eliashberg functions (which enter the Boltzmann kinetic equation) $$\begin{aligned}
B(\omega)_{tr\;ij}^{\alpha\beta}&=&\frac{1}{2N_{i}\left\langle v_{Fi}^{\alpha \text{ }2}\right\rangle }\sum_{{\bf k,k}^{\prime },\nu }\left| g_{{\bf k,k}^{\prime }}^{i,j,\nu }\right|
^{2}
\nonumber\\
&&\hskip-25pt\times(v_{Fi}^{\alpha }({\bf k})-v_{Fj}^{\beta }({\bf k}^{\prime }))^{2}\delta (\varepsilon _{{\bf k}}^{i})\delta (\varepsilon _{{\bf k}^{\prime }}^{j})\delta (\omega -\omega _{{\bf k-k}^{\prime }}^{\nu }),
\nonumber\end{aligned}$$ where $v_{Fi}^{\alpha }$ is the $\alpha $-th Cartesian component of the Fermi velocity on Fermi surface $i$. The average Fermi velocity is related to the plasma frequency by the standard expression $\omega _{pl\;i}^{2}=8\pi e^{2}N_{i}\left\langle v_{Fi}^{2}\right\rangle
=8\pi e^{2}\sum_{{\bf k}}v_{Fi}^{2}({\bf k})\delta (\varepsilon _{{\bf k}}^{i}).$ All Eliashberg functions satisfy the symmetry relations $M_{i}B _{ij}=M_{j}B _{ji},$ where $M_{i}=N_{i}$ and $M_{i}=\omega _{pl\,i}^{2}$ for the standard and transport Eliashberg functions, respectively.
Role of impurities and defects
------------------------------
Both normal and superconducting properties of a multiband superconductor significantly depend on impurity scattering. Unlike in conventional superconductivity, one has to distinguish between the intraband impurity scattering, which does not add any new physics (in the Born approximation) compared with single-band superconductivity, and *interband* scattering, which in many cases has an effect comparable to the pair-breaking effect of magnetic impurities (or of nonmagnetic impurities in superconductors with $p$- or $d$-wave pairing) [@PhysRevB.55.15146]. In this regard, the fact that no strong correlation has been observed between the residual resistivity (which indirectly characterizes the impurity scattering) and the critical temperature $T_{c}$ of the (nearly) optimally electron-doped BKFA (see Table \[table:residual\]) indicates that the level of *interband* impurities in the Born limit is very small. Thus one only needs to estimate the intraband scattering rates $\gamma_{\mathrm{A}},\ \gamma_{\mathrm{B}}$.
$T_{\textrm{c}}$, K Residual resistivity, $\textrm{m}\Omega\ \textrm{cm}$ Reference
--------------------- ------------------------------------------------------- ----------------------------
38.5 0.04 [@PhysRevLett.105.027003s]
38 0.075 [@PhysRevB.78.224512]
38 0.1 [@PhysRevLett.101.107006]
36.5 0.055 [@PhysRevB.79.174501]
: \[table:residual\]Superconducting transition temperature and residual resistivity $\rho_{40\textrm{K}}$ of (nearly) optimally hole-doped BKFA.
Theoretical model
-----------------
As a starting point we consider a $4-$band model based on the band-structure calculations with two hole bands and two electron bands crossing the Fermi level that has proven successful in accounting for the thermodynamical properties of BKFA [@PhysRevLett.105.027003s]. We use the same input parameters, namely, the densities of states $N_1=22\ \textrm{Ry-st}^{-1}$, $N_2=25\ \textrm{Ry-st}^{-1}$, and $N_3=N_4=7\ \textrm{Ry-st}^{-1}$, the first two having a hole while the other two an electron character. The main input, the spectral function of the intermediate boson, was taken following Ref. S in the form of a spin-fluctuation spectrum $\tilde{B}_{ij}(\Omega )=\lambda _{ij}f(\Omega /\Omega _{sf})$ with a linear $\omega$ dependence at low frequencies. Here $\lambda_{ij}$ is the coupling constant pairing band $i$ with band $j$ and $\Omega _{SF}$ is a characteristic spin-fluctuation frequency, the values of which correspond to those in Ref. : $\Omega_{sf}=13\ \textrm{meV}$ and$$\lambda _{ij}=\left(
\begin{array}{cccc}
0.2 & 0 & -1.7 & -1.7 \\
0 & 0.2 & -0.25 & -0.25 \\
-5.34 & -0.89 & 0.2 & 0 \\
-5.34 & -0.89 & 0 & 0.2\end{array}\right).\label{l4x4}$$Negative elements correspond to *interband* hole-electron repulsion, while the positive — to *intraband* attraction.
In order to apply this full 4-band model to description of the transport properties one has to take into account an additional set of 4 plasma frequencies and a 4x4 matrix of impurity scattering rates. The latter are difficult to determine theoretically and thus would have to be treated as free paremeters of the model. It would render the problem highly overparametrized. On the other hand, in the case of BKFA it is known that three larger gaps have approximately the same value $|\Delta_{\mathrm{A}}|\equiv|\Delta_1|\approx|\Delta_3|\approx|\Delta_4|\approx9\ \textrm{meV}$, while the smaller gap is $|\Delta_{\mathrm{B}}|\equiv|\Delta_2|\approx3\ \textrm{meV}$. One can assume this restriction exact and introduce it into the theory thus reducing the original 4-band model to a more tractable 2-band model as explained in the following section.
Reduction to a two-band model
-----------------------------
In general, the superconducting order parameters are a solution of a linear system of equations $$e_{i}=\sum_{j=1}^{4}B_{ij}(\omega)e_{j}.\label{eq:gapsystem}$$ The Eliashberg functions $B_{ij}(\omega)$ satisfy the symmetry relations $$N_iB_{ij}(\omega )=N_jB_{ji}(\omega)\label{eq:symmetryrelations}$$and, therefore, can be represented in the form $B_{ij}(\omega)=U_{ij}(\omega)N_j$, where $U_{ij}$ is a symmetrical matrix. Further, we can construct a functional $${\mathfrak{F}}\{e_i\}=\sum_{j=1}^{4}N_je_j^2-\sum_{i,j=1}^4N_ie_iU_{ij}N_je_j. \label{eq:functional}$$Equation (\[eq:gapsystem\]) then results from minimization of ${\mathfrak{F}}$ with respect to $e_i$. As mentioned above, BKFA has three gaps with very close absolute values (the first hole gap has the opposite sign with respect to the other two).
[ Real part of the optical conductivity (top left) and the optical scattering rate (top right) of $\textrm{Ba}_{0.68}\textrm{K}_{0.32}\textrm{Fe}_2\textrm{As}_2$ in the superconducting state at $10\ \textrm{K}$ (blue lines) and the normal state at $40\ \textrm{K}$ (red) as obtained experimentally (heavy) and from the effective two-band Eliashberg theory (thin). (bottom left) The spin-fluctuation spectrum used in the simulation in both the superconducting and the normal state. (bottom right) Interactive controls: adjust the sliders to select the physical parameters of the model.](FigS4.pdf "fig:"){width="\columnwidth"}](http://www.fkf.mpg.de/keimer/groups/optical/bkfafir.jar)
Minimizing functional Eq. (\[eq:functional\]) subject to the additional constraints $e_3=e_4=-e_1=\Delta
_{\mathrm{A}}$, and $e_2=-\Delta _{\mathrm{B}}$ one finds $$\left(
\begin{array}{c}
\Delta _{\mathrm{A}} \\
\Delta _{\mathrm{B}}\end{array}\right) =\left(
\begin{array}{cc}
\lambda_{AA}&\lambda_{AB}\\
\lambda_{BA}&\lambda_{BB}\end{array}\right)\left(
\begin{array}{c}
\Delta_{\mathrm{A}} \\
\Delta_{\mathrm{B}}\end{array}\right),$$where the matrix elements satisfy $$\begin{aligned}
\lambda_{AA}&=&\frac{N_1\left( \lambda_{11}-2\lambda_{13}-2\lambda_{14}\right) +N_3\lambda_{33}{+}N_4\lambda_44}{N_1+N_3+N_4}, \\
\lambda_{AB} &=&\frac{N_2\left(\lambda_{23}+\lambda_{24}\right)}{N_1+N_3+N_4}, \\
\lambda_{BA}&=&\lambda_{23}+\lambda_{24}, \\
\lambda_{BB}&=&\lambda_{22}.\end{aligned}$$Assuming the matrix elements Eq. (\[l4x4\]), the following coupling constants of the reduced 2-band model are obtained: $$\lambda_{IJ}=\left(
\begin{array}{cc}
4.36&-0.35 \\
-0.5&0.2\end{array}\right) ,\text{ }I,J=\{A,B\}.\label{l2x2}$$The partial densities of states on the Fermi level of effective band $A$ and band $B$ are$$\begin{aligned}
N_{\mathrm{A}}&=&N_1+N_3+N_4=36\ \textrm{Ry-st}^{-1},\nonumber\\ N_{\mathrm{B}}&=&N_2=25\ \textrm{Ry-st}^{-1}\label{dos2b}.\end{aligned}$$Interestingly, even though the main interactions in the 4-band model with the coupling constants Eq. (\[l4x4\]) come from the nondiagonal elements, in the reduced 2-band counterpart they are incorporated into the effective *intraband* $\lambda_{\mathrm{AA}}$ matrix element. Figure \[fig:interactive\] presents an interactive simulation of this effective two-band model (click on the figure to run the simulation and adjust the sliders to set the physical parameters of the system).
Verification of the $2\times 2$ model
-------------------------------------
The effective 2-band model closely reproduces all the predictions of the 4-band model such as the superconducting transition temperature $T_{\mathrm{c}}=38.4\ \textrm{K}$, superconducting gaps $\Delta_{\mathrm{A}}=9.7\ \textrm{meV}$ and $\Delta_{\mathrm{B}}=3.7\ \textrm{meV}$, free energy and superconducting gaps as functions of temperature, as shown in Figs. \[fig:verification\](a) and (b), respectively. The calculated densities of states $N_{\mathrm{A}}$ and $N_{\mathrm{B}}$ are very similar, in accordance with the partial Sommerfeld constants obtained in the treatment of the specific heat data in a phenomenological two-band $\alpha$-model [@PhysRevLett.105.027003s]: $\gamma_{\mathrm{A}}\simeq\gamma_{\mathrm{B}}$ (the Sommerfeld constant $\gamma$ is related to the density of states $N$ via $\gamma =\frac{2\pi}{3}N$).
![\[fig:verification\]Temperature depencence of the (a) free energy and (b) superconducting gaps the 4-band (lines) and reduced 2-band (symbol) models, with coupling matrices Eqs. (\[l4x4\],\[l2x2\]), respectively.](FigS3.pdf){width="\columnwidth"}
[1]{}
S. B. Nam, , 470 (1967).
S. B. Nam, , 487 (1967).
O. V. Dolgov [*et al.*]{}, , 024504 (2005).
P. Popovich [*et al.*]{}, , 027003 (2010).
A. A. Golubov and I. I. Mazin, , 15146 (1997).
M. Rotter, M. Tegel, and D. Johrendt, , 107006 (2008).
G. F. Chen [*et al.*]{}, , 224512 (2008).
G. Mu [*et al.*]{}, , 174501 (2009).
D. Parker [*et al.*]{}, , 134524 (2008).
|
{
"pile_set_name": "ArXiv"
}
|
=1200
**SPECTRAL FLOW AND LEVEL SPACING OF EDGE STATES**
0.1cm
**FOR QUANTUM HALL HAMILTONIANS**
Nicolas Macris
0.25cm
Institute for Theoretical Physics
Ecole Polytechnique Fédérale de Lausanne
CH-1015 Lausanne, Switzerland
**Abstract**
0.25cm
We consider a non relativistic particle on the surface of a semi-infinite cylinder of circumference $L$ submitted to a perpendicular magnetic field of strength $B$ and to the potential of impurities of maximal amplitude $w$. This model is of importance in the context of the integer quantum Hall effect. In the regime of strong magnetic field or weak disorder $B>>w$ it is known that there are chiral edge states, which are localised within a few magnetic lengths close to, and extended along the boundary of the cylinder, and whose energy levels lie in the gaps of the bulk system. These energy levels have a spectral flow, uniform in $L$, as a function of a magnetic flux which threads the cylinder along its axis. Through a detailed study of this spectral flow we prove that the spacing between two consecutive levels of edge states is bounded below by $2\pi\alpha L^{-1}$ with $\alpha>0$, independent of $L$, and of the configuration of impurities. This implies that the level repulsion of the chiral edge states is much stronger than that of extended states in the usual Anderson model and their statistics cannot obey one of the Gaussian ensembles. Our analysis uses the notion of relative index between two projections and indicates that the level repulsion is connected to topological aspects of quantum Hall systems.
Recently there has been mathematical progress concerning the spectral properties of disordered quantum Hall systems with boundaries. In the theory of the integer quantum Hall effect one considers non-interacting electrons confined on the surface of a finite cylinder \[1\] or on a corbino disk \[2\], submitted to a perpendicular uniform magnetic field of strength $B$ and to the potential of impurities of maximal amplitude $w$. In a classic paper on the subject \[2\] Halperin argued that, at least for strong magnetic field and weak disorder ($B>>w$ in appropriate units), there exist quantum mechanical states localised near and extended along the boundaries of the sample. These states carry a diamagnetic current contributing to the total Hall current. Halperin’s analysis applies to energies that lie in the gaps separating the Landau bands of the bulk disordered hamiltonian, i.e the hamiltonian of an infinite two dimensional planar system (with no boundaries). Here we will call this part of the spectrum the “pure edge spectrum”. Progress towards the characterisation of the nature of the pure edge spectrum has been made in recent works for systems with one smooth boundary \[3\], \[4\], \[5\]. In the present contribution we obtain new results for such systems, which are used in separate work on more realistic geometries involving two boundaries \[6\].
We consider the Hamiltonian of a particle on a cylinder of radius ${L\over 2\pi}$ thread by a flux line with flux $\Phi$ $$H(\Phi)={1\over 2}p_x^2+{1\over 2}(p_y-Bx+{\Phi\over L})^2 +W(x) +V(x,y)
\eqno(1.1)$$ where $x\in {\bf R}$, $-{L\over 2}\leq y\leq {L\over 2}$, with periodic boundary conditions in the $y$ direction $\Psi(x,-{L\over 2})=\Psi(x, {L\over 2})$. The particle is confined to the left half of the cylinder because of the external potential $W$ which models the boundary of a “semi-infinite cylinder”. We assume that it is continuous, and $W(x)=0$ for $x\leq 0$, $W^\prime(x)>0$ for $x\geq 0$, $W(x)\to +\infty$, $x\to +\infty$. For technical reasons we assume a growth of $W$ that is not too fast: we suppose that for $x\geq 0$, $u_1x^\gamma\leq W(x)\leq u_2 x^\gamma$, for some $0<u_1<u_2$ and $\gamma\geq 2$. The potential of impurities $V$ is piecewise continuous and bounded $|V(x,y)|\leq w$ with $0< w< {B\over 2}$. We also suppose that $V(x,y)=0$ for $x>0$, however our methods can be adapted to a more general model where the impurity potential extends inside the region of the boundary.
We will also use two other Hamiltonians: the “edge hamiltonian” $H_e(\Phi)$ obtained from (1.1) by removing $V$ and “the bulk hamiltonian” $H_b(\Phi)$ obtained from (1.1) by removing $W$.
The “semi-infinite planar” case corresponds to $L=+\infty$. In this limit the corresponding Hamiltonians become independent of $\Phi$ and we denote them $H_\infty$, $H_{e,\infty}$, $H_{b,\infty}$. It is easy to see that $H_{b,\infty}$ has gaps $G_n\supset ](n+{1\over 2})B+w, (n+{3\over 2})B-w[$, $n\in {\bf N}$. A basic fact is that for weak enough disorder the “pure edge spectrum” $\sigma(H_\infty)\cap G_n$, $n\in {\bf N}$ is continuous. This result is also proven for $W$ replaced by a Dirichlet boundary condition at $x=0$ and for smooth curved open boundaries (see \[3,4,5\]).
When $L$ is finite $G_n$ contains only discrete isolated eigenvalues. We formulate this result and all the subsequent ones in the special case $n=0$.
0.5cm
[**Lemma 1.**]{} Let $B>2w$. For any $0<\epsilon<{B\over 2}-w$ the set $\sigma(H(\Phi))\cap \tilde G_0$, $\tilde G_0=]{B\over 2}+w+\epsilon,
{3B\over 2}-w-\epsilon[$ contains only a finite number of isolated eigenvalues of finite multiplicity. We label the eigenvalues of $H(0)$ in $\tilde G_0$ as $E_1(0)\leq E_{2}(0)\leq...\leq E_{N}(0)$ for some finite $N$. Any $E_k(0)\in \tilde G_0$ can be continued into one or several analytic branches $E_k(\Phi)$ for $\Phi\in
[0,\Phi_k]$ for some small enough $\Phi_k>0$.
0.5cm
The discreteness of the spectrum in the specified interval is non trivial even if the circumference of the cylinder is finite because the impurity potential can extend to infinity in the direction $x\to -\infty$ where there is no confinement. In fact one can see that the rest of the spectrum may have dense parts. For example if $V$ is a typical realisation of a random potential the Landau bands $[(n+{1\over 2})B-w,
(n+{1\over 2})B+w]$ have dense spectrum. Now let $0<\delta<{B\over 2}-w-\epsilon$ and $\Delta=]B-\delta, B+\delta[$. For $L$ large enough, as long as an eigenvalue $E_k(\Phi)\in\Delta$ for some $\Phi$, then we are assured that it can be continued into an analytic branch for the whole interval $[0,2\pi]$. This comes from the fact (see inequality (3.15)) that the maximal variation of $E_k(\Phi)$ is $2\pi\sqrt{3B}L^{-1}$ so that it stays in $\tilde G_0$ and never merges in the Landau bands.
In the rest of this work we fix $\epsilon$ small and $0<\delta<{B\over 2}-w-\epsilon$, and look only at eigenvalues $E_k(\Phi)\in \Delta$. Note that as $\Phi$ varies from $0$ to $2\pi$ some of the branches may move in or out of $\Delta$. A reformulation of the analysis in \[3,4,5\] shows that there exists a spectral flow which is uniform in $L$. This is expressed by the following Lemma.
0.5cm
[**Lemma 2.**]{} Let $B>2w$. There exists $\delta$, $ w_0$ small enough, $ L_0$ large enough such that for $w< w_0$, $L> L_0$ all eigenvalues $E_k(\Phi)\in\Delta$ satisfy $$L{d\over d\Phi}E_k(\Phi)\geq \alpha
\eqno(1.2)$$ where $\alpha$ is strictly positive independent of $L$ and $k$, and depends only on $W$, $B$, $w$ and $\delta$.
0.5cm
The existence of a spectral flow is equivalent to the presence of a chiral diamagnetic current. Indeed by the Feynman-Hellman theorem $${d\over d\Phi}E_k(\Phi)=j_k(\Phi)
\eqno(1.3)$$ where $$j_k(\Phi)={1\over L}<\Psi_k(\Phi)|(p_y-Bx+{\Phi\over L})\Psi_k(\Phi)>
\eqno(1.4)$$ is the diamagnetic current (or edge current) associated to the eigenstate $|\Psi_k(\Phi)>$ corresponding to the level $E_k(\Phi)$.
The hamiltonians $H(\Phi)$ and $H(\Phi+2\pi)$ are unitarily equivalent, the unitary operator being multiplication by $\exp(2\pi i{y\over L})$. Thus for each $E_k(\Phi)$ which does not merge in the Landau bands there must exist some $k^\prime$ such that $E_k(2\pi)=E_{k^\prime}(0)$. From Lemma 2 it is clear that $k^\prime>k$, but this does not characterise completely the spectral flow. Our main new result states that $k^\prime=k+1$ and characterises the level spacing for the pure edge spectrum.
0.5cm
[**Theorem 1.**]{} Let $B>2w$. There exist $\delta$, $w_0$ small enough, $L_0$ large enough such that for $w< w_0$, $L> L_0$, the branches $E_k(\Phi)$ belonging to $\Delta$ for all $\Phi\in [0,2\pi]$ satisfy $$E_k(2\pi)=E_{k+1}(0)
\eqno(1.5)$$ Moreover the level spacing in $\Delta$ satisfies $${2\pi\alpha\over L}\leq |E_{k+1}(0)-E_k(0)|\leq {2\pi\sqrt{3B}\over L}
\eqno(1.6)$$
0.5cm
For the constant $\alpha$ in Lemma 2 and theorem 1 we can take the right hand side of (2.29). The important point is that in the lower bound of (1.6) $\alpha$ does not depend on the detailled configuration of the impurity potential but only on its maximal amplitude. So for a random potential the level spacing is random but our lower bound is non random.
0.5cm
For the usual Anderson model it is proven that the level spacing of localised states satisfies Poisson statistics \[7\], \[8\] and it is numerically established that extended states have a level repulsion satisfying the Wigner surmise \[9\]. Here we have a different situation: the states are extended, chiral and have a much stronger level repulsion which makes the level spacing very rigid. Let $\rho(E)$ denote the average density of edge states. We expect from (1.6) that, in the limit $L\to\infty$, the rescaled level spacing $s=L\rho(E_k)|E_{k+1}-E_k|$ has a histogram $p(s)$ which is a certain broadening of $\delta(s-1)$ with a finite support of $O({w^2\over B^2})$. The level statistics cannot follow the Gaussian ensembles and it would be worthwhile to investigate this question numericaly for an analogous model on a lattice. It is apparent from the proof of theorem 1 that the rigidity of the edge spectrum is related to the topological invariants of the quantum Hall effect. Also if the spectral flow would satisfy $E_k(2\pi)=E_{k+n}(0)$ with $n\geq 2$, it would not be forbidden to have $n$ consecutive levels arbitrarily close.
We wish to point out that all these features can be checked immediately for a simple toy Hamiltonian. Consider a one dimensional chiral particle on a circle of circumference $L$ thread by a flux $\Phi$ $$h(\Phi)=(-i\partial_y+{\Phi\over L}) + v(y)
\eqno(1.7)$$ The exact spectrum is $$e_m(\Phi)={2\pi m\over L}+{\Phi\over L}+{1\over L}
\int_{-{L\over 2}}^{L\over 2} dy v(y)
\eqno(1.8)$$ which satisfies (1.2), (1.5), (1.6) and has $p(s)=\delta(s-1)$. It is expected that (1.7) is a good approximation of (1.1) for distances to the boundary of the order of the magnetic length $x= O({1\over \sqrt B})$.
Finaly we recall how it follows from (1.5) that the “edge conductance” of the semi-infinite system is quantized (see \[1\], \[2\], \[4\] for similar discussions). Let $P_\Delta(\Phi)$ be the projector of $H(\Phi)$ on an energy range $\Delta$. The edge conductance may be defined as the total edge current per unit energy, $$\sigma_e=\lim_{L\to\infty}{1\over |\Delta|L}{\rm Tr}(p_y-Bx-{\Phi\over L})P_\Delta(\Phi)
\eqno(1.9)$$ We assume that for a suitable class of potentials $V$ this limit exists and is independent of $\Phi$ (the flux has no effect for the semi-infinite plane). We expect this assumption to be true for typical realisations of random potentials that are ergodic with respect to the translations along $y$. In this case the limit should be equal to ${1\over \Delta}{\rm Av}\int dx<x,0|(p_y-Bx)P_{\infty,\Delta}|x,0>$ where ${\rm Av}$ is the average over the disorder and $P_{\infty, \Delta}$ the projector of $H_\infty$ onto $\Delta$. The limit of the later quantity when $\Delta\to \mu$ has been shown to be an integer if $\mu$ is a point in the gap $G_0$, by non-commutative geometry techniques applied to the lattice case \[10\]. In the present situation it is easy to see that for $\Delta$ in the first gap of the bulk Hamiltonian $H_{b,\infty}$ $$\eqalign{
{1\over |\Delta|L} & ||(p_y-Bx-{\Phi\over L})P_\Delta(\Phi)||_1
\leq
{1\over |\Delta|L}||(p_y-Bx-{\Phi\over L})P_\Delta(\Phi)||.||P_\Delta(\Phi)||_1
\cr &
\leq {\sqrt 2\over |\Delta|L}{\rm sup}_{||\psi||=1}
(<\psi|P_\Delta(\Phi)(H(\Phi)-V)P_\Delta(\Phi)|\psi>)^{1/2} {\rm Tr} P_\Delta(\Phi)
\cr &
\leq {\sqrt{3B}\over |\Delta|L}{\rm Tr} P_\Delta(\Phi)=O(1)
\cr}
\eqno(1.10)$$ Here $||.||_1$ and $||.||$ are the trace and operator norms respectively and we used$||AB||_1\leq ||A||.||B||_1$ for $A$ bounded and $B$ trace class. In the last equality we used that there are $O(L)$ states in $\Delta$ because of (1.6) so that the final bound is uniform with respect to $L$. Since we have assumed that $\sigma_e$ is independent of $\Phi$, by averaging over $\Phi$ we get $$\eqalign{
\sigma_e & = \lim_{L\to\infty}{1\over|\Delta|}\int_0^{2\pi}{d\Phi\over 2\pi}
\sum_{E_k(\Phi)\in\Delta}{dE_k(\Phi)\over d\Phi} =
\lim_{L\to\infty}
{1\over |\Delta|}\sum_{k_{min}}^{k_{max}} \int_0^{2\pi} {d\Phi\over 2\pi}{dE_k(\Phi)\over d\Phi}
\cr &
=\lim_{L\to\infty}
{1\over 2\pi |\Delta|}\sum_{k_{min}}^{k_{max}} (E_{k+1}(0)-E_k(0))
=\lim_{L\to\infty}{1\over 2\pi|\Delta|}(E_{k_{max}}-E_{k_{min}}) ={1\over 2\pi}
\cr}
\eqno(1.11)$$ For the first equality we use (1.3), (1.4) and dominated convergence. To obtain the second equality we consider separately the contributions of the eigenvalues with $k_{min}\leq k\leq k_{max}$ such that $E_k(\Phi)\in \Delta$ for all $\Phi\in [0, 2\pi]$, and of a finite number of eigenvalues with $k<k_{min}$ (resp. $k>k_{max}$) which enter (resp. leave) $\Delta$ as $\Phi$ varies from $0$ to $2\pi$. From (1.6) and (3.15) this later contribution is $O(L^{-1})$. Finaly (1.5) is used in the third equality. Here the units are such that $e=\hbar=1$ so ${1\over 2\pi}={e^2\over h}$.
Section 2 contains the proofs of Lemmas 1 and 2 and a third Lemma that is needed for the proof of theorem 1 in section 3. The appendices A and B contain technical estimates.
0.5cm
0.5cm
[**Proof of Lemma 1**]{}
Let $D>0$ to be chosen later (large) and $V_D(x,y)=V(x,y)$ for $x\leq -D$, $V_D(x,y)=0$ for $x> -D$. Then $V(x,y)-V_D(x,y)$ has compact support and a standard argument using the resolvent identity implies that the essential spectra of $$H_D(\Phi)=H_e(\Phi)+V_D(x,y)
\eqno(2.1)$$ and $$H(\Phi)=H_D(\Phi)+V(x,y)-V_D(x,y)
\eqno(2.2)$$ coincide \[11\]. Therefore if we show that $\sigma(H_D(\Phi))\cap \tilde G_0$ contains only isolated eigenvalues of finite multiplicity, the same is true for $H(\Phi)$. This will be achieved below using a decoupling scheme \[12\], \[13\] which proves that $\sigma(H_D(\Phi))\cap\tilde G_0$ is a small perturbation of $\sigma(H_e(\Phi))\cap\tilde G_0$. The set $\sigma(H_e(\Phi))$ consists of non degenerate energy levels $\epsilon_n({2\pi m\over L}+{\Phi\over L})$, $n\in {\bf N}$ the Landau index and $m\in {\bf Z}$, where $\epsilon_n(k)$, $k\in {\bf R}$ the wavenumber conjugate to $y$, are the spectral branches of $H_{e,\infty}$. These spectral branches are monotone increasing entire functions of $k$ with $\epsilon_n(k)\to +\infty$ for $k\to+\infty$ and $\epsilon_n(k)\to (n+{1\over 2})B$ for $k\to -\infty$ (see for example \[3\]).
In order to set up the decoupling scheme we introduce the characteristic functions $\chi_e(x)$ of $-{D\over 2}\leq x<+\infty$ and $\chi_b(x)$ of $-\infty\leq x<-{D\over 2}$. Note that $\chi_e(x)+\chi_b(x)=1$ for all $x$. We also need the monotone and twice differentiable functions $J_e(x)$, $J_b(x)$ such that $J_e(x)=0$ for $-\infty<x<-{3D\over 4}-1$ and $J_e(x)=1$ for $-{3D\over 4}+1<x<\infty$; $J_b(x)=1$ for $-\infty<x<-{D\over 4}-1$, $J_b(x)=0$ for $-{D\over 4}+1<x<\infty$.
We introduce the Green functions $G_\alpha(z)=(H_\alpha(\Phi)-z)^{-1}$ for $\alpha=e, b, D$ and $z\in {\bf C}$ in the resolvent set of the corresponding hamiltonian. Since $$H_D(\Phi)J_\alpha=H_\alpha(\Phi)J_\alpha\qquad {\rm for}\qquad \alpha=e,b
\eqno(2.4)$$ following \[13\] we have $$\eqalign{
(H_D(\Phi)-z)&(J_eG_e(z)\chi_e+J_bG_b(z)\chi_b)
\cr &=(H_e(\Phi)-z)J_eG_e(z)\chi_e+(H_b(\Phi)-z)J_bG_b(z)\chi_b
\cr &
= J_e\chi_e+J_b\chi_b+{1\over 2}[p_x^2, J_e]G_e(z)\chi_e
+{1\over 2}[p_x^2, J_b]G_b(z)\chi_b
\cr &
=1+K_e(z)+K_b(z)
\cr}
\eqno(2.5)$$ where $K_\alpha(z)={1\over 2}[p_x^2, J_\alpha]G_\alpha(z)\chi_\alpha$, $\alpha=e, b$. Thus $$(H_D(\Phi)-z)^{-1}=(J_eG_e(z)\chi_e+J_bG_b(z)\chi_b)(1+K_e(z)+K_b(z))^{-1}
\eqno(2.6)$$ In Appendix A we prove the following estimates for the operator norms of $K_e(z)$ and $K_b(z)$ for ${B\over 2}+w<{\rm Re} z<{3B\over 2}-w$ (in what follows $c$ is a generic positive numerical constant) $$||K_e(z)||\leq{c B^{3\over 2}L\over \delta_e(z)}e^{-cBD^2}
\eqno(2.7)$$ $$||K_b(z)||\leq{cB^{3\over 2}L \over \delta_0(z)-cw}e^{-c \sqrt B D}
\eqno(2.8)$$ where $\delta_e(z)={\rm dist}(z, \sigma(H_e(\Phi))$ and where $\delta_0(z)=\min(|z-{B\over 2}|, |z-{3B\over 2}|)$. We have to take $w$ small enough so that the denominator in (2.8) stays positive. Later on we choose $z$ appropriately and $D$ large enough so that both terms become smaller than ${1\over 2}$. Thus $$(H_D(\Phi)-z)^{-1}=J_eG_e(z)\chi_e+J_bG_b(z)\chi_b +R(z)
\eqno(2.9)$$ where $$||R(z)||\leq \biggl(||G_e(z)||+||G_b(z)||\biggr)\biggl[(1-||K_e(z)||-||K_b(z)||)^{-1}-1\biggr]
\eqno(2.10)$$ Let $m\in {\bf Z}$ be such that $\epsilon_0({2\pi m\over L}+{\Phi\over L})$ is an eigenvalue belonging to $\sigma(H_e(\Phi))\cap \tilde G_0$. We can choose $\rho>0$ small enough independent of $m$ and $L$ such that the circle $C_m$ with center $\epsilon_0({2\pi m\over L}+{\Phi\over L})$ and radius ${\rho\over L}$ encloses only one such eigenvalue. By choosing $z$ in a sufficiently thin annulus around $C_m$ and $D$ large enough, (2.7) and (2.8) can be made smaller than ${cB^{3\over 2}L^2\over \rho}e^{-c\sqrt B D}<{1\over 2}$. At the same time from (2.10) we have $$||R(z)||\leq {cB^{3\over 2}L^3\over \rho^2}e^{-c\sqrt B D}
\eqno(2.11)$$ so that from (2.9) $(H_D(\Phi)-z)^{-1}$ is well defined for $z$ in a thin annulus surrounding $C_m$. Therefore we can compute the spectral projection $P_D(m, \Phi)$ of $H_D(\Phi)$ for the interval $I_m=]\epsilon_0({2\pi m\over L}+{\Phi\over L})-{\rho\over L},
\epsilon_0({2\pi m\over L}+{\Phi\over L})+{\rho\over L}[$ by Cauchy’s formula. Let $P_e(m, \Phi)$ be the projector of $H_e(\Phi)$ corresponding to the level $\epsilon_0({2\pi m\over L}+{\Phi\over L})$. Thanks to (2.9), (2.11) we obtain for $D$ large enough $$||P_D(m, \Phi)-P_e(m, \Phi)||\leq {cB^{3\over 2}L^2\over \rho}e^{-c\sqrt B D}<1
\eqno(2.12)$$ This estimate implies that $\sigma(H_D(\Phi))\cap I_m$ contains only one eigenvalue of multiplicity equal to one. Note that this conclusion holds for all $I_m\subset \tilde G_0$. Finally since $H_e(\Phi)$ and $H_b(\Phi)$ have no spectrum in $(\cup_m I_m)^{c}\cap \tilde G_0$ we deduce from (2.7), (2.8), (2.9) that $H_D(\Phi, L)$ has no spectrum in that same set. Therefore $\sigma(H_D(\Phi, L))\cap \tilde G_0$ consists of isolated eigenvalues of multiplicity one.
It remains to show that an eigenvalue $E_k(0)\in \tilde G_0$ can be continued into one or several analytic branches $E_k(\Phi)$ for $\Phi$ small enough. In the present case it is sufficient to show \[11\] that $(py-Bx)$ is relatively bounded with respect to $H(0)$. For any $\psi$ in the domain of $H(0)$ and any complex number $z$ with ${\rm Im}z\neq 0$ we have $$\eqalign{{1\over 2}
||&(p_y-Bx)\psi||^2 \leq <\psi|(H(0)-V)\psi>
\cr &
=<\psi|(H(0)-z)^{-1}(H(0)-z)|(H(0)-\bar z+z)\psi> -<\psi|V\psi>
\cr &
\leq ||(H(0)-z)^{-1}||.||H(0)\psi||^2 +|z|.||\psi||^2+|z|^2||(H(0)-z)^{-1}||.||\psi||^2
+w||\psi||^2
\cr &
\leq
{1\over |{\rm Im} z|} ||H(0)\psi||^2+(|z|+{|z|^2\over |{\rm Im} z|}+w)||\psi||^2
\cr}
\eqno(2.13)$$ This concludes the proof of the Lemma.
0.5cm
[**Remark:**]{} In (2.13) we can take $|{\rm Im} z|$ as large as we wish so the size of the interval of analyticity is not limited by the relative bound but rather by the fact that the branch $E_k(\Phi)$ may merge in the Landau bands (outside of $G_0$) where it may not be isolated anymore. Inequality (3.15) shows that for $L$ large enough the maximal variation of $E_k(\Phi)$ is $2\pi\sqrt{3B}L^{-1}$, so that if $E_k(\Phi)$ is contained in $\Delta$ for some $\Phi$ then it is contained in $\tilde G_0$ and it is analytic for all $\Phi\in [0,2\pi]$.
0.5cm
Before presenting the formal proof of Lemma 2 we would like to point out that in fact (1.2) is closely related to the ideas in \[3\] and \[4\]. Using the unitary translation operator $x\to x+{\Phi\over BL}$ and the Feynman-Hellman theorem it is easy to see that $$L{d\over d\Phi}E_k(\Phi)=<\Psi_k(\Phi)|(W^\prime+\partial_x V)\Psi_k(\Phi)>$$ where $|\Psi_k(\Phi)>$ is the eigenstate with eigenvalue $E_k(\Phi)$. Using the methods of \[3\] or \[4\] one may show that for $E_k(\Phi)\in \Delta$, $|\Psi_k(\Phi)>$ is mainly concentrated near the region where $W^\prime(x)$ is large so that (1.2) holds provided both $V$, $\partial_x V$ are small enough. Here we follow a different method which is closer to the original argument of Halperin \[2\] in that it uses directly the relation (1.4) instead of (2.13). Only the smallness of $V$ is required.
0.5cm
[**Proof of Lemma 2**]{}
The eigenstates $|u_{nm}(\Phi)>$ of $H_e(\Phi)$ with eigenvalues $\epsilon_n({2\pi m\over L}+{\Phi\over L})$ are of the form $$<xy|u_{nm}(\Phi)> = e^{i{2\pi m\over L}y} h_{nm}(x)
\eqno(2.14)$$ so that $<u_{nm}(\Phi)|(p_y-Bx-{\Phi\over L})u_{n^\prime m^\prime}(\Phi)>=0$ for $m\neq m^\prime$ and all $n$, $n^\prime$. Therefore writing $$|\Psi_k(\Phi)>=|\Psi_k^0(\Phi)>+|\Psi_k^1(\Phi)>
\eqno(2.15)$$ where $$|\Psi_k^0(\Phi)>=\sum_{m=-\infty}^{+\infty} c_{k}^{0m}|u_{0m}(\Phi)>
\eqno(2.16)$$ $$|\Psi_k^1(\Phi)>=\sum_{n\geq 1}\sum_{m=-\infty}^{+\infty} c_{k}^{nm}|
u_{nm}(\Phi)>
\eqno(2.17)$$ we obtain from (1.3), (1.4) $$\eqalign{
L{d\over d\Phi}E_k(\Phi)=&
\sum_{m=-\infty}^{+\infty} |c_{k}^{0m}|^2<u_{0m}(\Phi)|(p_y-Bx-{\Phi\over L})u_{0m}(\Phi)>
\cr &
+2{\rm Re}<\Psi_k^0(\Phi)|(p_y-Bx-{\Phi\over L})\Psi_k^1(\Phi)>
\cr &
+<\Psi_k^1(\Phi)|(p_y-Bx-{\Phi\over L})\Psi_k^1(\Phi)>
\cr}\eqno(2.18)$$ First we show that the last two terms on the right hand side of (2.18) are bounded by the norm $\sqrt{3B}||\Psi_k^1(\Phi)||$. The Schwartz inequality implies $$\eqalign{
|<\Psi_k^0(\Phi)| & (p_y-Bx-{\Phi\over L})\Psi_k^1(\Phi)>|
\leq ||\Psi_k^1(\Phi)||.||(p_y-Bx+{\Phi\over L})\Psi_k^0(\Phi)||
\cr &
\leq\sqrt{2}||\Psi_k^1(\Phi)||
(<\Psi_k^0(\Phi)|H_e(\Phi)\Psi_k^0(\Phi)>)^{1/2}
\cr &
\leq
\sqrt{2}||\Psi_k^1(\Phi)||
\biggl(<\Psi_k^0(\Phi)|H_e(\Phi)\Psi_k^0(\Phi)>
+ \cr & <\Psi_k^1(\Phi)|H_e(\Phi)\Psi_k^1(\Phi)>\biggr)^{1/2}
\cr &
=
\sqrt{2}||\Psi_k^1(\Phi)||
(<\Psi_k(\Phi)|H_e(\Phi)\Psi_k(\Phi)>)^{1/2}
\cr &
\leq
\sqrt{2}||\Psi_k^1(\Phi)||\bigl(E_k(\Phi)+w\bigr)^{1/2}
\leq \sqrt{3B}||\Psi_k^1(\Phi)||
\cr}
\eqno(2.19)$$ For the third matrix element on the right hand side of (2.18) the same method leads to an identical estimate. From the Feynman-Hellman formula we have $$\eqalign{
<u_{0m}(\Phi)|(p_y-Bx-{\Phi\over L})u_{0m}(\Phi)> &
=L{d\over d\Phi}\epsilon_0({2\pi m\over L}+{\Phi\over L})
\cr &
=
\epsilon_0^{\prime}({2\pi m\over L}+{\Phi\over L})
\cr}
\eqno(2.20)$$ where $\epsilon_0^\prime(k)$ is the derivative of the lowest monotone increasing spectral branch corresponding to the hamiltonian $H_{e,\infty}$. From (2.18), (2.19), (2.20) $$\eqalign{
L{d\over d\Phi}E_k(\Phi) & \geq
\sum_{m=-\infty}^{+\infty}|c_k^{0m}|^2\epsilon_0^{\prime}({2\pi m\over L}+{\Phi\over L})-2\sqrt{3B}||\Psi_k^1(\Phi)||
\cr &
\geq
v_F(M)
\sum_{|m-M|\leq \bar m}|c_k^{0m}|^2
-2
\sqrt{3B}||\Psi_k^1(\Phi)||
\cr}
\eqno(2.21)$$ with the Fermi velocity $$v_F(M)={\rm min}_{|m-M|\leq \bar m}\epsilon_0^\prime({2\pi m\over L}+{\Phi\over L})
\eqno(2.22)$$ The integers $M$ and $\bar m$ will be choosen conveniently below. Writting the Schrödinger equation in the form, $$\sum_{n=0}^{\infty}\sum_{m=-\infty}^{+\infty}
c_k^{nm}\biggl(\epsilon_n({2\pi m\over L}+{\Phi\over L})-E_k(\Phi)\biggr)|u_{nm}(\Phi)>=V(x,y)|\Psi_k(\Phi)>
\eqno(2.23)$$ and taking the norm on both sides $$\sum_{n=0}^{\infty}\sum_{m=-\infty}^{+\infty}
|c_k^{nm}|^2\biggl(\epsilon_n({2\pi m\over L}+{\Phi\over L})-E_k(\Phi)
\biggr)^2\leq
w^2
\eqno(2.24)$$ Dropping the term $n=0$, using $\bigl(\epsilon_n({2\pi m\over L}+{\Phi\over L})-E_k(\Phi)
\bigr)^2\geq ({B\over 2}-\delta)^2$ for $n\geq 1$ and $E_k(\Phi)\in \Delta$ we get $$\sum_{n\geq 1}\sum_{m=-\infty}^{\infty}|c_k^{nm}|^2=
||\Psi_k^1(\Phi)||^2\leq {w^2\over ({B\over 2}-\delta)^2}
\eqno(2.25)$$ From (2.24) one can also derive a lower bound for $\sum_{|m-M|<\bar m}|c_k^{0m}|^2$. Indeed retaining only the term $n=0$ and using the monotonicity of $\epsilon_0({2\pi m\over L}+{\Phi\over L})$ we have $$A(M, \bar m)^2
\sum_{|m-M|>\bar m}|c_k^{0m}|^2\leq w^2
\eqno(2.26)$$ where $A(M,\bar m)$ is the smallest of the two numbers $|\epsilon_0({2\pi\over L}(M\pm\bar m)+{\Phi\over L})|-E_k(\Phi)|$. Now we choose any $M$ such that $\epsilon_0({2\pi M\over L}+{\Phi\over L})\in\Delta$ and since $E_k(\Phi)\in \Delta$ we can take $\bar m$ such that $A(M,\bar m)\geq {B\over 2}-2\delta$. Thus $$\sum_{|m-M|>\bar m}|c_k^{0m}|^2\leq {w^2\over ({B\over 2}-2\delta)^2}
\eqno(2.27)$$ Finaly the normalisation condition for $|\Psi_k(\Phi)>$ combined with (2.25) and (2.27) imply $$\sum_{|m-M|\leq\bar m}|c_k^{0m}|^2\geq 1-{2w^2\over ({B\over 2}-2\delta)^2}
\eqno(2.28)$$ From (2.21), (2.25) and (2.28) we have $$L{d\over d\Phi}E_k(\Phi)
\geq v_F(M)\biggl[1-
2(1+{\sqrt{3B}\over v_F(M)}){w^2\over ({B\over 2}-2\delta)^2}\biggr]
\eqno(2.29)$$ Clearly $v_F(M)$ is a strictly positive number which does not depend on $V$ but only on $W$ and $B$. Therefore (2.29) implies the result of the Lemma for $w$ and $\delta$ small enough.
0.5cm
It will become clear in the next section that the proof of Theorem 1 requires the absence of crossings for the branches $E_k(\Phi)$ in $\Delta$. Since we do not know a priori if this is true for $H(\Phi)$, an intermediate step is to construct a suitable perturbation of $H(\Phi)$ for which the non-crossing property is satisfied. The perturbation that is added here has the effect to lift the degeneracy at each crossing in $\Delta$ in a way that (1.2) still holds for the perturbed branches. This is the content of the next Lemma.
0.5cm
[**Lemma 3.**]{} Fix $B$, $w$, $\delta$ and $L$ as in Lemma 2. Assume that $V(x,y)$ is such that the eigenvalues $E_l(0)$ are not degenerate. One can construct a finite rank perturbation $R(\Phi)$ with $||R(\Phi)||\leq L^{-10}$ such that the spectrum of $\tilde H(\Phi)
=H(\Phi)+R(\Phi)$ in $\Delta$ consists of non degenerate eigenvalues forming infinitely differentiable spectral branches which do not cross and are labeled as $\tilde E_l(\Phi)$ with $\tilde E_l(0)= E_l(0)$. Moreover the new branches satisfy $$L{d\over d\Phi}\tilde E_l(\Phi)\geq \tilde\alpha
\eqno(2.30)$$ where $\tilde \alpha$ is strictly positive and independent of $L$.
0.5cm
[**Proof of Lemma 3.**]{}
Let $P_\Delta(\Phi)$ be the eigenprojector of $H(\Phi)$ onto $\Delta$. Then we have $$P_\Delta(\Phi) H(\Phi)P_\Delta(\Phi)=\sum_{E_l(\Phi)\in \Delta} E_l(\Phi)
|\Psi_l(\Phi)><\Psi_l(\Phi)|
\eqno(2.31)$$ Since the branches $E_l(\Phi)$ are analytic and the eigenvalues are not degenerate for $\Phi=0$ the possible crossings are necessarily isolated. Indeed if two branches would coincide on a set with accumulation points they would concide over the whole interval $[0,2\pi]$ and therefore violate the non degeneracy assumption at $\Phi=0$. Therefore we can assume without loss of generality that there is at most a finite number of crossings in $\Delta$. Let us construct the perturbation $R(\Phi)$. First consider the set ${\cal C}$ of pairs of branches which cross in $\Delta$ (note that $n$ branches may cross at the same point and contribute as ${n(n-1)\over 2}$ pairs). Pick one pair of branches in ${\cal C}$ say $(ij)$ and assume $E_i(0)<E_j(0)$. Suppose they cross at points $\Phi_{ij}^\mu$ where the label $\mu$ takes into account the fact that the branches $i$ and $j$ may cross more than once, i.e $$E_i(\Phi_{ij}^\mu)=E_j(\Phi_{ij}^\mu)
\eqno(2.32)$$ Let $\lambda_{ij}^\mu(\Phi)$ be infinitely differentiable test functions centered at $\Phi_{ij}^\mu$, with a compact support of width $\beta_1$ and $\max_{0\leq \Phi\leq 2\pi} |\lambda_{ij}^\mu(\Phi)|\leq \lambda_1$. The real numbers $\delta_1$ and $\lambda_1$ will be adjusted in a suitable way below. Add to the Hamiltonian $H(\Phi)$ the perturbation $$R_1(\Phi)=\sum_{\mu} \lambda_{ij}^\mu(\Phi)
\bigl(|\Psi_i(\Phi)><\Psi_j(\Phi)|+
|\Psi_j(\Phi)><\Psi_i(\Phi)|\bigr)
\eqno(2.33)$$ We take $\beta_1$ small enough so that the supports of the test functions do not contain $\Phi=0$ and do not overlap. In order to diagonalise the new hamiltonian it is sufficient to work in the two dimensional subspace of the branches $i$ and $j$. The spectral branches of the new Hamiltonian do not change for $k\neq i,j$, whereas for $k=i,j$ they become $$E_i^{1}(\Phi)={1\over 2}\biggl(E_i(\Phi)+E_j(\Phi)-\sqrt{
(E_i(\Phi)-E_j(\Phi))^2+\lambda_{ij}^\mu(\Phi)^2}\biggr)
\eqno(2.34)$$ and $$E_j^{1}(\Phi)={1\over 2}\biggl(E_i(\Phi)+E_j(\Phi)+\sqrt{
(E_i(\Phi)-E_j(\Phi))^2+\lambda_{ij}^\mu(\Phi)^2}\biggr)
\eqno(2.35)$$ Since the difference $$E_j^{1}(\Phi)-E_i^{1}(\Phi)=\sqrt{
(E_i(\Phi)-E_j(\Phi))^2+\lambda_{ij}^\mu(\Phi)^2}
\eqno(2.36)$$ is always strictly positive the new pair $(ij)$ is non degenerate for all values of $\Phi$. Moreover by choosing $\lambda_1$ small enough we can make sure that we do not introduce more crossings. Therefore the perturbed hamiltonian $$H_1(\Phi)=H(\Phi)+ R_1(\Phi)
\eqno(2.37)$$ has a new set ${\cal C}_1$ of pairs of branches which cross, with one element less than ${\cal C}$. One can construct in the same way a perturbation $R_2(\Phi)$ of (2.37) (with $\delta_2$, $\lambda_2$ small enough) so that the new Hamiltonian $H_2(\Phi)=H_1(\Phi)+R_2(\Phi)$ has two less pairs of branches which cross than $H(\Phi)$. Since there is at most a finite number of such pairs by iterating this construction we end up with the Hamiltonian $$\tilde H(\Phi)=H(\Phi)+\sum_p R_p(\Phi)=H(\Phi)+R(\Phi)
\eqno(2.38)$$ of the Lemma, where the sum over $p$ contains a finite number of terms. Note that $\tilde H(0)=H(0)$ so that the labelling of the Lemma holds. The norm of the total perturbation is $$||R(\Phi)||\leq \sum_{p}||R_p(\Phi)||\leq
\sum_{p}\lambda_p
\eqno(2.39)$$ The condition $||R(\Phi)||\leq L^{-10}$ can always be achieved by choosing at each step $$\lambda_p\leq {\beta_p\over L^{10+p}}
\eqno(2.40)$$ and $\beta_p\leq {1\over 10}$.
It remains to check that (2.30) holds. From the formulas (2.34), (2.35) and Lemma 2, it is easy to check that at the first step of the construction the new branches have new derivatives satisfying $${d\over d\Phi}E_{i,j}^{1}(\Phi)\geq
{\rm min}\bigl({d\over d\Phi}E_i(\Phi),{d\over d\Phi}E_j(\Phi)\bigr)
- {1\over 2}|{d\over d\Phi}\lambda_{ij}^{\mu}(\Phi)|
\eqno(2.41)$$ for all $\Phi$. At each step of the construction it is possible to choose test functions such that $${\rm max}_{0\leq \Phi\leq 2\pi}|{d\over d\Phi}\lambda_{ij}^\mu(\Phi)|\leq {2
\over L^{10+p}}
\eqno(2.42)$$ in a way consistent with (2.40). So at the first step ($p=1$) $${d\over d\Phi}E_{i,j}^{1}(\Phi)\geq
{\alpha\over L}- {1\over L^{11}}
\eqno(2.43)$$ Of course (2.43) is also valid for the spectral branches of $H_1(\Phi)$ that correspond to $k\in {\cal N}$. Therefore it is valid for all eigenvalues of $H_1(\Phi)$. By iterating the construction we see that any branch of (2.38) satisfies $${d\over d\Phi}\tilde E_{l}(\Phi)\geq
{\alpha\over L}- \sum_p{1\over L^{10+p}}
\eqno(2.44)$$ which implies (2.30).
0.5cm
The main goal of this section is to prove Theorem 1. Let us first outline the strategy of the proof. Without loss of generality we can suppose that $V$ is such that $E_k(0)$ are non degenerate. Indeed if this is not the case one may find a sufficiently small perturbation $u(x,y)$, $||u||_\infty<L^{-10}$ such that this hypothesis is satisfied for $V+u$. If (1.5), (1.6) hold for $V+u$ then they hold for $V$ because the perturbation of the discrete levels separated by $O(L)$ is at most $O(L^{-10})$. From Lemma 2 we know that for $E_k(\Phi)\in \Delta$ there is a non trivial spectral flow: the branches are monotone increasing and since $H(0)$ and $H(2\pi)$ are unitarily equivalent we must have $E_k(2\pi)=E_{k^\prime}(0), k^\prime> k$. We want to show that in fact $k^\prime = k+1$. Let $E_F$ be a single “Fermi energy” lying between two consecutive levels of both Hamiltonians $H_D(0)$ and $\tilde H(0)$. Define the integers $Q_F^D$ and $\tilde Q_F$ to be the number of branches of the corresponding Hamiltonians which cross $E_F$ as $\Phi$ varies from $0$ to $2\pi$. We will show that $Q_F^D=\tilde Q_F=1$. We know from Lemma 3 that the branches of $\tilde H(\Phi)$ do not have crossings, and from the proof of Lemma 1 that the same is true for the branches of $H_D(\Phi)$. This enables us to relate $\tilde Q_F$ and $Q_F^D$ to the notion of relative index of a pair of projections introduced by Avron, Seiler and Simon \[14\]. Then by using the fact that the Fredholm index of an operator does not change under compact perturbations we deduce that $\tilde Q_F = Q_F^D$. By explicit computation we can check that $Q_F^D=1$ and therefore $\tilde Q_F=1$ which implies that $\tilde E_k(2\pi) = \tilde E_{k+1}(0)$. Since the branches of $\tilde H(\Phi)$ are a small perturbation of those of $H(\Phi)$ we deduce (1.5). Estimate (1.6) is then an immediate consequence.
In order to make the paper selfcontained we give a short summary of the mathematical tools used below, as developed in \[14\]. Let $P$ and $Q$ be orthogonal projections on a separable Hilbert space ${\cal H}$. The pair $(P;Q)$ is called Fredholm if $QP$ viewed as a map from $P{\cal H}$ to $Q{\cal H}$ is a Fredholm operator. The relative index ${\rm Ind}(P;Q)$ of the pair is the usual Fredholm index of $T=QP$, that is ${\rm dimKer}(T^\dagger T)-{\rm dimKer}(TT^\dagger)$. One proves that $(P;Q)$ is a Fredholm pair if and only if $1$ and $-1$ are isolated finitely degenerate eigenvalues of $P-Q$, when they belong to the spectrum. Moreover one has ${\rm Ind }(P,Q)={\rm dimKer}(P-Q-1) - {\rm dimKer}(P-Q+1)$. A useful formula (we use it for $m=0$) states that if $(P-Q)^{2m+1}$ is trace class for some integer $m$ then $(P;Q)$ is a Fredholm pair and ${\rm Ind}(P;Q)={\rm Tr}(P-Q)^{2n+1}$, for all ${n\geq m}$. A central result on which we rely is that if $(P;Q)$ and $(Q;R)$ are Fredholm pairs and either $P-Q$ or $Q-R$ is compact then $(P;R)$ is a Fredholm pair and $${\rm Ind}(P;R)={\rm Ind}(P;Q) + {\rm Ind}(Q;R)
\eqno(3.1)$$ Finaly we note that if $(P;Q)$ is Fredholm then so is $(UPU^\dagger;UQU^\dagger)$ for any unitary $U$ and the relative index remains invariant. Also ${\rm Ind}(P;Q)=-{\rm Ind}(Q;P)$.
0.5cm
[**Relation between $\tilde Q_F$, $Q_F^D$ and the relative index of a pair of projections.**]{}
We fix $E_F\in \Delta$ between two consecutive levels of $\tilde H(0)$ and $\tilde H(2\pi)$ (recall that they have the same spectrum). Let $\tilde P_{F,0}$ (resp. $\tilde P_{F,2\pi}$) be the projectors of $\tilde H(0)$ (resp. $\tilde H(2\pi)$) onto the energy range $]-\infty, E_F]$. We also need the projector on levels $\tilde E_k(0)$ whose spectral branch $\tilde E_k(\Phi)$ crosses $E_F$. Namely $$\tilde P_{F,0}^c=\sum_{\tilde E_k(0)<E_F {\rm s.t} \tilde E_k(\Phi)
{\rm crosses} E_F}
P(\tilde E_k(0))
\eqno(3.2)$$ where $P(\tilde E_k(0))$ is the eigenprojector of $\tilde H(0)$ corresponding to the discrete level $\tilde E_k(0)$. Since $E_F\in \Delta$ by taking $L$ large enough we are assured that this sum is finite and that the branches crossing $E_F$ remain in $\Delta$ for all $\Phi\in [0, 2\pi]$.
Setting $\tilde P_{F,0}^{n.c}=\tilde P_{F,0}-\tilde P_{F,0}^c$ we have
$$\tilde Q_F={\rm Tr} \tilde P_{F,0}^{c}={\rm Tr} (\tilde P_{F,0}-\tilde P_{F,0}^{nc})
={\rm Ind}(\tilde P_{F,0};\tilde P_{F,0}^{nc})
\eqno(3.3)$$ We introduce a smooth, monotone increasing function of time $\varphi(t)$, $0\leq t\leq T$, $\varphi(0)=0$ and $\varphi(T)=2\pi$, describing the adiabatic switching of a flux quantum through the axis of the cylinder. Let $U_t$ be the unitary time evolution associated to the time dependent Hamiltonian $\tilde H(\varphi(t))$. From Lemma 3, as $t$ varies the spectral branches in $\Delta$ do not cross, and are monotone increasing. So an application of the adiabatic theorem \[15\] assures that $U_T \tilde P_{F,0}^{nc}U_T^{\dagger}$ tends to $\tilde P_{F, 2\pi}$. Thus there exists some large enough $T_0$ such that for $T>T_0$, the pair of projections $(\tilde P_{F,0}^{nc};
U_T^\dagger \tilde P_{F,2\pi} U_T)$ satisfies $$||\tilde P_{F,0}^{nc} - U_T^\dagger \tilde P_{F,2\pi} U_T || < 1
\eqno(3.4)$$ Thus it is Fredholm and ${\rm Ind}(\tilde P_{F,0}^{nc};U_T^\dagger \tilde P_{F,2\pi} U_T)=0$. Since $\tilde P_{F,0}-\tilde P_{F,0}^{nc}$ is finite rank we can apply (3.1) to get $$\eqalign{
\tilde Q_F & = {\rm Ind}(\tilde P_{F,0};\tilde P_{F,0}^{nc})
\cr & = {\rm Ind}(\tilde P_{F,0};U_T^\dagger\tilde P_{F,2\pi}U_T)
+ {\rm Ind}(U_T^\dagger\tilde P_{F,2\pi}U_T; \tilde P_{F,0}^{nc})
\cr & = {\rm Ind}(\tilde P_{F,0};U_T^\dagger\tilde P_{F,2\pi}U_T)
\cr}
\eqno(3.5)$$ Finaly let $U$ be the multiplication operator by $e^{i {2\pi\over L} y}$. Since $U$ does not change the boundary conditions and $U^\dagger H(0)U=H(2\pi)$ we obtain the formula $$\tilde Q_F={\rm Ind}(\tilde P_{F,0}; U_T^\dagger U^\dagger \tilde P_{F,0} U U_T)
\eqno(3.6)$$ The same construction for $H_D(0)$ leads to $$Q_F^D={\rm Ind}(P_{F,0}^D; U_T^{D\dagger} U^\dagger P_{F,0}^D U U_T^D)
\eqno(3.7)$$ where $P_{F,0}^D$ is the projector of $H_D(0)$ onto $]-\infty, E_F]$ and $U_t^D$ is the time evolution associated to the Hamiltonian $H_D(\varphi(t))$. We remark that the identities of this paragraph can be checked by explicit computation for the simple toy Hamiltonian (1.7).
0.5cm
[**Remark:**]{} In \[18\] a different relative index for an infinite two dimensional system is studied and related to the Hall conductivity viewed as a Chern number. It would be interesting to investigate the analogous relationship in the present case with a boundary.
0.5cm
[**Equality of $\tilde Q_F$ and $Q_F^D$.**]{}
Since $V-V_D$ has a finite support $(z-H_D(0))^{-1}(V-V_D)$ is a compact operator for $z$ not in $\sigma(H_D(0))$. Therefore the resolvent identity and Cauchy’s formula imply that $\tilde P_{F,0}-P_{F,0}^D$ is compact. Thus the pair $(\tilde P_{F,0}; P_{F,0}^D)$ is Fredholm and we can apply (3.1) to get $$\eqalign{
{\rm Ind}(\tilde P_{F,0}; & U_T^\dagger U^\dagger\tilde P_{F,0}U U_T)
\cr & =
{\rm Ind}(\tilde P_{F,0}; P_{F,0}^D)+
{\rm Ind}( P_{F,0}^D;U_T^\dagger U^\dagger\tilde P_{F,0}U U_T)
\cr &
=
{\rm Ind}(\tilde P_{F,0}; P_{F,0}^D)+
{\rm Ind}( P_{F,0}^D;U_T^\dagger U^\dagger P_{F,0}^DU U_T)
\cr &
+
{\rm Ind}(U_T^\dagger U^\dagger P_{F,0}^DU U_T;
U_T^\dagger U \tilde P_{F,0}U^\dagger U_T)
\cr}
\eqno(3.8)$$ The first and third terms in the last equality of (3.8) cancel. Thus $$\eqalign{
\tilde Q_F & =
{\rm Ind}( P_{F,0}^D;U_T^\dagger U^\dagger P_{F,0}^DU U_T)
\cr &
= {\rm Ind}(P_{F,0}^D U U_T P_{F,0}^D| P_{F,0}^D{\cal H}\to
P_{F,0}^D{\cal H})
\cr}
\eqno(3.9)$$ where in the last line we introduced the Fredholm index of $P_{F,0}^D U U_T P_{F,0}^D$ viewed as a map from $P_{F,0}^D{\cal H}$ to itself (${\cal H}$ the Hilbert space of the cylinder). From Dyson’s equation $$P_{F,0}^D U U_T P_{F,0}^D - P_{F,0}^D U U_T^D P_{F,0}^D
= \int_0^T ds
P_{F,0}^D U U_{T-s}^D(V-V_D)U_s P_{F,0}^D
\eqno(3.10)$$ Therefore the Hilbert-Schmidt norm of the left hand side is smaller than $$\int_0^T ds ||P_{F,0}^D U U_{T-s}^D(V-V_D)||_{HS}
\eqno(3.11)$$ which is shown to be finite in Appendix B. Thus the difference (3.10) is compact and the two operators have the same Fredholm index $$\eqalign{
{\rm Ind}(P_{F,0}^D U U_T P_{F,0}^D & | P_{F,0}^D{\cal H}\to
P_{F,0}^D{\cal H})
\cr &
={\rm Ind}(P_{F,0}^D U U_T^D P_{F,0}^D| P_{F,0}^D{\cal H}\to
P_{F,0}^D{\cal H})
\cr}
\eqno(3.12)$$ which is equivalent to $\tilde Q_F= Q_F^D$.
0.5cm
[**End of Proof of (1.5) and (1.6).**]{}
From the analysis of section 2 we know that for $D$ large enough (say $D=O(L)$) the branches of $H_e(\Phi)$ and $H_D(\Phi)$ that belong to $\Delta$ lie close to each other within a distance $O(e^{-c\sqrt B L})$. Since the spacing of the branches of $H_e(\Phi)$ is $O(L^{-1})$ it follows that $Q_F^D= 1$ and therefore $\tilde Q_F=1$. Thus $\tilde E_k(2\pi)=\tilde E_{k+1}(0)$ and since there exists $0\leq \bar\Phi\leq 2\pi$ such that $$\tilde E_k(2\pi)-\tilde E_k(0)=2\pi {d \tilde E_k\over d\Phi}(\bar\Phi)
\eqno(3.13)$$ from (2.30) we get the lower bound $$| \tilde E_{k+1}(0)- \tilde E_k(0)|\geq {2\pi \tilde\alpha\over L}
\eqno(3.14)$$ Because $\tilde E_l(0)=E_l(0)$, this bound shows that the levels of $H(0)$ (or $H(2\pi)$) are spaced by $O(L^{-1})$. Using the spectral flow of $\tilde H(\Phi)$, together with the facts that the levels of $\tilde H(\Phi)$ and $H(\Phi)$ are separated by $O(L^{-10})$, and that ${dE_k(\Phi)\over d\Phi}$is strictly positive, one deduces that necessarily $E_k(2\pi)=E_{k+1}(0)$. Then proceeding as in (3.13) and (3.14) we obtain the lower bound (1.6). Finaly the upper bound is a consequence of $$\eqalign{
L|{dE_k\over d\Phi}(\bar\Phi)| & =<\Psi_k(\bar\Phi)|p_y-Bx+{\bar\Phi\over L}|
\Psi_k(\bar\Phi)>
\cr &
\leq
|| \Psi_k(\bar\Phi)||.||(p_y-Bx+{\bar\Phi\over L})\Psi_k(\bar\Phi)||
\cr &
\leq
(<\Psi_k(\bar\Phi)|2H(\bar\Phi)|
\Psi_k(\bar\Phi)> - <\Psi_k(\bar\Phi)|2V|
\Psi_k(\bar\Phi)>)^{1\over 2}
\cr &
\leq (2E_k(\bar\Phi)+2w)^{1\over 2}\leq (3B)^{1\over 2}
\cr}
\eqno(3.15)$$
0.5cm
We start with a sketch of preliminary estimates for the Green function of the pure magnetic problem on the cylinder of circumference $L$, $$H_0(\Phi)={1\over 2}p_x^2+{1\over 2}(p_y-Bx+{\Phi\over L})^2
\eqno(A.1)$$ Using the spectral decomposition of the Green function $G_0(z)=(H_0(\Phi)-z)^{-1}$ on a basis of eigenfunctions $$e^{i{2\pi m\over L}y }\varphi_{n,m}(x)
\eqno(A.2)$$ and the Poisson summation formula we obtain $$<x,y|G_0(\Phi)|x^\prime, y^\prime>=
\sum_{m=-\infty}^{+\infty} e^{i{\Phi\over L}(y-y^\prime-mL)}
<x,y-mL|G_{0,\infty}(z)|x^\prime,y^\prime>
\eqno(A.3)$$ where $G_{0,\infty}(z)$ is the Green function of the pure magnetic problem on the infinite two dimensional plane. In the Landau gauge (${\bf r}=(x,y)$) $$<{\bf r}|G_{0,\infty}(z)|{\bf r^\prime}>
={B\over 2}\Gamma({1\over 2}-{z\over B})
U({1\over 2}-{z\over B}, 1, {B\over 2}|{\bf r}-{\bf r^\prime}|^2)
\exp(-{B\over 4}|{\bf r}-{\bf r^\prime}|^2+{iB\over 4}(x+x^\prime)(y-y^\prime))
\eqno(A.4)$$ The presence of the Euler $\Gamma$ function indicates that the Landau levels remain unchanged on the cylinder, and $U$ is the Kummer function \[16\]. By using some technical estimates as in \[17\] one may show that for ${B\over 2}<Re z<{3B\over 2}$ the absolute value of (A.3) is bounded above by the simple expression $${cB\over \delta_0(z)}e^{-{B\over 8}|x-x^\prime|^2}\sum_{m=-1,0,+1}
S(x-x^\prime, y-y^\prime-mL)e^{-{B\over 8}(y-y^\prime-mL)^2}
\eqno(A.5)$$ where $c$ is a numerical constant independent of $B$ and $L$. The factor $S$ comes from the logarithmic divergence at coincident points $$\eqalign{
S(x-x^\prime, y-y^\prime)&=1 \qquad {\rm for} \qquad {B\over 2}|{\bf r}-{\bf r^\prime}|^2>1
\cr &
=\ln {B\over 2}|{\bf r}-{\bf r^\prime}|^2 \qquad {\rm otherwise}
\cr}
\eqno(A.6)$$ A bound similar to (A.5) holds for $|\partial_x<{\bf r}|G_{0,\infty}(z)|{\bf r^\prime}>|$, with $cB$ replaced by $cB^{3\over 2}$ and $S$ replaced by ${|x-x^\prime|\over
|{\bf r}-{\bf r^\prime}|^2}$ when ${B\over 2}|{\bf r}-{\bf r^\prime}|^2<1$. The important feature for the subsequent estimates is that all the above singularities are integrable. In what follows $c$ denotes a generic numerical positive constant.
0.5cm
[**Estimate of $||K_e||$.**]{}
From the resolvent identity $$K_e(z)={1\over 2}[p_x^2, J_e]G_0(z)\chi_e+{1\over 2}[p_x^2, J_e]G_0(z)WG_e(z)\chi_e
\eqno(A.7)$$ Evaluating the commutator, and using $||G_e(z)||\leq \delta_e(z)^{-1}$ we find $$||K_e(z)||\leq {1\over 2}||J_e^{\prime\prime}G_{0}(z)\chi_e||+
||J_e^{\prime}\partial_xG_{0}(z)\chi_e|| +
\delta_e(z)^{-1}(||J_e^{\prime\prime}G_{0}(z)W||+
||J_e^{\prime}\partial_xG_{0}(z)W||)
\eqno(A.8)$$ Estimate (2.7) follows from the fact that all norms on the right hand side of (A.8) involve matrix elements of $G_0(z)$ and $\partial_x G_0(z)$ separated by a distance at least equal to ${D\over 4}$. We use the estimate ($A$ an operator with kernel $A({\bf r}, {\bf r^\prime})$) $$||A||\leq {\rm max} \biggl(\sup_{\bf r^\prime}\int d{\bf r}|A({\bf r}, {\bf r^\prime})|;
\sup_{\bf r}\int d{\bf r^\prime}|A({\bf r}, {\bf r^\prime})|\biggr)
\eqno(A.9)$$ For the first norm we have $$\eqalign{
\int_{-{3D\over 4}-1}^{-{3D\over 4}+1} dx \int_{L\over 2}^{L\over 2} dy
& J_e^{\prime\prime}(x)
|<{\bf r}|G_0(z)|{\bf r^\prime}>|\chi_e(x^\prime)
\leq
{cBL\over \delta_0(z)}\int_{-{3D\over 4}-1}^{-{3D\over 4}+1} dx
e^{-{B\over 8}|x-x^\prime|^2}\chi_e(x^\prime)\cr &
\leq
{c \sqrt B L\over \delta_0(z)}e^{-cBD^2}
\cr}
\eqno(A.10)$$ In the first inequality we used (A.5) and in the last one we use the fact that $|x-x^\prime|\geq {D\over 4}$. On the other hand $$\eqalign{
J_e^{\prime\prime}(x) & \int_{-{D\over 2}}^{\infty} dx^\prime\int_{-{L\over 2}}^{{L\over 2}} dy^\prime
|<{\bf r}|G_0(z)|{\bf r^\prime}>|\chi_e(x^\prime)\cr &
\leq
{cBL\over \delta_0(z)} J_e^{\prime\prime}(x) \int_{-{D\over 2}}^{\infty} dx^\prime
e^{-{B\over 8}|x-x^\prime|^2}\chi_e(x^\prime)
\leq {c\sqrt BL\over \delta_0(z)}e^{-cBD^2}
\cr}
\eqno(A.11)$$ Thus $||J_e^{\prime\prime} G_0(z) \chi_e||\leq {CL^2\over \delta_0(z)}e^{-cBD^2}$. For the term involving $\partial_xG_0(z)$ the estimates are similar. The terms involving $W$ lead to the same estimates provided $$\int_{-{3D\over 4}-1}^{-{3D\over 4}+1}dx e^{-{B\over 8}|x-x^\prime|^2} U(x^\prime)
\qquad {\rm and} \qquad
J_e^{\prime\prime}(x)\int_{0}^{\infty} dx^\prime e^{-{B\over 8}|x-x^\prime|^2} U(x^\prime)
\eqno(A.12)$$ are bounded by $O(\exp(-cBD^2))$. This is the case for the class of functions $W(x)$ that grow polynomialy as $x\to +\infty$.
0.5cm
[**Estimate for $||K_b||$**]{}
First we sketch the derivation of an estimate for the kernel of $G_b(z)$ and its derivative for $z$ in the gap of $\sigma(H_b(\Phi))$. $$\eqalign{
&<{\bf r}|G_b(z)|{\bf r^\prime}>=<{\bf r}|G_0(z)|{\bf r^\prime}>
+ \sum_{m\geq 1}\int {\bf dr_1}...\int{\bf dr_m}
<{\bf r}|G_0(z)|{\bf r_1}>V({\bf r_1})\cr&\times<{\bf r_1}|G_0(z)|{\bf r_2}>V({\bf r_2})
...V({\bf r_m})<{\bf r_m}|G_0(z)|{\bf r^\prime}>\cr}
\eqno(A.13)$$ Here the range of the integrals over $x_1,...,x_m$ is $]-\infty, +\infty[$, and that of $y_1,...,y_m$ is $[-{L\over 2},{L\over 2}]$. In order to extract the decay for $|x-x^\prime|$ large from (A.13) and (A.5) we use, from $B|x-x^\prime|^2>2\sqrt B |x-x^\prime|-1$, $$\eqalign{
& e^{-{B\over 8}(|x-x_1|^2+|x_1-x_2|^2+...+|x_m-x^\prime|^2)}
\leq
e^{-{B\over 16}(|x-x_1|^2+|x_1-x_2|^2+...+|x_m-x^\prime|^2)}\cr &
\times e^{-{\sqrt B\over 8}(|x-x_1|+|x_1-x_2|+...+|x_m-x^\prime|)}
e^{m\over 16}
\cr & \leq e^{m\over 16}e^{-{\sqrt B\over 8}|x-x^\prime|}
e^{-{B\over 16}(|x-x_1|^2+|x_1-x_2|^2+...+|x_m-x^\prime|^2)}
\cr}
\eqno(A.14)$$ Thanks to (A.5), (A.13), (A.14) we obtain for ${B\over 2}|x-x^\prime|>1$ $$\eqalign{
|<{\bf r}|G_b(z)|{\bf r^\prime}>| & \leq {cB\over \delta_0(z)}e^{-{B\over 8}|x-x^\prime|^2}
+\sum_{m\geq 1} ({cB\over \delta_0(z)})^{m+1}({w\over B})^m
e^{-{\sqrt B\over 8}|x-x^\prime|}
\cr &
\leq
{cB\over \delta_0(z)-cw} e^{-{\sqrt B\over 8}|x-x^\prime|}
\cr}
\eqno(A.15)$$ This bound is valid as long as $w$ is small enough. Clearly from (A.13), following the same steps, we obtain a similar inequality, with $cB$ replaced by $cB^{3\over 2}$, for $|\partial_x<{\bf r}|G_b(z)|{\bf r^\prime}>|$ if ${B\over 2}|x-x^\prime|>1$.
To estimate $||K_b||$ we have to compute the norms on the right hand side of $$||K_b||\leq {1\over 2}||J_b^{\prime\prime}G_b(z)\chi_b||+||J_b^\prime\partial_xG_b(z)\chi_b||
\eqno(A.16)$$ This can be done easily using (A.9), (A.16) and the bound (A.15) together with that on the derivative. Then one finds $$||K_b||\leq {cB^{3\over 2}L\over \delta_0(z)-cw} e^{-c\sqrt B D}
\eqno(A.17)$$
0.5cm
By Cauchy’s formula, and the resolvent identity $$P_{F,0}^D=\int_{\Gamma_F} dz {1\over z-H_D(0)}=
\int_{\Gamma_F} dz {1\over z-H_0(0)}+\int_{\Gamma_F} dz {1\over z-H_D(0)}(W+V_D){1\over z-H_0(0)}
\eqno(B.1)$$ where the contour $\Gamma_F$ encloses the part of the spectrum of $H_D(0)$ lying below $E_F$. Setting $g=U U_{T-s}^D(V-V_D)$ we have for the Hilbert-Schmidt norm $$\eqalign{
&||P_{F,0}^Dg||_{HS}\leq |\Gamma_F|{\rm sup}_{z\in \Gamma_F}||{1\over z-H_0(0)}g||_{HS}
\cr& +{|\Gamma_F|\over {\rm dist}(E_F, \sigma(H_D(0)))}
\biggl({\rm sup}_{z\in \Gamma_F}||W{1\over z-H_0(0)}g||_{HS}
+w{\rm sup}_{z\in \Gamma_F}||{1\over z-H_0(0)}g||_{HS}\biggr)
\cr}
\eqno(B.2)$$ Here $|\Gamma_F|$ is the length of the contour which is finite because the spectrum is bounded below. Since $V-V_D$ has compact support, $g$ is a square integrable function on the cylinder. Therefore from the bound (A.5), (A.6) on the kernel of $(z-H_0)^{-1}$ it is easily seen that all the Hilbert-Schmidt norms in (B.2) are finite. These norms can be bounded above uniformly in $0\leq s\leq T$, and the supremum over $z$ stays finite as long as the contour does not touch a Landau level. Therefore (3.11) is finite.
0.5cm
[**ACKNOWLEDGEMENTS.**]{} I wish to thank Jürg Fröhlich for drawing my attention on the spectral flow.
\[1\] R. B. Laughlin, “Quantized Hall conductivity in two dimensions”, Phys. Rev. B[**23**]{}, 5632-5633 (1981)
\[2\] B. I. Halperin, “Quantized Hall conductance, current carrying edge states, and the existence of extended states in a two dimensional disordered potential”, Phys. Rev B[**25**]{}, 2185-2190 (1982)
\[3\] N. Macris, Ph. A. Martin, J. V. Pulé, “On edge states in semi-infinite quantum Hall systems”, J. Phys. A [**32**]{}, 1985-1996 (1999)
\[4\] J. Fröhlich, G. M. Graf, J. Walcher, “On the extended nature of edge states of quantum Hall hamiltonians”, Ann. H. Poincaré [**1**]{}, 405 (2000)
\[5\] S. De Bievre, J. V. Pulé, “Propagating edge states for a magnetic hamiltonian”, Elect. J. Math. Phys. [**5**]{} (1999); http://mpej.unige.ch/mpej/MPEJ.html
\[6\] N. Macris, Ch. Ferrari, preprint EPFL.
\[7\] S. Molchanov, “The local structure of the spectrum of the one dimensional Schroedinger operator”, Comm. Math. Phys [**78**]{}, 429-446 (1981)
\[8\] N. Minami, “Local fluctuation of the spectrum of a multidimensional Anderson tight binding model”, Comm. Math. Phys [**177**]{}, 709-725 (1996)
\[9\] B. I. Shklovskii, B. Shapiro, B. R. Sears, P. Lambrianides, H. B. Shore, “Statistics of spectra of disordered systems near the metal insulator transition”, Phys. Rev. B [**47**]{} 11487-11490 (1993)
\[10\] J. Kellendonk, T. Richter, H. Schulz-Baldes, “Edge versus bulk currents in the integer quantum Hall effect”, J. Phys. A: Math Gen [**33**]{}, 27-32 (2000); and “Edge current channels and Chern numbers in the integer quantum Hall effect”, mp-arc/00-266
\[11\] T. Kato, “Perturbation theory of linear operators”, Springer Verlag, Berlin (1980)
\[12\] P. Briet, J. M. Combes, P. Duclos, “Spectral stability under tunneling”, Comm. Math. Phys [**1206**]{}, 133 (1989)
\[13\] F. Bentosela, V. Grechi, “Stark Wannier ladders”, Comm. Math. Phys [**142**]{}, 169 (1991)
\[14\] J. E. Avron, R. Seiler, B. Simon, “The index of a pair of projections”, J. Funct. Anal, 220-237 (1994)
\[15\] A. Messiah “Quantum mechanics”, volume II, North Holland publishing company, Amsterdam (1961)
\[16\] M. Abramovitz, I. A. Stegun, “Handbook of mathematical functions”, Dover Publications, New York (1965); see page
\[17\] T. Dorlas, N. Macris, J. V. Pulé, “Characterisation of the spectrum of the Landau hamiltonian with delta impurities”, Comm. Math. Phys [**204**]{}, 367-396 (1999)
\[18\] J. E. Avron, R. Seiler, B. Simon, “Charge deficiency, charge transport and comparison of dimensions”, Comm. Math. Phys. [**159**]{}, 399-422 (1994)
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Exoplanets, and in particular hot ones such as hot Jupiters, require a very significant quantities of molecular spectroscopic data to model radiative transport in their atmospheres or to interpret their spectra. This data is commonly provided in the form of very extensive transition line lists. The size of these line lists is such that constructing a single model may require the consideration of several billion lines. We present a procedure to simplify this process based on the use of cross sections. Line lists for water, H$_3^+$, HCN /HNC and ammonia have been turned into cross sections on a fine enough grid to preserve their spectroscopic features. Cross sections are provided at a fixed range of temperatures and an interpolation procedure which can be used to generate cross sections at arbitrary temperatures is described. A web-based interface (www.exomol.com/xsec) has been developed to allow astronomers to download cross sections at specified temperatures and spectral resolution. Specific examples are presented for the key water molecule.'
address: 'Department of Physics and Astronomy, University College London, Gower Street, WC1E 6BT London, UK'
author:
- 'Christian Hill, Sergei N. Yurchenko and Jonathan Tennyson'
title: 'Temperature-dependent molecular absorption cross sections for exoplanets and other atmospheres'
---
Atmospheres, composition ,Extra-solar planets ,Infrared observations ,Radiative transfer
Introduction {#intro}
============
With the growing realization that many, probably most, stars support exoplanets, developing the means to systematically characterize the atmospheres of these planets has become a major scientific priority [@jt523]. Given the likely complex chemistry of these atmospheres and the elevated temperature that is found in the most observable planets, there is a significant demand for spectroscopic data on the probable exoplanet atmospheric constituents.
Recently we have launched a new project, called ExoMol (see www.exomol.com), with the aim of providing molecular transition data appropriate for exoplanet models which is reliable over a wide range of temperatures [@jt529]. The ExoMol project involves constructing line lists of spectroscopic transitions for key molecules which are valid over the entire temperature and wavelength domain that is likely to be astrophysically important for these species. Especially for polyatomic molecules, these line lists can become very large with hundreds of millions [@jt374; @jt378; @jt469; @11TaPe.CO2] or even billions [@jt500] of individual transitions needing to characterized and stored. A complete linelist for methane, for which so far only a preliminary version is available [@09WaScSh.CH4], can be expected to be even larger. Indeed potential line lists for larger species, such as higher hydrocarbons, for which spectroscopic data is needed for exoplanetary research, are likely to be so large as to potentially make their use impractical.
Molecular line lists are being actively used to model the spectra of exoplanets (eg @jt495) and cool brown dwarfs with similar temperatures (eg @jt484 [@ckg11]). However, sampling billions of individual transitions to model relatively low resolution astronomical spectra is probably not necessary in many cases. An alternative approach is to represent the molecular absorptions as cross sections generated at an appropriate resolution and temperature. The advantage of this approach is that the data handling issues related to dealing with large data sets largely disappear. The disadvantage is that cross sections are inflexible - a particular cross section set is only valid for a single state of temperature and pressure. Cross sections are therefore often regarded a second choice compared to maintaining a full line list [@jt453].
In this paper we develop a strategy whereby cross sections are provided for the user in a flexible fashion without loss of accuracy or the specificity of using a complete line list. To this end we have provided a web application which, starting from very high resolution cross sections generated for each molecule at a range of temperatures, can provide cross sections at a temperature and resolution specified by the user. Of course this approach is based on the implicit assumption of local thermodynamic equilibrium (LTE) and any non-LTE treatment will continue to have to rely on the explicit use of transition line lists. So far, these cross sections do not consider collisional broadening effects and are therefore, at their highest resolution, appropriate for the zero pressure limit only.
The line lists for water [@jt378; @jt469], H$_3^+$ [@jt181; @jt478], HCN /HNC [@jt298; @jt374; @jt447] and ammonia [@jt500] were used to generate cross sections for these species. For concreteness, this work uses the main water isotopologue, [$\mathrm{H_2}\mathrm{^{16}O}$]{}, as its working example. Water is known to be a key species in exoplanetary atmospheres and the BT2 line list has been used in studies of exoplanets [@jt400; @Swain09; @tinetti10a; @baraffe10; @tinetti10b; @Shabram11; @Barman11; @jt516] as well in a large variety of planetary [@jt431; @jt463; @jeremy09], astrophysical [@Warren07; @jt330; @jt349; @Burgasser08; @jt452; @jt417; @jt357] and, indeed, engineering [@Kr07; @Lindermeir20121575] studies which generally focus on the radiative transport by hot water. The BT2 line list was used as part of the recently updated HITEMP database [@jt480]. In that work the size of the line list was reduced using a technique based upon importance sampling at a range of key temperatures. In practice the number of water lines in HITEMP remains large, over 100 million.
The calculation of opacities and other spectral properties due to water vapour at these elevated temperatures can therefore become onerous, and so we present here pre-calculated absorption cross sections for a range of temperatures between 296 K and 3000 K, binned to different resolutions. The highest resolution cross sections are suitable for modelling low-density environments where only Doppler broadening contributes to the line width whereas by binning to a wavenumber grid spacing significantly larger than the pressure-broadened half-width, higher-density environments are described well by the calculated cross sections. However, no attempt is made to include contributions to the opacity from the water vapour continuum or water dimer absorption.
Method
======
The high-resolution cross section is calculated on an evenly-spaced wavenumber grid, ${\ensuremath{\tilde{\nu}_i}}$, defining bins of width ${\ensuremath{\Delta\tilde{\nu}}}$. Only Doppler broadening is considered so each absorption line has a Gaussian shape: $$\begin{aligned}
f_\mathrm{G}(\tilde{\nu}; {\ensuremath{\tilde{\nu}_{0;j}}}, \alpha_j) = \sqrt{\frac{\ln 2}{\pi}}\frac{1}{\alpha_j}\exp\left( -\frac{(\tilde{\nu}-{\ensuremath{\tilde{\nu}_{0;j}}})^2 \ln 2}{\alpha_j^2} \right)\end{aligned}$$ where the line centre position is ${\ensuremath{\tilde{\nu}_{0;j}}}$ and the Doppler half-width at half-maximum, $$\begin{aligned}
\alpha_j = \sqrt{\frac{2k_\mathrm{B}T\ln 2}{m}}\frac{{\ensuremath{\tilde{\nu}_{0;j}}}}{c},\end{aligned}$$ at temperature $T$ for a molecule of mass $m$.
![The calculation of the absorption cross section in a wavenumber bin centered on ${\ensuremath{\tilde{\nu}_i}}$ due to a single line. The integrated line intensity within the shaded region, of width ${\ensuremath{\Delta\tilde{\nu}}}$, contributes to $\sigma_{ij}$.[]{data-label="fig:line_fig"}](line_fig.png){width="14cm"}
The contribution to the cross section within each bin is a sum over contributions from individual lines: $$\begin{aligned}
\sigma_i = \sum_j \sigma_{ij}\end{aligned}$$ where $$\begin{aligned}
\label{eqn:sig_ij}
\sigma_{ij} &= \frac{S_j}{{\ensuremath{\Delta\tilde{\nu}}}} \int_{{\ensuremath{\tilde{\nu}_i}}-{\ensuremath{\Delta\tilde{\nu}}}/2}^{{\ensuremath{\tilde{\nu}_i}}+{\ensuremath{\Delta\tilde{\nu}}}/2} f_\mathrm{G}(\tilde{\nu}; {\ensuremath{\tilde{\nu}_{0;j}}}, \alpha_j) \,\mathrm{d}\tilde{\nu}\\
&= \frac{S_j}{2{\ensuremath{\Delta\tilde{\nu}}}} \left[ {\ensuremath{\mathrm{erf}(x_{ij}^+)}} - {\ensuremath{\mathrm{erf}(x_{ij}^-)}} \right],\end{aligned}$$
where erf is the error function and $$\begin{aligned}
x_{ij}^\pm = \frac{\sqrt{\ln 2}}{\alpha_j}\left[ {\ensuremath{\tilde{\nu}_i}}\pm \frac{{\ensuremath{\Delta\tilde{\nu}}}}{2} - {\ensuremath{\tilde{\nu}_{0;j}}}\right]\end{aligned}$$ are the scaled limits of the wavenumber bin centred on ${\ensuremath{\tilde{\nu}_i}}$ relative to the line centre, ${\ensuremath{\tilde{\nu}_{0;j}}}$, and the line intensity in units of $\mathrm{cm^{-1}/(molecule\,cm^{-2})}$ is $$\begin{aligned}
S_j = \frac{A_j}{8\pi c}\frac{g_j'\mathrm{e}^{-c_2E_j{''}/T}}{{\ensuremath{\tilde{\nu}_{0;j}}}^2 Q(T)}\left( 1 - \mathrm{e}^{-c_2{\ensuremath{\tilde{\nu}_{0;j}}}/T} \right).\end{aligned}$$ Here, $g_j'$ and $E_j{''}$ are the upper-state degeneracy and lower-state energy respectively, $A_j$ is the Einstein $A$ coefficient for the transition and $c_2 \equiv hc/k_\mathrm{B}$ is the second radiation constant. For [$\mathrm{H_2}\mathrm{^{16}O}$]{}, the molecular partition function, $Q(T)$, was obtained from the tabulated values of @jt263.
Note that in the limit of ${\ensuremath{\Delta\tilde{\nu}}}\gg \alpha_j$, eqn (\[eqn:sig\_ij\]) reduces to $$\begin{aligned}
\sigma_{ij} &\approx \frac{S_j}{{\ensuremath{\Delta\tilde{\nu}}}} \int_{-\infty}^{+\infty} f_\mathrm{G}(\tilde{\nu}; {\ensuremath{\tilde{\nu}_{0;j}}}, \alpha_j) \,\mathrm{d}\tilde{\nu} = \frac{S_j}{{\ensuremath{\Delta\tilde{\nu}}}},\end{aligned}$$ whereas for ${\ensuremath{\Delta\tilde{\nu}}}\ll \alpha_j$, $$\begin{aligned}
\sigma_{ij} &\approx S_j f_\mathrm{G}({\ensuremath{\tilde{\nu}_i}}; {\ensuremath{\tilde{\nu}_{0;j}}}, \alpha_j).\end{aligned}$$ However, the exact expression in all calculations of the cross sections presented in this work.
Results
=======
The absorption cross section of [$\mathrm{H_2}\mathrm{^{16}O}$]{} was calculated between $10\;{\ensuremath{\mathrm{cm^{-1}}}}$ and $30000\;{\ensuremath{\mathrm{cm^{-1}}}}$ across the temperature range 296 K – 3000 K (Table \[tab:T-summary\]), using the wavenumber grid-spacing given in Table \[tab:dnu-summary\].
-------- -------- -------- --------
296 K 400 K 500 K 600 K
700 K 800 K 900 K 1000 K
1200 K 1300 K 1400 K 1600 K
1800 K 2000 K 2200 K 2400 K
2600 K 2800 K 3000 K
-------- -------- -------- --------
: Temperatures at which calculated [$\mathrm{H_2}\mathrm{^{16}O}$]{} cross sections are provided.[]{data-label="tab:T-summary"}
wavenumber range $/{\ensuremath{\mathrm{cm^{-1}}}}$ ${\ensuremath{\Delta\tilde{\nu}}}/{\ensuremath{\mathrm{cm^{-1}}}}$
----------------------------------------------------- --------------------------------------------------------------------
10 – 100 $10^{-5}$
100 – 1000 $10^{-4}$
1000 – 10000 $10^{-3}$
10000 – 30000 $10^{-2}$
: Summary of the grid spacings, ${\ensuremath{\Delta\tilde{\nu}}}$ for the cross sections calculated in different wavenumber regions[]{data-label="tab:dnu-summary"}
For comparison with experimental spectra, low-resolution cross sections were produced by binning the high-resolution cross sections to the following fixed grid spacing across the entire wavenumber range: ${\ensuremath{\Delta\tilde{\nu}}}= 0.01$, 0.1, 1, 10, $100\;{\ensuremath{\mathrm{cm^{-1}}}}$. At these resolutions, the structure due to individual lines is lost and direct comparison can be made with, for example, the experimental water vapour cross sections of the PNNL database [@sjs04]. Such a comparison for the ${\ensuremath{\Delta\tilde{\nu}}}= 10\;{\ensuremath{\mathrm{cm^{-1}}}}$ resolution spectra is shown in Figure \[fig:PNNL-comparison\].
![Comparison of the calculated [$\mathrm{H_2}\mathrm{^{16}O}$]{} cross section presented in this work (blue) with the experimental cross section of the PNNL database [@sjs04] (green) in the region of the fundamental $\nu_2$ bending mode, at 296 K, both binned to a $10\;{\ensuremath{\mathrm{cm^{-1}}}}$ wavenumber grid. Also shown is the difference (this work - PNNL) between the two spectra (red).[]{data-label="fig:PNNL-comparison"}](PNNL-comparison.png){width="14cm"}
Interpolation of cross sections between temperatures
====================================================
For use in the web-based application described below, cross sections for the molecules given in Table \[tab:species\] have been calculated using a wavenumber grid spacing of $0.01\;\mathrm{cm^{-1}}$ at a range of temperatures in 100 K intervals below 1000 K and 200 K intervals above 1000 K. A cross section at some intermediate temperature between the values at which the stored cross sections have been calculated may be obtained by interpolation. Suppose that $\sigma_i(T_1)$ and $\sigma_i(T_2)$ are calculated cross sections at temperatures which bracket the temperature of the desired cross section: $T_1 < T < T_2$ (we consider interpolation using only $\sigma_i$ calculated at the two temperatures closest to $T$). One possible interpolation strategy is the linear interpolation $$\begin{aligned}
\sigma_i(T) = \sigma_i(T_1) + m(T-T_1), \; \mathrm{where} \; m = \frac{\sigma_i(T_2)-\sigma_i(T_1)}{T_2-T_1}.\end{aligned}$$ However, we find a more accurate approach is to estimate the temperature dependence to be of the form $$\begin{aligned}
\sigma_i(T) = a_i\mathrm{e}^{-b_i/T},\end{aligned}$$ where the coefficients $a_i$ and $b_i$ at each wavenumber bin may be calculated from $$\begin{aligned}
b_i = \left(\frac{1}{T_2}-\frac{1}{T_1}\right)^{-1}\ln \frac{\sigma_i(T_1)}{\sigma_i(T_2)}\;\;\;\mathrm{and} \;\;\; a_i = \sigma_i(T_1)\mathrm{e}^{b_i/T_1}.\end{aligned}$$ The largest values of the interpolation residual error in the region 1000 - 20000 [$\mathrm{cm^{-1}}$]{}, calculated as $\delta\sigma_i = \sigma_{i,\mathrm{calc}} - \sigma_{i,\mathrm{interp}}$, are found to be associated with the $\nu_2$ band - as an illustration, this is plotted in Figure \[fig:interp-error\] at 350 K. The maximum value of the interpolation residual across this wavenumber region at a range of temperatures and wavenumber binning intervals is given in Table \[tab:interp-error\], expressed as a percentage of the corresponding absorption cross section: $$\begin{aligned}
\delta\sigma_\mathrm{max}^\% = \max \left(\frac{|\sigma_{i,\mathrm{calc}} - \sigma_{i,\mathrm{interp}}|}{\sigma_{i,\mathrm{calc}}} \right) \times 100.\end{aligned}$$ In all cases, $\delta\sigma_\mathrm{max}^\%$ is found to be less than the estimated uncertainty in the *ab initio* line intensities that the cross section calculation is based on. Interpolation is performed on the 0.01 [$\mathrm{cm^{-1}}$]{} grid before binning to a coarser wavenumber grid, if required.
![Calculated absorption cross section (upper pane) and interpolation residual (lower pane) in the region of the fundamental $\nu_2$ bending mode, at 350 K, on a wavenumber grid spacing of 0.01 [$\mathrm{cm^{-1}}$]{}. The interpolation residual error does not exceed 1.34%.[]{data-label="fig:interp-error"}](interp-error.png){width="14cm"}
$\Delta\tilde{\nu}$ 0.01 [$\mathrm{cm^{-1}}$]{} 0.1 [$\mathrm{cm^{-1}}$]{} 1.[$\mathrm{cm^{-1}}$]{} 10\. [$\mathrm{cm^{-1}}$]{}
--------------------- ----------------------------- ---------------------------- -------------------------- -----------------------------
350 K 1.64 % 1.34 % 1.07 % 1.10 %
1100 K 1.00 % 0.50 % 0.49 % 0.46 %
2500 K 0.66 % 0.40 % 0.37 % 0.36 %
: Maximum interpolation errors in the [$\mathrm{H_2}\mathrm{^{16}O}$]{} cross section as a function of wave number grid spacing and temperature[]{data-label="tab:interp-error"}
Finally we note that [@12HaLiBe.NH3] recently presented an experimental ammonia spectrum recorded at a range of temperatures at 100 K intervals. We suggest that our proposed interpolation scheme would also be appropriate for interpolating their data.
Web based application
=====================
Calculated absorption cross sections can be obtained from the interface at the url http://www.exomol.com/xsecs. The user of this web-based interface can select a wavenumber range, temperature and wavenumber grid spacing; using these parameters the interface software first obtains a high-resolution cross section at the desired temperature by the interpolation procedure described in the previous section on the pre-calculated spectra, and then bins this interpolated cross section to the requested wavenumber grid.
Cross sections are returned as a list of floating point numbers in a text file, separated by the Unix-style newline character, `LF` (‘\\n’, `0x0A`). The wavenumber grid can be generated from the linear sequence $$\tilde{\nu}_i = \tilde{\nu}_\mathrm{min} + i\Delta\tilde{\nu}; \quad i = 0,1,2,\cdots,n-1$$ where the total number of points in the requested cross section is $$n = \frac{\tilde{\nu}_\mathrm{max} - \tilde{\nu}_\mathrm{min}}{\Delta\tilde{\nu}} + 1.$$ We also provide an XML file in XSAMS format [@xsams01], compatible with the standards of the VAMDC project [@jt481]. This file may be thought of as a ‘wrapper’ to the cross section data, providing contextual metadata such as the molecular identity and structure, temperature of the calculation, and wavenumber limits and grid spacing. An example of the format is given in Table \[tab:xsams-example\].
<AbsorptionCrossSection envRef="EEXOMOL-1" id="PEXOMOL-XSC-1">
<Description>
The absorption cross section for H2O at 296.0 K, calculated at
Sun Mar 11 19:50:45 2012, retrieved from www.exomol.com/xsecs
</Description>
<X parameter="nu" units="1/cm">
<LinearSequence count="901" initial="1000." increment="10."/>
</X>
<Y parameter="sigma" units="cm2">
<DataFile>H2O_1000-20000_296K-10.0.sigma</DataFile>
</Y>
<Species>
<SpeciesRef>XEXOMOL-XLYOFNOQVPJJNP-FNDQEIABSA-N</SpeciesRef>
</Species>
</AbsorptionCrossSection>
Species $\ensuremath{\tilde{\nu}_{\mathrm{max}}}$/[$\mathrm{cm^{-1}}$]{} $T_{\mathrm{max}}$/K Reference
------------------------- ------------------------------------------------------------------ ---------------------- -----------------
H$_3^+$ 10 000 4000 @jt181
H$_2$D$^+$ 10 000 4000 @jt478
H$_2$O 20 000 3000 @jt378
HDO 17 000 3000 @jt469
HCN / HNC 10 000 4000 @jt298 [@jt374]
H$^{13}$CN / H$^{13}$NC 10 000 4000 @jt447
NH$_3$ 12 000 1500 @jt500
: Summary of species for which are cross sections currently available. Also given for each species is the maximum wavenumber ($\ensuremath{\tilde{\nu}_{\mathrm{max}}}$), the maximum temperature ($T_{\mathrm{max}}$) and the reference to the original line list.[]{data-label="tab:species"}
Cross section files have been generated for the polyatomic line lists currently available on the ExoMol website. These are listed in Table \[tab:species\]. The table also specifies the maximum wavenumber ($\ensuremath{\tilde{\nu}_{\mathrm{max}}}$) and maximum temperature ($T_{\mathrm{max}}$) for each species; we strongly caution against relying on the cross sections or indeed the underlying line lists at temperatures greater than those given. Further cross sections will be provided as line lists for new species as they become available.
Conclusion
==========
High resolution absorption cross sections have been calculated for a number of molecules likely to be important in the atmospheres of exoplanets. The online interface provided at the ExoMol website (www.exomol.com) allows customized cross sections for a given molecular species to be returned at a specified temperature and resolution. Cross sections are only available for those species for which extensive line lists exist. New cross sections will be provided as further species are added to the ExoMol database, see @jt529 for example.
It is our intention to make the cross section facility in ExoMol fully interoperable with other spectroscopic databases as part of the VAMDC (Virtual Atomic and Molecular Data Centre) project [@jt481]. Work in this direction will be reported in due course.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Giovanna Tinetti and Bob Barber for many helpful discussions during the course of this work. This work was performed as part of ERC Advanced Investigator Project 267219 and the project VAMDC which is funded by the European Union INFRA-2008-1.2.2 Scientific Data Infrastructure program under Grant Agreement number 239108.
[42]{} natexlab\#1[\#1]{}url \#1[`#1`]{}urlprefix
Bailey, J., 2009. A comparison of water vapor line parameters for modeling the venus deep atmosphere. Icarus 201, 444 – 453.
Banerjee, D. P. K., Barber, R. J., Ashok, N. K., Tennyson, J., 2005. [Near-infrared water lines in V838 Monocerotis]{}. Astrophys. J. 672, L141–L144.
Baraffe, I., Chabrier, G., Barman, T., [2010]{}. [The physical properties of extra-solar planets]{}. Rep. Prog. Phys. [73]{}, 016901.
Barber, R. J., Miller, S., [Dello Russo]{}, N., Mumma, M. J., Tennyson, J., Guio, P., 2009. [Water in the near IR spectrum of comet 8P/Tuttle]{}. Mon. Not. R. Astr. Soc. 398, 1593–1600.
Barber, R. J., Tennyson, J., Harris, G. J., Tolchenov, R. N., 2006. [A high accuracy computed water line list]{}. Mon. Not. R. Astr. Soc. 368, 1087–1094.
Barman, T. S., Macintosh, B., Konopacky, Q. M., Marois, C., [2011]{}. Clouds and chemistry in the atmosphere of extrasolar planet hr8799b. Astrophys. J. [733]{}, 65.
Beaulieu, J. P., Tinetti, G., Kipping, D., Ribas, I., Barber, R. J., Cho, J. Y.-K., Polichtchouk, I., Tennysson, J., Yurchenko, S. N., Griffith, C. A., Waldmann, I., Miller, S., Carey, S., Mousis, O., Fossey, S. J., Aylward, A., 2011. [Probing the atmosphere of the transiting hot Neptune GJ436b]{}. Astrophys. J. 731, 18.
Burgasser, A. J., Liu, M. C., Ireland, M. J., Cruz, K. L., Dupuy, T. J., [2008]{}. Subtle signatures of multiplicity in late-type dwarf spectra: The unresolved m8.5+t5 binary 2mass j03202839-0446358. Astrophys. J. [681]{}, 579–593.
Bykov, A. D., Lavrientieva, N. N., Mishina, T. P., Sinitsa, L. N., Barber, R. J., Tolchenov, R. N., Tennyson, J., 2008. [Water vapor line width and shift calculations with accurate vibration-rotation wave functions]{}. J. Quant. Spectrosc. Radiat. Transf. 109, 1834–1844.
Chesnokova, T. Y., Voronin, B. A., Bykov, A. D., Zhuravleva, T. B., Kozodoev, A. V., Lugovskoy, A. A., Tennyson, J., 2009. [Calculation of solar radiation atmospheric absorption with different H$_2$O spectral line data banks]{}. J. Mol. Spectrosc. 256, 41–44.
Cushing, M. C., Kirkpatrick, J. D., Gelino, C. R., Griffith, R. L., Skrutskie, M. F., Mainzer, A., Marsh, K. A., Beichman, C. A., Burgasser, A. J., Prato, L. A., Simcoe, R. A., Marley, M. S., Saumon, D., Freedman, R. S., Eisenhardt, P. R., Wright, E. L., [2011]{}. The discovery of y dwarfs using data from the wide-field infrared survey explorer [(WISE)]{}. Astrophys. J. [743]{}.
, N., Bonev, B. P., DiSanti, M. A., Gibb, E. L., Mumma, M. J., Magee-Sauer, K., Barber, R. J., Tennyson, J., 2005. [Water production rates, rotational temperatures and spin temperatures in comets C/1999 H1 (Lee), C/1999 S4 and C/2001 A2]{}. Astrophys. J. 621, 537–544.
, N., DiSanti, M. A., Magee-Sauer, K., Gibb, E. L., Mumma, M. J., Barber, R. J., Tennyson, J., 2004. [Accurate rotational temperature retrievals from hot-band line near 2.9 $\mu$m]{}. Icarus 168, 186–200.
Dubernet, M. L., Boudon, V., Culhane, J. L., Dimitrijevic, M. S., Fazliev, A. Z., Joblin, C., Kupka, F., Leto, G., Le Sidaner, P., Loboda, P. A., Mason, H. E., Mason, N. J., Mendoza, C., Mulas, G., Millar, T. J., Nunez, L. A., Perevalov, V. I., Piskunov, N., Ralchenko, Y., Rixon, G., Rothman, L. S., Roueff, E., Ryabchikova, T. A., Ryabtsev, A., Sahal-Brechot, S., Schmitt, B., Schlemmer, S., Tennyson, J., Tyuterev, V. G., Walton, N. A., Wakelam, V., Zeippen, C. J., 2010. [Virtual Atomic and Molecular Data Centre]{}. J. Quant. Spectrosc. Radiat. Transf. 111, 2151–2159.
Dubernet, M. L., Humbert, D., Ralchenko, Y., Roueff, E., Schultz, D. R., Chung, H.-K., Braams, B., 2009. XSAMS: XML Schema for Atomic, Molecular and Solid Data, version 0.1. Internation Atomic Energy Authority, <http://www-amdis.iaea.org/xsams/documents/>.
Hargreaves, R. J., Li, G., Bernath, P. F., [2012]{}. [Ammonia line lists from 1650 to 4000 cm$^{-1}$]{}. J. Quant. Spectrosc. Radiat. Transf. [113]{}, 670–679.
Harris, G. J., Larner, F. C., Tennyson, J., Kaminsky, B. M., Pavlenko, Y. V., Jones, H. R. A., 2008. [A H$^{13}$CN/HN$^{13}$C linelist, model atmospheres and synthetic spectra for carbon stars]{}. Mon. Not. R. Astr. Soc. 390, 143–148.
Harris, G. J., Polyansky, O. L., Tennyson, J., 2002. [Opacity data for HCN and HNC from a new [*ab initio*]{} linelist]{}. Astrophys. J. 578, 657–663.
Harris, G. J., Tennyson, J., Kaminsky, B. M., Pavlenko, Y. V., Jones, H. R. A., 2006. [Improved HCN/HNC linelist, model atmsospheres syntheic spectra for WZ Cas]{}. Mon. Not. R. Astr. Soc. 367, 400–406.
Kranendonk, L. A., An, X., Caswell, A. W., Herold, R. E., Sanders, S. T., Huber, R., Fujimoto, J. G., Okura, Y., Urata, Y., 2007. High speed engine gas thermometry by fourier-domain mode-locked laser absorption spectroscopy. Opt. Express 15, 15115–15128.
Lindermeir, E., Beier, K., 2012. Hitemp derived spectral database for the prediction of jet engine exhaust infrared emission using a statistical band model. J. Quant. Spectrosc. Radiat. Transf. 113, 1575 – 1593. [http://www.sciencedirect.com/science/article/pii/S00224%
07312001847](http://www.sciencedirect.com/science/article/pii/S00224%
07312001847)
Lucas, P. W., Tinney, C. G., Burningham, B., Leggett, S. K., Pinfield, D. J., Smart, R., Jones, H. R. A., Marocco, F., Barber, R. J., Yurchenko, S. N., Tennyson, J., Ishii, M., Tamura, M., Day-Jones, A. C., Adamson, A., Allard, F., Homeier, D., 2010. [The discovery of a very cool, very nearby brown dwarf in the Galactic plane]{}. Mon. Not. R. Astr. Soc. 408, L56–L60.
Lyubchik, Y., Jones, H. R. A., Pavlenko, Y. V., Martin, E., McLean, I., Prato, L., Barber, R. J., Tennyson, J., 2007. [Spectral analysis of high resolution near-infrared spectra of ultracool dwarfs]{}. Astron. Astrophys. 473, 257–264.
Neale, L., Miller, S., Tennyson, J., 1996. Spectroscopic properties of the [H$_3^+$]{} molecule: a new calculated linelist. Astrophys. J. 464, 516–520.
Rothman, L. S., Gordon, I. E., Barbe, A., Benner, D. C., Bernath, P. E., Birk, M., Boudon, V., Brown, L. R., Campargue, A., Champion, J. P., Chance, K., Coudert, L. H., Dana, V., Devi, V. M., Fally, S., Flaud, J. M., Gamache, R. R., Goldman, A., Jacquemart, D., Kleiner, I., Lacome, N., Lafferty, W. J., Mandin, J. Y., Massie, S. T., Mikhailenko, S. N., Miller, C. E., Moazzen-Ahmadi, N., Naumenko, O. V., Nikitin, A. V., Orphal, J., Perevalov, V. I., Perrin, A., Predoi-Cross, A., Rinsland, C. P., Rotger, M., Simeckova, M., Smith, M. A. H., Sung, K., Tashkun, S. A., Tennyson, J., Toth, R. A., Vandaele, A. C., Vander Auwera, J., 2009. [The [*HITRAN*]{} 2008 molecular spectroscopic database]{}. J. Quant. Spectrosc. Radiat. Transf. 110, 533–572.
Rothman, L. S., Gordon, I. E., Barber, R. J., Dothe, H., Gamache, R. R., Goldman, A., Perevalov, V. I., Tashkun, S. A., Tennyson, J., 2010. [HITEMP, the High-Temperature Molecular Spectroscopic Database]{}. J. Quant. Spectrosc. Radiat. Transf. 111, 2139–2150.
Shabram, M., Fortney, J. J., Greene, T. P., Freedman, R. S., [2011]{}. Transmission spectra of transiting planet atmospheres: Model validation and simulations of the hot neptune [GJ 436b]{} for the [James Webb Space Telescope]{}. Astrophys. J. [727]{}, 65.
Sharpe, S. W., Johnson, T. J., Sams, R. L., Chu, P. M., Rhoderick, G. C., Johnson, P. A., 2004. Gas-phase databases for quantitative infrared spectroscopy. Appl. Spectrosc. 58, 1452–1461.
Sochi, T., Tennyson, J., 2010. [A computed line list for the H$_2$D$^+$ molecular ion]{}. Mon. Not. R. Astr. Soc. 405, 2345–2350.
Swain, M. R., Tinetti, G., Vasisht, G., Deroo, P., Griffith, C., Bouwman, J., Chen, P., Yung, Y., Burrows, A., Brown, L. R., Matthews, J., Rowe, J. F., Kuschnig, R., Angerhausen, D., [2009]{}. Water, methane, and carbon dioxide present in the dayside spectrum of the exoplanet [HD 209458b]{}. Astrophys. J. [704]{}, 1616–1621.
Tashkun, S. A., Perevalov, V. I., JUN 2011. Cdsd-4000: High-resolution, high-temperature carbon dioxide spectroscopic databank. J. Quant. Spectrosc. Radiat. Transf. 112 (9), 1403–1410.
Tennyson, J., Yurchenko, S. N., 2012. [ExoMol: molecular line lists for exoplanet and other atmospheres]{}. Mon. Not. R. Astr. Soc.
Tessenyi, M., Ollivier, M., Tinetti, G., Beaulieu, J. P., du Foresto, V. C., Encrenaz, T., Micela, G., Swinyard, B., Ribas, I., Aylward, A., Tennyson, J., Swain, M. R., Sozzetti, A., Vasisht, G., Deroo, P., 2012. [Characterising the Atmospheres of Transiting Planets with a Dedicated Space Telescope]{}. Astrophys. J. 746, 45.
Tinetti, G., Aylward, A., [van Boekel]{}, R., Coradini, A., Encrenaz, T., Snellen, I., Zapatero-Osorio, M. R., Bouwman, J., Cho, J. Y.-K., [Coudé du Foresto]{}, V., Guillot, T., Lopez-Morales, M., Mueller-Wodarg, I., Palle, E., Selsis, F., Sozzetti, A., Tennyson, J., [et al.]{}, 2012. [Exoplanet Characterisation Observatory]{}. Experimental Astronomy.
Tinetti, G., Deroo, P., Swain, M. R., Griffith, C. A., Vasisht, G., Brown, L. R., Burke, C., McCullough, P., [2010b]{}. Probing the terminator region atmosphere of the hot-jupiter [XO-1b]{} with transmission spectroscopy. Astrophys. J. [712]{}, [L139–L142]{}.
Tinetti, G., Griffith, C. A., Swain, M. R., P. Deroo, J. P. B., Vasisht, G., Kipping, D., Waldmann, I., Tennyson, J., Barber, R. J., Bouwman, J., Allard, N., Brown, L., [2010a]{}. [Exploring extrasolar worlds: from gas giants to terrestrial habitable planets]{}. Faraday Discuss. 147, 369–377.
Tinetti, G., Vidal-Madjar, A., Liang, M.-C., Beaulieu, J.-P., Yung, Y., Carey, S., Barber, R. J., Tennyson, J., Ribas, I., Allard, N., Ballester, G. E., Sing, D. K., Selsis, F., 2007. Water vapour in the atmosphere of a transiting extrasolar planet. Nature 448, 169–171.
Vidler, M., Tennyson, J., 2000. Accurate partition function and thermodynamic data for water. J. Chem. Phys. 113, 9766–9771.
Voronin, B. A., Tennyson, J., Tolchenov, R. N., Lugovskoy, A. A., Yurchenko, S. N., 2010. [A high accuracy computed line list for the HDO molecule]{}. Mon. Not. R. Astr. Soc. 402, 492–496.
Warmbier, R., Schneider, R., Sharma, A. R., Braams, B. J., Bowman, J. M., Hauschildt, P. H., 2009. Ab initio modeling of molecular ir spectra of astrophysical interest: application to ch4. Astron. Astrophys. 495, 655–661.
Warren, S. J., Mortlock, D. J., Leggett, S. K., Pinfield, D. J., Homeier, D., Dye, S., Jameson, R. F., Lodieu, N., Lucas, P. W., Adamson, A. J., Allard, F., Barrado y Navascues, D., Casali, M., Chiu, K., Hambly, N. C., Hewett, P. C., Hirst, P., Irwin, M. J., Lawrence, A., Liu, M. C., Martin, E. L., Smart, R. L., Valdivielso, L., Venemans, B. P., 2007. A very cool brown dwarf in [UKIDSS DR1]{}. Mon. Not. R. Astr. Soc. [381]{}, 1400–1412.
Yurchenko, S. N., Barber, R. J., Tennyson, J., 2011. [A variationally computed hot (up to T=1500 K) line list for NH$_3$]{}. Mon. Not. R. Astr. Soc. 413, 1828–1834.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We construct an independent increments Gaussian process associated to a class of multicolor urn models. The construction uses random variables from the urn model which are different from the random variables for which central limit theorems are available in the two color case.'
address: |
$^{1}$Department of Mathematics, University of Bristol, University Walk, Bristol, BS8 1TW, UK\
$^{2}$Stat-Math Unit, Indian Statistical Institute, 203 B.T. Road, Kolkata 700 108, India\
E-mail: [email protected] and [email protected]; [email protected]
author:
- 'GOPAL K BASAK$^{1,2}$ and AMITES DASGUPTA$^{2}$'
title: A functional central limit theorem for a class of urn models
---
\[theore\][**Theorem**]{} \[theore\][*Lemma*]{} \[theore\][PROPOSITION]{}
Introduction
============
Consider a four-color urn model in which the replacement matrix is actually a stochastic matrix $\mathbf{R}$ as in ref. [@gouet2]. That is, we start with one ball of any color, which is the trial. Let $\mathbf{W}_n$ denote the column vector of the number of balls of the four colors up to the $n$-th trial, where the components of $\mathbf{W}_n$ are nonnegative real numbers. Then a color is observed by random sampling from a multinomial distribution with probabilities $(1/(n+1)) \mathbf{W}_n$. Depending on the color that is observed, the corresponding row of $\mathbf{R}$ is added to $\mathbf{W}_n^\prime$ and this gives $\mathbf{W}_{n+1}^\prime$. A special case of the main theorem of Gouet [@gouet2] is that if the stochastic matrix $\mathbf{R}$ is irreducible, then $(1/(n+1))
\mathbf{W}_n^\prime$ converges almost surely (a.s.) to the stationary distribution $\mathbf{\pi}$ of the irreducible stochastic matrix $\mathbf{R}$ (it should be carefully noted that the multicolor urn model is vastly different from the Markov chain evolving according to the transition matrix equal to the stochastic matrix $\mathbf{R}$, also notice that $\pi$ is a row vector). Suppose the nonprincipal eigenvalues of $\mathbf{R}$ satisfy $\lambda_1 < 1/2, \lambda_2 = 1/2, \lambda_3
> 1/2$ respectively, which are assumed to be real (and hence lie in $(-1, 1)$), and $\xi_1, \xi_2, \xi_3$ be the corresponding eigenvectors. Using $\mathbf{\pi} \xi_i = \mathbf{\pi} \mathbf{R}
\xi_i = \lambda_i \mathbf{\pi} \xi_i$ it is seen that $(1/(n+1))
\mathbf{W}_n^\prime \xi_i \rightarrow 0$.
Central and functional central limit theorems for $\mathbf{W}_n^\prime \xi_i$ have been the subject of several papers in the literature [@freedman; @gouet; @smythe] especially for two-color models and also some multicolor models. The norming in the central limit theorems in the two color urn models depends on the nonprincipal eigenvalue as follows: for $\lambda < 1/2$ the rate is $\sqrt{n}$, for $\lambda = 1/2$ the rate is $\sqrt{n \log
n}$ and the limits are normal in these two cases. However for $\lambda > 1/2$ the rate is $\Pi_0^{n - 1} (1 + (\lambda/(j+1))$ and in this case the limit exists almostsurely.
Functional central limit theorems (FCLT) for a class of two-color urn models have been considered by Gouet [@gouet]. These FCLT’s of Gouet [@gouet] use the same norming, as stated in the previous paragraph, under which central limit theorems have been proved in [@freedman] and [@gouet]. Ref. [@janson] contains a survey of the literature on such FCLT’s. In this article we prove a different FCLT that uses random variables with the norming $\Pi_0^{n - 1} (1 + (\lambda/(j+1))$ irrespective of whether $\lambda$ is less than 1/2, equal to 1/2 or greater than 1/2. This is the main result of the paper. For the sake of convenience we restrict ourselves to real eigenvalues only. We state the result for the above four-color model but it can be seen from the proof that it can be extended to urn models with any number ofcolors.
The article is organized as follows. In §2 we develop the notation, state the main result and give its proof. Some of the calculations have been done separately in§3.
Main result
===========
We write $$Z_{i,n} = \frac{\mathbf{W}_n^\prime \xi_i}{\Pi_0^{n - 1} \left(1
+ \frac{\lambda_i}{j+1}\right)},\label{fera}$$ where $\xi_i$ is the eigenvector corresponding to the eigenvalue $\lambda_i$. From the description of the urn model we have $\mathbf{W}_{n+1}^\prime \xi_i = \mathbf{W}_n^\prime \xi_i +
\chi_{n+1}^\prime \mathbf{R} \xi_i = \mathbf{W}_n^\prime \xi_i +
\lambda_i \chi_{n+1}^\prime \xi_i$, where $\chi_{n+1}$ is the column vector consisting of the indicator functions of balls of the four colors respectively. We also have $$E \{ \chi_{n+1}^\prime \xi_i |{\cal F}_n \} = \frac{1}{n+1}
\mathbf{W}_n^\prime \xi_i,\label{hir}$$ where ${\cal F}_n$ is the $\sigma$-field of observations up to the $n$-th trial. From this it follows that $Z_{i,n}$ is a martingale. From §3, it follows that $Z_{3,n}$ is $L^2$-bounded so that it converges almost surely. However in the two color case, for $\lambda < 1/2, \mathbf{W}_n^\prime \xi/\sqrt{n}$ and for $\lambda
= 1/2, \mathbf{W}_n^\prime \xi/ \sqrt{n \log n}$ converge to normal distributions and the FCLT’s proved in [@gouet] use such normalizations. Thus the question of using the same norming $\Pi_0^{n - 1} (1 + (\lambda/(j+1)) \sim n^{\lambda}$, to get an FCLT irrespective of $\lambda < 1/2, \lambda = 1/2$ or $\lambda >
1/2$, is of interest. Our main result, Proposition 2.1, is a step in this direction using the tails of the sequence $(Z_{1,n}, Z_{2,
n}, Z_{3, n})$ whereas the FCLT’s in the literature are based on partial sums starting from thebeginning.
$\left.\right.$
The sequence of processes $\mathbf{G}_n(t) =
(G_{1,n}(t), G_{2,n}(t), G_{3,n}(t))$ where $$G_{i, n}(t) = \sum_{m = n}^{[ne^t]} m^{\lambda_i - 1/2} (Z_{i,
m+1} - Z_{i, m}),\quad i = 1, 2, 3, \ t \geq 0,$$ converges to an independent increments Gaussian process $\mathbf{G}(t)$ with covariance function $c_{i,j}(t) = t \lambda_i
\lambda_j \langle \xi_i \xi_j, \pi^\prime \rangle, i, j = 1, 2,
3$[,]{} where the vector of the coordinate-wise product of the components of the two vectors $\xi_i$ and $\xi_j$ is denoted by $\xi_i \xi_j$ and the Euclidean inner product of the two vectors is denoted by $\langle . , .\rangle $.
Note that the process $\mathbf{G}$ can be viewed as a multidimensional Wiener process with covariance function $c_{i,
j}(\cdot)$.
From eq. (\[fera\]) we have the following expansion: $$\begin{aligned}
Z_{i, m+1} - Z_{i,m} &\sim - \frac{\lambda_i}{m} Z_{i,m} +
\lambda_i \frac{\chi_{m+1}^\prime \xi_i}{\Pi_{k=0}^{m} \left(1 +
\frac{\lambda_i}{k+1}\right)}\nonumber\\[.2pc]
&\sim - \frac{\lambda_i}{m} Z_{i,m} + \lambda_i
\frac{\chi_{m+1}^\prime \xi_i}{m^{\lambda_i}}.\label{hi}\end{aligned}$$ Since the components of $\mathbf{G}_n(t)$ are martingales, an independent increments Gaussian process as a limiting process is expected. In particular we follow Theorem 1.4, p. 339 of [@ek], by which it is enough to show that the joint characteristics of the martingales converge to a joint covariance function. Note that from (\[hi\]) $m^{\lambda_i -
\frac{1}{2}}(Z_{i,m+1} - Z_{i, m}) = O(1/\sqrt{m})$, as $\mathbf{W}_m^\prime \xi_i/m$ and $\chi_{m+1}^\prime \xi_i$ are bounded. This takes care of continuity of the paths and cross quadratic variations which is condition (b) of that theorem. Thus it remains to show that the cross quadratic variations converge to $c_{i,j}(t)$. We first do this for $i = 1, j = 2$. From (\[hi\]) we have $$\begin{aligned}
m^{\lambda_1 - 1/2}(Z_{1,m+1} - Z_{1,m}) &\sim - \lambda_1
m^{\lambda_1 - 1/2} \frac{\mathbf{W}_m^\prime \xi_1}{m^{\lambda_1
+ 1}} + \lambda_1 m^{\lambda_1 - 1/2} \frac{(\chi_{m+1}^\prime
\xi_1)}{m^{\lambda_1}},\nonumber\\[.2pc]
m^{\lambda_2 - 1/2}(Z_{2,m+1} - Z_{2,m}) &\sim - \lambda_2
m^{\lambda_2 - 1/2} \frac{\mathbf{W}_m^\prime \xi_2}{m^{\lambda_2
+ 1}} + \lambda_2 m^{\lambda_2 - 1/2} \frac{(\chi_{m+1}^\prime
\xi_2)}{m^{\lambda_2}}.\label{fi}\end{aligned}$$ We want to show that in computing the cross quadratic variation, which is the limit of $$\sum_n^{[n e^t]} E\{ m^{\lambda_1 - 1/2} m^{\lambda_2 - 1/2}
(Z_{1, m+1} - Z_{1, m})(Z_{2, m+1} - Z_{2, m})|{\cal F}_m \},$$ only the second term from the right-hand side of each of eqs (\[fi\]) contributes. Since $\chi_{n+1}$ consists of indicator functions, which implies that $$\left(\sum_k \xi_{1,k} \chi_{n+1, k}\right)\left(\sum_l \xi_{2,l}
\chi_{n+1, l}\right) = \sum_k \xi_{1,k} \xi_{2, k} \chi_{n+1, k},$$ this contribution is the limit of $$\lambda_1 \lambda_2 \sum_n^{[ne^t]} \frac{1}{m} \left\langle \xi_1
\xi_2, \frac{\mathbf{W}_m}{m+1} \right\rangle,$$ which is $t \lambda_1 \lambda_2 \langle \xi_1 \xi_2, \pi^\prime
\rangle$, since from [@gouet2] we know ${\mathbf{W}^\prime_m}/{(m+1)} \rightarrow \pi$ a.s. Also notice that this part of the argument does not depend on whether $\lambda_1$ or $\lambda_2$ are less than or equal to 1/2.
To see why the contribution to the cross quadratic variation from the first terms of (\[fi\]) goes to $0$, by Cauchy–Schwarz inequality it is enough to show that the sum of squares over $n$ to $[n e^t]$ of the first terms in each line of (\[fi\]) goes to $0$. This part of the argument will depend on the value of $\lambda_i$. Note the following which have been proved in §3: $$\begin{aligned}
\hbox{For} \ \lambda_1 &< 1/2,\quad \frac{\mathbf{W}_m^\prime
\xi_1}{\sqrt{m}} \mbox{ is } L^2\hbox{-bounded},\\[.3pc]
\hbox{For} \ \lambda_2 &= 1/2,\quad \frac{\mathbf{W}_m^\prime
\xi_2}{\sqrt{m\log m}} \mbox{ is } L^2\hbox{-bounded}.\end{aligned}$$ Consider the case $\lambda_1 < 1/2$. We need to show $$\sum_n^{[n e^t]} \frac{( \mathbf{W}_m^\prime \xi_1)^2}{m^3}
\rightarrow 0\ \hbox{a.s}.$$ We know that for $\lambda_1 < 1/2$, $\mathbf{W}_m^\prime \xi_1
/\sqrt{m}$ is $L^2$-bounded, so that $$E \sum_n^{[n e^t]} \frac{( \mathbf{W}_m^\prime \xi_1)^2}{m^3} \leq
E \sum_n^\infty \frac{( \mathbf{W}_m^\prime \xi_1)^2}{m^3} \leq
\hbox{const.} \sum_n^\infty \frac{1}{m^2} \rightarrow
0.\label{fii}$$ Since the sum inside the expectation in the middle is decreasing in $n$, it converges to $0$ a.s. For $\lambda_2 = 1/2$, $\mathbf{W}_m^\prime \xi_2 /\sqrt{m\log m}$ is $L^2$-bounded, and one can proceed similarly. Thus we have proved that $c_{1, 2}(t) =
t \lambda_1 \lambda_2 \langle \xi_1 \xi_2, \pi^\prime \rangle$. Similarly $c_{i, j}(t), i, j = 1, 2$, can be computed as given in Proposition 2.1.
Now consider as to what will happen if we were computing say $c_{1,3}(t)$. For $\lambda_3 > 1/2$, the expansion (\[fi\]) is similar, and in the cross quadratic variation the contribution of the second term from the right-hand side of $$m^{\lambda_3 - 1/2}(Z_{3,m+1} - Z_{3,m}) \sim - \lambda_3
m^{\lambda_3 - 1/2} \frac{\mathbf{W}_m^\prime \xi_3}{m^{\lambda_3
+ 1}} + \lambda_3 m^{\lambda_3 - 1/2} \frac{(\chi_{m+1}^\prime
\xi_3)}{m^{\lambda_3}}$$ is similar to what we had before. For $\lambda_3 > 1/2$, $\mathbf{W}_m^\prime \xi_3 / \Pi_0^{m - 1} (1 + (\lambda_3/(j+1))$ is a martingale and from Appendix 3.3, $$\frac{\mathbf{W}_m^\prime \xi_3}{m^{\lambda_3}} \hbox{ is } L^2
\hbox{-bounded}.$$ So $\mathbf{W}_m^\prime \xi_3 / m^{\lambda_3}$ converges almost surely. This implies that the contribution of the first term $$\sum_n^{[ne^t]} \frac{(\mathbf{W}_m^\prime \xi_3)^2}
{m^{2\lambda_3}} \frac{1}{m^{3 - 2\lambda_3}} \rightarrow 0 \hbox{
a.s. }$$ since $2 \lambda_3 < 2$. Thus $c_{i, j}(t), i = 1,2,3, j = 3$, can be computed as given in the statement of Proposition 2.1. This completes the proof.$\Box$
Appendix
========
Suppose real eigenvalues satisfy $\lambda_1 < 1/2, \lambda_2 =
1/2, \lambda_3 > 1/2$ and $\xi_1, \xi_2, \xi_3$ be the corresponding eigenvectors. In this section we prove that $X_n$, $Y_n$ and $Z_n$ are $L^2$-bounded where $$X_n = \frac{\mathbf{W}_n^\prime \xi_1}{\sqrt{n}},\quad Y_n =
\frac{\mathbf{W}_n^\prime \xi_2}{\sqrt{n \log n}},\quad Z_n =
\frac{\mathbf{W}_n^\prime \xi_3}{\Pi_0^{n - 1} \left(1 +
\frac{\lambda_3}{j+1}\right)},$$ a fact which has been used in the proof of Proposition 2.1. For $X_n$ and $Y_n$ verification of $L^2$-boundedness is through Lemma 2.1 of [@kersting]. This is done on a case by case basis depending on $\lambda_1$ and $\lambda_2$ in the next two subsections. For the reader’s convenience we state Kersting’s lemma from [@kersting] here:
-.4pc[[@kersting].]{} Let $\alpha_n,
\beta_n\ (n \geq 1)$ be nonnegative numbers such that $\alpha_n
\rightarrow 0$[,]{} $\sum_{n = 1}^\infty \alpha_n = \infty$, and for large $n$[,]{} $$\beta_{n+1} \leq \beta_n ( 1 - c \alpha_n) + d \alpha_n$$ with $c, d > 0$. Then $\limsup_{n \rightarrow \infty} \beta_n \leq
d/c$.
*$L^2$-boundedness of $X_n$*
----------------------------
Using $\mathbf{W}_{n+1}^\prime \xi_1 = \mathbf{W}_n^\prime \xi_1 +
\lambda_1 \chi_{n+1}^\prime \xi_1$ and the definition of $X_n$, we get $$X_{n+1} = X_n \sqrt{\frac{n}{n+1}} + \lambda_1
\frac{\chi_{n+1}^\prime \xi_1}{\sqrt{n+1}}.$$ Taking conditional expectation and using (\[hir\]) we get $$\begin{aligned}
E \{ X_{n+1}^2|{\cal F}_n \} = X_n^2 \left(1 - \frac{1}{n+1}
\right) \left(1 + \frac{2\lambda_1}{n+1} \right) +
\frac{\lambda_1^2}{n+1} \left\langle \frac{\mathbf{W}_n}{n+1},
\xi_1^2\right\rangle,\end{aligned}$$ from which taking further expectation we get $$E X_{n+1}^2 = E X_n^2 \left(1 - \frac{1}{n+1}\right) \left(1 +
\frac{2\lambda_1}{n+1}\right) + \frac{\lambda_1^2}{n+1}
\left\langle E \frac{\mathbf{W}_n}{n+1}, \xi_1^2 \right\rangle.$$ The last vector $E ({\mathbf{W}_n}/{(n+1)})$ consists of bounded components. Thus if $\lambda_1 < 0$, then $$E X_{n+1}^2 \leq E X_n^2 \left(1 - \frac{1}{n+1}\right) +
\frac{{\rm const}}{n+1},$$ and Kersting’s lemma applies. If $\lambda_1
> 0$ then we still have $\lambda_1 < 1/2$ i.e. $2\lambda_1 < 1$. In this case $$\begin{aligned}
\left(1 - \frac{1}{n+1}\right) \left(1 + \frac{2\lambda_1}{n+1}
\right) &\leq 1 + \frac{2\lambda_1}{n+1} - \frac{1}{n+1}\\[.2pc]
&= 1 - \frac{1 - 2\lambda_1}{n+1},\end{aligned}$$ i.e. Kersting’s lemma applies.
*$L^2$-boundedness of $Y_n$*
----------------------------
Using $\mathbf{W}_{n+1}^\prime \xi_2 = \mathbf{W}_n^\prime \xi_2 +
\lambda_2 \chi_{n+1}^\prime \xi_2$ and the definition of $Y_n$, we get $$Y_{n+1} = Y_n \sqrt{\frac{n \log n}{(n+1) \log (n+1)}} + \lambda_2
\frac{\chi_{n+1}^\prime \xi_2}{\sqrt{(n+1) \log (n+1)}}.$$ Taking conditional expectation we get (recall $\lambda_2 = 1/2$) $$\begin{aligned}
E \{ Y_{n+1}^2|{\cal F}_n \} &= Y_n^2 \frac{n \log n}{(n+1) \log
(n+1)} \left(1 + \frac{1}{n+1} \right)\\[.2pc]
&\quad\, + \frac{\lambda_2^2}{(n+1) \log (n+1)} \left\langle
\frac{\mathbf{W}_n}{n+1}, \xi_2^2 \right\rangle,\end{aligned}$$ from which taking further expectation we get $$\begin{aligned}
E Y_{n+1}^2 &= E Y_n^2 \left( 1 - \frac{(n+1) \log (n+1) - n \log
n}{(n+1) \log (n+1)} \right) \left(1 + \frac{1}{n+1} \right)\\[.2pc]
&\quad\, + \frac{\lambda_2^2}{(n+1) \log (n+1)} \left\langle E
\frac{\mathbf{W}_n}{n+1}, \xi_2^2 \right\rangle.\end{aligned}$$ The vector $E ({\mathbf{W}_n}/{(n+1)})$ consists of bounded components. Now we apply the second trick of the previous subsection to apply Kersting’s lemma. The following calculation does the rest of the work.
We show that $$(n+1)\log (n+1) \left\{ \frac{(n+1)\log (n+1) - n\log n}{(n+1)\log
(n+1)} - \frac{1}{n+1} \right\} \rightarrow c > 0.$$ We approximate $\log (n+1)$ by $\log n + \frac{1}{n}$. This gives $$\begin{aligned}
&(n+1) \log (n+1) - n\log n \sim (n+1) \log n + (n+1)\frac{1}{n} -
n\log n\\[.2pc]
&\quad\, = \log n + 1 + \frac{1}{n}.\end{aligned}$$ Hence $$\begin{aligned}
\left(\log n + 1 + \frac{1}{n} \right) - \log (n+1) \sim
\left(\log n + 1 + \frac{1}{n} \right) - \left(\log n +
\frac{1}{n}\right) \rightarrow 1.\end{aligned}$$
*$L^2$-boundedness of $Z_n$*
----------------------------
The proof follows Lemma 3.1 of [@freedman]. We have earlier proved the approximation $$Z_{n+1} - Z_n \sim - \frac{\lambda_3}{n} Z_n + \lambda_3
\frac{\chi_{n+1}^\prime \xi_3}{n^{\lambda_3}}.$$ Now with the martingale property of $Z_n$, $Z_{n+1} = Z_n +
(Z_{n+1} - Z_n)$, and by the above approximation we have $$E (Z_{n+1}^2|{\cal F}_n) \sim Z_n^2 +
\left(\frac{\lambda_3^2}{n^2} Z_n^2 - 2 \frac{\lambda_3^2}{n^2}
Z_n^2 + \lambda_3^2 \frac{\left\langle \frac{\mathbf{W}_n}{n+1},
\xi_3^2 \right\rangle}{n^{2\lambda_3}} \right),\nonumber$$ implying $$E Z_{n+1}^2 \leq E Z_n^2 \left( 1 - \frac{\lambda_3^2}{n^2}
\right) + \lambda_3^2 \frac{{\rm const}}{n^{2\lambda_3}}.
\label{hii}$$ Since $2 \lambda_3 > 1$, by iteration of (\[hii\]) it follows that $Z_n$ is $L^2$-bounded.
[9]{} Ethier S N and Kurtz T G, Markov processes: Characterization and convergence (New York: John Wiley & Sons) (1986)
Freedman D A, Bernard Friedman’s urn, [*Ann. Math. Stat.*]{} [**36(3)**]{} (1965) 956–970
Gouet R, Martingale functional central limit theorems for a generalized Polya urn, [*Ann. Probab.*]{} [**21(3)**]{} (1993) 1624–1639
Gouet R, Strong convergence of proportions in a multicolor Polya urn, [*J. Appl. Probab.*]{} [**34(2)**]{} (1997) 426–435
Janson S, Functional limit theorems for multitype branching processes and generalized Polya urns, [*Stoch. Proc. Appl.*]{} [**110(2)**]{} (2004) 177–245
Kersting G D, A weak convergence theorem with application to the Robbins-Monro process, [*Ann. Probab.*]{} [**6(6)**]{} (1978) 1015–1025
Smythe R T, Central limit theorems for urn models, [*Stoch. Proc. Appl.*]{} [**65**]{} (1996) 115–137
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This paper reports turbulent boundary layer measurements made over open-cell reticulated foams with varying pore size [[[and thickness, but constant porosity ($\epsilon \approx 0.97$)]{}]{}]{}. The foams were flush-mounted into a cutout on a flat plate. A Laser Doppler Velocimeter (LDV) was used to measure mean streamwise velocity and turbulence intensity immediately upstream of the porous section, and at [[[multiple]{}]{}]{} measurement stations along the porous substrate. The friction Reynolds number upstream of the porous section was $Re_\tau \approx 1690$. For all [[[but the thickest]{}]{}]{} foam tested, the internal boundary layer was fully developed by $<10 \delta$ downstream from the porous transition, where $\delta$ is the boundary layer thickness. Fully developed mean velocity profiles showed the presence of a substantial slip velocity at the porous interface ($>30\%$ of the free stream velocity) and a mean velocity deficit relative to the canonical smooth-wall profile further from the wall. While the magnitude of the mean velocity deficit increased with average pore size, the slip velocity remained approximately constant. Fits to the mean velocity profile suggest that the logarithmic region is shifted relative to a smooth wall, and that this shift increases with pore size until it becomes comparable to substrate thickness $h$. For all foams, the turbulence intensity was found to be elevated further into the boundary layer to $y/ \delta \approx 0.2$. An outer peak in intensity was also evident for the largest pore sizes. Velocity spectra indicate that this outer peak is associated with large-scale structures resembling Kelvin-Helmholtz vortices that have streamwise length scale $2\delta-4\delta$. Skewness profiles suggest that these large-scale structures may have an amplitude-modulating effect on the interfacial turbulence.'
author:
- 'Christoph Efstathiou$^{1}$'
- 'Mitul Luhar$^{1}$[^1]'
bibliography:
- 'Efstathiou2017.bib'
title: 'Mean turbulence statistics in boundary layers over high-porosity foams'
---
Introduction {#sec:intro}
============
Turbulent flows of scientific and engineering interest are often bounded by walls that are not smooth, solid, or uniform. Manufacturing techniques, operational requirements and natural evolution often lead to non-uniform, rough, and porous boundaries. Examples include flows over heat exchangers, forest canopies, bird feathers and river beds [e.g., @finnigan2000turbulence; @jimenez2004turbulent; @ghisalberti2009obstructed; @favier2009passive; @manes2011turbulent; @jaworski2013aerodynamic; @chandesris2013direct; @kim2016experimental]. Porous boundaries also enable active flow control through suction and blowing for use in drag reduction and delaying transition from laminar to turbulent flow [e.g., @parikh2011passive].
Despite the potential applications, relatively little is known about the relationship between turbulent flows and porous substrates. For example, it is unclear how established features of smooth-wall flows, such as the self-sustaining near-wall cycle, the logarithmic region in the mean profile, the larger-scale structures found further from the wall, and the interaction between the inner and outer regions of the flow [e.g. the amplitude modulation phenomenon; @marusic2010predictive] are modified over porous surfaces. As a result, there are few models that can predict how a porous substrate of known geometry will influence the mean flow field and turbulent statistics. A key limitation is that few experimental and numerical datasets exist for turbulent flows over porous media. Further, most previous experimental datasets (motivated primarily by flows over packed sediment beds or canopies) include limited near-wall measurements, while previous numerical simulations have been restricted to relatively low Reynolds numbers. [[[In addition, almost all previous studies on turbulent flow over porous substrates have employed relatively thick media, such that flow penetration into the substrate depends only on pore size or permeability [see e.g., @manes2011turbulent]. However, in many natural and engineered systems, the porous substrate can be of finite thickness and bounded by a solid boundary. Examples include feathers or fur in natural locomotion [@itoh2006turbulent; @jaworski2013aerodynamic], and heat exchangers employing metal foams [@mahjoob2008synthesis]. In such systems, substrate thickness can also have an important influence on flow development and the eventual equilibrium state. At the limit where porous medium thickness becomes comparable to pore size, the substrate can essentially be considered a rough wall. The transition from this rough-wall limit to more typical porous medium behaviour is not fully understood. The experimental study described in this paper seeks to address some of the limitations described above.]{}]{}]{}
Previous studies {#sec:literature}
----------------
Previous numerical efforts investigating the effect of porous boundaries include the early direct numerical simulations (DNS) performed by @jimenez2001turbulent, which employed an effective admittance coefficient linking the wall-pressure and wall-normal velocity to model the porous wall. More recent DNS efforts have either employed the volume-averaged Navier-Stokes equations, where the porous substrate is modeled as a resistive medium [@breugem2006influence], or explicitly modeled the porous medium as an array of cubes [@chandesris2013direct].
In particular, the results of @breugem2006influence showed important deviations from turbulent channel flow over smooth walls. The mean velocity was significantly reduced across much of the channel for the cases with high porosity ($\epsilon = 0.8$ and $0.95$) and this was accompanied by a skin friction coefficient increase of up to $30\%$. The presence of the porous substrate also led to substantial changes in flow structure. Large spanwise rollers with streamwise length scale comparable to the channel height were observed at the porous interface. Given the presence of an inflection point in the mean velocity profile near the porous interface, @breugem2006influence attributed the emergence of these large-scale structures to a Kelvin-Helmholtz type of instability mechanism [see also @jimenez2001turbulent; @chandesris2013direct]. In addition, the near-wall cycle comprising streaks and streamwise vortices, a staple of smooth wall turbulent boundary layers, was substantially weakened over the porous substrate. This weakening was linked to a reduction in the so-called wall-blocking effect and enhanced turbulent transport across the interface. @breugem2006influence also found that the root mean square (rms) of the wall-normal velocity fluctuations did not exhibit outer layer similarity and suggested two potential causes for this (i) more vigorous sweeps and ejections due to the absence of an impermeable wall and (ii) insufficient scale separation between the channel half-height and the penetration distance into the porous substrate. @chandesris2013direct noted broadly similar trends in their DNS, which also considered thermal transport.
[[[@rosti2015direct performed DNS studies of turbulent channel flow at $Re_\tau = 180$ using a VANS formulation that allowed for the porosity and permeability to be decoupled. These simulations showed that even relatively low wall permeabilities led to substantial modification of the turbulent flow in the open channel. Further, despite a substantial variation in the porosities tested ($\epsilon = 0.3-0.9$), the flow was found to be much more sensitive to permeability. @motlagh2016pod performed a VANS Large Eddy Simulation (LES) of channel flow over porous substrates with $\epsilon=0, 0.8$ and $0.95$ at bulk Reynolds number 5500. Based on proper orthogonal decomposition (POD) of the flow field, they showed that large-scale features prevalent in smooth-walled flows were still present but beginning to break down for the lower porosity case. Consistent with the results of @breugem2006influence, for the higher porosity case, these structures were replaced by spanwise-elongated counter-rotating vortices.]{}]{}]{}
[[[Recently, @kuwata2016lattice and @kuwata2017direct have performed Lattice-Boltzman simulations over rough walls, staggered arrays of cubes, as well as anisotropic porous media. In particular, @kuwata2017direct considered a model system in which the streamwise, spanwise, and wall-normal permeabilities could be altered individually. These simulations showed that the streamwise permeability is instrumental in preventing the high- and low-speed streaks associated with the near-wall cycle. The simulations also indicated that, unlike the streamwise and spanwise permeabilities, the wall-normal permeability does not significantly enhance turbulence intensities.]{}]{}]{}
Similar to turbulent flows over smooth and rough walls, previous studies suggest that a logarithmic region in the mean profile $U(y)$ can also be expected in turbulent flows over porous media. This logarithmic region is often parametrized as: $$\label{eq:log-law}
U^+ = \frac{1}{\kappa} \ln{\left(\frac{y + y_d}{k_0} \right)} = \frac{1}{\kappa} \ln{(y^+ + y_d^+)} + B -\Delta U^+,$$ where $\kappa$ and $B$ are the von Karman and additive constants, $y$ is the wall-normal distance from the porous interface, $y_d$ is the shift of the logarithmic layer (or zero-plane displacement height), $k_0$ is the equivalent roughness height, and $\Delta U^+$ is the roughness function [@jimenez2004turbulent]. A superscript $+$ indicates quantities normalized by the friction velocity $u_\tau$ and the kinematic viscosity $\nu$. Fits to the mean velocity profiles obtained in DNS by @breugem2006influence suggest that the von Karman constant decreases from $\kappa \approx 0.4$ for the smooth wall case to $\kappa = 0.23$ for the most porous substrate. However, further tests were recommended at higher Reynolds number to confirm this effect.
Experimental efforts in this realm have considered flows over beds of packed spheres, perforated sheets, foams, as well as seal fur [e.g., @ruff1972turbulent; @zagni1976channel; @kong1982turbulent; @itoh2006turbulent; @suga2010effects; @manes2011turbulent; @kim2016experimental]. Interestingly, the seal fur experiments show a reduction in skin friction though the exact mechanism behind this remains to be understood [@itoh2006turbulent]. @kong1982turbulent investigated the effect of small scale roughness over smooth, rough, and porous surfaces on turbulent boundary layers over bluff bodies. The porous boundaries consisted of perforated sheets and mesh screens. These experiments showed that the turbulent Reynolds stresses increased near the interface, as did the skin friction. However, it is difficult to separate roughness effects from permeability effects for these experiments. The mean velocity profile was shifted by $\Delta U^+ \approx 3-4$, which was similar to the shift obtained over an impermeable rough wall of similar geometry.
@suga2010effects studied laminar and turbulent channel flow over foamed ceramics with porosity $\epsilon \approx 0.8$ and varying pore sizes via Particle Image Velocimetry (PIV). These experiments were carried out at relatively low Reynolds numbers (bulk Reynolds number $Re_b \le 10,200$). The measurements indicated that the transition to turbulence occurs at lower Reynolds number over the porous media. Further, turbulence intensities were generally enhanced over the porous medium, and the displacement and roughness heights in the modified logarithmic law were found to increase with increasing pore size and permeability. [[[@detert2010synoptic also used PIV to study the flow over packed spheres and gravel from the river Rhine in an open channel flow with friction Reynolds numbers $Re_\delta = 1.88-14.7\times 10^3$. At the relatively low porosities tested ($\epsilon = 0.26-0.33$), the flow exhibited many of the large-scale structures observed in turbulent boundary layers over smooth walls, including hairpin vortex packages. The observed flow patterns were also found to be relatively insensitive to Reynolds number.]{}]{}]{}
[@manes2011turbulent] made open channel turbulent flow measurements over very porous ($\epsilon > 0.96$) polyurethane foam mattresses with 10-60 pores per inch (ppi) at Reynolds number $Re_\tau>2000$ using an LDV. In contrast to [@breugem2006influence], the results obtained by @manes2011turbulent supported the outer layer similarity hypothesis. This suggests that the lack of outer-layer similarity in the DNS carried out by @breugem2006influence resulted from insufficient scale separation between the inner and outer layers of the flow, which is analogous to the breakdown of outer-layer similarity in shallow boundary layers over rough walls [@jimenez2004turbulent]. Further, @manes2011turbulent reported a reduction in the streamwise fluctuation intensity near the wall and an increase in the intensity of wall-normal fluctuations, which they attributed to reduced wall blocking. For the logarithmic region in the mean profile, a fitting procedure similar to the one employed by @breugem2006influence and @suga2010effects yielded $\kappa \approx 0.3$ over the porous media and an equivalent roughness height $k_0$ that generally increased with increasing pore size. However, the mean profiles did not exhibit a clear logarithmic region over the 10 ppi foam. Since the flow is expected to penetrate further into the foam as pore size increases, @manes2011turbulent suggested that a lack of scale separation between the penetration distance and the water depth may have influenced the results.
Finally, note that there are substantial similarities in turbulent flows over porous media and vegetation or urban canopies [e.g., @finnigan2000turbulence; @poggi2004effect; @white2007shear]. For example, large-scale structures resembling Kelvin-Helmholtz vortices play an essential role in dictating mass and momentum transport at the interface [see e.g., @finnigan2000turbulence] and a shifted log law of the form shown in equation (\[eq:log-law\]) provides a reasonable fit for the velocity profile above the canopy. In a comparative study of obstructed shear flows, @ghisalberti2009obstructed suggested that an inflection point exists in the mean profile if the distance to which the flow penetrates into the porous medium or canopy is much smaller than the height of the medium. This penetration distance is expected to scale as $\sqrt{k}$ for porous substrates [e.g., @battiato2012self], where $k$ is the permeability, and $(C_D a)^{-1}$ for canopies where, $C_D$ is a drag coefficient and $a$ is the frontal area per unit volume of the canopy. For densely packed or tall substrates where an inflection point is observed, the interfacial dynamics are dominated by structures resembling Kelvin-Helmholtz vortices. In such cases, the slip velocity at the interface depends primarily on the friction velocity, with little dependence on substrate geometry. Further, the interfacial turbulence also tends to be more isotropic, such that intensity of the wall-normal velocity fluctuations is comparable to the intensity of the streamwise fluctuations.
Contribution and outline {#sec:outline}
------------------------
The present study builds on the experiments pursued by @manes2011turbulent to provide further insight into the near-wall flow physics over high-porosity surfaces. An LDV was used to measure streamwise velocity profiles over commercially-available foams with systematically varying pore sizes [[[and thicknesses]{}]{}]{} at moderately high Reynolds number. For reference, the friction Reynolds number over the smooth wall upstream of the porous section was $Re_\tau = u_\tau \delta/\nu \approx 1690$, where $\delta$ is the $99\%$ boundary layer thickness. The velocity profiles include measurements very close to the porous interface (corresponding to 2-3 viscous units over the smooth wall). [[[Unfortunately, wall-normal velocities were not measured due to instrumentation limitations.]{}]{}]{}
Morphologically, the foams tested in this study are similar to those considered by [@manes2011turbulent]. However, one important distinction is the thickness of the foam layer, $h$. In an effort to isolate the effects of pore size, $s$, and permeability, $k$, @manes2011turbulent considered very thick porous media with $h \gg \sqrt{k}$ and $h \gg s$. The scale separation is more limited in the present study, with [[[$h/s \approx 4.3-44$ and $h/\sqrt{k} \approx 36-160$]{}]{}]{}. As a result, porous layer thickness does influence the flow for foams with the largest pore sizes and permeabilities. In other words, the present study provides insight into the transition between conditions in which the shear flow penetrates across the entire porous domain and conditions in which the shear layer only reaches a small distance into the porous medium. In a sense, this is analogous to the transition between sparse and dense canopy behaviour observed in vegetated shear flows [@luhar2008interaction; @ghisalberti2009obstructed].
In recent years, the amplitude modulation phenomenon observed in smooth- and rough-walled turbulent flows has led to a promising class of predictive models [e.g., @mathis2009large; @marusic2010predictive; @mathis2013estimating; @pathikonda2017inner]. Specifically, it has been observed that the so-called very-large-scale motions (VLSMs) prevalent in the logarithmic region of the flow at high Reynolds number [@smits2011high] have a modulating influence on the intensity of the near-wall turbulence. Further, it has been shown that there is an intrinsic link between this phenomenon and the skewness of the streamwise velocity fluctuations [@mathis2011relationship; @duvvuri2015triadic]. Therefore, by considering the skewness of the streamwise velocity, we also evaluate whether such interactions between the inner and outer region persist in turbulent flows over porous media that may be dominated by a different class of large-scale structure, i.e. Kelvin-Helmholtz vortices.
The remainder of this paper is structured as follows: §\[sec:expts\] describes the experiments, providing details on the flow facility, porous substrates, and diagnostic techniques; §\[sec:development\] shows results on flow development over the porous foams; §\[sec:poresize\] illustrates the effect of pore size on mean turbulence statistics; [[[§\[sec:thickness\] explores the effect of substrate thickness; §\[sec:spectra\] presents velocity spectra over all the different foams tested; ]{}]{}]{}§\[sec:discussion\] tests whether a shifted logarithmic region exists over the porous surfaces, evaluates the relative effect of pore size and substrate thickness, and also considers how porous substrates may affect the amplitude modulation phenomenon. Conclusions are presented in §\[sec:conclusions\].
Experimental methods {#sec:expts}
====================
Flow facility and flat plate apparatus {#sec:facility}
--------------------------------------
![Schematics showing flat plate apparatus in the wall normal-spanwise plane ($y-z$, top) and the streamwise-wall normal plane ($x-y$, bottom). Measurement positions marked 1-4 were located at $x/h = 11, 21, 42, 53$, where $h=12.7$ mm is the baseline substrate thickness, and $x=0$ is defined as the smooth-porous wall transition. The smooth wall reference measurements were made at a location $x/h=-21.5$. Note that the wall-normal coordinate is $y$.[]{data-label="fig:diag"}](Figures/Fig1-Exp.pdf){width="\textwidth"}
All experiments were conducted in the USC water channel, a free-surface, recirculating facility with glass along both sidewalls and at the bottom to allow for unrestricted optical access. The water channel has a test section of length 762 cm, width 89 cm, and height 61 cm, and is capable of generating free-stream velocities up to 70 cm/s with background turbulence levels $<1\%$ at a water depth of 48 cm. The temperature for all experiments was $23\pm 0.5^\circ C$ for which the kinematic viscosity is $\nu = 0.93 \times 10^{-2}$ cm$^2$/s.
Figure \[fig:diag\] provides a schematic of the experimental setup in the wall normal-spanwise (top) and streamwise-wall normal (bottom) planes. A 240 cm long flat plate was suspended from precision rails at a height $H=30$ cm above the test section bottom. To avoid free-surface effects, measurements were made below the flat plate. The nominal free-stream velocity was set at $U_e = 58$ cm/s for all the experiments. The confinement between the flat plate and bottom of the channel naturally led to a slightly favorable pressure gradient and slight free stream velocity increase along the plate $(\leq 3\%)$, however the non-dimensional acceleration parameter, $\Lambda = \frac{\nu}{U_e^2} \frac{dU_e}{dx}$ was on the order of $10^{-7}$ suggesting any pressure gradient effects are likely to be mild [@PatelPreston1965; @de2000reynolds; @schultz2007rough].
A cutout of length 89 cm and width 60 cm was located 130 cm downstream of the leading edge. Smooth and porous surfaces were substituted into the cutout, flush with the smooth plate around it. The porous test specimens, described in further detail below, were bonded to a solid Garolite^TM^ sheet to provide a rigid structure and prevent bleed through. [[[The setup was designed to accommodate porous substrates of thicknesses $h=6.35, 12.7$ and $25.4$ mm.]{}]{}]{} Care was taken to minimize gaps and ensure a smooth transition from solid to porous substrate. Velocity profiles were measured over the smooth section upstream of the cutout and at four additional locations over the porous walls. The flow was tripped by a wire of 0.5 mm diameter located 15 cm downstream of the leading edge.
Velocity measurement {#sec:diagnostics}
--------------------
Measurements of [[[streamwise velocity, $u$,]{}]{}]{} were made at the channel centerline using a Laser Doppler Velocimeter (LDV, MSE Inc.) with a 50 cm standoff distance and a measurement volume of $300 \mu$m by $150 \mu$m by $1000 \mu$ m ($x$ by $y$ by $z$). The LDV was mounted on a precision traverse capable of 16 $\mu$m resolution. Polyamide seeding particles (PSP) with an average size of 5 $\mu$m were used to seed the flow. While the large standoff distance enabled measurements at the channel centerline, it also limited data rates to 50 Hz in the free-stream and less than 1 Hz at the stations closest to the wall. A minimum of 800 data points in the regions with lowest velocity and 4000-25000 data points in regions of higher velocity were collected to ensure fully converged statistics. In all cases, a preliminary coarse velocity profile was measured to determine the boundary layer thickness and the approximate location of the wall. This preliminary profile was used to generate a finer logarithmically-spaced vertical grid for the actual measurements. The nominal smooth-wall location ($y^\prime = 0$) was identified as the position where the data rate dropped to zero. For all the profiles, the vertical grid resolution was reduced to 30 $\mu$m in the near-wall region. Given the 150$\mu$m measurement volume and the $30\mu$m vertical resolution, the nominal estimate for wall-normal location suffers from an uncertainty of $\le 75\mu$m.
LDV measurements suffer from two significant distortions: velocity gradients across the measurement volume and velocity biasing [@durst1976principles; @degraaff2001high]. The former is significant in regions where large velocity gradients are present across the measurement volume, as is the case near solid walls. The latter occurs because, in turbulent flows, more high velocity particles move through the measurement volume in a given period compared to low-velocity particles (assuming uniform seeding density). To correct for this bias, an inverse velocity weighting factor was used to correct mean statistics [@mclaughlin1973biasing]. The mean streamwise velocity was estimated using the following relationship: $$\label{eq:weight-U}
U = \frac{\sum\limits_{i=1}^{N}b_i u_i}{\sum\limits_{i=1}^{N} b_i},$$ where $u_i$ is an individual velocity sample, $N$ is the total number of samples, and $b_i = 1/|u_i|$ is the weighting factor. Similarly, the weighted streamwise turbulence intensity was estimated as: $$\label{eq:weight-u}
\overline{u^2}=\frac{\sum\limits_{i=1}^{N} b_i \left[u_i-U\right]^2}{\sum\limits_{i=1}^{N}b_i}.$$ Alternative methods of correcting for velocity biasing are discussed and evaluated in [@herrin1993investigation]. These methods include using the inter-arrival time ($b_i = t_i-t_{i-1}$) as the weighting factor, or the sample-and-hold technique ($b_i=t_{i+1}-t_i$). Figure \[fig:weights\] shows how these different correction techniques affect a representative mean velocity ($U$) and streamwise turbulent intensity ($\overline{u^2}$) profile. Inverse velocity weighting leads to the largest correction relative to the raw data in the near-wall region where the sampling is more intermittent (n.b. this is also consistent with previous studies). However, the overall trends remain very similar in all cases and the correction is minimal in the outer region of the flow (see e.g. $y/\delta \ge 0.1$ in figure \[fig:weights\]).
![Mean velocity $U$ and streamwise turbulence intensity $\overline{u^2}$ profiles corrected by inverse velocity magnitude and inter-arrival time weighting. The wall-normal coordinate is normalized using the local 99% boundary layer thickness, $\delta$.[]{data-label="fig:weights"}](Figures/Fig2-Weights.pdf)
While the corrections accounting for velocity biasing do introduce uncertainty, this uncertainty is likely to be correlated across all cases. Therefore, for comparison across cases, the uncertainty in mean velocity is taken to be the larger of the instrument uncertainty ($0.1$%, MSE Inc.) and the standard error, given by $\overline{\sigma_u}= \sigma_u/\sqrt{N}$, where $\sigma_u$ is the standard deviation of the measurements. In all cases, the standard error was the larger contributer to uncertainty in the near-wall region due to the limited number of samples acquired. As shown in figure \[fig:uncertainty\], the standard error was typically $\overline{\sigma_u} \le 1$% across the entire boundary layer.
![Profile showing standard error of mean velocity (in $\%$) over upstream smooth wall section. The discontinuity in $\overline{\sigma_u}$ between $y/\delta = 0.01$ and $0.02$ corresponds to the location where the measurement duration for each point was reduced from 45 minutes to 10 minutes.[]{data-label="fig:uncertainty"}](Figures/Fig3-StandardError.pdf)
The friction velocity, $u_\tau$, over the smooth-wall section upstream of the porous cutout was estimated by fitting the following relationship to the near-wall mean velocity measurements: $U(y) = (u_\tau^2/\nu)(y^\prime + y_0)$, in which $y_0$ represents an offset from the nominal wall location where the data rate falls to zero, $y^\prime = 0$. The smooth-wall velocity profiles reported below correct for this offset. In other words, the true wall-normal distance is assumed to be $y = y^\prime + y_0$, such that the near-wall velocity profile is consistent with the theoretical relation $U^+ = y^+$ (see figure \[fig:dev-u\]). No such correction was made for the porous substrate. For reference, the friction velocity upstream of the porous section was estimated to be $u_\tau = 2.3 \pm 0.05$ cm/s. The uncertainty is estimated by fitting the relationship in the viscous sublayer to different ranges of 5-10 points near the wall. The offset was $y_0 = 40\mu$m, which translates into approximately one viscous length scale. Note that the above estimate for the friction velocity is also consistent with estimates obtained from fits to the logarithmic region in the mean velocity profile using $\kappa =0.39$ [@marusic2013logarithmic]. The estimates for $u_\tau$ derived above also agreed with the third method utilizing the velocity gradient. See figure \[fig:log-law\] and related discussion in §\[sec:log-shift\] for further detail.
Reynolds number ranges and spatio-temporal resolution {#sec:Re}
-----------------------------------------------------
The 99% boundary layer thickness, estimated via interpolation, was $\delta = 6.83 \pm 0.07$ cm for the smooth-wall profile, and so the friction Reynolds number was $\delta^+ = u_\tau \delta/\nu = 1690 \pm 70$ upstream of the porous section. Over the porous sections, the boundary layer thickness increased; the maximum measured value was $\delta = 11.2 \pm 0.1$ cm. As a result, the Reynolds number based on the nominal free-stream velocity ranged from $Re = U_e \delta/\nu = 42600-69900$. The estimated friction velocity $u_\tau = 2.3 \pm 0.05$ cm/s translates into a viscous length-scale $\nu/u_\tau \approx 40 \mu$m and viscous time-scale $\nu/u_\tau^2 \approx 1.7$ ms. Thus, the 16 $\mu$m precision of the traverse provides adequate vertical resolution for profiling purposes. However, the LDV measurement volume ($150 \mu$m in $y$) extends across 3 viscous units in the wall-normal direction, which means that the near-wall measurements reported below suffer from distortion due to velocity gradients. In the near-wall region, the LDV sampling frequency was approximately 0.5 Hz, which corresponds to an average sampling time of 1200 viscous units. In the outer region of the flow ($y/\delta \gtrsim 0.05$), the sampling frequencies were as high as 50 Hz, which translates into an average sampling time of 12 viscous units. In other words, time-resolved velocity measurements are only expected in the outer region of the flow.
Porous Substrates {#sec:foams}
-----------------
Boundary layer measurements were made adjacent to four different types of open-cell reticulated polyurethane foams. Per the manufacturer, the porosity of all the foams was $\epsilon \approx 0.97$. This was confirmed to within $0.5$% via measurements that involved submerging the foams in water to measure solid volume displacements. The nominal pore sizes corresponded to 10, 20, 60, and 100 pores per inch (ppi, see figure \[fig:foam\]). Pore size distributions for each foam were estimated from photographs of thin foam sheets via image analysis routines in Matlab (Mathworks Inc.). The measured average pore sizes ranged from $s = 2.1 \pm 0.3$ mm for the 10 ppi foam to $s = 0.29 \pm 0.02$ mm for the 100 ppi foam (see Table \[tab:foam\]). These measurements are generally within $\pm 20\%$ of the nominal pore sizes.
Pore size measurements for the 10 and 20 ppi foam were also made using precision calipers. These caliper-based measurements were consistent with the imaging-based estimates to within uncertainty ($s = 2.2 \pm 0.1$mm for 10 ppi and $s = 1.7 \pm 0.1$ mm for 20 ppi, where uncertainties correspond to standard error).
Note that all of the pore sizes discussed above and listed in Table. \[tab:foam\] correspond to the *exposed* streamwise-spanwise plane of the foam. Caliper-based measurements suggest that the pore structure may be anisotropic. For the 10 ppi foam, average pore sizes were approximately 14% larger than the listed values in the spanwise-wall normal plane of the foam ($2.5 \pm 0.1$ mm) and 21% larger in the streamwise-wall normal plane ($2.7 \pm 0.1$ mm). Similarly, for the 20 ppi foam, average pore sizes were approximately 10% larger in the spanwise-wall normal plane and 23% larger in the streamwise-wall normal plane.
Another important length scale arises from the permeability, $k$, of the porous medium. Specifically, $\sqrt{k}$ determines the distance to which the shear penetrates into the porous medium [@battiato2012self]. [[[Permeabilities were estimated from pressure drop experiments using Darcy’s law. These estimates ranged from $k = 6.6 \pm 0.6 \times 10^{-9}$ m$^2$ for the 100 ppi foam to $k = 46 \pm1 \times 10^{-9}$ for the 10 ppi foam.]{}]{}]{}
Based on the friction velocity measured upstream of the plate, the inner-normalized pore sizes range from $s^+ \approx 7$ for the 100 ppi foam to $s^+ \approx 52$ for the 10 ppi foam. [[[Similarly, the Reynolds number based on permeability varies between $Re_k = \sqrt{k}u_\tau / \nu \approx 2.0 $ for the 100 ppi foam to $Re_k \approx 5.3 $ for the 10 ppi foam. The baseline thickness tested for all foams was $h=12.7$ mm. For the 20 ppi foam, two additional thicknesses, $h = 6.35$ mm and $h = 25.4$ mm, were also considered. This means that the ratio of foam thickness to average pore size ranged between $h/s = 4.3$ for the thin 20 ppi foam to $h/s = 44$ for the 100 ppi foam. Finally, keep in mind that despite having the same nominal pore sizes, the 10 and 60 ppi foams tested here are not identical to those tested by @manes2011turbulent. For example, the 10 ppi foam tested by @manes2011turbulent had a pore size ($s=3.9$ mm) approximately twice that of the 10ppi foam used here, and a permeability ($k = 160 \times 10^-9$ m$^2$) almost four times higher.]{}]{}]{} Table \[tab:foam\] lists physical properties for all the foams tested in the experiments, along with related dimensionless parameters.
![Photographs showing thin sheets of the 10, 20, 60, and 100 ppi foams (from left to right). Each image represents a 2 cm $\times$ 2 cm cross section.[]{data-label="fig:foam"}](Figures/Fig4-Foams.jpg){width="\textwidth"}
------------------- --------------------- ------------------- ----------------- -------- ------- -------------- ------- -- -- --
Foam $k$($10^{-9}$m$^2$) $\epsilon$ $s$ (mm) $Re_k$ $s^+$ $h/\sqrt{k}$ $h/s$
\[0.1cm\] 10 ppi $46\pm1$ $0.976 \pm 0.003$ $2.1 \pm 0.3$ $5.3$ 52 59 6
(160) (3.9)
\[0.1cm\] 20 ppi 73 8.5
20 ppi thin $30\pm2$ $0.972 \pm 0.003$ $1.5 \pm 0.2$ $4.3$ 37 37 4.3
20 ppi thick 147 17
\[0.1cm\] 60 ppi $7.9\pm0.6$ $0.965 \pm 0.005$ $0.40 \pm 0.03$ $2.2$ 10 143 32
(6) (0.5)
\[0.1cm\] 100 ppi $6.6\pm 0.6$ $0.967 \pm 0.005$ $0.29 \pm 0.02$ $2.0$ 7 156 44
------------------- --------------------- ------------------- ----------------- -------- ------- -------------- ------- -- -- --
: [[[Permeability ($k$), porosity ($\epsilon$), average pore sizes ($s$), and related dimensionless parameters for tested foams. Permeability and pore size values from [@manes2011turbulent] are noted in parenthesis for the 10 and 60 ppi foams. Note that $s^+ = s u_\tau /\nu$ and $Re_k = \sqrt{k} u_\tau/\nu$ are defined using the friction velocity upstream of the porous section. Porosity ($\epsilon$) was estimated from solid volume displacement in water and permeability was estimated from pressure drop experiments.]{}]{}]{}[]{data-label="tab:foam"}
Results {#sec:results}
=======
Boundary layer development {#sec:development}
--------------------------
First, we consider boundary layer development over the porous foams. The results presented below correspond to the 20 ppi foam [[[of thickness $h=12.7$mm]{}]{}]{}. Similar trends were observed for all the porous substrates.
As illustrated schematically in figure \[fig:diag\], the transition from the smooth wall to the porous substrate leads to the development of an internal layer, which starts at the transition point and grows until it spans the entire boundary layer thickness. When this internal layer reaches the edge of the boundary layer, a new equilibrium boundary layer profile is established. This new profile reflects the effects of the porous substrate. Internal layers have been studied extensively in the context of smooth to rough wall transitions in boundary layers [e.g., @antonia1971response; @jacobi2011new]. In particular, previous literature suggests that turbulent boundary layers adjust relatively quickly for transitions from smooth to rough walls; the adjustment occurs over a streamwise distance of $O(10\delta)$. The profiles of mean velocity ($U/U_e$) and streamwise intensity ($\overline{u^2}/U_e^2$) shown in figure \[fig:dev-u\] suggest that the adjustment from smooth to porous wall velocity profiles occurs over a similarly short streamwise distance.
![Streamwise mean velocity (a) and turbulence intensity (b) measured at streamwise locations $x/h={-21.5,11,21,42,53}$ relative to the transition from smooth wall to porous substrate. These data correspond to the 20 ppi foam with $h=12.7$mm. The dashed lines correspond to a linear profile of the form $U^+ = y^+$ in the near-wall region and a logarithmic profile of the form $U^+ = (1/\kappa)\ln(y^+)+B$ in the overlap region, with $\kappa = 0.39$ and $B = 4.3$. The friction velocity was estimated from fitting a linear slope to the near-wall measurements.[]{data-label="fig:dev-u"}](Figures/Fig5-Development.pdf)
Upstream of the porous section, the measured profiles are consistent with previous smooth wall literature. In the near-wall region ($y/\delta < 5\times 10^{-3}$), the measurements agree reasonably well with the fitted linear velocity profile $U^+ = y^+$. The measured mean velocities are higher at the first two measurement locations, which can be attributed to the bias introduced by velocity gradients across the LDV measurement volume. In the overlap region ($0.02 \le y/\delta \le 0.2$), the mean velocity profile is consistent with the logarithmic law: $U^+ = (1/\kappa)\ln (y^+) + B$. The streamwise intensity profile shows the presence of a distinct inner peak at $y/\delta \approx 0.006$, or $y^+ \approx 10$, which is also consistent with previous studies.
The presence of the porous substrate substantially modifies the mean velocity and streamwise intensity profiles. Figure \[fig:dev-u\] shows clear evidence of substantial slip at the porous interface ($U(y\approx 0) > 0.3U_e$). Farther from the interface, there is a velocity deficit relative to the upstream, smooth wall profile. This velocity deficit increases with distance along the porous substrate, and appears to saturate for the final two profiles measured at $x/h \ge 42$. The mean velocity profiles collapse together for $y/\delta \ge 0.5$, suggesting that the outer part of the wake region remains unchanged over the porous substrate.
Consistent with the mean velocity profiles, the streamwise intensity profiles also show a substantial departure from the smooth case. Although there is some scatter in the measurements closest to the interface, the inner peak is replaced by an elevated plateau at $\overline{u^2}/U_e^2 \approx 0.01$, which extends from the porous interface to $y/\delta \approx 0.01$. Further, an outer peak appears near $y/\delta \approx 0.1$. The origin of this outer peak is discussed further in §\[sec:poresize\] below. In general, the streamwise intensity profiles also converge for $x/h \ge 42$.
![Normalized boundary layer thickness $\delta/h$ for the 20 ppi foam measured at streamwise locations $x/h={-21.5,11,21,42,53}$ relative to the substrate transition point.[]{data-label="fig:dev-delta"}](Figures/Fig6-Delta.pdf)
The streamwise evolution of the boundary layer thickness is presented in figure \[fig:dev-delta\]. Boundary layer growth over the first two measurement locations ($x/h \le 21$) past the smooth to porous transition is relatively slow and appears unchanged from the smooth wall boundary layer, suggesting that the internal layer does not yet span the boundary layer thickness at these locations. For the last two measurement locations, $x/h \geq 42$, the boundary layer thickness grows much faster, suggesting that the flow adjustment is complete and that the effects of the porous substrate extend across the entire boundary layer. These observations are consistent with the mean velocity and streamwise intensity profiles shown in figure \[fig:dev-u\]. As an example, the profiles at measurement location $x/h = 11$ show that the internal layer only extends to $y/\delta \approx 0.2$. For $y/\delta \ge 0.2$, the mean velocity and streamwise intensity collapse onto the smooth-wall profiles.
Development data for the remaining foams with are not presented here for brevity. [[[For all the foams of thickness $h \le 12.7$ mm]{}]{}]{}, the velocity measurements and boundary layer thickness data suggest that flow adjustment is complete by the measurement station at $x/h = 42$. Since the incoming boundary layer thickness is $\delta \approx 5.5 h$, the streamwise adjustment happens over $x \approx 8\delta$, which is consistent with previous literature on the transition from smooth to rough walls. [@antonia1971response] [[[However, this analogy to flow adjustment over rough-wall flows breaks down for the the thick 20 ppi foam with $h = 25.4$ mm. For the thick foam, flow adjustment was not complete even at the last measurement location (see results presented in §\[sec:thickness\]). This observation is still in broad agreement with the results presented above since the last measurement location only yields a dimensionless development length of $x/h \approx 26$ for the thick foam, while figure \[fig:dev-delta\] suggests that $x/h \ge 30$ is required for adjustment.]{}]{}]{}
[[[The specific setup considered here may also be seen as flow over a backward facing step, with the region beyond the step filled with a highly porous ($\epsilon > 0.96$) and permeable ($Re_k>1$) material. Step Reynolds numbers in the present experiment range from $Re_h=U_e h/\nu = 4-16\times 10^3$. DNS by @le1997direct and LDV measurements by @jovic1994backward at $Re_h=5100$ for a canonical (i.e., unfilled) backward facing step indicate that the mean velocity profile behind the step does not return to a log-law for $x/h=20$ beyond the step. This is comparable to the development length observed here over the flush-mounted highly porous substrates ($x/h>30$). Thus, it appears that both $\delta$ and $h$ play a role in dictating flow adjustment for the system considered here.]{}]{}]{} Unfortunately, since the present study was limited to four measurement locations along the porous substrate, there is insufficient spatial resolution in the streamwise direction to provide further insight into the scaling behaviour of boundary layer adjustment and growth over porous substrates.
Effect of pore size {#sec:poresize}
-------------------
![Mean velocity (a) and turbulence intensity (b) profiles for the smooth wall and for all the porous foams at $x/h=42$.[]{data-label="fig:ppi-u"}](Figures/Fig7-PoreSizeU.pdf)
Next, we consider the effect of varying pore size on the fully-developed boundary layer profiles measured at $x/h = 42$ [[[for the foams of thickness $h = 12.7$ mm]{}]{}]{}. Figure \[fig:ppi-u\] shows the measured mean velocity and streamwise intensity for each of the foams tested, together with the smooth-wall profile measured upstream of the porous section.
The mean velocity profile over the porous foams is modified in two significant ways with respect to the smooth wall profile taken upstream. First, there is a substantial slip velocity near the porous substrate. This slip velocity is approximately 30$\%$ of the external velocity across all substrates, with little dependence on pore size. The observed slip velocity is consistent with the DNS results of @breugem2006influence, who observed a slip velocity of approximately $30\%$ for their highest porosity case ($\epsilon=0.95$, $Re_k=9.35$). Note that the properties for the foams tested here (listed in table \[tab:foam\]) are similar to the properties of the porous substrate considered in the simulations. Second, there is a mean velocity deficit relative to the smooth-wall case from $0.004\leq y/\delta \leq 0.4$. This deficit generally increases with increasing pore size, though there is some non-monotonic behaviour. The maximum deficit relative to the smooth-wall profile is approximately 15$\%$ for the 100 ppi foam ($s^+ = 7$) and this increases to almost 50$\%$ for the 20 ppi foam ($s^+ = 37$). However, the deficit for the foam with the largest pore sizes (10 ppi, $s^+ = 52$) is smaller than the deficit over the 20 ppi foam. In the outer wake region ($y/\delta > 0.5$), all mean velocity profiles collapse onto the canonical smooth-wall profile again.
The streamwise turbulence intensity profiles, plotted in figure \[fig:ppi-u\]b, show that the inner peak disappears for all the porous foams. Instead, there is a region of elevated but roughly constant intensity that extends from the porous interface to $y/\delta \approx 0.01$. For $y/\delta > 0.01$, the intensity profiles show a strong dependence on pore size. For the smallest pore sizes, the streamwise intensity either decreases slightly (100 ppi) or stays approximately constant until $y/\delta \approx 0.1$ (60 ppi). For foams with larger pore sizes, the intensity increases and a distinct outer peak appears near $y/\delta \approx 0.1$. All the profiles collapse towards smooth-wall values for $y/\delta \ge 0.5$. Consistent with the mean velocity measurements, the streamwise intensity profiles also show non-monotonic behaviour with pore size. While the streamwise intensity in the outer region of the flow generally increases with pore size, the magnitude of the outer peak is higher for the 20 ppi foam compared to the 10 ppi foam. Velocity spectra, presented in §\[sec:spectra\] below, provide further insight into the origin of this outer peak.
![Wall-normal profiles of skewness ($Sk$) for the smooth wall and for all the porous foams at $x/h = 42$.[]{data-label="fig:skew"}](Figures/Fig8-PoreSizeSkew.pdf)
Figure \[fig:skew\] shows skewness profiles over the smooth wall and porous foams. Consistent with previous measurements at comparable Reynolds number [e.g., @mathis2011relationship], the skewness over the smooth wall is negative or close to zero across much of the boundary layer ($0.004 < y/\delta < 1$). In contrast, over the foam substrates, the sign of the skewness is positive until $y/\delta \approx 0.1$. Interestingly, the location of this change in sign for the skewness corresponds to the location of the outer peak in streamwise intensity profiles. Further, the magnitude of the skewness generally increases with pore size, which is consistent with the intensity measurements (the 10 and 20 ppi cases again show non-monotonic behaviour). These observations are particularly interesting given the intrinsic link between skewness and amplitude modulation [@schlatter2010quantifying; @mathis2011relationship], and suggest that structures responsible for the outer peak in streamwise intensity over the foams may have a modulating effect on the interfacial turbulence. This possibility is discussed in greater detail in §\[sec:amplitude-mod\] below.
Effect of substrate thickness {#sec:thickness}
-----------------------------
![Mean velocity (a) and turbulence intensity (b) profiles for the smooth wall and for the 20ppi foam of varying thickness at the same physical location.[]{data-label="fig:thick"}](Figures/Fig9-ThicknessU.pdf)
The thickness of the porous substrate, $h$, is another important parameter that dictates flow behaviour. Figure \[fig:thick\] shows the measured mean velocity and streamwise intensity profiles over 20 ppi foams of varying thickness, $h = 6.35, 12.7$ and $25.4$ mm. In dimensionless terms, these thicknesses correspond to $h/s = 4.3, 8.5$ and $17$, respectively, where $s = 1.5 \pm 0.2$ mm is the average pore size (Table \[tab:foam\]). Note that the measurements shown in figure \[fig:thick\] are for the same physical location, $x = 53.3$cm, which yields normalized distances increasing from $x/h = 21$ for the thickest foam to $x/h = 84$ for the thinnest foam. The foam with thickness $h=12.7$mm is considered the baseline case, since velocity measurements made over this foam have already been discussed in §\[sec:development\] and §\[sec:poresize\]. The foams of thickness $h = 6.35$mm and $h = 25.4$mm are referred to as the thin and the thick foam, respectively. As noted in §\[sec:development\], the flow was fully developed for the thin and baseline foams, but still developing for the thick foam.
Mean velocity profiles for both the thick and thin foam show similar slip velocities at the interface compared to the baseline foam of thickness $h=12.7$ mm ($\approx 0.3 U_e$), though the velocity measured over the thin foam is slightly higher ($\approx 0.35U_e$). In general, the mean profile for the thin foam is consistently higher than that for the baseline foam, and collapses onto the smooth-wall profile above $y/\delta > 0.2$ (black triangles in figure \[fig:thick\]a). In contrast, the mean profile for the thicker foam (white triangles in figure \[fig:thick\]a) is closer to the baseline case in the near-wall region $y/\delta < 0.005$. However, a little farther from the wall, the mean profile for the thick foam diverges from that for the baseline foam. The thick foam mean profile shows a smaller velocity deficit in the region $y/\delta \approx 0.01$ to $y/\delta \approx 0.1$ compared to the baseline case, and verges on the smooth-wall profile for $y/\delta \ge 0.2$.
The streamwise turbulence intensity profiles plotted in figure \[fig:thick\]b show that $\overline{u^2}/U_e^2$ is similar near the interface for all three foams. In fact, the baseline and thin foams show very similar intensity profiles through most of the boundary layer, barring two minor differences. First, the outer peak in streamwise intensity appears closer to the interface for the thin foam. Second, the thin foam profile shows a better collapse onto the smooth-wall profile for $y/\delta \ge 0.3$. The turbulence intensity profile for the thick foam also collapses onto the smooth-wall profile for $y/\delta \ge 0.3$. However, closer to the wall, $\overline{u^2}$ is much higher over the thick foam, and the outer peak in intensity also appears to move slightly closer to the interface. These features are also seen in the velocity spectra presented below.
In summary, for the thin foam, the mean velocity and streamwise intensity measurements both collapse onto the smooth-wall profile for $y/\delta \ge 0.3$. This suggests that the outer-layer similarity hypothesis proposed by Townsend for rough-walled flows [see e.g., @schultz2007rough] also holds for thin porous media. The mean profile for the thick foam is similar to that for the baseline foam close to the interface, but moves closer to the thin foam profile further from the interface. However, the streamwise intensities are significantly higher for the thick foam compared to the thin foam for $y/\delta \le 0.1$, even in regions where the mean profiles show agreement. These discrepancies between the mean velocity and streamwise intensity profiles for the thick foam are consistent with a developing flow.
Velocity spectra {#sec:spectra}
----------------
The premultiplied velocity spectra shown in figure \[fig:spectra\] provide further insight into the origin of the outer peak in streamwise intensity observed over the porous substrates. [[[As is customary in the boundary layer literature, the premultiplied spectrum is defined as $f E_{uu}$, where $f$ is the frequency and $E_{uu}$ is the power spectral density, normalized by $U_e^2$. This quantity is computed for each wall-normal location, and the results are compiled into the contour plots shown in figure \[fig:spectra\].]{}]{}]{} Due to the low data rates obtained near the interface, spectra are only shown for $y/\delta \ge 0.04$. Note that the spectra are expressed in terms of a normalized streamwise wavelength estimated using Taylor’s hypothesis, $U/(f\delta)$.
![Contour maps showing variation in premultiplied frequency spectra [[[(normalized by $U_e^2$)]{}]{}]{} for streamwise velocity as a function of wall-normal distance $y/\delta$ over the smooth wall (a), the 100 ppi foam (b), the 10 ppi foam (c), the thin 20 ppi foam (d), the baseline 20 ppi foam (e), and the thick 20 ppi foam (f). The spectra are plotted against a normalized streamwise length scale, $U/f\delta$, computed using Taylor’s hypothesis. The white box in (a) denotes the region typically associated with VLSMs while the markers ($*$) represent the nominal frequency $f_{KH}$ for structures resembling Kelvin-Helmholtz vortices. [[[The spectra refer to the same physical measurement location, corresponding to $x/h=42$ for the foams with $h = 12.7$mm and $x/h=21$ and $84$ for the thick and thin 20 ppi foams, respectively.]{}]{}]{}[]{data-label="fig:spectra"}](Figures/Fig10-Spectra.pdf)
For the smooth wall case, there is evidence of weak very-large-scale motions (VLSMs), which is similar to results obtained in previous studies at comparable Reynolds number [@hutchins2007evidence]. The box in the top left panel encompasses the spectral region typically associated with VLSMs, i.e. structures of length $6\delta - 10\delta$ ($U/f\delta = 6-10$) located between $y/\delta = 0.06$ and $y^+ = 3.9 \sqrt{Re_\tau}$ [see @hutchins2007evidence; @marusic2010predictive; @smits2011high]. This box coincides with a region of elevated spectral density for the measurements.
Spectra for the porous substrates are different in several ways. For all the foams, the spectra are elevated over the frequency range that corresponds to structures with streamwise length scale $1\delta-5\delta$. The spectra are most energetic [[[at, or below,]{}]{}]{} wall-normal location $y/\delta \approx 0.1$ and remain elevated until $y/\delta \approx 0.3$. There is a marked increase in the spectral energy density from the 100 ppi foam to the 20 ppi foam, and little difference between the spectra for the 20 ppi and 10 ppi foams. Together, these features suggest that the outer peak in streamwise intensity observed over the porous foams [[[in figure \[fig:ppi-u\]b]{}]{}]{} is associated with large-scale structures of length $1\delta-5\delta$ that are distinct from VLSMs.
[[[Spectra for the 20ppi foam of different thickness, plotted in figure \[fig:spectra\]e-f, show a substantial increase in energy for the thickest foam, which is consistent with the elevated streamwise intensity profiles shown in figure \[fig:thick\]b. For the thick foam, the energy is also concentrated closer to the interface compared to the thin and baseline foams with identical pore sizes.]{}]{}]{}
Note that the spectral features described above are consistent with previous experiments and simulations. @manes2011turbulent showed that the premultiplied frequency spectra for streamwise velocity peak at wavenumber $k_x\delta \approx 2-4$ over porous substrates, i.e. corresponding to structures of length $1.5\delta-3\delta$. Similarly, the DNS carried out by @breugem2006influence indicated the presence of spanwise rollers with streamwise length-scale comparable to the total channel height. Since these structures have been linked to a Kelvin-Helmholtz shear instability arising from the inflection point in the mean profile, it is instructive to consider the characteristic frequency associated with this mechanism. Drawing an analogy to mixing layers, @white2007shear and @manes2011turbulent suggested that the characteristic frequency for the Kelvin-Helmholtz instability can be estimated as $f_{KH} = 0.032 \overline{U}/\theta$ [@ho1984perturbed], in which $\overline{U} = (U_p + U_e)/2$ is the *average* velocity in the shear layer with $U_p$ being the velocity deep within the porous medium, and $\theta$ is the momentum thickness. Estimates for this characteristic frequency, converted to streamwise length-scale based on the mean velocity at $y/\delta = 0.1$, $\lambda_{KH}/\delta = U(0.1)/f_{KH}\delta$, are shown in figure \[fig:spectra\]. These estimates assume $U_p \approx 0$, so that $\overline{U} \approx 0.5U_e$, and that the momentum thickness $\theta$ can be approximated based on the measured velocity profile in the fluid domain, i.e. for $y \ge 0$. With these assumptions, the predicted length scales associated with the instability range from $\lambda_{KH}/\delta \approx 4.3$ for the 100 ppi foam to $\lambda_{KH}/\delta \approx 3.6$ for the 10 ppi foam, which is in reasonable agreement with the location of the peaks in the spectra.
There is an evident overprediction of length scale for both the 10 ppi foam (figure \[fig:spectra\]c) and the thick 20 ppi foam (figure \[fig:spectra\]c). This overprediction can be attributed to greater flow penetration into the porous medium for thicker foams with larger pore sizes. Greater flow penetration would create higher velocities inside the porous medium, $U_p$. This would increase $\overline{U}$ and $f_{KH}$, resulting in lower $\lambda_{KH}$. Greater flow penetration into the porous medium may also result in the inflection point moving closer to the interface. In this scenario, conversion from $f_{KH}$ to $\lambda_{KH}$ should be based on a lower mean velocity from $y/\delta < 0.1$. This would also result in lower $\lambda_{KH}$ compared to the estimates shown in figure \[fig:spectra\].
To test whether the observed peaks in streamwise intensity and velocity spectra tracked the location of inflection points in the mean profile, $y_i$, finite-difference approximations of the second derivative of mean velocity, $(d^2U/dy^2)_{y_i} = 0$, were considered. This yielded inflection point locations ranging from $y/\delta \approx 0.03$ to $y/\delta \approx 0.12$ for the 10 and 20 ppi foams, which is broadly consistent with the location of energetic peaks in figure \[fig:spectra\]. However, the second derivatives estimated from experimental data were very noisy, and the inflection point locations were highly sensitive to the accuracy (i.e., order) of the finite-difference approximation and the lower limit in $y$ used for the estimates. As a result, exact inflection point locations are not presented here.
Discussion {#sec:discussion}
==========
Amplitude modulation {#sec:amplitude-mod}
--------------------
As noted earlier, the skewness profiles shown in figure \[fig:skew\] strongly suggest that the amplitude modulation phenomenon that has received significant attention in recent smooth and rough wall literature [e.g., @mathis2009large; @marusic2010predictive; @mathis2013estimating; @pathikonda2017inner] may also be prevalent in turbulent flows over permeable walls.
For smooth wall flows at high Reynolds number, it has been shown that the VLSMs prevalent in the logarithmic region of the flow can have a modulating effect on the small-scale fluctuations found in the near-wall region. Specifically, it has been suggested that the turbulent velocity field in the near-wall region can be decomposed into a large-scale (or low-frequency) component $u_L$ that represents the near-wall signature of the VLSMs, and a ‘universal’ (i.e. Reynolds number independent) small-scale component $u_S$ that is associated with the local near-wall turbulence.
Analysis of the resulting signals indicates that the filtered envelope of the small-scale activity, $E_L(u_S)$, obtained via a Hilbert transform [for details, see @mathis2011relationship], is strongly correlated with the large-scale signal. In other words, the single-point correlation coefficient: $$\label{eq:R}
R = \frac{\overline{u_L E_L(u_S)}}{\sqrt{\overline{u_L^2}}\sqrt{\overline{E_L(u_S)^2}}}$$ tends to be positive in the near-wall region, which suggests that the small-scale signal $u_S$ is modulated by the large-scale signal $u_L$.
Since the local large-scale signal $u_L$ arises from VLSM-type structures centered further from the wall,
these observations have given rise to predictive models of the form $$\label{eq:modulation}
u = u^*(1+\beta u_{OL}) + \alpha u_{OL},$$ where $u(y)$ is the predicted velocity at a specified location in the near-wall region, $u^*(y)$ is a statistically universal small-scale signal at that wall-normal location, $u_{OL}$ is the large-scale velocity measured in the outer region of the flow, and $\alpha(y)$ and $\beta(y)$ are superposition and modulation coefficients, respectively [@marusic2010predictive]. Assuming that the universal small-scale signal can be obtained via detailed experiments or simulations carried out at low Reynolds numbers, equation (\[eq:modulation\]) allows for near-wall predictions at much higher Reynolds numbers based only on measurements of $u_{OL}$ in the outer region of the flow. [[[Note that the modulation coefficient $\beta$ is similar to the single-point correlation coefficient $R$, but also accounts for changes in phase and amplitude of the large-scale signal from the measurement location to the near-wall region. In other words, $\beta$ also accounts for the relationship between $u_{OL}$ and $u_L(y)$, which is thought to be Reynolds-number independent [@marusic2010predictive].]{}]{}]{}
[[[Subsequent studies have shown rigorously that the correlation coefficient $R$ in (\[eq:R\]) is intrinsically linked to the skewness of the velocity [@schlatter2010quantifying; @mathis2011relationship; @duvvuri2015triadic], whereby positive correlations between scales ($R > 0$) translate into increased skewness. Thus, the increase in skewness observed in figure \[fig:skew\] near the porous interface suggests that $R>0$ in this region.]{}]{}]{}
Near-wall measurements made in the present study do not have sufficient time resolution to allow for a quantitative evaluation of the amplitude modulation phenomenon over porous substrates (i.e., a decomposition into small- and large-scale components). However, the increase in the skewness in the near-wall region over the porous substrates suggests that the large-scale structures responsible for the outer peak in streamwise intensity at $y/\delta \approx 0.1$ may have a modulating effect on the turbulence near the interface. This is particularly interesting given that the large-scale structures found over porous substrates are distinct from the VLSMs found over smooth walls, and that the small-scale turbulence near the porous interface may also be modified from the near-wall cycle [@robinson1991coherent; @schoppa2002coherent] found over smooth walls.
Logarithmic region {#sec:log-shift}
------------------
![(a) Scaled velocity gradient $(y/U_e)\partial U/\partial y$ plotted as a function of $y/\delta$ for the smooth wall profile and porous foam data. Equation (\[eq:log-law-fit\]) was fitted to these data to estimate the normalized displacement height, $y_d/h$, and the friction velocity weighted by the von Karman constant, $u_\tau/(\kappa U_e)$, shown in (b) and (c), respectively. Using these estimates for $y_d$ and $u_\tau/\kappa$, the roughness height $k_0/h$, shown in (d), was evaluated from the velocity profiles using equation (\[eq:log-law\]). Dotted lines in (a) represent the upper limits of $y/\delta=0.16,0.20$ and $0.25$ employed in the fitting procedure. The dashed line represents the minimum lower limit, $y/\delta > 0.02$. Larger marker sizes in (b,c,d) denote higher values for the upper limit. The red cross in (c) represents the friction velocity estimated via a linear fit to the near-wall velocity measurements over the smooth wall. [[[Horizontal error bars in panels (b)-(d) represent uncertainty in pore size, $s$.]{}]{}]{}[]{data-label="fig:log-law"}](Figures/Fig11-loglaw.pdf)
Previous studies have devoted considerable effort to testing whether a modified logarithmic region of the form shown in equation (\[eq:log-law\]) exists in turbulent flows over porous substrates. While there is no definitive consensus on how the von Karman constant, $\kappa$, displacement height, $y_d$, and equivalent roughness height, $k_0$, vary with substrate properties, experiments and simulations generally show that the displacement and roughness heights increase with increasing permeability (or pore size). @manes2011turbulent showed that empirical relationships of the form: $$\label{eq:ys}
y_d^+ = 15.1 Re_k - 13.5$$ and $$\label{eq:k0}
k_0^+ = 6.28 Re_k - 9.82,$$ where $Re_k = u_\tau \sqrt{k}/\nu$ is the permeability Reynolds number, led to reasonable fits for the experimental data obtained by @suga2010effects and @manes2011turbulent, but underestimated $y_d$ and $k_0$ for the simulations performed by @breugem2006influence. The von Karman constant decreased relative to smooth wall values but demonstrated a complex dependence on both the permeability Reynolds number and the ratio of displacement height to boundary layer thickness, $y_d/\delta$ [@manes2011turbulent].
Unfortunately, the present study did not involve independent measurements of the shear stress at the interface (or Reynolds’ shear stress in the near-wall region) and so it is not possible to estimate the friction velocity and von Karman constant independently. However, the velocity measurements can still be used to test whether a modified logarithmic region exists, and to estimate $u_\tau/\kappa$, $y_d$ and $k_0$. Taking the partial derivative of equation (\[eq:log-law\]) with respect to $y$ and rearranging yields: $$\label{eq:log-law-fit}
(y+y_d)\frac{\partial U}{\partial y} = \frac{u_\tau}{\kappa}.$$ Following @breugem2006influence and @suga2010effects, we estimated $y_d$ as the value that forces $(y+y_d)\partial U/\partial y$ to be constant over specified ranges of $y/\delta$. Based on equation (\[eq:log-law-fit\]), the resulting constant value for the weighted velocity gradient was assumed to be the friction velocity divided by von Karman constant, $u_\tau/\kappa$. Using these estimates for displacement height and friction velocity, the roughness height was estimated directly from the velocity measurements using equation (\[eq:log-law\]).
Since the fitting procedure described above relies on noisy velocity gradient data (see figure \[fig:log-law\]a), the uncertainty in the fitted values and sensitivity to fitting ranges was evaluated as follows. First, the fitting procedure was carried out over the range $0.02 < y/\delta < 0.16$. The fit was then repeated with the lower limit sequentially increased by one to four measurement points. The estimates of $y_d$, $u_\tau/\kappa$, and $k_0$ reported in figure \[fig:log-law\](b-d) represent an average of the five different values obtained via this process and the error bars represent the standard error. This entire procedure was then repeated for outer limits $y/\delta = 0.20$ and $y/\delta = 0.25$. A similar process was used to evaluate $u_\tau/\kappa$ and $k_0$ for the smooth wall case with the displacement height constrained to be zero, $y_d = 0$. Estimates for $y_d$, $k_0$, and $u_\tau/\kappa$ are reported in table \[tab:log-law\].
Note that the fitting procedure employed by @manes2011turbulent was also considered, where $y_d$ is estimated as the value that minimizes residuals between the measured velocity profile and equation (\[eq:log-law\]) over specified ranges of $y/\delta$. The resulting fitted coefficients are then used to estimate $u_\tau/\kappa$ and $k_0$. This process led to fitted values within the uncertainty ranges shown in figures \[fig:log-law\]b-d.
Assuming $\kappa = 0.39$, the fitting procedure described above led to a friction velocity estimate of $u_\tau = 2.31 \pm 0.02$ cm/s for the smooth wall profiles with outer limit $y/\delta = 0.15$. This is consistent with the value obtained via a linear fit to the near-wall mean velocity measurements, $u_\tau = 2.3 \pm 0.05$ cm/s. The additive constant $B = -(1/\kappa)\ln k_0^+$ in equation (\[eq:log-law\]) was estimated to be $B = 4.8 \pm 0.2$, which is slightly higher than the value, $B = 4.3$, reported in @marusic2013logarithmic, but still broadly consistent with previous literature.
As expected, the fitted log law constants for the porous substrates show a strong dependence on average pore size. Figures \[fig:log-law\](b-d) show that the displacement height, friction velocity, and roughness height generally increase with increasing pore size, though here is some evidence of saturation at the largest pore sizes. Specifically, figure \[fig:log-law\]b suggests that the displacement height levels out above $y_d/h \approx 1$ for the [[[baseline 20 ppi foam, the 10 ppi foam, and the thin 20 ppi foam, for which $s/h > 0.1$]{}]{}]{}. This saturation in $y_d/h$ as a function of [[[normalized]{}]{}]{} pore size is accompanied by saturation, or perhaps slight decreases, in normalized friction velocity $u_\tau/\kappa U_e$ (figure \[fig:log-law\]c) and roughness height $k_0/h$ (figure \[fig:log-law\]d) [[[above $s/h > 0.1$]{}]{}]{}.
While the exact values of the log law constants shown in figure \[fig:log-law\] and table \[tab:log-law\] must be treated with some caution due to the uncertainty associated with the fitting procedure, the overall trends suggest the following physical interpretation. The displacement height $y_d$ represents the level at which momentum is extracted within the porous medium [@jackson1981displacement], or alternatively, the distance to which the turbulence penetrates into the medium [@luhar2008interaction] or the effective plane at which attached eddies are initiated [@poggi2004effect]. For the least permeable substrates tested in the present study (i.e. the 100 ppi and 60 ppi foams), $y_d$ increases approximately linearly with average pore size. In other words, at this low permeability or thick substrate limit with $h/s \gg 1$, the distance to which turbulence penetrates into the porous medium increases with increasing pore size or permeability. Since the flow does not interact with the entire porous medium, the foam thickness $h$ does not play a role. However, with further increases in pore size [[[or decreases in porous medium thickness]{}]{}]{}, at some point the displacement height becomes comparable to the foam thickness $y_d \approx h$. At this high permeability or thin substrate limit, turbulence penetrates the entire porous medium and the foam essentially acts as a roughness or obstruction. Results shown in figure \[fig:log-law\] suggest that [[[the baseline 20 ppi foam, the thin 20 ppi foam, and the 10 ppi foam, for which $h/s \le 10$]{}]{}]{}, may be approaching this thin substrate limit.
Figures \[fig:log-law\]b-d show that the normalized friction velocity and roughness height are strongly correlated with each other as well as the displacement height. Physically, the friction velocity is a measure of momentum transfer into the porous medium while the roughness height is a measure of momentum loss, or friction increase, due to the presence of the complex substrate. As a result, correlation between $k_0$ and $u_\tau$ is unsurprising for boundary layer experiments carried out at constant free-stream velocity (n.b., for channel or pipe flow experiments, the friction velocity can be controlled independently by setting the pressure gradient).
The link between the displacement and roughness heights can be explained by considering the rough-wall literature. For flows over conventional *K*-type roughness, $k_0$ has been shown to depend on both the height of the roughness elements and the solidity $\lambda$, which is defined as the total projected frontal area per unit wall-parallel area [@jimenez2004turbulent]. Similarly, for flows over porous media, $k_0$ can be expected to depend on the displacement height, which represents the thickness of porous medium that interacts with the flow, as well as the porous medium microstructure [see also @jackson1981displacement; @manes2011turbulent]. In other words, a relationship of the form $k_0/y_d = f(\lambda)$ may be appropriate for turbulent flows over porous media. Note that the physical link between the roughness and displacement heights is also evident in the empirical relationships for $y_d$ and $k_0$ shown in equations (\[eq:ys\]-\[eq:k0\]).
------------------- ------------------------ ------------ --------- --------------------- -----------
Substrate $u_\tau/\kappa$ (cm/s) $y_d$ (mm) $y_d/h$ $k_0$ (mm) $k_0/y_d$
\[0.1cm\] smooth 5.8 0 0 $5.3\times 10^{-3}$ $\infty$
\[0.1cm\] 10 ppi 27.5 14.2 1.1 7.8 0.55
\[0.1cm\] 20 ppi 32.9 15.6 1.2 9.8 0.63
20 ppi thin 22.2 8.7 1.4 3.9 0.44
20 ppi thick 26.9 10.2 0.4 5.5 0.52
\[0.1cm\] 60 ppi 16.5 5.3 0.42 1.5 0.28
\[0.1cm\] 100 ppi 11.9 3.2 0.25 0.49 0.15
------------------- ------------------------ ------------ --------- --------------------- -----------
: Fitted values for log-law parameters in dimensional and dimensionless form. Listed values of $u_\tau/\kappa$, $y_d$, and $k_0$ are averages of the three estimates shown in figure \[fig:log-law\], which were obtained for three different outer limits in the fitting procedure. The displacement height is assumed to be zero for the smooth wall flow.[]{data-label="tab:log-law"}
Non-monotonic behaviour with pore size {#sec:nonmonotonic}
--------------------------------------
In many ways, the aforementioned transition from thick substrate behaviour, where turbulence penetration into the porous medium is limited and $y_d$ increases with permeability, to thin substrate behaviour, where turbulence penetrates the entire porous medium and $y_d \approx h$, is analogous to the $\lambda$-dependent transition from dense to sparse canopy behaviour proposed in the vegetated flow literature [@belcher2003adjustment; @luhar2008interaction; @nepf2012flow].
As noted in the introduction, for vegetated flows the distance to which the flow penetrates into the canopy is dependent on the drag length-scale $(C_Da)^{-1}$, where $C_D$ is a representative drag coefficient and $a$ is the frontal area per unit volume. As a result, the ratio of shear penetration to canopy height, $h$, is given by the dimensionless parameter $C_D a h = C_D \lambda$, where $\lambda$ is the solidity as before. For dense canopies with $C_D \lambda \ge O(1)$, the shear layer does not penetrate the entire canopy and so an inflection point is expected in the mean profile. This gives rise to large-scale structures resembling Kelvin-Helmholtz vortices. However, this instability mechanism is expected to weaken in sparse canopies. For $C_D \lambda < O(0.1)$, turbulence penetrates the entire canopy and there is no inflection point in the mean profile [@nepf2012flow].
The non-monotonic behaviour in mean velocity and turbulence intensity observed for the 10 ppi foam in the present experiments could be attributed to a similar weakening of the shear layer instability as the turbulence penetrates the entire porous medium, i.e. as $y_d \approx h$. Consistent with this hypothesis, the reduced magnitude of the outer peak in streamwise intensity for the 10 ppi foam relative to the 20 ppi foam (figure \[fig:ppi-u\]b) suggests that the large-scale structures resembling Kelvin-Helmholtz vortices are weaker over the 10 ppi foam. Since these structures contribute substantially to vertical momentum transfer, a reduction in their strength also translates into a smaller mean velocity deficit (figure \[fig:ppi-u\]a).
Note that there is evidence of non-monotonic behaviour as a function of solidity $\lambda$ in the rough-wall literature as well. Based on a compilation of experimental data, @jimenez2004turbulent showed that the normalized roughness height (i.e. ratio of $k_0$ to roughness dimension) depends on the solidity $\lambda$, and that there are two regimes of behaviour. For sparse roughness with solidity less than $\lambda \approx 0.15$, the normalized roughness increases with increasing $\lambda$. In other words, an increase in frontal area leads to an increase in roughness drag. However, for densely packed roughness with $\lambda \gtrsim 0.15$, the normalized roughness decreases with increasing $\lambda$ because the roughness elements shelter each other. Assuming that the relevant vertical dimension for turbulent flows over porous media is the displacement height $y_d$, we may expect similar non-monotonic behaviour for the normalized roughness $k_0/y_d = f(\lambda)$. The values for $k_0/y_d$ listed in table \[tab:log-law\] provide some support for this hypothesis: the normalized roughness height increases from $k_0/y_d = 0.15$ for the 100 ppi foam to $k_0/y_d = 0.63$ for the baseline 20 ppi foam, before decreasing to $k_0/y_d = 0.55$ for the 10 ppi foam. Bear in mind that, for identical $h$, the solidity is expected to increase with decreasing pore size from the ‘sparsely packed’ 10 ppi foam to the ‘densely packed’ 100 ppi foam.
Although the solidity is a difficult parameter to measure for porous media, it may be estimated for the foams employed here based on simple geometric assumptions. Consider, for instance, a cubic lattice comprising thin rectangular ligaments of cross-section $d \times d$ and length $s$ (i.e., the pore size). Each unit cell of volume $s^3$ in this lattice comprises three orthogonal filaments aligned in the $x,y,$ and $z$ directions intersecting in a three-dimensional cross. Neglecting the overlapping volume at the center of the cross, the porosity is approximately $\epsilon \approx 1 - 3d^2s/s^3 = 1-3(d/s)^2$ for this geometry. For the foams tested here, the porosity is constant, $\epsilon \approx 0.97$. So, the above equation implies that the lattice must be geometrically similar with $d/s \approx \sqrt{(1-\epsilon)/3} = 0.1$. In other words, the ligament width $d$ increases linearly with pore size $s$ to maintain constant $\epsilon$. For this assumed geometry, the frontal area per unit volume for flow in either the $x$, $y$, or $z$ directions is $a \approx 2ds/s^3 = 2d/s^2$, since there are always 2 ligaments with area $ds$ normal to the flow. This results in solidity $\lambda = ah \approx 2(d/s)(h/s) = 0.2(h/s)$. This estimate suggests that the pore size threshold above which non-monotonic behavior is observed in the log-law constants ($s/h \ge 0.12$ in figure \[fig:log-law\]) corresponds to solidity $\lambda \sim O(1)$. Of course, since the foam pore structures do not resemble a cubic lattice (figure \[fig:foam\]), the numerical factors appearing in the equations above are unlikely to be accurate. However, the linear relationship between $\lambda$ and $h/s$ is expected to hold.
Finally, keep in mind that the non-monotonic behavior with solidity and pore size described above does not preclude the possibility of monotonic behavior with permeability Reynolds number $Re_k=u_\tau\sqrt{k}/\nu$. Although $y_d$ and $k_0$ are shown to decrease with increasing pore size, and hence permeability, from the 20 ppi foam to the 10 ppi foam, this decrease in the displacement and roughness heights is also accompanied by a decrease in $u_\tau/\kappa$ (Table \[tab:log-law\]). In other words, $Re_k$ may be lower for the 10 ppi foam compared to 20 ppi foam, even if $\sqrt{k}$ is higher. Unfortunately, this hypothesis cannot be tested further without independent estimates of $u_\tau$.
A note on scaling {#sec:scaling}
-----------------
Previous studies on turbulent flows over porous media indicate that shear penetration into the porous medium depends on the permeability length scale $\sqrt{k}$, which determines the effective flow resistance within the porous medium per the well-known Darcy-Forchheimer equation [@breugem2006influence; @suga2010effects]. This is also evident in the empirical relationship shown in equation (\[eq:ys\]), which indicates that $y_d \approx 15.1\sqrt{k}$ for sufficiently high permeability Reynolds number $Re_k \gg 1$. Although the permeability is related to geometric parameters such as pore size [e.g. $\sqrt{k}/s \approx 0.08$ for the foams tested by @suga2010effects] and frontal area per unit volume, it is essentially a dynamic parameter that is typically estimated from fitting the Darcy-Forchheimer law: $$-\frac{1}{\rho}\frac{\Delta P}{\Delta x} = \frac{\nu U_v}{k} + \frac{C_f}{\sqrt{k}}U_v^2,$$ to experimental pressure drop measurements ($\Delta P/ \Delta x$) across porous media. In the equation above, $U_v$ is the volume-averaged velocity and $C_f$ is the Forchheimer coefficient. Since such pressure drop measurements are usually carried out at low Reynolds number in steady pipe or channel flow with uniformly distributed porous media (essentially a one-dimensional system), there are inherent risks in employing the resulting permeability values for unsteady, three-dimensional, spatially varying flows at higher Reynolds number. Further, the non-linear Forchheimer term that becomes increasingly important at higher speeds requires an additional coefficient $C_f$ that is not a universal constant.
To avoid these issues, the present study presents results primarily as a function of the normalized pore size, $s/h$. For example, figure \[fig:log-law\]b suggests that $y_d \propto s$ until it becomes comparable to foam thickness. Another alternative would be to use the frontal area per unit volume, $a$, and solidity, $\lambda$, as the relevant scales. [[[The frontal area per unit volume is a difficult quantity to measure for complex porous media. However, the discussion presented in the previous section suggests that the solidity $\lambda$ increases linearly with $(h/s)$ for geometrically-similar porous media with constant porosity. As a result, the use of $s/h$ for scaling purposes also allows for greater reconciliation with the canopy flow and rough wall literature.]{}]{}]{}
Conclusions {#sec:conclusions}
===========
The experimental results presented in this paper show that turbulent boundary layers over high-porosity foams are modified substantially compared to canonical smooth wall flows. Development data in §\[sec:development\] suggest that the boundary layer adjusts relatively quickly to the presence of the porous substrate. Specifically, [[[for most of the foams tested,]{}]{}]{} the mean velocity profile adjusts to a new equilibrium over a streamwise distance $<10\delta$, which is similar to the adjustment length observed in previous literature for the transition from smooth to rough walls.[[[However, this rough-wall analogy does not hold for the thickest foam tested, which suggests that the foam thickness may also provide a bound on development length.]{}]{}]{} Fully-developed mean velocity profiles presented in §\[sec:poresize\] show the presence of substantial slip velocity ($>0.3U_e$) that is relatively insensitive to pore size [[[for foams of constant thickness. Profiles in §\[sec:thickness\] also show a near constant slip velocity over substrates with constant pore size and varying thickness.]{}]{}]{} These observations remain to be explained fully.
Profiles of streamwise intensity show the emergence of an outer peak at $y/\delta \approx 0.1$ over the porous substrates, which is associated with large-scale structures of length $2\delta - 4\delta$. Such structures have also been observed in previous simulations and experiments, and are thought to arise from a Kelvin-Helmholtz instability associated with an inflection point in the mean profile. Although the magnitude of the outer peak in streamwise intensity generally increases with pore size, there is some evidence of weakening for the foam with the largest pore size. The log-law fits presented in §\[sec:log-shift\] provide further insight into this non-monotonic behaviour. Specifically, the displacement height increases with [[[normalized pore size, $s/h$,]{}]{}]{} until it becomes comparable to the foam thickness. Further increases in pore size beyond this point do not lead to an increase in $y_d$. In other words, there is a transition from thick substrate behaviour, in which the thickness of the porous medium interacting with the flow is determined by pore size ($y_d \propto s$), to thin substrate behaviour, in which the flow penetrates the entire porous medium ($y_d \approx h$). The weakening in the outer layer structures may be attributed to this transition from thick to thin substrate behaviour. Drawing an analogy to sparse canopy behaviour for vegetated flows, at the thin substrate limit, the mean velocity profile becomes fuller with increasing pore size and ultimately loses the inflection point. This results in a weakening of the Kelvin-Helmholtz instability. [[[For canopy flows, the transition from dense-canopy behavior to sparse-canopy behavior occurs as the solidity parameter becomes small, $\lambda \ll 1$. Simple geometric arguments show that $\lambda \propto h/s$ for the foams tested here, and that the transition from thick- to thin-substrate behavior occurs around $\lambda \sim O(1)$.]{}]{}]{}
Interestingly, the skewness of the near-wall velocity measurements increases substantially over the porous substrates relative to smooth wall values. Further, this increase in skewness is correlated with an increase in the magnitude of the outer peak in streamwise intensity. Given the link between skewness and the amplitude modulation phenomenon, these observations suggest that the large-scale structures that are energetic over porous media may have a modulating influence on the interfacial turbulence. This is analogous to the interaction between VLSMs and near-wall turbulence in smooth wall flows at high Reynolds number. Unfortunately, the near-wall velocity measurements collected as part of this study were not time resolved, and so did not allow for a quantitative evaluation of this effect. However, given the substantial similarities between turbulent flows over porous media and vegetation or urban canopies, further studies into such scale interactions could lead to the development of promising wall models for a variety of flows.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the Air Force Office of Scientific Research under AFOSR grant No. FA9550-17-1-0142 (Program Manager: Dr. Douglas Smith).
[^1]: Email address for correspondence: [email protected]
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Several recent works have aimed to explain why severely overparameterized models, generalize well when trained by Stochastic Gradient Descent (SGD). The emergent consensus explanation has two parts: the first is that there are “no bad local minima”, while the second is that SGD performs implicit regularization by having a bias towards low complexity models. We revisit both of these ideas in the context of image classification with common deep neural network architectures. Our first finding is that there exist bad [*global*]{} minima, [*i.e.*]{}, models that fit the training set perfectly, yet have poor generalization. Our second finding is that given only *unlabeled* training data, we can easily construct initializations that will cause SGD to quickly converge to such bad global minima. For example, on CIFAR, CINIC10, and (Restricted) ImageNet, this can be achieved by starting SGD at a model derived by fitting random labels on the training data: while subsequent SGD training (with the correct labels) will reach zero training error, the resulting model will exhibit a test accuracy degradation of up to $40\%$ compared to training from a random initialization. Finally, we show that regularization seems to provide SGD with an escape route: once heuristics such as data augmentation are used, starting from a complex model (adversarial initialization) has no effect on the test accuracy.'
author:
- |
Shengchao Liu, Dimitris Papailiopoulos\
University of Wisconsin–Madison\
Dimitris Achlioptas\
University of California, Santa Cruz\
bibliography:
- 'bad\_habbits\_arxiv.bib'
title: Bad Global Minima Exist and SGD Can Reach Them
---
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We discuss the progress of the SCUBA Local Universe Galaxy Survey (SLUGS), the first large, statistical sub-mm survey of the local universe. Since our original survey of a sample of 104 *IRAS*-selected galaxies we have recently completed a sample of 78 Optically-Selected galaxies. Since SCUBA is sensitive to the large proportion of dust too cold to be detected by *IRAS* the addition of this optically-selected sample allows us for the first time to determine the amount of cold dust in galaxies of different Hubble types. We detect 6 ellipticals in the sample and find them to have dust masses in excess of $10^{7}$ $M_{\odot}$. We derive local sub-mm luminosity functions, both directly for the two samples, and by extrapolation from the *IRAS* PSCz, and find excellent agreement.'
author:
- 'Catherine Vlahakis, Stephen Eales and Loretta Dunne'
---
Introduction
============
Relatively little is known about the sub-mm properties of “normal” galaxies in the local universe – prior to SCUBA there existed only a handful of sub-mm flux measurements or maps. SLUGS is the first, large, systematic survey of the local sub-mm universe. It consists of a sample selected from the *IRAS* Bright Galaxy Sample (Dunne et al., 2000), and a sample selected from the CfA optical redshift survey (Vlahakis et al., in prep., Fig. 2). With the optically-selected sample we seek to understand for the first time key questions such as how the amount of cold dust varies with Hubble type.
The $850\mu$m Luminosity Function
=================================
Using the *IRAS*-selected sample Dunne et al. produced the first direct estimate of the sub-mm luminosity function (LF). However, since the *IRAS* sample is biased toward galaxies with larger amounts of warmer dust its LF may also be subject to bias. Conversely, the optically-selected sample should, by definition, be free from temperature selection effects.
Here, we derive a direct $850\mu$m LF for the optically-selected sample. However, in order to better constrain the LF at the lower luminosity end we need more data points, spanning a greater range of luminosities. We do this by determining an $850\mu$m LF using $\sim10000$ galaxies from the *IRAS* PSCz survey. We predict their $850\mu$m luminosities by extrapolating their *IRAS* fluxes using the colour-colour relation from SLUGS (Serjeant and Harrison, 2003). In Fig. 1 we compare the extrapolated (PSCz) and direct LFs and find excellent agreement between all 3 LFs.
.2in
Ellipticals in the Optically-Selected SLUGS
===========================================
It was once thought that ellipticals were entirely devoid of dust and gas, but optical absorption studies now show that dust is usually present. Dust masses for the $\sim15\%$ of ellipticals detected by *IRAS* have been found to be as much as a factor of 10–100 higher when estimated from their FIR emission compared to estimates from optical absorption, suggesting a diffuse cold dust component (Goudfrooij and de Jong, 1995 and refs. therein, Bregman et al., 1998). At $850\mu$m we detect 6 ellipticals, from a total of 11 ellipticals in the optically-selected sample, and find them to have dust masses in excess of $10^{7}$ $M_{\odot}$. We will investigate this further with SCUBA observations of a larger sample of ellipticals.
[1]{}
Bregman Joel N. et al., 1998, ApJ, 499, 670
Dunne Loretta et al., 2000, MNRAS, 315, 115
Goudfrooij P. and de Jong T., 1995, A&A, 298, 784
Serjeant Stephen and Harrison Diana, 2003, astro-ph/0309629
Vlahakis C.E. et al., in prep.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The ground states of Na$_x$CoO$_2$ ($0.0<x<1.0$) is studied by the LDA+Gutzwiller approach, where charge transfer and orbital fluctuations are all self-consistently treated [*ab-initio*]{}. In contrast to previous studies, which are parameter-dependent, we characterized the phase diagram as: (1) Stoner magnetic metal for $x>0.6$ due to $a_{1g}$ van-Hove singularity near band top; (2) correlated non-magnetic metal without $e_g^{\prime}$ pockets for $0.3<x<0.6$; (3) $e_g^{\prime}$ pockets appear for $x<0.3$, and additional magnetic instability involves. Experimental quasi-particle properties is well explained, and the $a_{1g}$-$e_g^{\prime}$ anti-crossing is attributed to spin-orbital coupling.'
author:
- 'Guang-Tao Wang$^1$, Xi Dai$^1$, Zhong Fang$^{1}$'
title: 'Phase Diagram of Na$_x$CoO$_2$ Studied by Gutzwiller Density Functional Theory'
---
Transition-metal oxides have complex phase diagrams due to the interplay between the charge, spin and orbital degrees of freedom. Among them Na$_{x}$CoO$_{2}$ is a typical system showing doping-dependent phase control [@NaCoO_phase]. It has been found experimentally as non-magnetic (NM) metal for Na poor side, while Curie-Weiss metal for Na rich side [@NaCoO_phase; @Spin], and A-type (layered) anti-ferromagnetic (AF) state for $x\sim
0.75$ [@A-AF]. In addition, superconductivity is discovered for hydrated Na$_{x}$CoO$_{2}$$ \cdot y$H$_{2}$O ($x\sim
0.35$) [@SC], and charge-spin-orbital ordered states are suggested for $x=0.5$ due to Na ordering [@x=0.5]. The rich properties of Na$_{x}$CoO$_{2}$ attract much of the research interests due to not only its potential applications, but also the challenging theoretical issues in this system generated by both the *multi-orbital* nature and strong e-e correlation.
Na$_{x}$CoO$_{2}$ is crystallized in planar triangle lattice, with each Co site being coordinated by edge-shared oxygen-octahedron. The $e_{g}$ states are about 2eV higher than $t_{2g}$, and the Fermi level is located within the Co-$t_{2g}$ multiplet, which splits again into one $a_{1g}$ and two $e_{g}^{\prime }$ orbitals under trigonal crystal field. For Na concentration $x$, the effective number of $t_{2g}$ electrons per Co is given as 5+$x$, and thus the low energy physics here is dominated by the multiple orbits ($a_{1g} $+$e_{g}^{\prime }$), where charge, spin and orbital degrees of freedom are all active. The rigorous computational tools for such systems are still lacking, and the observed rich phenomena remains far from even qualitative being understood. The main controversial issues are: (1) for $x=0.3$, are there any Fermi surface pockets for the $e_{g}^{\prime }$ band? These pockets are predicted by LDA (local density approximation) calculations [@Singh] but not observed by ARPES [@ARPES]. (2) Is the $x>0.5$ side more correlated than $x<0.5$ side, as suggested by the Curie-Weiss behavior for the Na rich side? It is expected from simple band picture that $x=1.0$ end compound is a band insulator rather than Mott insulator.
Both LDA and LDA+$U$ methods fail for such system due to the insufficient treatment of electron correlation. This is why many issues look controversial following the LDA pictures. For instances, LDA predicts ferromagnetic metal as the ground state for the whole doping region, and LDA+$U$ even enhances the tendency to be ferromagnetic [@Louie]; the band width obtained by LDA or LDA+$U$ is about two times larger than what observed by ARPES [@ARPES2]; the $e_g^\prime$ pockets problem as mentioned above; and etc. To treat the electron correlation more precisely, the Gutzwiller [@ZQWang] and DMFT (dynamic mean field theory) [@Ishida] approaches has been adopted, where fluctuation effects are included. However, those studies are only focused on the $x=0.3$ compound using tight-binding Hamiltonian extracted from LDA, and conflicting results are drawn due to different parameters [@DMFT-new].
In this paper, we show that the above mentioned theoretical challenging issues of this multi-orbital correlated electron systems can be well studied by using the recently developed LDA+Gutzwiller method [@LDA+G], which keeps the parameter-free character of density functional theory (DFT) and includes all possible charge transfer and orbital fluctuation effects self-consistently. As the results, a phase diagram for the whole doping region is constructed, which removes away most of the above mentioned controversial issues, and suggests that physics here dominates by the doping-dependent orbital, charge and spin fluctuations.
To overcome the problem of LDA, the common procedure of LDA+$U$ and DMFT schemes [@DMFT] are to draw out from the LDA Hamiltonian the interaction terms for the localized orbitals, such as the $3d$ or $4f$ states, and then treat the interaction Hamiltonian explicitly in a proper way (beyond LDA). The total Hamiltonian reads:
$$\begin{aligned}
&&H^{LDA+G}=H^{LDA}+H_{int}-H_{dc} \nonumber \\
&&H_{int}=U\sum_{i\alpha} n_{i\alpha}^{\uparrow} n_{i\alpha}^{\downarrow}
+\frac{U^\prime}{2}\sum_{\stackrel{\alpha\neq\beta}{i,\sigma,\sigma^\prime}}
n_{i\alpha}^\sigma n_{i\beta}^{\sigma^\prime}-\frac{J}{2}
\sum_{i\sigma,\alpha\neq\beta} n_{i\alpha}^\sigma n_{i\beta}^\sigma \nonumber \\
&&<H_{dc}>^{LDA}=\overline{U}N(N-1)-\frac{\overline{J}}{2}
\sum_{\sigma}[N^\sigma(N^\sigma-1)]\end{aligned}$$
where $\left\vert i\alpha \right\rangle^\sigma $ are a set of local orbitals with spin index $\sigma$ and occupation number $n_{i\alpha}$ for lattice site $i$; the $U, U^\prime$ and $J$ gives the intra-, inter-orbital repulsive interaction and Hund’s exchange coupling, respectively. The $H_{dc}$ is the double counting term from LDA, where interaction strength $\overline{U}$ and $\overline{J}$ are averaged over orbitals.
If the interaction term is treated by the Hartree-like scheme (LDA+$U$), the correction over LDA is a set of energy shift of the local orbitals, leaving the kinetic part unchanged. This is fine if the fluctuation effect is not strong, but will fail in opposite case, such as the Na$_x$CoO$_2$ system studied here. In this sense, the DMFT method, in which frequency-dependent self-energy is properly computed, is much better than LDA+$U$. However, due to the heavy computational cost for multi-orbtial systems, the current DMFT studies [@Ishida; @DMFT-new] are all applied to tight-binding Hamiltonians extracted from LDA without full charge density self-consistency. This is insufficient if the charge, spin and orbtial degrees of freedom are all active as discussed above for Na$_x$CoO$_2$.
In the LDA+Gutzwiller approach [@LDA+G], the Gutzwiller wave function $|\Psi _{G}\rangle =\hat{P}|\Psi _{0}\rangle $ ($\hat{P}$ is a projection to many-body configuration) is used instead of single slater determinate wave function $|\Psi _{0}\rangle $. The orbital, charge and spin fluctuations can be included by the multi-configuration nature of the Gutzwiller wave function. As the results, a set of orbital-dependent kinetic energy renormalization factor $Z_{\alpha }$ are obtained for the correlated states in addition to the on-site energy shift. Unlike the previous Gutzwiler or DMFT studies [@ZQWang; @Ishida; @DMFT-new], here all charge transfer processes, crystal field and orbital fluctuations are self-consistently treated within the framework of DFT, which allows for the accurate computation of ground state total energy.
We use the plane-wave pseudo-potential method, and choose the Co-$3d$ wannier functions as the correlated local orbitals. The atomic-limit convention $U=U^\prime+2J$ is followed. How to determine the value of $U$ and $J$ is a common problem for LDA+$U$, DMFT and present methods, and no unified way is established yet. Nevertheless, reasonable estimations have been done for $U=3.0\sim 5.0 eV$ and $J\sim 1.0 eV$ for Na$_x$CoO$_2$ system following the literatures [@ZQWang; @Ishida]. Instead of using single fixed $U$, various values have been studied, and our qualitative results are not changed, as shown below. In addition, since our main purpose is to establish a general picture for the physics of Na$_x$CoO$_2$, the structure differences among different doping $x$ are neglected, and the Na doping is treated by virtual crystal approximation (therefore the charge-ordered states with Na ordering at particular doping, which is interesting but not our purpose here, is out of the phase diagram).
Fig.1 shows the phase diagram computed for the whole doping range 0.0$<$$x$$<$1.0. The solid lines and the dashed lines represent the stabilization energies of FM state and layer-type AF state relative to NM solution, respectively. The intra-plane AF state is hard to be stabilized due to geometrical fluctuation of triangle lattice. The LDA (shown in Fig.1(a)) gives magnetic ground states for all doping $x$, which are inconsistent with experiments. In contrast, the phase diagram by LDA+Gutzwiller has three distinct regions, which can be understood as the consequence of competition among crystal field splitting, inter-orbtial charge and spin fluctuation, as discussed in the following parts. As shown in Fig.1(b) and (c), the features of phase diagram are qualitatively the same using $U$=3.0 eV or $U$=5.0 eV.
[*Phase II: correlated non-magnetic metal (0.3$<$$x$$<$0.6)*]{}.
First of all, the NM state is now correctly predicted for this region. It is known that LDA overestimates the tendency to be FM for several systems, such as ruthenates [@Singh-Ru], and this artifact is even enhanced by LDA+$U$. The physical reason is that correlation effect is not properly treated in LDA and LDA+$U$, but it is well included in the present formalism. In Fig.2 we summarize the properties of the NM solutions for the whole doping range. For $x$ larger than $0.3$, as shown in Fig.2 (a), the $e_g'$ bands are fully occupied and the inter-orbital fluctuation (defined as F and S, see caption of Fig.2) is weak, indicating a effective single band system. However with $x$ approaching the phase boundary around $x_c=0.3$, a crossover to multi-band behavior has been detected from the strong inter-orbital spin and charge fluctuation as shown in Fig.2(b). Such fluctuations in low doping area is induced by the Hund’s rule coupling, which favors even distribution of electrons among different orbitals with same spin. Due to the presence of the correlation effect, the quasi-particle band width (kinetic energy) is renormalized by factor $Z_\alpha^2$ which is about 0.5 (0.7) for the $a_{1g}$ ($e_g^\prime$) state at $x$=0.3 (as shown in Fig.2). The same amplitude of renormalization is reported by ARPES [@ARPES].
Secondly, it has long been a controversial issue whether the $e_g^\prime$ states cross the Fermi level? To answer this question, the correlation renormalized level shift is crucial (as shown in Fig.2(c) the $a_{1g}$-$e_g^\prime$ level splitting $\Delta$ is much renormalized compared to LDA results). Unfortunately after including the correlation effect, two studies have been done, and conflicting results are drawn [@ZQWang; @Ishida] for $x$=0.3. It was recently pointed out by Marianetti [*et al.*]{} that the controversial is due to the different choice of crystal-field splitting in their tight-binding model [@DMFT-new]. To go further, in our studies, not only the crystal-field is treated parameter-free, but also the full charge self-consistency is achieved. To see the difference, we performed one-loop calculations (i.e. the charge-density is fixed to the LDA value and only Gutzwiller wave functions are optimized), then Marianetti’s results are recovered, i.e, the $e_g^\prime$ pockets are not present for $x$=0.3. In addition, we also found that even for $x$=0.2 the $e_g^\prime$ pockets are not present in this one-loop calculation. However, after including the charge density self-consistency, the renormalization of level splitting are suppressed (see Fig.2(c)). As the results, we found that $x$=0.3 is the critical point, i.e. $e_g^\prime$ pockets are absent for $x>$0.3 but present for $x<$0.3. Our calculated quasi-particle bands (as shown in Fig.3 and discussed below) can be well compared with ARPES. Since with the interactions we used ($U$=5.0eV and $J$=1.0eV) the static Hartree Fock shift is zero [@Ishida], the renormalized energy level shift here is fully contributed by the fluctuation effect. From Fig.2(c), we found that the $\Delta$ is peaked in the crossover region indicating that the inter-orbital fluctuation is the main reason for the renormalization effect on the energy levels.
[*Phase III: weakly correlated Stoner metal for $x>0.6$.*]{} Clearly seen from the calculated density of states (Fig.3(d)), a sharp van-Hove singularity (VHS) is present near the $a_{1g}$ band top-edge (due to the flat dispersion). For the Na rich side, the Fermi level is shifted close to the VHS, the Stoner instability make the system FM (in-plane). This conclusion is supported by the following facts for this region: (1) LDA works qualitatively well; (2) the magnetic solution only weakly depends on interaction strength $U$ and $J$ (see Fig.1); (3) calculated spin and charge fluctuation are all weak (Fig.2(b)). It is therefore suggested that strong correlation is not the driving force for the magnetic state, instead the VHS is responsible. Furthermore, we correctly predict that the A-type AF state is more stable than FM state (see the difference between solid and dashed lines in Fig.1). From Fig.3(d) for $x$=0.8, it is seen that spin moment mostly comes from the $a_{1g}$ states, which aligns towards the $z$ direction (inter-layer) of the crystal. According to the Goodenough-Kanamori rule, the exchange-coupling along $z$ direction is dominates by AF super-exchange, which stabilizes the A-type AF state. In fact, the estimated inter-layer exchange coupling (from the total energy difference between FM and A-AF solutions) is about $J_c$=3.0 meV, in good agreement with experimental results [@A-AF]. From itinerant FM theory, the size of spin moment may change with raising temperature, but in the presence of strong VHS near band edge, this possibility is prevented by the sharp density barrier. Indeed we found that the calculated moments of FM and A-type AF solutions are the same. Therefore, the $a_{1g}$ moment will behavior like localized spin as observed experimentally. In other words, the Curie-Weiss behavior in this region does not necessarily suggest the [*enhanced correlation*]{}.
[*Phase I: magnetic correlated metal:*]{} The main difference between phase I and II is that the $e_g^\prime$ band start to go cross the Fermi level and the $e_g^\prime$ hole pockets are present. Therefore phase I is effectively a multi-band system. In the meanwhile, the magnetic instability is recovered, and the system is stabilized in FM ground state. Several points should be addressed here to understand this phase: (1) in contrast to phase III, the FM state is slightly more stable than A-type AF state; (2) both $e_g^\prime$ and $a_{1g}$ contribute to the spin moment, in other words, the stabilization of FM state is due to the enhancement of inter-orbit (rather than intra-orbit) spin fluctuation (Fig.2(b)), which is induced by the Hund’s rule coupling; (3) strong correlation (rather than VHS) is responsible for the magnetic instability. It is interesting to note that the critical point $x\sim 0.3$ (boundary between phase I and II) is close to the doping level where superconductivity was observed. The experimental information for this region is quite limited and not conclusive due to the difficulty of sample preparation, our prediction should be evaluated by future experiments.
Finally, we show systematically in the Fig.3(a)-(c) the calculated quasi-particle band dispersion for $x$=0.35, 0.5, and 0.7 after including the spin-orbital-coupling (SOC) effect. The overall picture can be nicely compared to ARPES data [@ARPES]. In particular, (1) the band width renormalization around factor of 2 is now obtained; (2) the $e_g^\prime$ Fermi surface pocket is absent; (3) the $a_{1g}$-$e_g^\prime$ anti-crossing along $\Gamma-K$ line is nothing but a effect of SOC, and the gap around 0.1 eV is comparable to experimental data [@Anti].
In summary, using the recently developed LDA+Gutzwiller method, we are now able to calculate the ground state total energy of correlated multi-orbital systems from [*ab-initio*]{} after taking into account the orbital fluctuation. The calculated phase diagram of Na$_x$CoO$_2$ establishes a general understanding for the physics behind. Most of the discrepancies between experiments and previous theories, such as the $e_g^\prime$ pocket, $a_{1g}$-$e_g^\prime$ anti-crossing, Curie-weiss behavior for $x>$0.5, are self-consistently understood. Three distinct phase regions are identified, which is instructive to future experiments.
We acknowledge valuable discussions with H. Ding, and the supports from NSF of China (No.10334090, 10425418, 60576058), and that from the 973 program of China (No.2007CB925000).
M. L. Foo and,[*et al.*]{}, Phys. Rev. Lett [**92**]{}, 247001 (2004).
Y. Wang, N. S. Rogado, R. J. Cava, N. P. Ong, Nature, [**423**]{}, 425 (2003).
J. Sugiyama et al., Phys. Rev. B 67, 214420 (2003); A.T. Boothroyd, et al., Phys. Rev. Lett. [**92**]{}, 197201 (2004); S. P. Bayrakci et al., Phys. Rev. Lett. 94, 157205 (2005); L. M. Helme, Phys. Rev. Lett. [bf 94]{}, 157206 (2005).
K. Takada, H. Sakurai, E. T. Muromachi, F. Izumi, R. A. Dilanian, and T. Sasaki, Nature, [**422**]{}, 53 (2003).
K.-W. Lee and W. E. Pickett, Phys. Rev. Lett. [**96**]{}, 096403 (2006); G. Gasparovic, and et.al, Phys. Rev. Lett. [**96**]{}, 046403 (2006).
D. J. Singh, Phys. Rev. B [**61**]{}, 13397 (2000).
H. B. Yang et al., Phys. Rev. Lett. 95, 146401 (2005); D. Qian et al., Phys. Rev. Lett. 96, 046407 (2006); D. Qian et al., Phys. Rev. Lett. 96, 216405 (2006); D. Qian, L. Wray, D. Hsieh, L. Viciu, R. J. Cava, J. L. Luo, D. Wu, N. L. Wang, and M. Z. Hasan, Phys. Rev. Lett. 97, 186405 (2006).
P. Zhang and [*et al.*]{} Phys. Rev. Lett [**93**]{},236402 (2004); P. Zhang and [*et al*]{}, Phys. Rev. B [**70**]{}, 085108 (2004).
M.Z. Hasan and [*et al.*]{} Phys. Rev. Lett [**92**]{},246402 (2004); H.B.Yang and [*et al.*]{} Phys. Rev. Lett [**92**]{},246403 (2004)
S. Zhou and [*et al.*]{} Phys. Rev. Lett [**94**]{},206401 (2005).
H. Ishida, M. D. Johannes, and A. Liebsch3, Phys. Rev. Lett.[**94**]{}, 196401 (2005).
C. A. Marianetti, and [*et al.*]{}, Phys. Rev. Lett. [**99**]{}, 246404 (2007).
X. Y. Deng, X. Dai, and Z. Fang, cond-mat/0707.4606.
J. Geck, et. al. Phys. Rev. Lett. [**99**]{}, 046403 (2007).
G. Kotliar, et al. Rev. Mod. Phys. [**78**]{}, 865 (2006).
I. I. Mazin, and D. J. Singh, Phys. Rev. B [**56**]{}, 2556 (1997); Z. Fang, and Terakura, [**64**]{}, 020509 (2001).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Deformation quantization for any Grassmann scalar free field is described via the Weyl-Wigner-Moyal formalism. The Stratonovich-Weyl quantizer, the Moyal $\star$-product and the Wigner functional are obtained by extending the formalism proposed recently in [@imelda] to the fermionic systems of infinite number of degrees of freedom. In particular, this formalism is applied to quantize the Dirac free field. It is observed that the use of suitable oscillator variables facilitates considerably the procedure. The Stratonovich-Weyl quantizer, the Moyal $\star$-product, the Wigner functional, the normal ordering operator, and finally, the Dirac propagator have been found with the use of these variables.'
author:
- 'I. Galaviz$^2$, H. García-Compeán$^{1,2}$'
- 'M. Przanowski'
- 'F.J. Turrubiates'
title: Deformation Quantization of Fermi Fields
---
-1truecm
Introduction
============
Deformation quantization is commonly regarded as an alternative approach to quantization. In principle, it can be implemented into any classical system of particles, fields, strings or string fields. This is based on the philosophy of deforming suitable standard mathematical structures. Such a philosophy was introduced in 1978 by Bayen, Flato, Fronsdal, Lichnerowicz and Sternheimer [@bffls] (for a recent account, see [@disq; @reviews]). The deformation quantization formalism has a firm mathematical basis. However, its application to quantize an arbitrary physical system has still great challenges (see the third reference from [@reviews]).
The canonical realization of deformation quantization is done with the use of the Weyl-Wigner-Moyal (WWM) formalism introduced originally in Refs. [@wwmoriginal], which establishes a one-one correspondence between the operator algebra (on certain Hilbert space) and the algebra of symbols of operators via the so called WWM correspondence. The product of operators is mapped into an associative and noncommutative product of the respective symbols called the $\star$-product. The simplest example is here the famous Moyal $\star$-product. Quantum theory in terms of deformation quantization has been intensively studied for physical systems with a finite number of degrees of freedom.
Recently, much more interesting and difficult case of classical fields (with an infinite number of degrees of freedom) has been also considered by some authors. For example, deformation quantization of scalar field is done in Refs. [@dito; @ditothree; @ditodos; @cfz; @zachosone]. In some of these references it has been shown that the $\star$-product can be modified to absorb the divergences of the field theory with the $\lambda \phi^4$-interaction. This modification is accomplished by a $\lambda$ dependent redefinition of the $\star$-product consistent with a cocycle condition. Free electromagnetic field in deformation quantization within the Coulomb gauge has been studied in [@campos]. Deformation quantization of gravitational field as a constrained system has been discussed in [@antonu; @antond]. Then the linearized gravitational field from the point of view of deformation quantization has been described in [@hquevedo]. It is worth while to note that also perturbation quantum field theory can be given the form of deformation quantization formalism [@paftadq; @ditofour; @pqft]. Moreover, as has been shown in [@strings] deformation quantization theory can be applied to quantize the classical bosonic strings in the light-cone gauge. On the other hand, string theory uses also the deformation quantization formalism (WWM-formalism) to describe the spacetime worldvolume of a D-brane. The presence of a non-zero constant $B$-field on the worldvolume deforms the product of functions (or classical fields) on the D-brane into a Moyal $\star$-product, and the ordinary effective field theory on the D-brane turns into a noncommutative field theory with this Moyal $\star$-product [@sw].
Most of the cases studied by the deformation quantization for systems with a finite number of degrees of freedom deal with bosonic variables. However, the analysis of some classical physical systems requires the description of fermionic degrees of freedom which involve Grassmann variables. These systems have been discussed in the literature for several years [@berezinbook; @casa; @marinov; @sberezin; @dewittbook; @susyqmb]. (Some extension of the WWM formalism to infinite degrees of freedom for fermions can also be found in [@bereone; @beretwo; @schmutz]). The canonical quantization of the fermionic systems by using the Grassmann variables has been studied in all detail in [@marnelius]. Very recently some authors started to apply the machinery of deformation quantization to quantize the classical fermionic systems [@bordemann; @bordemannII; @hirshfeld; @duetsch; @zachosfermions; @clifford]. In particular, in [@hirshfeld; @imelda] it has been shown concretely how the deformation quantization program can be carried over to specific physical fermionic systems. The same techniques have been applied recently for the noncommutative superspace [@seiberg].
In the present paper we study the canonical approach to deformation quantization for fermionic fields by employing the traditional Weyl-Wigner-Moyal correspondence [@wwmformalism; @tata; @hillery]. To this end we extend the results obtained in the previous paper [@imelda] in the context of fermionic systems with a finite number of degrees of freedom. We deal with the Dirac free field and reproduce several results about its quantization and its Green’s functions.
Our paper is organized as follows. Sec. \[scalar\_field\] is devoted to deformation quantization of a generic Fermi field. First, we construct the Stratonovich-Weyl (SW) quantizer for this Fermi field. Using a modified notion of the [*trace*]{} of an operator one finds that the properties of the fermionic SW quantizer have a similar form as the corresponding properties in the bosonic case. Then the Moyal $\star$-product and the Wigner functional are found. Finally, the normal ordering for the generic Fermi field is briefly discussed. In Sec. \[dirac\] the deformation quantization formalism developed in Sec. \[scalar\_field\] is applied to the case of the Dirac free field. Using the oscillator variables we were able to simplify the considerations. In particular, with the use of these variables we have found the Wigner functional corresponding to the ground state and also the Wigner functional of an arbitrary excited state. Finally, in this section the Dirac propagator within the deformation quantization is computed. Some conclusions and final remarks in Sec. \[remarks\] close the paper.
2truecm
Deformation Quantization of Grassmann Scalar Field {#scalar_field}
==================================================
Consider a scalar Grassmann field on the Minkowski spacetime $M^{d+1}$ of signature $(-,-,\dots,-,+).$ By a Grassmann scalar field we will understand a smooth function $\Theta$ over $M^{d+1}$ which takes values in the [*field*]{} of (anti-commuting) Grassmann numbers $\mathbb{G}$, [*i.e.*]{}, $\Theta$ is the map $\Theta: M^{d+1} \to \mathbb{G}$. Canonical variables of this classical Grassmann field will be denoted by $\Theta(\vec{x},t)$ and $\pi_{\Theta} (\vec{x},t)$ with $(\vec{x},t) \in M^{d+1}=
\mathbb{R}^d \times \mathbb{R}.$
We deal with fields at the instant $t=0$ and we denote $\Theta(\vec{x},0) \equiv \Theta(\vec{x})$ and $\pi_{\Theta}
(\vec{x},0) \equiv \pi_{\Theta}(\vec{x})$. It is worth while to mention that some of the functional formulas and their manipulations are formal. It is also important to notice that since we will deal with Grassmann variables all the computations and results obtained in the present paper are valid under the specified conventions and ordering of factors given in this section. In the present section we study the deformation quantization of the Grassmann fields, including: the Stratonovich-Weyl quantizer, the Moyal $\star$-product, the Wigner functional [@wwmformalism; @tata; @hillery], and the normal ordering. In this paper we follow the conventions and notation used in Ref. [@imelda].
1truecm
The Stratonovich-Weyl Quantizer
-------------------------------
Let $F[\pi_{\Theta},\Theta]$ be a functional on the phase space ${\cal
Z}_{\mathbb{G}}\equiv \{ (\pi_{\Theta}, \Theta)\}$ and let $\widetilde{F}[\lambda, \mu]$ be its Fourier transform $$\widetilde{F}[\lambda,\mu] = \int F[\pi_{\Theta}, \Theta ]
\exp\bigg\{-i \int d^d x \bigg( \pi_{\Theta}(\vec{x}) \cdot
\lambda(\vec{x}) +
\Theta(\vec{x}) \cdot \mu(\vec{x}) \bigg) \bigg\} \prod d \pi_ {\Theta } d \Theta,
\label{uno}$$ where $\lambda(\vec{x})$ and $\mu(\vec{x})$ are the Fourier transformed Grassmann fields corresponding to $\pi_{\Theta}(\vec{x})$ and $\Theta(\vec{x})$ respectively, $\pi_\Theta(\vec{x}) \cdot \lambda(\vec{x}) = \sum_{\alpha =1}^N
\pi_{\Theta \alpha}(\vec{x}) \lambda_\alpha(\vec{x})$ and the functional measure $\prod d \pi_ {\Theta } d \Theta :=
\prod_{\vec{x},\alpha} d \pi_{\Theta \alpha}(\vec{x}) d
\Theta_\alpha(\vec{x}).$ By the analogy to quantum mechanics we can define the Weyl quantization rule as follows $$\widehat{F}= W(F[\pi_{\Theta},\Theta]) := \int
\widetilde{F}[\lambda,\mu] \ \exp\bigg\{i\int d^d x \bigg(
\widehat{\pi}_{\Theta}(\vec{x}) \cdot \lambda(\vec{x}) +
\widehat{\Theta}(\vec{x}) \cdot \mu(\vec{x}) \bigg)\bigg \} \prod
d{\lambda} d \mu, \label{tres}$$ with $\widehat{\pi}_{\Theta}$ and $\widehat{\Theta}$ being the field operators given by $\widehat{\pi}_{\Theta}(\vec{x})
|\pi_{\Theta}\rangle = \pi_{\Theta}(\vec{x}) |\pi_{\Theta}
\rangle$ and $\widehat{\Theta}(\vec{x}) |\Theta\rangle =
\Theta(\vec{x}) |\Theta \rangle$.
The coherent state $|\Theta \rangle$ can be defined in terms of the vacuum state $|0\rangle$ [@wein; @imelda] $$|\Theta\rangle = \exp \left\{-{i\over\hbar} \int d^d x
\widehat{\pi}_\Theta (\vec{x}) \cdot \Theta (\vec{x}) \right\} | 0
\rangle, \label{defini}$$ with $|\Theta\rangle\in {\cal F}_F$, where ${\cal F}_F= {\cal H}_F
\oplus {\cal H}_F\otimes {\cal H}_F \oplus \cdots$ is the Fock space and ${\cal H}_F$ is a Hilbert space of fermions. By using Eq. (\[defini\]) it is easy to see that the functional state satisfies the following property $$\exp \left\{-{i\over\hbar}\int d^d x \widehat{\pi}_\Theta
(\vec{x}) \xi (\vec{x}) \right\} | \Theta \rangle = |\Theta + \xi
\rangle. \label{transla}$$
It is also known that $(|\Theta \rangle)^* \neq \langle \Theta|$, [@wein; @imelda]. The dual Fock space ${\cal F}^*_F$ is constructed from the dual vacuum state $\langle 0|$ as follows $$\begin{aligned}
\langle \Theta | &=& \langle 0 | \prod_{\vec{x},\alpha}
\widehat{\Theta}_\alpha (\vec{x}) \exp \left\{-{i\over\hbar} \int
d^d x \Theta
(\vec{x}) \cdot \widehat{\pi}_\Theta (\vec{x}) \right\}\nonumber \\
&=& \langle 0 | \prod_{\vec{x},\alpha} \widehat{\Theta}_\alpha
(\vec{x}) \exp \left\{{i\over\hbar} \int d^d x
\widehat{\pi}_\Theta (\vec{x}) \cdot \Theta (\vec{x}) \right\}.
\label{bra}\end{aligned}$$ This bra satisfies $\langle \Theta | \widehat{\Theta}_\alpha
(\vec{x}) = \langle \Theta | \Theta_\alpha (\vec{x})$ and $$\langle \Theta' | \Theta \rangle = \prod_{\vec{x},\alpha}
\bigg(\Theta_\alpha (\vec{x}) - \Theta'_\alpha (\vec{x})\bigg) =:
\delta (\Theta - \Theta').$$ Analogously for the momentum representation we define $$|\pi_\Theta \rangle = \exp \left\{ -{i\over\hbar} \int d^d x
\widehat{\Theta} (\vec{x}) \cdot \pi_\Theta (\vec{x}) \right\}
\prod_{\vec{x},\alpha} \widehat{\pi}_{\Theta_\alpha} (\vec{x}) |0
\rangle, \ \ \ \ \ \ \langle \pi_\Theta| = \langle 0 | \exp
\left\{ {i\over\hbar} \int d^d x \widehat{\Theta} (\vec{x}) \cdot
\pi_\Theta (\vec{x}) \right\}.$$ Then one gets $$\langle \pi'_\Theta | \pi_\Theta \rangle = \prod_{\vec{x},\alpha}
\bigg(\pi'_{\Theta_\alpha} (\vec{x}) - \pi_{\Theta_\alpha}
(\vec{x})\bigg) =: \delta (\pi'_{\Theta} - \pi_\Theta).$$ From the above relations it is easy to compute $$\langle \pi_\Theta | \Theta \rangle = \exp \left\{ -{i\over\hbar}
\int d^d x \pi_\Theta (\vec{x}) \cdot \Theta (\vec{x}) \right\},
\hspace{1.0cm} \langle \Theta | \pi_\Theta \rangle =
(i^\infty\hbar)^\infty\exp \left\{ {i\over\hbar} \int d^d x
\pi_\Theta (\vec{x}) \cdot \Theta (\vec{x}) \right\}.
\label{relat}$$ With the use of (\[relat\]) one can obtain the following completeness relations $$\int |\pi_{\Theta} \rangle (-1)^\infty {\cal D} \pi_{\Theta}
\langle \pi_{\Theta} | = \widehat{1}, \hspace{2.0cm} \int | \Theta
\rangle {\cal D} \Theta \langle \Theta | = \widehat{1},
\label{complete}$$ where ${\cal D}\Theta := \widetilde{\prod}_{\vec{x},\alpha} d
\Theta_\alpha(\vec{x}),$ ${\cal D} \pi_{\Theta} :=
\widetilde{\prod}_{\vec{x},\alpha} d\pi_{\Theta \alpha} (\vec{x})$ and $\widetilde{\prod}_{\vec{x},\alpha}$ means that the differentials are ordered oppositely to the ordering of the Grassmann variables in the integrated function.
Return to our description of the Weyl correspondence. Substituting (\[uno\]) into (\[tres\]) one gets $$\widehat{F} = W(F[\pi_{\Theta},\Theta])= \int
F[\pi_{\Theta},\Theta] \widehat{\Omega} [\pi_{\Theta},
\Theta] \prod d\pi_ {\Theta} d\Theta ,
\label{trece}$$ where $\widehat{\Omega}$ is the Stratonovich-Weyl quantizer (see Refs. [@imelda; @wwmformalism]) $$\widehat{\Omega} [\pi_{\Theta},\Theta] = \int \exp \bigg \{ -i
\int d^d x \bigg( \pi_{\Theta}(\vec{x})\cdot \lambda (\vec{x}) +
{\Theta} (\vec{x})\cdot \mu (\vec{x}) \bigg) \bigg \}
\exp\bigg\{i\int d^d x \bigg( \widehat{\pi}_{\Theta}(\vec{x})
\cdot \lambda(\vec{x}) + \widehat{\Theta}(\vec{x})\cdot
\mu(\vec{x}) \bigg)\bigg\}\prod d\lambda d\mu .
\label{catorce}$$
Making use of the well known Campbell-Baker-Hausdorff formula for Grassmann variables in the appropriate ordering, the commutation rules for $\widehat{\pi}_{\Theta}$ and $\widehat{\Theta}$ and the relations (\[transla\]) and (\[complete\]), it is possible to reexpress $\widehat{\Omega}[\pi_{\Theta},\Theta]$ in a very useful form $$\widehat{\Omega}[\pi_{\Theta},\Theta] = (-i)^\infty \int {\cal D}
\mu \exp \bigg \{ -i \int d^d x \ \Theta(\vec{x}) \cdot
\mu(\vec{x}) \bigg \} \bigg{|} \pi_{\Theta} - {\hbar\mu \over 2}
\bigg\rangle \bigg\langle \pi_{\Theta} + {\hbar\mu \over
2}\bigg{|}$$ $$= (i)^\infty \int {\cal D} \lambda \exp \bigg \{ - i \int d^d x
\pi_{\Theta}(\vec{x}) \cdot \lambda(\vec{x}) \bigg\} \bigg{|}
\Theta - {\hbar \lambda \over 2} \bigg\rangle \bigg\langle \Theta
+ {\hbar \lambda \over 2} \bigg{|}. \label{dnueve}$$ Now let us define the “trace” as follows $$\begin{aligned}
{\rm tr}\{ \widehat{\cal O}\} &:=& (i\hbar)^{-\infty}\int{\cal D}
\Theta
\big\langle \Theta \big| \widehat{\cal O} \big| \Theta \big\rangle\nonumber\\
&=&(i\hbar)^{-\infty}\int{\cal D} \pi_{\Theta} \big\langle
\pi_{\Theta} \big| \widehat{\cal O} \big| \pi_{\Theta}
\big\rangle, \nonumber\end{aligned}$$ for any operator $\widehat{\cal O}$ (compare with [@imelda]).
With this definition one can check also that the Stratonovich-Weyl operator satisfies the following properties
$${\rm tr} \big \{ \widehat{\Omega}[\pi_{\Theta},\Theta] \big \} =
1, \label{dseis}$$
$${\rm tr} \bigg \{ \widehat{\Omega}[\pi_{\Theta},\Theta]
\widehat{\Omega}[\pi_{\Theta} ', \Theta '] \bigg \} =
\prod_{\vec{x},\alpha} \big(\Theta_\alpha(\vec{x}) -
\Theta'_\alpha(\vec{x})\big) \big(\pi_{{\Theta}_\alpha}(\vec{x}) -
\pi'_{{\Theta}_\alpha}(\vec{x})\big) =: \delta (\Theta - \Theta' ,
\pi_{\Theta} - \pi_{\Theta} '). \label{dsiete}$$
Finally, multiplying (\[trece\]) by $\widehat{\Omega}[\pi_{\Theta},\Theta]$ and taking into account the property (\[dsiete\]) one easily gets the fundamental relation $$F[\pi_{\Theta},\Theta] = {\rm tr} \bigg \{ \widehat{\Omega}
[\pi_{\Theta},\Theta] \widehat{F} \bigg\}. \label{docho}$$
.5truecm
The Moyal $\star$-Product {#SP}
-------------------------
We are at the position to define the Moyal $\star$-product in field theory involving Grassmann scalar fields. Let $F=F[\pi_{\Theta},\Theta]$ and $G=G[\pi_{\Theta},\Theta]$ be some functionals on $\cal{Z}_{\mathbb{G}}$ that correspond to the field operators $\widehat{F}$ and $\widehat{G}$ respectively, i.e. $F[\pi_{\Theta},\Theta]=W^{-1}(\widehat{F})={\rm tr} \bigg \{
\widehat{\Omega} [\pi_{\Theta},\Theta] \widehat{F} \bigg \}$ and $G[\pi_{\Theta},\Theta]=W^{-1}(\widehat{G}) = {\rm tr} \bigg \{
\widehat{\Omega}[\pi_{\Theta},\Theta] \widehat{G} \bigg \}$. The functional which corresponds to the operator product $\widehat{F}
\widehat{G}$ will be denoted by $(F \star
G)[\pi_{\Theta},\Theta]$. So we have $$(F \star G)[\pi_{\Theta},\Theta]:= W^{-1}(\widehat{F}
\widehat{G})= {\rm tr} \bigg \{
\widehat{\Omega}[\pi_{\Theta},\Theta] \widehat{F} \widehat{G}
\bigg \}. \label{vsiete}$$ Using Eqs. (\[trece\]) and (\[vsiete\]) and performing some simple but rather lengthy calculations one gets $$(F \star G) [\pi_{\Theta},\Theta] = \bigg({ i\hbar \over 2}
\bigg)^{2\infty}\int \ F[\pi_{\Theta} ', \Theta '] G[\pi_{\Theta}
'', \Theta '']$$ $$\times \exp \bigg \{ -{2i \over \hbar} \int d^d x \bigg(
\pi_{\Theta} (\Theta' - \Theta'' ) + \pi_{\Theta}' (\Theta'' -
\Theta ) + \pi_{\Theta}''(\Theta - \Theta ') \bigg) \bigg \} \prod
d\pi_{\Theta} ' d\Theta ' \prod d\pi_{\Theta}'' d \Theta '' .
\label{vocho}$$
Introduce new variables: $\Psi ' = \Theta ' -
\Theta,$ $\Psi '' = \Theta '' - \Theta,$ $\Pi ' = \pi_{\Theta} '
- \pi_{\Theta}$, $\Pi ''= \pi_{\Theta}'' - \pi_{\Theta} .$ Using the expansion of $F[\pi_{\Theta} ', \Theta '] = F[\pi_{\Theta} +
\Pi ', \Theta + \Psi ']$ and $G[\pi_{\Theta} '', \Theta ''] =
G[\pi_{\Theta} + \Pi '', \Theta + \Psi ''] $ into the Taylor series and performing some manipulations we obtain $$\big(F \star G\big)[\pi_{\Theta},\Theta] = F[\pi_{\Theta},\Theta]
\exp\bigg\{{i\hbar\over 2} \buildrel{\leftrightarrow}\over {\cal
P}_{\mathbb{G}}\bigg\} G[\pi_{\Theta},\Theta], \label{treinta}$$ with $$\buildrel{\leftrightarrow}\over {\cal P}_{\mathbb{G}} := \int d^d
x \bigg({{\buildrel{\leftarrow}\over {\delta}}\over \delta
\Theta(\vec{x})} {{\buildrel{\rightarrow}\over {\delta}}\over
\delta \pi_{\Theta}(\vec{x})} + {{\buildrel{\leftarrow}\over
{\delta}}\over \delta \pi_{\Theta}(\vec{x})}
{{\buildrel{\rightarrow}\over {\delta}}\over \delta
\Theta(\vec{x})}\bigg), \label{tuno}$$ where $\overleftarrow{\delta}$ and $\overrightarrow{\delta}$ are the right and left functional derivatives, respectively. The last formula corresponds exactly to the Poisson bracket for two functionals $F$ and $G$ which is given by $$F\buildrel{\leftrightarrow}\over {\cal P}_{\mathbb{G}} G :=\{
F[\pi_{\Theta},\Theta],G[\pi_{\Theta},\Theta] \}_{P}$$ $$= (-1)^{\epsilon_{F}} \int d^d x \bigg\{ {\delta
F[\pi_{\Theta},\Theta] \over \delta\Theta(\vec{x})}{\delta
G[\pi_{\Theta},\Theta] \over \delta \pi_{\Theta}(\vec{x})}
+ {\delta
F[\pi_{\Theta},\Theta] \over \delta \pi_{\Theta}(\vec{x})} {\delta
G[\pi_{\Theta},\Theta] \over \delta \Theta(\vec{x})}\bigg\},
\label{poissonsusy}$$ with $\epsilon_F=0,1$ depending on whether the corresponding functional $F$ is even or odd, respectively. This Poisson bracket has been reported in the literature for many years [@casa; @marinov; @sberezin; @dewittbook; @bordemann; @bordemannII; @duetsch; @zachosfermions; @hirshfeld; @clifford; @imelda].
The associated symplectic form $$\omega_{\mathbb{G}} = \int d^d x\ \delta \Theta(\vec{x}) \wedge
\delta \pi_{\Theta}(\vec{x}), \label{symstruct}$$ provides ${\cal Z}_{\mathbb{G}}$ with the structure of a symplectic supermanifold.
.5truecm
The Wigner Functional
---------------------
If $\widehat{\rho}$ is the density operator of a quantum state then the functional $\rho_{_W}[\pi_{\Theta},\Theta]$ defined by $$\rho_{_W}[\pi_{\Theta},\Theta] = {\rm tr} \bigg \{
\widehat{\Omega}[\pi_{\Theta},\Theta] \widehat{\rho}\bigg \} =
2^{-\infty} \int {\cal D} {\eta}
\exp \bigg\{ - {2i \over \hbar} \int d^d x \ \pi_{\Theta}(\vec{x})
\cdot \eta(\vec{x}) \bigg\} \langle \Theta +\eta| \widehat{\rho} | \Theta - \eta
\rangle
\label{ttres}$$ is called the Wigner functional corresponding to this state. In particular, the Wigner functional $\rho_{_W}[\pi_{\Theta},\Theta]$ corresponding to the quantum state $\widehat{\rho} = |\Phi\rangle
\langle \Phi^\dag|$ , where $\langle \Phi^\dag| := (|\Phi\rangle)^\dag $ , is given by [@wwmoriginal; @tata; @hillery; @cfz; @imelda] $$\rho_{_W}[\pi_{\Theta},\Theta] =2^{-\infty} \int {\cal D} {\eta}
\exp \bigg\{ - {2i \over \hbar} \int d^d x \ \pi_{\Theta}(\vec{x})
\cdot \eta(\vec{x}) \bigg\} \Phi[\Theta + \eta ] \Phi^\dag[\Theta
- \eta ], \label{tcinco}$$ where $ \Phi[\Theta + \eta ] = \langle \Theta +\eta| \Phi \rangle$ and $\Phi^\dag[\Theta - \eta ] = \langle \Phi^\dag | \Theta - \eta
\rangle$.
.5truecm
Normal Ordering
---------------
Let $F[\pi_{\Theta},\Theta]$ be a functional on ${\cal
Z}_\mathbb{G}$. We define $$F_{\widehat{\cal N}} := \widehat{\cal N}F,$$ The operator $\widehat{\cal N}$ is the relativistic field generalization of the respective operator given in [@imelda] by the formula (64), and in the present case it reads $$\widehat{\cal N} := \exp \left\{ {1\over 2} \sum_{i=1}^N \int d^d
p f(E_{\vec{p}}) \bigg( {\overrightarrow{\delta}^2 \over \delta
b_i(\vec{p}) \delta b_i^*(\vec{p})} - {\overrightarrow{\delta}^2
\over \delta d_i^*(\vec{p}) \delta d_i(\vec{p})} \bigg) \right\},$$ where $f(E_{\vec{p}})$ is a function to be determined, $b_i(\vec{p})$, $b_i^*(\vec{p})$ stand for the $W^{-1}-images$ of the annihilation or creation operators (respectively) for particles, and $ d_i(\vec{p})$, $d_i^*(\vec{p})$ denote the $W^{-1}-images $ of annihilation or creation operators for antiparticles.
Then the Berezin-Wick or normal quantization is given by $$\widehat{F}_{\widehat{\cal N}} = \int F_{\widehat{\cal N}}
\widehat\Omega[\pi_\Theta,\Theta] \prod d\pi_{\Theta} d\Theta,$$ and the Weyl mapping of $F_{\widehat{\cal N}}$ gives the normal ordering of the Weyl image of $F_{\widehat{\cal N}}$ i.e. $$: \widehat{F} :=: W(F):\ \buildrel{df}\over {=} W(F_{\widehat
{\cal N}}) \ \buildrel{df}\over {=} W_{\widehat {\cal N}} (F).$$ In the next section the normal ordering in the case of the free Dirac field will be considered in more detail.
1truecm
Deformation Quantization of the Dirac Free Field {#dirac}
================================================
The aim of this section is to provide an example of the application of the deformation quantization of Grassmann fields to the Dirac free field. In addition we compute the propagator of the Dirac field in this context.
In the next subsection we briefly survey the Dirac field in order to introduce the notation and conventions that we use in this paper. We stress the uses of the oscillator variables ${\bf b}^*$ and ${\bf b}$ which allowed us to perform the construction [@hatfield].
The Dirac Free Field
--------------------
In this section we discuss the Dirac free field $\psi(x)$ over Minkowski spacetime $M^{3+1}=\mathbb{R}^3 \times \mathbb{R}$ with the signature $(-,-,-,+)$ and $x=(\vec{x},t) \in M$. The action is given by $$I_D[\psi] = \int d^3xdt {\cal L}_D$$ $$= \int d^3xdt \overline{\psi}(\vec{x},t)\cdot \hbar c \big(i
\not \!
\partial -{mc\over \hbar} \big) \psi(\vec{x},t),
\label{diraclagran}$$ where $\not \! \partial=\gamma^{\mu}\partial_{\mu}$, $\gamma^{\mu}$ are the Dirac matrices ($\mu = 0, \dots, 3$), $\overline{\psi}({x}) \equiv \psi^{\dag}({x}) \gamma^0$, $m$ is the mass parameter, $c$ is the speed of light and $\hbar$ the Planck constant. Thus, the field $\psi({x})$ fulfills the Dirac equation $$\big(i\not \! \partial - {mc\over\hbar} \big) \psi (\vec{x},t) =
0. \label{diraceqn}$$ Its [*conjugate momentum*]{} is given by $\pi_{\psi}(\vec{x},t) =
{-\partial {\cal L}_D \over \partial (
\partial \dot{\psi}) } = i \hbar {\psi}^{\dag}(\vec{x},t)$, where $
\dot{\psi}(\vec{x},t) \equiv {\partial \psi(\vec{x},t) \over
\partial t}$ [^1]. Then the Hamiltonian reads $$H_D[\pi_{\psi}, \psi] = \int d^3x \ \overline{\psi}(\vec{x},t)
\cdot \hbar c \big(-i\gamma^j\partial_j + {mc\over\hbar}
\big)\psi(\vec{x},t)$$ $$= \int d^3x \ \psi^{\dag}(\vec{x},t) \cdot i \hbar
{\partial \over
\partial t} \psi(\vec{x},t)$$ $$= \int d^3x \ \pi_{\psi}(\vec{x},t) \cdot {\partial \over
\partial t} \psi(\vec{x},t),
\label{dhamilton}$$ where as usual $ \pi_\psi \cdot \partial_t \psi \equiv
\sum_{\alpha=1}^4 \pi_{\psi_\alpha} \partial_t \psi_\alpha, $ and $\gamma^i = \beta \alpha^i$, $\gamma^0 = \beta,$ with $i=1,2,3$. Note that we use the Weyl (or chiral) representation of the Dirac matrices $\gamma^\mu,$ $\mu=0,1,2,3$ $$\begin{array}{ccc}
\gamma^0 = \left(\begin{array}{cc}
1 & 0 \\
0 & -1 \\
\end{array}\right), & \hspace{0.5cm} \gamma^j = \left(\begin{array}{cc}
0 & \sigma_j \\
-\sigma_j & 0 \\
\end{array}\right), &\hspace{0.5cm} j=1,2,3 \\
\end{array}$$ with $\sigma_j$ being Pauli’s matrices. The Dirac matrices have the following properties: $ {\gamma^j}^\dag = -\gamma^j,$ ${\gamma^0}^\dag=\gamma^0,$ $ \gamma^\mu \gamma^\nu + \gamma^\nu
\gamma^\mu = 2\eta^{\mu\nu}.$ According to the definition of the Poisson bracket for Grassmann fields given by (\[poissonsusy\]), the Poisson bracket corresponding to the Dirac free field takes the form $$\{F,G\}_P = F \ {\buildrel \leftrightarrow \over {\cal P}}\ G$$ where $$\begin{aligned}
{\buildrel \leftrightarrow \over {\cal P}} &=& \int d^3x \bigg(
{{\buildrel \leftarrow \over \delta}\over \delta \psi(\vec{x},t)}
{{\buildrel \rightarrow \over \delta}\over \delta
\pi_{\psi}(\vec{x},t)} + {{\buildrel \leftarrow \over \delta}\over
\delta \pi_{\psi}(\vec{x},t)} {{\buildrel
\rightarrow \over \delta}\over \delta \psi(\vec{x},t)} \bigg)\nonumber\\
&=&-{i \over \hbar} \int d^3x \bigg( {{\buildrel \leftarrow \over
\delta}\over \delta \psi(\vec{x},t)} {{\buildrel \rightarrow \over
\delta}\over \delta \psi^\dag(\vec{x},t)} + {{\buildrel \leftarrow
\over \delta}\over \delta \psi^\dag(\vec{x},t)} {{\buildrel
\rightarrow \over \delta}\over \delta \psi(\vec{x},t)} \bigg),
\label{PB}\end{aligned}$$ and ${{\buildrel \leftarrow \over \delta}\over \delta
\psi(\vec{x},t)} {{\buildrel \rightarrow \over \delta}\over \delta
\pi_{\psi}(\vec{x},t)} \equiv \sum_{\alpha =1}^4 {{\buildrel
\leftarrow \over \delta}\over \delta \psi_{\alpha}(\vec{x},t)}
{{\buildrel \rightarrow \over \delta}\over \delta
\pi_{\psi_{\alpha}}(\vec{x},t)},$ etc.
Consequently, for $\psi_{\alpha}(\vec{x},t)$ and ${\pi_{\psi}}_{\alpha} (\vec{x},t)$ one gets $$\{ \psi_{\alpha} (\vec{x},t), {\pi_{\psi}}_{\beta} (\vec{y},t)
\}_{P} = \delta (\vec{x}-\vec{y}) \delta_{\alpha\beta},$$ $$\{ \psi_{\alpha} (\vec{x},t), \psi_{\beta} (\vec{y},t) \}_{P} = 0,
\hspace{1.0cm} \{
{\pi_{\psi}}_{\beta}(\vec{x},t),{\pi_{\psi}}_{\beta}(\vec{y},t)
\}_{P}=0. \label{comm}$$ (Remember that $\alpha=1,2,3,4$ runs over the components of the Dirac spinor).
Usually quantization is done by the substitution $
{\{\cdot,\cdot\}}_P \mapsto {1\over i\hbar} {[\cdot,\cdot]}_+$. Thus we obtain $$\begin{array}{l}
{[\widehat{\psi}_\alpha(\vec{x},t),
\widehat{\psi}_\beta(\vec{y},t)]}_+ =
{[\widehat{\pi}_{\psi_\alpha}(\vec{x},t),
\widehat{\pi}_{\psi_\beta}
(\vec{y},t)]}_+ = 0, \\
{[\widehat{\psi}_\alpha(\vec{x},t), \widehat{\pi}_{\psi_\beta}
(\vec{y},t)]}_+
= i\hbar \delta_{\alpha\beta} \delta(\vec{x}-\vec{y}),\\
{[\widehat{\psi}_\alpha(\vec{x},t),
\widehat{\psi}_\beta^\dag(\vec{y},t)]}_+ =
\delta_{\alpha\beta} \delta(\vec{x}-\vec{y}).\\
\end{array}$$
The field variables $\psi_{\alpha}(\vec{x},t)$ and ${\pi_{\psi}}_{\alpha} (\vec{x},t)$ can be expanded into the plane waves $$\psi_{\alpha}(\vec{x},t) = \sum_{i=1}^2 \int {d^3p
\over (2 \pi\hbar)^{3}} {m c^2\over E_{\vec{p}}} \bigg[ {b}_i
(\vec{p},t) u_{\alpha}^i(\vec{p}) e ^{i\vec{p}\cdot\vec{x}/ \hbar
} + {d}_i^* (\vec{p},t) {v_{\alpha}^i}(\vec{p})
e^{-i\vec{p}\cdot\vec{x}/\hbar} \bigg], \label{fieldone}$$ $${\pi_{\psi}}_{\alpha} (\vec{x},t) = i \hbar
{\psi}_{\alpha}^{\dag}(\vec{x},t) = \sum_{i=1}^2 \int {d^3p \over
(2 \pi\hbar)^{3}} \ i \hbar {mc^2 \over E_{\vec{p}}}
\bigg[{b}_i^{*}(\vec{p},t) {{u}_{\alpha}^i}^{\dag}(\vec{p})
e^{-i\vec{p}\cdot\vec{x}/\hbar } + {d}_i(\vec{p},t)
{{v}_{\alpha}^i}^{\dag}(\vec{p}) e^{i\vec{p}\cdot\vec{x}/ \hbar
}\bigg], \label{fieldtwo}$$ where $b_i(\vec{p},t)=b_i(\vec{p})\exp{(-iE_{\vec{p}}t/\hbar)},$ $d_i^*(\vec{p},t)=d_i^*(\vec{p})\exp{(iE_{\vec{p}}t/\hbar)}$ and $E_{\vec{p}} =\sqrt{c^2\vec{p}^2 + m^2c^4}$. Here the index $i=1,2$ stands for the spin degrees of freedom. Therefore, $u^i$ and $v^i$ are solutions of the Dirac equation of momentum $p$ with positive or negative energies, respectively $$(\not \! \! p-mc)u^i=0,\ \ \ \ \ \ \ \ \ \ (\not \! \!
p+mc)v^i=0, \label{pnspinor}$$ with $\not \! \! p= \gamma^{\mu}p_{\mu}$; and they are restricted to satisfy the following relations $${u^i}^{\dag}(\vec{p})u^j(\vec{p})=
{v^i}^{\dag}(\vec{p})v^j(\vec{p})={E_{\vec{p}}
\over mc^2}\delta_{ij},$$ $$\overline{v}^i(\vec{p})u^j(\vec{p})=0.
\label{relations}$$ Classically Dirac fields are described by Grassmann variables since they are functions of the anti-commuting Grassmann variables $b_i(\vec{p}), b_i^*(\vec{p}) ,d_i(\vec{p})$ and $d_i^*(\vec{p})$. Poisson brackets for these variables are
$$\{b_i(\vec{p}),b_j^*(\vec{p}\ ') \}_P =
\{d_i(\vec{p}),d_j^*(\vec{p}\ ')\}_P= -{i \over \hbar} (2
\pi\hbar)^3\bigg({E_{\vec{p}}\over mc^2}
\bigg)\delta_{ij}\delta(\vec{p}-\vec{p}\ '). \label{grass}$$
.5truecm [*Oscillator Variables*]{}
The variables $b$ and $d$ are quite asymmetric and it is not easy to express these variables in terms of the field variables $\psi$ and $\pi_{\psi}$. To carry out the deformation quantization in the simplest possible way we introduce the variables ${\bf
b}(\vec{p},t,r)$ and ${\bf b}^{*}(\vec{p},t,r)$ [@hatfield] which are related to the variables $b$ and $d$ as follows $${\bf b}(\vec{p},t,r)= (2 \pi\hbar)^{-3/2}\sqrt{ mc^2
\over E_{\vec{p}}} b_r(\vec{p},t), \hspace{1.0cm} {\rm with\ }
r=1,2,$$ and $${\bf b}(\vec{p},t,r)= (2 \pi\hbar)^{-3/2}\sqrt{ mc^2 \over
E_{\vec{p}}} d^*_{r-2}(-\vec{p},t), \hspace{1.0cm} {\rm with\ }
r=3,4;$$ and ${\bf b}^*$ is determined by the complex conjugate of these equations.
These variables suggest to redefine the solutions to the Dirac equation in the following form $$w_{\alpha}(\vec{p},r)={u_{\alpha}}^r(\vec{p}), \ \ \ {\rm for } \
\ r=1,2,$$ $$w_{\alpha}(\vec{p},r)= v_{\alpha}^{r-2}(-\vec{p}), \
\ \ {\rm for } \ \ r=3,4. \label{orelat}$$ From (\[relations\]) one gets the relations $${w}^{\dag}(\vec{p},r) w(\vec{p},r')=
{E_{\vec{p}} \over mc^2} \delta_{rr'} ,$$ $$\bar{w}(\varepsilon_r \vec{p},r)
w(\varepsilon_r' \vec{p},r')=\varepsilon_r \delta_{rr'},$$ $$\sum_{r=1}^4 w_{\alpha}(\vec{p},r) w_{\alpha'}^{\dag}(\vec{p},r)=
{E_{\vec{p}}\over mc^2 }\delta_{\alpha \alpha'}. \label{ident}$$
In terms of these variables Eqs. (\[fieldone\]) and (\[fieldtwo\]) read $$\psi_{\alpha}(\vec{x},t) = \sum_{r=1}^4 \int
{d^3p \over (2 \pi\hbar)^{3/2}} \sqrt{mc^2 \over E_{\vec{p}}} {\bf
b}(\vec{p},t,r)w_{\alpha}(\vec{p},r) \exp \big( i
\vec{p}\cdot\vec{x}/\hbar \big), \label{fpone}$$ $${\pi_{\psi}}_{\alpha} (\vec{x},t) = i \hbar
{\psi}_{\alpha}^{\dag}(\vec{x},t) = \sum_{r=1}^4 \int {d^3p \over
(2 \pi\hbar)^{3/2}} \ i \hbar \sqrt{mc^2 \over E_{\vec{p}}} {\bf
b}^{*}(\vec{p},t,r) {w}_{\alpha}^{\dag}(\vec{p},r)
\exp\big(-i\vec{p}\cdot\vec{x}/\hbar \big), \label{fptwo}$$ where ${\bf b}(\vec{p},t,r) = {\bf b}(\vec{p},r)\exp \big\{ -i
\varepsilon_r E_{\vec{p}}\ t /\hbar\big\}$ with $\varepsilon_r=1$ for $r=1,2$ and $\varepsilon_r=-1$ for $r=3,4$.
Substituting Eqs. (\[fpone\]) and (\[fptwo\]) into (\[comm\]) we find the Poisson brackets for the variables ${\bf b}$ y ${\bf
b}^{*}$ $$\{ {\bf b}(\vec{p},r), {\bf b}^{*}(\vec{p}\ ',r') \}_{P} = -{i
\over \hbar} \delta (\vec{p} - \vec{p}\ ') \delta_{rr'},$$ $$\{ {\bf b}(\vec{p},r), {\bf b}(\vec{p}\ ',r')
\}_{P} = 0, \ \ \ \ \ \{ {\bf b}^{*}(\vec{p},r), {\bf
b}^{*}(\vec{p}\ ',r') \}_{P}= 0. \label{ocommrela}$$ Multiplying Eq. (\[fpone\]) by $w_{\beta}^{\dag}(\vec{p}\ ',r')
\exp \big(-i\vec{p}\ '\cdot\vec{x}/\hbar\big)$ and integrating over $\mathbb{R}^3$ we get $$\int d^3x \ w_{\beta}^{\dag}(\vec{p\
}',r')\psi_{\alpha}(\vec{x},t) \exp \big( -i\vec{p}\
'\cdot\vec{x}/\hbar \big)$$ $$= \sum_{r=1}^4 \int {d^3x \ d^3p \over (2 \pi\hbar)^{3/2}}
\sqrt{mc^2 \over E_{\vec{p}}} w_{\beta}^{\dag} (\vec{p}\ ',r')
{\bf b}(\vec{p},t,r)w_{\alpha}(\vec{p},r) \exp \big(
i(\vec{p}-\vec{p}\ ')\cdot\vec{x}/\hbar \big). \label{comput}$$ Using now the relations (\[ident\]), after some computations we obtain $${\bf b}(\vec{p},t,r)=\sqrt{m c^2 \over E_{\vec{p}}}
\sum_{\alpha}\int {d^3x \over (2\pi\hbar)^{3/2}}\
w_{\alpha}^{\dag}(\vec{p},r) \psi_{\alpha}(\vec{x},t) \exp
\big(-i\vec{p}\cdot\vec{x}/\hbar \big), \label{annihi}$$ $${\bf b}^{*}(\vec{p},t,r)=\sqrt{mc^2 \over E_{\vec{p}}}
\sum_{\alpha} \int {d^3x \over (2\pi\hbar)^{3/2}}\
\psi_{\alpha}^{\dag}(\vec{x},t) w_{\alpha}(\vec{p},r)
\exp \big(i\vec{p}\cdot\vec{x}/\hbar \big).
\label{creat}$$ Thus the Grassmann variables ${\bf b}$ and ${\bf b}^{*}$, determine precisely the canonical conjugate variables which we will use to describe the Dirac field in the WWM formalism.
Substituting (\[fpone\]) and (\[fptwo\]) into the Hamiltonian (\[dhamilton\]) yields $$H_D[{\bf b}^*,{\bf
b}]= \sum_{r=1}^4 \int {d^3p} \varepsilon_r E_{\vec{p}} {\bf
b}^*(\vec{p},r) {\bf b}(\vec{p},r), \label{hamione}$$ or in terms of the standard variables, $$H_D[b,b^*,d,d^*]= \sum_{i=1}^2 \int {d^3p \over (2\pi\hbar)^{3}}
mc^2 \bigg[ {b}_i^*(\vec{p}) {b}_i(\vec{p}) - {d}_i(\vec{p})
{d}_i^*(\vec{p}) \bigg]. \label{hamitwo}$$
.5truecm
The Stratonovich-Weyl Quantizer
-------------------------------
The Weyl-Wigner-Moyal correspondence for the Dirac field in the field variables reads $$\widehat{F} = W(F[\pi_{\psi},\psi])= W(F[{\psi}^{\dag},\psi])=
\big({i\hbar}\big)^{-\infty} \int F[{\psi}^{\dag},\psi]
\widehat{\Omega} [{\psi}^{\dag},\psi] \prod d{\psi}^{\dag} d\psi ,
\label{operatorone}$$ where $$\widehat{\Omega}[{\psi}^{\dag},\psi] = (-i)^\infty\int {\cal D}
\mu \exp \bigg \{ -{i } \int d^3x \psi(\vec{x}) \cdot \mu(\vec{x})
\bigg \} \bigg|i \hbar {\psi}^{\dag} - {\hbar\mu \over 2}
\bigg\rangle \bigg\langle i \hbar {\psi}^{\dag} + {\hbar\mu \over
2}\bigg|$$ $$= (i)^\infty \int {\cal D} \lambda \exp \bigg \{\hbar \int d^3x
{\psi}^{\dag}(\vec{x}) \cdot \lambda(\vec{x}) \bigg \} \bigg|\psi
-{\hbar \lambda \over 2} \bigg\rangle \bigg\langle \psi + {\hbar \lambda \over 2}\bigg|.
\label{stratono}$$ In terms of the oscillator variables we have $$\widehat{F} = W(F[{\bf b}^*,{\bf b}])=
\big({i\hbar}\big)^{-\infty} \int \prod {d\bf b}^* {d\bf b} F[{\bf
b}^*,{\bf b}] \widehat{\Omega} [{\bf b}^*,{\bf b}],
\label{operatorsb}$$ and $$\widehat{\Omega}[{\bf b}^*,{\bf b}] =(-i)^\infty \int {\cal D}
\chi \exp \bigg \{ -i\sum_{r=1}^4 \int {d^3p} {\bf b}(\vec{p},r)
\chi(\vec{p},r) \bigg \} \bigg|i \hbar{\bf b}^* - {\hbar\chi \over
2} \bigg\rangle \bigg\langle i \hbar {\bf b}^* + {\hbar\chi \over
2}\bigg|$$ $$= (i)^\infty \int {\cal D} \xi \exp \bigg \{ \hbar \sum_{r=1}^4\int
{d^3p}{\bf b}^*(\vec{p},r) \xi(\vec{p},r) \bigg
\} \bigg|{\bf b} - {\hbar\xi \over 2} \bigg\rangle \bigg\langle
{\bf b} + {\hbar\xi \over 2}\bigg|, \label{stratobs}$$ where $\chi$ and $\xi$ are Dirac spinors.
\[Observe that here $\infty$ is in fact $4\infty$ as $r=4$. Consequently, one can put 1 instead of $i^\infty$ or $(-i)^\infty$\].
.5truecm
The Moyal $\star$-Product {#the-moyal-star-product}
-------------------------
According to the WWM correspondence, the symbol of an operator $\widehat{F}$ is given by $$F[{\bf b}^*,{\bf b} ] = W^{-1}(\widehat{F}) = {\rm tr}
\bigg\{\widehat{\Omega}[{\bf b}^*,{\bf b}] \widehat{F} \bigg\}.
\label{invmap}$$ The Moyal $\star$-product in this case can be defined similarly as it has been done in (\[treinta\]). Let $F_1[{\bf b}^*,{\bf b}]$ and $F_2[{\bf b}^*,{\bf b}]$ be functionals over the Dirac phase space defined by: ${\cal Z}_D = \{(
{\pi_{\psi}}_{\alpha}(\vec{x}),\psi_{\alpha}(\vec{x}))_{\vec{x}\in
\mathbb{R}^3}\} = \{(i \hbar {\bf b}^*(\vec{p},r), ({\bf
b}(\vec{p},r))_{r=1,\cdots,4}\}$ and let $\widehat{F}_1$ and $\widehat{F}_2$ be their corresponding operators. Then by a similar computation to that done in Sec. [\[SP\]]{} we finally get $$\big(F_1 \star F_2\big)[{\bf b}^*,{\bf b}] = F_1[{\bf b}^*,{\bf
b}] \exp\bigg({i\hbar\over 2} \buildrel{\leftrightarrow} \over
{\cal P}_D\bigg) F_2[{\bf b}^*,{\bf b}], \label{fmoyalp}$$ where $$\buildrel{\leftrightarrow}\over {\cal P}_D := -{i \over \hbar}
\sum_{r=1}^{4}\int {d^3p }\bigg( {{\buildrel{\leftarrow}\over
{\delta}}\over \delta {\bf b}(\vec{p},r)}
{{\buildrel{\rightarrow}\over {\delta}}\over \delta {\bf
b}^*(\vec{p},r)} + {{\buildrel{\leftarrow}\over {\delta}}\over
\delta {\bf b}^*(\vec{p},r)} {{\buildrel{\rightarrow}\over
{\delta}}\over \delta {\bf b}(\vec{p},r)} \bigg),
\label{dpoisson}$$ is the Poisson operator. In terms of the standard Grassmann variables $b$, $b^*$, $d$ and $d^*$, it can be rewritten as $$\buildrel{\leftrightarrow}\over {\cal P}_D = - {i\over \hbar}
(2\pi\hbar)^3 \sum_{i=1}^{2}\int {d^3p } \bigg( {E_{\vec{p}} \over
mc^2} \bigg) \bigg \{ \bigg( {{\buildrel{\leftarrow}\over
{\delta}}\over \delta {b}_i(\vec{p})}
{{\buildrel{\rightarrow}\over {\delta}}\over \delta
{b}_i^*(\vec{p})} + {{\buildrel{\leftarrow}\over {\delta}}\over
\delta {b}_i^*(\vec{p})} {{\buildrel{\rightarrow}\over
{\delta}}\over \delta {b}_i(\vec{p})} \bigg)
+ \bigg( {{\buildrel{\leftarrow}\over {\delta}}\over
\delta {d}_i^*(\vec{p})} {{\buildrel{\rightarrow}\over
{\delta}}\over \delta {d}_i(\vec{p})} +
{{\buildrel{\leftarrow}\over {\delta}}\over \delta {d}_i(\vec{p})}
{{\buildrel{\rightarrow}\over {\delta}}\over \delta
{d}_i^*(\vec{p})} \bigg)\bigg\}. \label{dpoissonbs}$$ The operator $\buildrel{\leftrightarrow}\over {\cal P}_D$ determines the Poisson bracket $\{F,G\}_P=F\buildrel{\leftrightarrow} \over {\cal P}_D G$ defined by the symplectic structure $$\omega_D=\int d^3x \sum_{\alpha} \delta \psi_{\alpha}(\vec{x})
\wedge \delta \pi_{\psi}(\vec{x})$$ $$= i \hbar \int d^3x \sum_{\alpha} \delta \psi_{\alpha}(\vec{x})
\wedge \delta \psi_{\alpha}^{\dag}(\vec{x}). \label{dsymplstru}$$ This symplectic structure defines the symplectic supermanifold $({\cal Z}_D,\omega_D)$ which represents the phase space. In terms of the original field variables $\psi_{\alpha}^{\dag},\psi_{\alpha}$ one has $$\buildrel{\leftrightarrow}\over {\cal P}_D := - {i \over \hbar}
\sum_{\alpha}\int d^3x \bigg({{\buildrel{\leftarrow}\over
{\delta}}\over \delta \psi_{\alpha}(\vec{x})}
{{\buildrel{\rightarrow}\over {\delta}}\over \delta
\psi_{\alpha}^{\dag}(\vec{x})} + {{{\buildrel{\leftarrow}\over
{\delta}}\over \delta \psi_{\alpha}^{\dag}(\vec{x})}
{{\buildrel{\rightarrow}\over {\delta}}\over \delta
\psi_{\alpha}(\vec{x})}}\bigg). \label{anotherpoiss}$$ .5truecm
The Wigner Functional
---------------------
The Wigner functional for the Dirac free field is defined in analogy to the scalar field case. Let $\widehat{\rho}^{phys}$ be the density operator corresponding to the quantum physical state of the Dirac field. Then the Wigner functional corresponding to this state is given by $$\rho_{_W}[{\bf b}^*,{\bf b}] = {\rm tr} \bigg\{
\widehat{\Omega}[{\bf b}^*,{\bf b}] \widehat{\rho}^{phys}\bigg\}$$ $$= 2^{-\infty} \int {\cal D} \xi \exp
\bigg\{ -{2}\sum_{r=1}^4 \int d^3p \ {\bf b}^*(\vec{p},r)
\xi(\vec{p},r) \bigg\} \langle {\bf b} - \xi |
\widehat{\rho}^{phys}| {\bf b} + \xi \rangle. \label{wignerf}$$ In the case when $\widehat{\rho}^{phys} = |\Phi \rangle \langle
\Phi^\dag|$ the above equation gives $$\rho_{_W}[{\bf b}^*,{\bf b}] = 2^{-\infty} \int {\cal D} \xi \exp
\bigg\{-{2} \sum_{r=1}^4 \int d^3p \ {\bf b}^*(\vec{p},r)
\xi(\vec{p},r) \bigg\} \Phi[{\bf b} - \xi] \Phi^{\dag}[{\bf b} +
\xi], \label{wignerftwo}$$ where $\Phi[{\bf b}]=\langle {\bf b} | \Phi \rangle$ and $\Phi^\dag[{\bf b}]=\langle \Phi^\dag |{\bf b} \rangle$.
\[Note that since $r$ runs through $1,2,3$ and $4$ i.e., through an even number of indices the Jacobian of transformation leading from (\[ttres\]) or (\[tcinco\]) to (\[wignerf\]) or (\[wignerftwo\]), respectively, is equal to $1$.\]
1truecm [*Example: The Ground State*]{}
In the Schrödinger representation the ground state corresponding to the system described by the Dirac equation can be easily found. In order to do that we use the wave functional defined by $\Phi[\psi] = \langle \psi | \Phi \rangle$.Then the canonical conjugate momentum to $\psi$ is represented by $i \hbar
\widehat{\psi}^{\dag}= i \hbar{\delta \over \delta \psi}$. Substituting this momentum into the Hamiltonian (\[dhamilton\]) we get the time independent Schrödinger equation $$\int d^3x \left(\hbar {\delta \over \delta \psi(\vec{x})}c(-i{\bf
{\buildrel \rightarrow \over \alpha} \cdot {\buildrel \rightarrow
\over \nabla} } + \beta {mc\over\hbar}) \psi(\vec{x})\right)
\Phi[\psi]= E\Phi[\psi]. \label{fdiraceq}$$
It is an easy matter to show that Eqs. (\[fpone\]) and (\[fptwo\]) at $t=0$ can be rewritten in terms of respective operators in the Schrödinger representation as follows $$\widehat{\psi}(\vec{x}) = \sum_{r=1}^4 \int {d^3p \over (2
\pi\hbar)^{3/2}} \sqrt{mc^2 \over E_{\vec{p}}}\widehat{\bf
b}(\vec{p},r)w(\vec{p},r) \exp \big(i\vec{p} \cdot \vec{x} / \hbar
\big),$$ $$\widehat{\psi}^{\dag}(\vec{x})= {\delta \over \delta
\psi(\vec{x})} = \sum_{r=1}^4 \int {d^3p \over (2 \pi\hbar)^{3/2}}
\ \sqrt{mc^2 \over E_{\vec{p}}} {\delta \over \delta {\bf
b}(\vec{p},r) } w^{\dag}(\vec{p},r) \exp\big(-i \vec{p} \cdot
\vec{x} /\hbar \big).$$ Inserting these formulae into (\[fdiraceq\]) one gets $$\sum_{r=1}^4 \int d^3p \varepsilon_r E_{\vec{p}} {\delta \over
\delta {\bf b}(\vec{p},r)} {\bf b}(\vec{p},r)\Phi[{\bf b}] = E
\Phi[{\bf b}], \label{functdeq}$$ where ${\bf b}(\vec{p},r)$ is the eigenvalue of $\widehat{\bf
b}(\vec{p},r)$.
The stable vacuum $|0 \rangle$ for the Dirac free field is the ground state in which all the negative energy states are filled. To be precise consider the case of having only one fermion state and remind that ${\bf b}$ and $\hbar {\delta \over \delta {\bf
b}}$ are the annihilation and creation operators of a single fermion. In addition, let $\Omega_0({\bf b})$ and $\Omega_1({\bf
b})$ be the (one fermion) wave functions of the states $|0\rangle$ and $|1\rangle$, respectively. Hence one has $${\bf b}\Omega_0({\bf b})=0, \ \ \ \ \ \ \ \ \ {\delta \over
\delta {\bf b}}\Omega_1({\bf b})=0,$$ $${\bf b}\Omega_1({\bf b})=\Omega_0({\bf b}), \ \ \ \ \ \ \ \
{\delta \over \delta {\bf b}} \Omega_0({\bf b}) =\Omega_1({\bf
b}). \label{groundone}$$
Since ${\bf b}$ is a Grassmann variable it is easy to see that $\Omega_0({\bf b})={\bf b}$ and $\Omega_1({\bf b})=1$. In order to be consistent with the structure of the stable vacuum the positive energy states ($r=1,2$) are determined by the function ${\bf b}$, and therefore they are empty. While for the negative energy states ($r=3,4$), the function is $1$, and consequently they are occupied.
The Hamiltonian is a sum over all momenta $p$’s and states $r$’s. This implies that the ground state is given by an infinite product $$\Phi_0[{\bf b}]=\eta \prod_{r=1}^2\prod_{\vec{p}} {\bf
b}(\vec{p},r), \label{infprod}$$ where $\eta$ is a normalization factor. Thus the adjoint functional is given by $$\Phi_0^{*}[{\bf b}]=\eta^*\prod_{r=3}^4\prod_{\vec{p}} {\bf
b}(\vec{p},r). \label{adjprod}$$
Now we are at the position to compute the form of the Wigner functional of the ground state corresponding to the Dirac field. Substituting the wave functionals (\[infprod\]) and (\[adjprod\]) into (\[wignerftwo\])and taking $|\eta|^2 =1$ one finally gets $${\rho_{W}}_{0}[{\bf b}^*,{\bf b}] = 2^{-\infty} \int {\cal D} \xi
\exp \bigg\{ -2 \sum_{r=1}^4 \int {d^3p } \ {\bf b}^*(\vec{p},r)
\xi(\vec{p},r) \bigg\} \Phi_0[{\bf b} - \xi] \Phi^\dag_0[{\bf b}
+ \xi]$$ $$= 2^{-\infty} \int {\cal D} \xi \exp \bigg\{ -2 \sum_{r=1}^4 \int
{d^3p } \ {\bf b}^*(\vec{p},r) \xi(\vec{p},r) \bigg\}
\prod_{r=1}^4 \prod_{\vec{p}}({\bf b}(\vec{p},r) - \varepsilon_r
\xi). \label{wigfgs}$$ Recall that $\xi$ is also a Dirac spinor and ${\bf b}(\vec{p},r) +
\varepsilon_r \xi$ is proportional to the Dirac delta function. Then the Wigner functional (including the normalization factor) for the ground state is given by $${\rho_{W}}_{0}[{\bf b}^*,{\bf b}] = 2^{-\infty} \exp \bigg\{- {2}
\int {d^3p } \sum_{r=1}^4 \varepsilon_r{\bf b}^*(\vec{p},r) {\bf
b}(\vec{p},r) \bigg\}. \label{wfgsb}$$ In terms of the standard variables it yields $${\rho_{W}}_{0}[b,b^*,d,d^*] = 2^{-\infty} \exp \bigg\{ -{2} \int
{d^3p \over (2\pi\hbar)^{3}}\bigg( {mc^2 \over E_{\vec{p}}} \bigg)
\sum_{i=1}^2\bigg( b_i^*(\vec{p}) b_i(\vec{p}) -
d_i(\vec{p})d_i^*(\vec{p}) \bigg)\bigg\}. \label{wfgsbd}$$
Remark: One can quickly show that the ground state Wigner functional ${\rho_{W}}_{0}[{\bf b}^*,{\bf b}]$ can also be found within purely deformation quantization language from the following equations $${\bf b}(\vec{p},r=1,2) \star {\rho_{W}}_{0}[{\bf b}^*,{\bf b}]= 0
\ \ \ \ {\rm and} \ \ \ \ {\bf b}^*(\vec{p},r=3,4) \star
{\rho_{W}}_{0}[{\bf b}^*,{\bf b}]= 0.$$ 1truecm [*Wigner Functional for Excited States*]{}
The Winger functionals for excited states can be constructed as follows: If one wants to get the wave functional of an electron (of positive energy) of momentum $\vec{p}$ and spin $s=\pm {1
\over 2}$, one have to remove from the ground state wave functional $\Phi_0[{\bf b}]$ a corresponding factor ${\bf
b}(\vec{p},r=1,2)$, where $r=1$ corresponds to spin up and $r=2$ to spin down, and substitute it by the factor 1. Thus, the wave functional for a state with one electron of momentum $\vec{p}$ and spin up is: $\Phi_{1e}[{\bf b}]=\eta \prod_{r=1}^2\prod_{\vec{q}}
{\bf b}(\vec{q},r),$ except the factor ${\bf b}(\vec{p},1)$. If you prefer to add a positron of momentum $\vec{p}$ and spin $s=\pm
{1 \over 2}$, it is necessary to annihilate an electron of negative energy. In order to do that one needs to multiply $\Phi_0[{\bf b}]$ by ${\bf b} (-\vec{p},r=4,3)$, where $r=4$ corresponds to the positron spin up and $r=3$ to spin down (for details, see [@hatfield]). For example, the wave functional for the positron of momentum $\vec{p}$ and spin up is given by: $\Phi_{1p}[{\bf b}]=\eta {\bf b} (-\vec{p},r=4)
\prod_{r=1}^2\prod_{\vec{q}} {\bf b}(\vec{q},r)$. The Wigner functionals corresponding to excited states can be obtained in this context. Consider the higher quantum state of the Dirac field describing $n$ electrons of the momentum $p_{1},\dots,p_{n}$ with some definite spin, and $n'$ positrons of the momentum $p_{1'},\dots,p_{n'}$ and a definite spin: $|(\vec{p}_1,r_1),
\dots (\vec{p}_n,r_n), (\vec{p}_{1'},r_{1'}), \dots
(\vec{p}_{n'},r_{n'}) \rangle := \widehat{\bf
b}^{\dag}(\vec{p}_1,r_1) \cdots \widehat{\bf
b}^{\dag}(\vec{p}_n,r_n) \widehat{\bf b}(-\vec{p}_{1'},r_{1'})
\cdots
\widehat{\bf b}(-\vec{p}_{n'},r_{n'}) |0\rangle$, where $r_{1},\dots,r_{n}$ take the values $1$ or $2$ and $r_{1'},\dots,r_{n'}$ take 3 or 4. The respective density operator is given by $$\widehat{\rho} = \widehat{\bf b}^{\dag}(\vec{p}_1,r_1) \cdots
\widehat{\bf b}^{\dag}(\vec{p}_n,r_n) \widehat{\bf
b}(-\vec{p}_{1'},r_{1'}) \cdots
\widehat{\bf b}(-\vec{p}_{n'},r_{n'})$$ $$\times |0\rangle \langle 0| \widehat{\bf
b}^{\dag}(-\vec{p}_{n'},r_{n'}) \cdots \widehat{\bf
b}^{\dag}(-\vec{p}_{1'},r_{1'}) \widehat{\bf b}(\vec{p}_n,r_n)
\cdots \widehat{\bf b}(\vec{p}_1,r_1).$$
Thus the Wigner functional corresponding to our excited quantum state reads $${\rho_{W}}[{\bf b}^*,{\bf b}] ={\bf b}^{*}(\vec{p}_1,r_1)\star
\cdots \star {\bf b}^{*} (\vec{p}_n,r_n) \star {\bf
b}(-\vec{p}_{1'},r_{1'})\star \cdots \star {\bf
b}(-\vec{p}_{n'},r_{n'}) \star {\rho_{W}}_{0}$$ $$\star{\bf b}^{*}(-\vec{p}_{n'},r_{n'}) \star \cdots \star {\bf
b}^{*}(-\vec{p}_{1'},r_{1'}) \star {\bf b}(\vec{p}_n,r_n)\star
\cdots \star {\bf b}(\vec{p}_1,r_1), \label{fexcststes}$$ where ${\rho_{W}}_{0}$ is given by (\[wfgsb\]). Due to the form of the Moyal product (\[fmoyalp\]) and (\[dpoisson\])this Wigner functional can be rewritten in the following form $${\rho_{W}}[{\bf b}^*,{\bf b}] ={\bf b}^{*}(\vec{p}_1,r_1) \cdots
{\bf b}^{*} (\vec{p}_n,r_n) {\bf b}(-\vec{p}_{1'},r_{1'})\cdots
{\bf b}(-\vec{p}_{n'},r_{n'}) \star {\rho_{W}}_{0}$$ $$\star {\bf b}^{*}(-\vec{p}_{n'},r_{n'})\cdots {\bf
b}^{*}(-\vec{p}_{1'},r_{1'}) {\bf b}(\vec{p}_n,r_n) \cdots {\bf
b}(\vec{p}_1,r_1).
\label{fexcststesII}$$ From this last formula one can find the Wigner functional for any excited state.
.5truecm
[*Normal Ordering*]{}
The normal ordering of the Dirac field operators can also be defined in the deformation quantization formalism. This can be done by using the operator $\widehat{\cal N}$ acting on the phase space ${\cal Z}_D$ $$\widehat{\cal N} := \exp \bigg\{ {1\over 2} \int {d^3p}
\sum_{r=1}^4 {\varepsilon_r \stackrel{\to}{\delta}^2 \over \delta
{\bf b}(\vec{p},r) \delta {\bf b}^*(\vec{p},r)} \bigg \}.
\label{normalop}$$ In terms of the standard variables $$\widehat{\cal N} = \exp \bigg\{ {1\over 2}{(2\pi\hbar)^3} \int
{d^3p} \bigg( {E_{\vec{p}} \over mc^2} \bigg) \sum_{i=1}^2 \bigg(
{\stackrel{\to}{\delta}^2 \over \delta b_i(\vec{p}) \delta
b_i^*(\vec{p})} - {\stackrel{\to}{\delta}^2 \over \delta
d_i^*(\vec{p}) \delta d_i(\vec{p})}\bigg) \bigg \}.
\label{normalobds}$$
Let $F[{\bf b}^*,{\bf b}]$ be a functional defined on the phase space ${\cal Z}_D$. Then $F_{\widehat{\cal N}}$ is defined as follows $$F_{\widehat{\cal N}}[{\bf b}^*,{\bf b}]=\widehat{\cal
N}F[{\bf b}^*,{\bf b}]. \label{sepone}$$ Consequently, the Weyl image of $F_{\widehat{\cal N}}[{\bf b}^*,{\bf b}]$ gives the normal ordering of the Weyl image of $F[{\bf b}^*,{\bf b}]$ $$: \widehat{F} : = : W(F[{\bf b}^*,{\bf b}]):\ \buildrel{df}\over {=} W(F_{\widehat {\cal
N}}[{\bf b}^*,{\bf b}]) \ \buildrel{df}\over {=} W_{\widehat {\cal
N}} (F[{\bf b}^*,{\bf b}]). \label{septwo}$$
1.5truecm
[*Example: The Hamiltonian*]{}
The normal ordering of the Hamiltonian can be obtained by applying (\[septwo\]) with $\widehat{\cal N}$ given by Eq. (\[normalop\]) to $H_D[{\bf b}^*,{\bf b}]$ given by Eq. (\[hamione\]). Simple calculations yield $$H_{D_{\widehat{\cal N}}} = H_D + 2 \int {d^3p } E_{\vec{p}}
\delta(0). \label{sephamiltwo}$$ Then the normal ordered Hamiltonian operator reads $$:\widehat H_D:\ = \sum_{i=1}^2 \int {d^3p \over (2\pi \hbar)^3}
mc^2 \bigg(\widehat{b}^{\dag}_i(\vec{p}) \widehat{ b}_i(\vec{p}) +
\widehat{d}^{\dag}_i(\vec{p}) \widehat{d}_i(\vec{p}) \bigg).
\label{hamiltlast}$$
1truecm
Dirac Propagator
----------------
In order to compute the propagator of the Dirac field we need to find $$iS_F({x}-{y}) = \langle
0|\widehat{\psi}_{\alpha}({x}) \widehat{\overline{\psi}}_{\beta}({y})|0\rangle \cdot
\theta(t-t') - \langle 0|
\widehat{\overline{\psi}}_{\beta}({y}) \widehat{\psi}_{\alpha}({x})|0\rangle \cdot
\theta(t'-t). \label{propa}$$ So we first compute the quantities $\langle 0|\widehat{\psi}_{\alpha}({x})
\widehat{\overline{\psi}}_{\beta}({y})|0\rangle $ and $\langle 0|
\widehat{\overline{\psi}}_{\beta}({y})\psi_{\alpha}({x})|0\rangle$. In terms of deformation quantization these expectation values are given by (compare with [@campos])
$$\langle 0 |\widehat{\psi}_{\alpha}({x}) \widehat{\overline{\psi}}_{\beta}({y})| 0
\rangle =
{\int \prod d{\bf b}^* d {\bf b} \ \psi_{\alpha}({x})
\star
\overline{\psi}_{\beta}({y}) \ \rho_{W_0} [{\bf b}^*,{\bf b}]
\over \int \prod d{\bf b}^* d {\bf b} \ \rho_{W_0} [{\bf b}^*,{\bf
b}]}, \label{promedio}$$
and the analogous formula for the second expectation value.
Carrying out the corresponding integrations and making use of the following relations $$\sum_{r=1}^2 w_{\alpha}(\vec{p},r) \overline{w}_{\beta}(\vec{p},r)
= {(\not \! \! p + mc)_{\alpha \beta} \over 2mc}, \hspace{1.5cm}
\sum_{r=3}^4 \overline{w}_{\alpha}(\vec{p},r)
{w}_{\beta}(\vec{p},r) = {(\not \! \! p - mc)_{\alpha \beta} \over
2mc}.
\label{secondfor}$$ after straightforward calculations we arrive at the results which are well known in quantum field theory $$\langle 0 |\widehat{\psi}_{\alpha}({x})
\widehat{\overline{\psi}}_{\beta}({y})| 0 \rangle = \int {d^3p
\over (2 \pi\hbar)^3}{c(\not \! \! p + mc)_{\alpha \beta} \over
2E_{\vec{p}}} \exp \big(-i p\cdot(x-y)/\hbar\big);
\label{cienvcinco}$$ and $$\langle 0|
\widehat{\overline{\psi}}_{\beta}({y}) \widehat{\psi}_{\alpha}({x})|0\rangle =
\int {d^3p \over (2 \pi\hbar)^3}{ c(\not \! \! p -
mc)_{\beta \alpha} \over 2E_{\vec{p}}} \exp \big(i
p\cdot(x-y)/\hbar\big). \label{otro}$$ The above formulas reproduce exactly the propagator of the Dirac field in the deformation quantization formalism.
Final Remarks {#remarks}
=============
In this paper we have carried over the deformation quantization program via the Weyl-Wigner-Moyal formalism to fermionic fields. In a sense it is an extension of the results for a finite number of degrees of freedom given in Ref. [@imelda] to field theory. To see how our construction explicitly works we have applied it to the quantization of the Dirac free field. The quantization of this field is possible since the Dirac free field can be regarded as an infinite number of decoupled Fermi oscillators. Consequently, the prescription worked out in Refs. [@hirshfeld; @imelda] can be employed to each oscillator separately. The oscillator variables ${\bf b}$ and ${\bf b}^*$ greatly facilitates the procedure.
In the next step it is natural to consider a formalism which involves both bosonic and fermionic fields, and also their interactions in a similar way to the one given in Ref. [@imelda] for the case of supersymmetric quantum mechanics. Then, gathering the results of Ref. [@campos] concerning deformation quantization of the electromagnetic field with those of the present paper (in addition of some other considerations), a version of QED in the context of deformation quantization can be quickly implemented [@qed].
We are aware that the quantization by deformation seems to be at most equivalent to those of canonical or path integral quantizations. However, there is some evidence that deformation quantization would be more general. For instance, recently it was found that one of the intriguing features of deformation quantization is well defined for spaces with orbifold and conical singularities [@fronsdal]. The results obtained in the present paper would be important for the description of fermions on noncommutative orbifolds [@martinec; @belhaj]. This problem will be reported elsewhere.
2truecm
**Acknowledgments**
It is a pleasure to thank G. Dito and C. Maldonado-Mercado for useful discussions. This work was supported in part by CONACyT México Grant 45713-F. I.G. wish to thank CINVESTAV, Unidad Monterrey, for its hospitality where part of this work was done. The research of I.G. is supported by a CONACyT graduate fellowship.
F. Bayen, M. Flato, C. Fronsdal, A. Lichnerowicz and D. Sternheimer, [*Ann. Phys.*]{} [**111**]{}, 61 (1978); F. Bayen, M. Flato, C. Fronsdal, A. Lichnerowicz and D. Sternheimer, [*Ann. Phys.*]{} [**111**]{}, 111 (1978).
D. Sternheimer, “Quantization is Deformation”, talk http://www.u-bourgogne.fr/monge/d.sternh/papers/DSoberwolfach.pdf.
C.K. Zachos, “Deformation Quantization: Quantum Mechanics Lives and Works in Phase Space”, Int. J. Mod. Phys. A [**17**]{} (2002) 297, hep-th/0110114; A.C. Hirshfeld and P. Henselder, “Deformation Quantization in the Teaching for Quantum Mechanics”, Am. J. Phys. [**70**]{} (2002) 537; G. Dito and D. Sternheimer, “Deformation Quantization: Genesis, Developments and Metamorphoses, [*Deformation Quantization*]{} (Strasbourg 2001) Lect. Math. Theor. Phys. [**1**]{} Ed. de Gruyter, Berlin, IRMA (2002) pp. 9-54.
H. Weyl, [*Group Theory and Quantum Mechanics*]{}, (Dover, New York, 1931); E.P. Wigner, [*Phys. Rev.*]{} [**40**]{}, 749 (1932); A. Groenewold, Physica [**12**]{} (1946) 405-460; J.E. Moyal, [*Proc. Camb. Phil. Soc.*]{} [**45**]{}, 99 (1949).
G. Dito, “Star Product Approach to Quantum Field Theory: The Free Scalar Field”, [*Lett. Math. Phys.*]{} [**20**]{}, 125 (1990).
G. Dito, “Star Products and Nonstandard Quantization for Klein-Gordon Equation”, J. Math. Phys. [**33**]{} (1992) 791.
G. Dito, “An Example of Cancellation of Infinities in the Star Quantization of Fields”, [*Lett. Math. Phys.*]{} [**27**]{}, 73 (1993).
T. Curtright D. Fairlie and C.K. Zachos, [*Phys. Rev.*]{} [**D58**]{}, 025002 (1998).
T. Curtright and C.K. Zachos, J. Phys. A [**32**]{}, 771 (1999).
H. García-Compeán, J.F. Plebański, M. Przanowski and F.J. Turrubiates, “Deformation Quantization of Classical Fields”, Int. J. Mod. Phys. A [**16**]{} (2001) 2533.
F. Antonsen, [*Phys. Rev.*]{} [**D56**]{}, 920 (1997).
F. Antonsen, “Deformation Quantization of Constrained Systems”, gr-qc/9710021.
H. Quevedo and J. G. Tafoya, “Towards the deformation quantization of linearized gravity,” Gen. Rel. Grav. [**37**]{}, 2083 (2005), \[arXiv:gr-qc/0401088\].
M. Duetsch and K. Fredenhagen. “Perturbative Algebraic Field Theory, and Deformation Quantization”, hep-th/0101079.
G. Dito, “Deformation Quantization of Covariant Fields”, math.qa/0202271.
A.C. Hirshfeld and P. Henselder, “Star Products and Perturbative Quantum Field Theory”, Annals Phys. [**298**]{} (2002) 382, hep-th/0208194.
H. García-Compeán, J.F. Plebański, M. Przanowski and F.J. Turrubiates, “Deformation Quantization of Bosonic Strings”, J. Phys. A: Math. Gen. [**33**]{} (2000) 7935.
N. Seiberg and E. Witten, “String theory and Noncommutative Geometry,” JHEP [**9909**]{}, 032 (1999) \[arXiv:hep-th/9908142\].
F.A. Berezin, [*The Method of Second Quantization*]{}, (Academic Press, New York, 1966).
R. Casalbuoni, “The Classical Mechanics for Bose- Fermi Systems”, Nuovo Cimmento [**33**]{} (1976) 389.
F.A. Berezin and M.S. Marinov, “‘Particle Spin Dynamics as the Grassmann Variant of Classical Mechanics’’, Ann. Phys. [**104**]{} (1977) 336.
F.A. Berezin, [*Introduction to Superanalysis*]{}, D. Reidel Publishing Company, Dordrecht (1987).
B. deWitt, [*Supermanifolds*]{}, Second Edition Cambridge University Press, Cambridge (1992).
A. Lahiri, P. Kumar Roy and B. Bagchi, “Supersymmetry in Quantum Mechanics”, Int. J. Mod. Phys. A [**5**]{} (1990) 1383.
F.A. Berezin, Theor. Math. Phys. (USSR) [**6**]{} (1971) 94.
F.A. Berezin, Proc. Moscow Math. Soc. [**17**]{} (1967) 117.
M. Schmutz, Nuovo Cimento B [**25**]{} (1975) 337.
R. Marnelius, “Half-Integer Ghost States and Simple BRST Quantization”, Nucl. Phys. B [**294**]{} (1987) 671; “Fermionic Quantum Mechanics and Superfields”, Int. J. Mod. Phys. A [**5**]{} (1990) 329.
M. Bordemann, “The Deformation Quantization of Certain Super-Poisson Brackets and BRST-Cohomology”, in [*Quantization, Deformations, and Symmetries*]{}, Vol. II, Eds. G. Dito and D. Sternheimer, Kluwer Academic Publishers, Dordrecht (2000).
M. Bordemann, H.-C. Herbig and S. Walmann, Commun. Math. Phys. [**210**]{} (2000) 107.
M. Duetsch and K. Fredenhagen, “Deformation Stability of BRST-quantization,” AIP Conf. Proc. [**453**]{}, 324 (1998) \[arXiv:hep-th/9807215\].
C. Zachos, J. Math. Phys. [**41**]{} (2000) 5129.
A.C. Hirshfeld and P. Henselder, “Deformation Quantization for Systems with Fermions”, Ann. Phys. (N.Y.) [**302**]{} (2002) 59.
A.C. Hirshfeld and P. Henselder, “Clifforddization, Spin and Fermionic Star Products”, Annals Phys. [**314**]{} (2004) 75, quant-ph/0404168.
I. Galaviz, H. García-Compeán, M. Przanowski and F. J. Turrubiates, “Weyl-Wigner-Moyal Formalism for Fermi Classical Systems,” arXiv:hep-th/0612245.
N. Seiberg, “‘Noncommutative Superspace, ${\cal
N}={1 \over 2}$ Supersymmetry, Field Theory and String Theory’’, JHEP [**0306**]{}, 010 (2003), \[arXiv:hep-th/0305248\].
R.L. Stratonovich, [*Sov. Phys. JETP*]{} [**31**]{}, 1012 (1956); A. Grossmann, [*Commun. Math. Phys.*]{} [**48**]{}, 191 (1976) ; J.M. Gracia-Bondía, “Hydrogen Atom in the Phase-space Formulation of Quantum Mechanics”, Phys. Rev. A [**30**]{} (1984) 691; J.M. Gracia Bondía and J.C. Varilly, [*J. Phys. A: Math. Gen.*]{} [**21**]{}, L879 (1988), [*Ann. Phys.*]{} [**190**]{}, 107 (1989); J.F. Cariñena, J.M. Gracia Bondía and J.C. Varilly, [*J. Phys. A: Math. Gen.*]{} [**23**]{}, 901 (1990): M. Gadella, M.A. Martín, L.M. Nieto and M.A. del Olmo, [*J. Math. Phys.*]{} [**32**]{}, 1182 (1991); J.F. Plebański, M. Przanowski and J. Tosiek, [*Acta Phys. Pol.*]{} [**B27**]{} 1961 (1996).
W.I. Tatarskii, [*Usp. Fiz. Nauk*]{} [**139**]{}, 587 (1983).
M. Hillery, R.F. O’Connell, M.O. Scully and E.P. Wigner, [*Phys. Rep.*]{} [**106**]{}, 121 (1984).
S. Weinberg, [*The Quantum Theory of Fields*]{} Vol. I, (Cambridge University Press, Cambridge, 1995).
B. Hatfield, [*Quantum Field Theory of Point Particles and Strings*]{}, Addison-Wesley Publishing Company, 1992.
I. Galaviz, H. García-Compeán, M. Przanowski and F.J. Turrubiates, “Deformation Quantization of Spinor Electrodynamics”, to appear (2007).
C. Fronsdal and M. Kontsevich, “Quantization on Curves,” arXiv:math-ph/0507021.
E. J. Martinec and G. W. Moore, “Noncommutative solitons on orbifolds,” arXiv:hep-th/0101199.
A. Belhaj, J. J. Manjarin and P. Resco, “On non-commutative orbifolds of K3 surfaces,” J. Math. Phys. [**44**]{}, 2507 (2003) \[arXiv:hep-th/0207160\].
[^1]: The conjugate momentum $\widetilde{\pi}_{\psi}$ of the field $\psi$ is defined by [@hirshfeld; @imelda] $$\widetilde{\pi}_{\psi_{\alpha}} = {\partial{\cal L} \over
\partial(\partial_t \psi_{\alpha})} = -i \hbar \psi_{\alpha}^\dag,$$ where the sign $(-)$ appears as one consider $\overline{\psi}$ and $\psi$ as Grassmann variables. In field theory is usual to write down it as $\pi = i \hbar \psi^\dag$, see for instance, [@wein; @hatfield]. Therefore to be consistent with these references on quantum fields we choice $ \psi \mathrm{\ \ and\ \
} \pi_{\psi}= i \hbar \psi^\dag,$ as our fundamental variables, instead of $\psi$ and $\widetilde{\pi}_{\psi}$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'This work is developing, and we will include many additions in the near future. Our purpose here is to highlight that there is plenty of space for a topological development of the Fermat Real Line.'
author:
- 'Kyriakos Papadopoulos, [email protected]\'
title: '[**A Few Questions on the Topologization of the Ring of Fermat Reals.**]{}'
---
v Ł ł Ø I̋ H Ł Ø
[**Keywords:**]{} Ring of Fermat Reals, GO-space, LOTS, Nest, Interlocking
[**2010 AMS Subject Classification.**]{} 54C35
\[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\] \[section\]
Introduction
============
The idea of the ring of Fermat Reals ${{{}^\bullet}\mathbb{R}}$ has come as a possible alternative to Synthetic Differential Geometry (see e.g. \[11, 12, 13\]) and its main aim is the development of a new foundation of smooth differential geometry for finite and infinite-dimensional spaces. In addition, ${{{}^\bullet}\mathbb{R}}$ could play a role of a potential alternative in some certain problems in the field ${}^\star \mathbb{R}$ in Nonstandard Analysis (NSA), because the applications of NSA in differential geometry are very few. One of the “weak” points of ${{{}^\bullet}\mathbb{R}}$ at the moment is the lack of a natural topology, carrying the strong topological properties of the line.
P. Giordano and M. Kunzinger have recently done brave steps towards the topologization of the ring ${}^\bullet \mathbb{R}$ of Fermat Reals. In particular, they have constructed two topologies; the Fermat topology and the omega topology (see [@ref3]). The Fermat topology is generated by a complete pseudo-metric and is linked to the differentiation of non-standard smooth functions. The omega topology is generated by a complete metric and is linked to the differentiation of smooth functions on infinitesimals. Although both topologies are very useful in developing infinitesimal instruments for smooth differential geometry, none of these two topologies aims to characterize the Fermat real line from an order-theoretic perspective. In fact, neither makes the space $T_1$, while an appropriate order-topology would equip the Fermat Real Line with the structure of a monotonically normal space, at least. The possibility to define a linear order relation on ${{{}^\bullet}\mathbb{R}}$, so that it can be viewed as a LOTS (linearly ordered topological space) can be considered important, because ${{{}^\bullet}\mathbb{R}}$ is an alternative mathematical model of the real line, having some features with respect to applications in smooth differential geometry and mathematical physics. It is therefore natural to ask whether for ${{{}^\bullet}\mathbb{R}}$ peculiar characteristics of $\mathbb{R}$ hold or not.
In this paper we will focus in the order relation which is introduced in [@ref4] (which is linear, but it generates the discrete topology on the space and also if considered minus the diagonal, i.e. as a strict order, the topology when restricted to the set of proper reals is again the discrete topology) and we will add properties to it, so that it will both extend the natural order of the real line and it will also give a stronger topology than the Fermat topology and the omega topology. We aim to do this using interlocking nests.
As we shall see in Definition \[definition-equivalens relation Fermat Real\], the idea of the formation of ${{{}^\bullet}\mathbb{R}}$ starts with an equivalence relation in the little-oh polynomials, where ${{{}^\bullet}\mathbb{R}}$ is the quotient space under this relation. This treatment permits us to view these little-oh polynomials as numbers.
Preliminaries.
==============
LOTS and GO-spaces via Nests.
-----------------------------
The notions nest and order are closely related. J. van Dalen and E. Wattel gave a complete characterization of LOTS (linearly ordered topological spaces) and of GO-spaces (generalized ordered spaces, i.e. subspaces of LOTS) using properties of nests (see [@ref1]). In this paper we will use tools from [@ref2], where the authors improved the techniques of van Dalen and Wattel in order to characterize ordinals in topological terms and from [@ref5], where the author studies further properties of order relations via nests.
\[definition-T0-T1\] Let $X$ be a set.
1. A collection $\mathcal{L}\subset \mathcal{P}(X)$ of subsets of $X$ $T_0$-[*separates*]{} $X$, if and only if for every $x,y \in X$, such that $x \neq y$, there exists $L \in
\mathcal{L}$, such that $x \in L$ and $y \notin L$ or $y \in L$ and $x \notin L$.
2. A collection $\mathcal{L} \subset \mathcal{P}(X)$ of subsets of $X$ $T_1$-[*separates*]{} $X$, if and only if for every $x,y \in X$, such that $x \neq y$, there exist $L,L' \in
\mathcal{L}$, such that $x \in L$ and $y \notin L$ and also $y \in
L'$ and $x \notin L'$.
Let $X$ be a set and let $\mathcal{L}$ be a family of subsets of $X$. $\mathcal{L}$ is a [*nest*]{} on $X$ if, for every $M, N \in \mathcal{L}$, either $M \subset N$ or $N \subset M$.
\[definition-order-reflexive\] Let $X$ be a set and $x, y \in X$. We declare $x {\triangleleft_{\mathcal{L}}}y$, if and only if there exists $L \in {\mathcal{L}}$, such that $x \in L$ and $y \notin L$.
It follows that $x {\trianglelefteq_{\mathcal{L}}}y$, if and only if either $x {\triangleleft_{\mathcal{L}}}y$ or $x = y$. One can easily see that ${\trianglelefteq_{\mathcal{L}}}$ is a linear order, if ${\mathcal{L}}$ is a $T_0$-separating nest.
\[theorem - preliminary to Lemma for Theorem Dalen-Wattel\] Let $X$ be a set and suppose that ${\mathcal{L}}$ and ${\mathcal{R}}$ are two nests on $X$. Then, ${\mathcal{L}}\cup {\mathcal{R}}$ is $T_1$-separating, if and only if ${\mathcal{L}}$ and ${\mathcal{R}}$ are both $T_0$-separating and ${\triangleleft_{\mathcal{L}}}= {\triangleright_{\mathcal{R}}}$.
\[definiton-interlocking\] Let $X$ be a set and let ${\mathcal{L}}\subset {\mathcal{P}}(X)$. We say that ${\mathcal{L}}$ is interlocking if and only if, for each $L \in {\mathcal{L}}$, $L = \bigcap\{N \in {\mathcal{L}}: L \subsetneq N\}$ implies $L = \bigcup \{N \in {\mathcal{L}}: N \subsetneq L\}$.
\[theorem-interlocking\] Let $X$ be a set and let ${\mathcal{L}}$ be a $T_0$-separating nest on $X$. The following are equivalent:
1. ${\mathcal{L}}$ is interlocking;
2. for each $L \in {\mathcal{L}}$, if $L$ has a ${\triangleleft_{\mathcal{L}}}$-maximal element, then $X-L$ has a ${\triangleleft_{\mathcal{L}}}$-minimal element;
\[theorem - them Dalen-Wattel\] Let $(X,\mathcal{T})$ be a topological space. Then:
1. If $\mathcal{L}$ and $\mathcal{R}$ are two nests of open sets, whose union is $T_1$-separating, then every $\triangleleft_{\mathcal{L}}$-order open set is open, in $X$.
2. $X$ is a GO-space, if and only if there are two nests $\mathcal{L}$ and $\mathcal{R}$ of open sets, whose union is $T_1$-separating and forms a subbasis for ${\mathcal{T}}$.
3. $X$ is a LOTS, if and only if there are two interlocking nests ${\mathcal{L}}$ and ${\mathcal{R}}$ of open sets, whose union is $T_1$-separating and forms a subbasis for ${\mathcal{T}}$.
The Ring ${}^\bullet \mathbb{R}$ of Fermat Reals.
-------------------------------------------------
The material in this subsection can be found in [@ref7], [@ref6] and also in [@ref4].
A little-oh polynomial $x_t$ (or $x(t)$) is an ordinary set-theoretical function, defined as follows:
1. $x : \mathbb{R}_{\ge 0} \to \mathbb{R}$ and
2. $x_t = r + \sum_{i=1}^k \alpha_i t^{a_i} + o(t)$, as $t \to 0^+$, for suitable $k \in \mathbb{N}$, $r,\alpha_1,\cdots,\alpha_k \in \mathbb{R}$ and $a_1,\cdots,a_k \in \mathbb{R}_{\ge 0}$.
The set of all little-oh polynomials is denoted by the symbol $\mathbb{R}_o[t]$. So, $x \in \mathbb{R}_o(t)$, if and only if $x$ is a polynomial function with real coefficients, of a real variable $t \ge 0$, with generic positive powers of $t$ and up to a little-oh function $o(t)$, as $t \to 0^+$.
\[definition-equivalens relation Fermat Real\] Let $x,y \in \mathbb{R}_o[t]$. We declare $x\sim y$ (and we say $x=y$ in ${}^\bullet \mathbb{R}$), if and only if $x(t) = y(t) + o(t)$, as $t \to 0^+$.
The relation $\sim$ in Definition \[definition-equivalens relation Fermat Real\] is an equivalence relation and ${}^\bullet \mathbb{R} :=\mathbb{R}_o[t]/\sim$.
A first attempt to define an order in ${}^\bullet \mathbb{R}$ has come from Giordano.
\[definition-order Giordano\] Let $x,y \in {}^\bullet \mathbb{R}$. We declare $x \le y$, if and only if there exists $z \in {}^\bullet \mathbb{R}$, such that $z = 0$ in ${}^\bullet \mathbb{R}$ (i.e. $\lim_{t \to 0^+} z_t/t = 0$) and for every $t \ge 0$ sufficiently small, $x_t \le y_t + z_t$.
For simplicity, one does not use equivalence relation but works with an equivalent language of representatives. If one chooses to use the notations of [@ref4], one has to note that Definition \[definition-order Giordano\] does not depend on representatives.
As the author describes in [@ref4], the order relation in NSA admits all formal properties among all the theories of (actual) infinitesimals, but there is no good dialectic of these properties with their informal interpretation. In particular, the order in ${}^\star \mathbb{R}$ inherits by transfer all the first order properties but, on the other hand, in the quotient field ${}^\star \mathbb{R}$ it is difficult to interpret these properties of the order relation as intuitive properties of the corresponding representatives. For example, a geometrical interpretation like that of ${{{}^\bullet}\mathbb{R}}$ seems not possible for ${}^ \star \mathbb{R}$. Definition \[definition-order Giordano\] provides a clear geometrical representation of the ring ${{{}^\bullet}\mathbb{R}}$ (see, for instance, section 4.4 of [@ref4]).
The Fermat Topology and the omega-topology on ${{{}^\bullet}\mathbb{R}}$.
-------------------------------------------------------------------------
A subset $A \subset {{{}^\bullet}\mathbb{R}}^n$ is open in the Fermat topology, if it can be written as $A = \bigcup \{{}^\bullet U \subset A: U \textrm{ is open in the natural topology in }\mathbb{R}^n\}$. Giordano and Kunzinger describe this topology as the best possible one for sets having a “sufficient amount of standard points”, for example ${}^\bullet U$. They add that this connection between the Fermat topology and standard reals can be glimpsed by saying that the monad $\mu(r) := \{x \in {{{}^\bullet}\mathbb{R}}: \textrm{ standard part of }x = r\}$ of a real $r \in \mathbb{R}$ is the set of all points which are limits of sequences with respect to the Fermat topology. However it is obvious that in sets of infinitesimals there is a need for constructing a (pseudo-)metric generating a finer topology that the authors call the omega-topology (see [@ref3]). Since neither the Fermat nor the omega-topology are Hausdorff when restricted to ${{{}^\bullet}\mathbb{R}}$ and since each of them describes sets having a “sufficient amount” of standard points or infinitesimals, respectively, there is a need for defining a natural topology on ${{{}^\bullet}\mathbb{R}}$ describing sufficiently all Fermat reals and carrying the best possible properties.
Interlocking Nests on ${}^\bullet \mathbb{R}$.
==============================================
A first disadvantage of the construction in Definition \[definition-order Giordano\] is that the order $\le$ in ${{{}^\bullet}\mathbb{R}}$ does not generate interlocking nests, missing points from the Fermat real line. In particular, the nest ${\mathcal{L}}$ consisting of sets $L=\{k \in {{{}^\bullet}\mathbb{R}}: k \le l\}$, for some $l \in {{{}^\bullet}\mathbb{R}}$, has as maximal element the fermat real $l$, but the complement of $L$, i.e. $L^c = \{k \in {{{}^\bullet}\mathbb{R}}: k> l\}$, for some $l \in {{{}^\bullet}\mathbb{R}}$, does not have a minimal element. Thus, we first remark that the order of Definition \[definition-order Giordano\] makes ${}^\bullet \mathbb{R}$ a GO-space, a subspace of a particular LOTS. So we will now have to construct an appropriate order in ${}^\bullet \mathbb{R}$ which makes it LOTS, by completing the missing minimal elements from complements of sets with maximal elements. Even the fact that $\le$ is linear, it generates the discrete topology on ${{{}^\bullet}\mathbb{R}}$ and, if considered as a strict order, the restriction of its topology in $\mathbb{R}$ will be again the discrete topology.
Order Relations and an Order Topologies on ${{{}^\bullet}\mathbb{R}}$.
----------------------------------------------------------------------
The pair $({}^\bullet \mathbb{R},<_F)$, where $<_F$ is defined as follows:
$$x<_Fy \Leftrightarrow \begin{cases}
\exists\,\{k \in {{{}^\bullet}\mathbb{R}}: k \le l\}, \textrm{ some } l \in {{{}^\bullet}\mathbb{R}}, \textrm{ such that } x \in \{k \in {{{}^\bullet}\mathbb{R}}: k \le l\} \not\ni y, l \in {{{}^\bullet}\mathbb{R}}\\
or\\
x = \max\{k \in {{{}^\bullet}\mathbb{R}}: k \le l\}, \textrm{ some } l \in {{{}^\bullet}\mathbb{R}}\textrm{ and } \exists h \in {{{}^\bullet}\mathbb{R}}: h>0,\,y = x + h\\
or\\
y = \min \{k \in {{{}^\bullet}\mathbb{R}}: l \le k\}, \textrm{ some } l \in {{{}^\bullet}\mathbb{R}}\textrm{ and } \exists h \in {{{}^\bullet}\mathbb{R}}: h>0,\,x = y - h
\end{cases}$$ where $x,y$ are distinct Fermat reals, is a linearly ordered set.
The order of Definition \[definition-order Giordano\] gives two nests, namely the nest ${\mathcal{L}}$, which consists of all sets $L = \{k \in {{{}^\bullet}\mathbb{R}}: k \le l\}$, some $l \in {{{}^\bullet}\mathbb{R}}$ and the nest ${\mathcal{R}}$, which consists of all sets $R = \{k \in {{{}^\bullet}\mathbb{R}}: l \le k\}$, some $l \in {{{}^\bullet}\mathbb{R}}$. In addition, Theorem \[theorem - preliminary to Lemma for Theorem Dalen-Wattel\] implies that ${\trianglelefteq_{\mathcal{L}}}= \trianglerighteq_{{\mathcal{R}}} = \le$.
We remark that, for any $L \in {\mathcal{L}}$ (respectively for any $R \in {\mathcal{R}}$), $L$ (resp. $R$) has a ${\trianglelefteq_{\mathcal{L}}}$-maximal element (resp. ${\trianglelefteq_{\mathcal{R}}}$-maximal element for $R$), such that $X-L$ has no ${\trianglelefteq_{\mathcal{L}}}$-minimal element (resp. $X-R$ has no ${\trianglelefteq_{\mathcal{R}}}$-minimal element). So, according to Theorem \[theorem-interlocking\], neither ${\mathcal{L}}$ nor ${\mathcal{R}}$ are interlocking. Now, for all $L=\{k \in {{{}^\bullet}\mathbb{R}}: k \le l\} \in {\mathcal{L}}$, some $l \in {{{}^\bullet}\mathbb{R}}$, let $x_L$ denote the ${\trianglelefteq_{\mathcal{L}}}$-maximal element of $L$ and for all $R=\{k \in {{{}^\bullet}\mathbb{R}}: l \le k\} \in {\mathcal{R}}$, some $l \in {{{}^\bullet}\mathbb{R}}$ let $y_R$ denote the ${\trianglelefteq_{\mathcal{L}}}$-minimal element of $R$.
Furthermore, for each $L \in {\mathcal{L}}$ choose $x_L^+ \in {{{}^\bullet}\mathbb{R}}$ and for each $R \in {\mathcal{R}}$ choose $y_R^- \in {{{}^\bullet}\mathbb{R}}$, where $x_L^+$ and $y_R^-$ are distinct points in ${}^\bullet \mathbb{R}$, and define a map $p : {}^\bullet \mathbb{R} \to {}^\bullet \mathbb{R}-(\{x_L^+ : L \in {\mathcal{L}}\} \cup \{y_R^- : R \in {\mathcal{R}}\})$, as follows: $$p(x)=\begin{cases}
x,&\text{if }x\in {}^\bullet \mathbb{R}-(\{x_L^+ : L \in {\mathcal{L}}\} \cup \{y_R^- : R \in {\mathcal{R}}\})\\
x_L,&\text{if }x=x_L^+\\
y_R,&\text{if }x=y_R^-
\end{cases}$$
Now, define an order $<_F$ on ${}^\bullet \mathbb{R}$, so that: $$x<_Fy \Leftrightarrow \begin{cases}
p(x) {\triangleleft_{\mathcal{L}}}p(y) \\
or\\
x=x_L \textrm{ and } y = x_L^+\\
or\\
x=y_R^- \textrm{ and } y = y_R
\end{cases}$$
Obviously, $<_F$ is a linear order and the restriction of $<_F$ to ${}^\bullet \mathbb{R}-(\{x_L^+ : L \in {\mathcal{L}}\} \cup \{y_R^- : R \in {\mathcal{R}}\})$ equals ${\trianglelefteq_{\mathcal{L}}}$, the order in Definition \[definition-order Giordano\]. In addition, we can set $x_L^+ = x_L + h$, where $h$ is not zero in ${{{}^\bullet}\mathbb{R}}$ and $h>0$, that is, $\lim_{t \to 0^+} h_t/t \neq 0$ and, respectively, we set $x_R^- = x_R-h$, and this completes the proof.
${{{}^\bullet}\mathbb{R}}$ equipped with the order topology from $<_F$ is a LOTS.
We will now show that the topology ${\mathcal{T}}$ on ${}^\bullet \mathbb{R}-(\{x_L^+ : L \in {\mathcal{L}}\} \cup \{y_R^- : R \in {\mathcal{R}}\})$ coincides with the subspace topology on ${}^\bullet \mathbb{R}-(\{x_L^+ : L \in {\mathcal{L}}\} \cup \{y_R^- : R \in {\mathcal{R}}\})$ that is inherited from the $<_F$-order topology on ${}^\bullet \mathbb{R}$.
But, since ${\mathcal{L}}\cup {\mathcal{R}}$ forms a subbasis for ${\mathcal{T}}$, that consists of two nests, every set in ${\mathcal{T}}$ can be written as a union of sets of the form $L \cap R$, where $L \in {\mathcal{L}}$ and $R \in {\mathcal{R}}$. It suffices therefore to show that every $L \in {\mathcal{L}}$ and $R \in {\mathcal{R}}$ can be written as the intersection of an order-open set with ${}^\bullet \mathbb{R}-(\{x_L^+ : L \in {\mathcal{L}}\} \cup \{y_R^- : R \in {\mathcal{R}}\})$. But this is always true, since if $L \in {\mathcal{L}}$, with ${\trianglelefteq_{\mathcal{L}}}$-maximal element $x_L$, then $L = {}^\bullet \mathbb{R}-(\{x_L^+ : L \in {\mathcal{L}}\} \cup \{y_R^- : R \in {\mathcal{R}}\}) \cap \{ x \in {{{}^\bullet}\mathbb{R}}: x<_F x_L^+\}$.
The argument for $R \in {\mathcal{R}}$ is similar, and this completes the proof.
Remarks.
--------
1. The order topology ${\mathcal{T}}_{<_F}$ equals the topology ${\mathcal{T}}_{{\mathcal{L}}_{<_F} \cup {\mathcal{R}}_{<_F}}$, where ${\mathcal{L}}_{<_F}=\{k \in {{{}^\bullet}\mathbb{R}}: k <_F l\}$, some $l \in {{{}^\bullet}\mathbb{R}}$ and ${\mathcal{R}}_{<_F}=\{k \in {{{}^\bullet}\mathbb{R}}: l <_F k\}$, some $l \in {{{}^\bullet}\mathbb{R}}$. This is because ${\mathcal{L}}_{<_F} \cup {\mathcal{R}}_{<_F}$ $T_1$-separates ${{{}^\bullet}\mathbb{R}}$ and both ${\mathcal{L}}_{<_F}$ and ${\mathcal{R}}_{<_F}$ are interlocking nests. So, unlike the GO-space topology ${\mathcal{T}}_\le$ on ${{{}^\bullet}\mathbb{R}}$, where ${\mathcal{T}}_\le \subset {\mathcal{T}}_{{\mathcal{L}}\cup {\mathcal{R}}}$, $<_F$ provides a natural extension of the natural linear order of the set of real numbers to the Fermat real line and the order topology from $<_F$ can be completely described via the nests ${\mathcal{L}}_{<_F}$ and ${\mathcal{R}}_{<_F}$.
2. Viewing the Fermat real line as a LOTS and working with nests ${\mathcal{L}}_{<_F}$ and ${\mathcal{R}}_{<_F}$, one can now define the product topology for ${{{}^\bullet}\mathbb{R}}^n$, some positive integer $n$, or even more generaly for $\Pi_{i \in I} {{{}^\bullet}\mathbb{R}}_i$, some arbitrary indexing set $I$, in the usual way via the subbasis $\pi_{j_0}^{-1}(A_{j_0}) = \Pi_{i \in I}\{{{{}^\bullet}\mathbb{R}}_i : i \neq j_0\} \times A_{j_0}$, where $A_{j_0}$ is an open subset in the coordinate space ${{{}^\bullet}\mathbb{R}}_{j_0}$ in the order topology ${\mathcal{T}}_{<_F}$ and $\pi_i : \Pi_{i \in I}{{{}^\bullet}\mathbb{R}}_i \to {{{}^\bullet}\mathbb{R}}_i$ the projection.
3. In this way one can define continuity for any function $f$ from a topological space $Y$ into the product space $\Pi_{i \in I} {{{}^\bullet}\mathbb{R}}_i$ via the continuity of $\pi_i \circ f :Y \to {{{}^\bullet}\mathbb{R}}_i$.
4. The neight of ${{{}^\bullet}\mathbb{R}}$ is $2$ and the neight of ${{{}^\bullet}\mathbb{R}}^n = n+1$ (see [@ref10]). Using the product topology, as stated in Remark (2), we use four nests in order to define -for example- the topology in ${{{}^\bullet}\mathbb{R}}^2$, but since the neight of ${{{}^\bullet}\mathbb{R}}^2$ is $3$, one can define a topology using three nests exclusively.
Questions.
----------
1. As a LOTS, $({{{}^\bullet}\mathbb{R}},<_F)$ has rich topological properties. It is, for example, a monotone normal space. It would be interesting though to have an extensive study on the metrizability of this space. It is known that in a GO-space the terms metrizable, developable, semistratifiable, etc. are equivalent (see [@ref8] and [@ref9]). The real line (i.e. the set of all standard reals, from the point of view of ${{{}^\bullet}\mathbb{R}}$) is a developable LOTS and this is equivalent to say that it is also a metrizable LOTS. Is $({{{}^\bullet}\mathbb{R}},{\mathcal{T}}_{<_F})$ developable?
2. Which of the subspaces of $({{{}^\bullet}\mathbb{R}},{\mathcal{T}}_{<_F})$ are developable?
Since any sequence $x_1,x_2,\cdots$ of points in $\Pi_{i \in I} {{{}^\bullet}\mathbb{R}}_i$ will converge to a point $x \in \Pi_{i \in I}{{{}^\bullet}\mathbb{R}}_i$, iff for every projection $\pi_i: \Pi_{i \in I} {{{}^\bullet}\mathbb{R}}_i \to {{{}^\bullet}\mathbb{R}}_i$ the sequence $\pi_i(x_1), \pi_i(x_2),\cdots$ converges to $\pi_i(x)$ in the coordinate space ${{{}^\bullet}\mathbb{R}}_i$, any answer to the above questions will be foundamental towards our understanding of convergence in the ring of Fermat Reals.
[99]{} J. van Dalen and E. Wattel, *A Topological Characterization of Ordered Spaces*, General Topology and Appl., 3:347-354 (1973).
C. Good and K. Papadopoulos, *A Topological Characterization of Ordinals: van Dalen and Wattel revisited*, Topology Appl., 159:1565-1572 (2012).
Paolo Giordano and Michael Kunzinger, *Topological and Algebraic Structures on the Ring of Fermat Reals*, Israel Journal of Mathematics, 193 (2013), 459-505.
Paolo Giordano, *Fermat Reals: Nilpotent Infinitesimals and Infinite Dimensional Spaces*, Prebook (http://arxiv.org/abs/0907.1872).
Kyriakos Papadopoulos, *On the Orderability Problem and the Interval Topology*, chapter to appear in the Volume “Topics in Mathematical Analysis and Applications”, in the Optimization and Its Applications Springer Series, T. Rassias and L. Toth Eds, Springer Verlag, 2014.
Giordano P., *Fermat reals: infinitesimals without Logic*, Miskolc Mathematical Notes, Vol. 14 (2013), No. 3, pp. 65–80.
Giordano P., *The ring of fermat reals*, Advances in Mathematics 225 (2010), pp. 2050-2075.
K.P. Hart, J.-I. Nagata, J.E. Vaughan, *Encyclopedia of General Topology*, Elsevier Science and Technology Books (2014).
M.J. Faber, *Metrizability in Generalized Ordered Spaces*, Matematisch Centrum, Amsterdam 1974.
W.R. Brian *Neight: The Nested Weight of a Topological Space*, to appear in Topology Proceedings.
A. Kock, *Synthetic Differential Geometry*, volume 51 of London Math, Soc. Lect. Note Series, Cambridge Univ. Press, 1981.
R. Lavendhomme, *Basic Concepts of Synthetic Differential Geometry*, Kluwer Academic Publishers, Dordrecht, 1996.
I. Moerdijk, G.E. Reyes, *Models for Smooth Infinitesimal Analysis*, Springer, Berlin, 1991.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Spatial heterogeneity in the elastic properties of soft random solids is examined via vulcanization theory. The spatial heterogeneity in the *structure* of soft random solids is a result of the fluctuations locked-in at their synthesis, which also brings heterogeneity in their *elastic properties*. Vulcanization theory studies semi-microscopic models of random-solid-forming systems, and applies replica field theory to deal with their quenched disorder and thermal fluctuations. The elastic deformations of soft random solids are argued to be described by the Goldstone sector of fluctuations contained in vulcanization theory, associated with a subtle form of spontaneous symmetry breaking that is associated with the liquid-to-random-solid transition. The resulting free energy of this Goldstone sector can be reinterpreted as arising from a phenomenological description of an elastic medium with quenched disorder. Through this comparison, we arrive at the statistics of the quenched disorder of the elasticity of soft random solids, in terms of residual stress and Lamé-coefficient fields. In particular, there are large residual stresses in the equilibrium reference state, and the disorder correlators involving the residual stress are found to be long-ranged and governed by a universal parameter that also gives the mean shear modulus.'
author:
- Xiaoming Mao
- 'Paul M. Goldbart'
- Xiangjun Xing
- Annette Zippelius
title: Soft random solids and their heterogeneous elasticity
---
Introduction {#SEC:Intr}
============
Random solids, such as chemical gels, rubber, glasses and amorphous silica, are characterized by their [*structural*]{} heterogeneity, which results from the randomness locked-in at the time they are synthesized. The mean positions of the constituent particles exhibit no long-range order, and every particle inhabits a unique spatial environment. The [*elasticity*]{} of random solids also inherits heterogeneity from this locked-in randomness. For example, the Lamé coefficients and the residual stress vary from point to point throughout the elastic medium. The central goal of this paper is to develop a statistical characterization of random elastic media, via the mean values of the Lamé coefficients and the residual stress as well as the two-point spatial correlations amongst the fluctuations of these quantities, which we shall name as the disorder correlator. These mean values and correlations are to be thought of as averages taken over realizations of the sample fabrication for a given set of fabrication parameters. We expect these characteristic quantities to coincide with the volume-averages of their single-sample counterparts.
Our focus will be on [*soft*]{} random solids. These are network media that include chemical gels [@Addad1996], which are formed by the permanent random chemical bonding of small molecules, as well as rubber [@Treloar1975], which is formed via the introduction of permanent random chemical cross-links between nearby monomers in melts or solutions of flexible long-chain polymers. Soft random solids are characterized by their *entropic elasticity*. These are media in which the shear modulus originates in the strong thermal fluctuations of the configurations of the constituent particles and is much smaller than the bulk modulus, which is energetic in nature and originates in the excluded-volume interactions between the particles. The concept of entropic elasticity forms the basis of the classical theory of rubber elasticity, developed long ago by Kuhn, Flory, Wall, Treloar and others (see Ref. [@Treloar1975]).
As we discuss soft random solids we shall take chemical gels as our prototype media. When the density of the introduced links exceeds the percolation threshold, an infinite cluster of linked molecules forms, spanning the system, and the network acquires a thermodynamic rigidity with respect to shear deformations [^1]. This event is often called the gelation transition or the vulcanization transition [@Goldbart1996].
The [*geometrical*]{} or [*architectural*]{} aspects of the gelation/vulcanization transition can be well captured by the theory of percolation [@Stauffer1994]. However, to study the [*elasticity*]{} that emerges at the gelation/vulcanization transition, and especially its heterogeneity, one needs a theory that incorporates not only the geometrical aspects, but also the equilibrium thermal fluctuations of the particle positions and the strong, qualitative changes that they undergo at the gelation/vulcanization transition. In the setting of rubber elasticity, although the classical theory is successful in explaining the entropic nature of the shear rigidity of rubber, it is essentially based on a single-chain picture and, as such, is incapable of describing the consequences of the long scale random structure of rubbery media, e.g., the random spatial variations in their local elastic parameters and the resulting nonaffinity of their local strain-response to macroscopic applied stresses.
The general problem of heterogeneous elasticity and nonaffine deformations has been studied in the setting of flexible polymer networks [@Rubinstein1997; @Glatting1995; @Holzl1997; @Svaneborg2005], and also semi-flexible polymer networks [@Head2003; @Heussinger2007], glasses [@Wittmer2002], and granular materials [@Utter2008]. Particularly noteworthy is the recent investigation by DiDonna and Lubensky of the general relationship between the spatial correlations of the nonaffine deformations and those of the underlying quenched random elastic parameters [@DiDonna2005].
The mission of the present work is to develop a statistical characterization of the heterogeneous elasticity of soft random solids by starting from a semi-microscopic model and applying a body of techniques that we shall call vulcanization theory to it. In particular, we aim to obtain the mean values and disorder correlators of elastic parameters, such as the Lamé coefficients and the residual stress, in terms of the parameters of the semi-microscopic model, such as the density of cross-links, the excluded-volume interactions, etc. One of our key findings is that the disorder correlator of the residual stress is long ranged, as are all cross-disorder correlators between the residual stress and the Lamé coefficients. We also find that these disorder correlators are controlled by a universal scale parameter—independent of the microscopic details—that, moreover, controls the scale of the mean shear modulus. In addition, we characterize the nonaffininity of the deformations in terms of these parameters.
The strategy we adopt for accomplishing our goals involves a handshaking between two different analytical schemes. [^2] The first scheme follows a well-trodden path. We begin with a semi-microscopic model, the Randomly Linked Particle Model (RLPM) [@Mao2007; @Ulrich2006; @Broderix2002; @Mao2005], involving particle coordinates and quenched random interactions between them that represent the randomness that is locked-in at the instant of cross-linking. In order to account for this quenched randomness, as well as the thermal fluctuations in particle positions which are the origin of the entropic elasticity, we adopt the framework of vulcanization theory [@Goldbart1996]. This framework includes the use of the replica method to eliminate the quenched randomness, followed by a Hubbard-Stratonovich transformation to construct a field-theoretic representation in terms of an order parameter field—in this case, the random solidification order parameter. We analyze this representation at the stationary-point level of approximation, and then focus on the gapless excitations around the stationary point—the Goldstone fluctuations—observing that these excitations can be parameterized in terms of a set of replicated shear deformation fields [@Mao2007; @Ulrich2006; @Goldbart2004]. The second prong of our approach is less conventional. It begins with our introduction of a phenomenological model free energy of an elastic continuum, characterized by a nonlocal kernel of quenched random attractions between mass-points. To obtain statistical information about this quenched random kernel, we use the replica method to eliminate the randomness, and obtain a pure model of replicas of the deformation field with couplings controlled by the disorder-moments of the kernel; these disorder-moments are then treated as unknown quantities to be determined. This pure model has precisely the same structure as the Goldstone theory mentioned in the previous paragraph has. Thus, by comparing the two models we can learn the moments of the quenched random kernel from the (already-computed) coupling functions of the Goldstone model. Then we analyze a particular realization of the phenomenological model having a fixed value of the quenched randomness. We observe that the natural reference state (i.e., the state of vanishing displacement field) of this model is not in fact an equilibrium state for any given realization of disorder, due to the random attractive interactions embodied by the kernel. We analyze how these attractions compete with the near-incompressibility of the medium to determine the displacement to the new, equilibrium configuration, which we shall term the relaxed state.(This process can be understood in the setting of a hypothetical, instant process of preparing a sample of rubber: the cross-links introduce attractions and random stresses, and the system then undergoes relaxation, including global contraction and local deformation.) We then explore shear deformations around this relaxed state, pass to the local limit, and arrive at the standard form of continuum elasticity theory, expressed in terms of the strain around the relaxed state, but with coefficients that are explicit functions of the quenched random kernel. Thus, using the information about the statistics of the quenched random kernel obtained via the comparison with the RLPM, we are able to infer statistical information about the elastic properties of the random elastic medium in the relaxed state, which is of experimental relevance.
Why is it legitimate to identify the Goldstone theory arising from the microscopic model with the replica theory of the phenomenological model? The reason is that, within the schemes that we have chosen to analyze them, both models describe shear deformations not of the equilibrium state of the system but, rather, of the system immediately after cross-linking has been done but before any cross-linking-induced relaxation has been allowed to occur. [[ The equivalence between these two schemes is not based solely on the equality of free energies of the two models; it is also based on the identity of the physical meaning of the (replicated) deformation fields in the two theories, and thus the way that these fields couple to externally applied forces. Both schemes are descriptions of the elasticity of soft random solids displaying heterogeneous elastic properties, one from a semi-microscopic viewpoint, the other invoking phenomenological parameters. Thus, the equivalence of the two schemes provides the values of the phenomenological parameters as functions of the semi-microscopic parameters. ]{}]{}
Before concluding this introduction, let us emphasize that this work is a first attempt to [*derive*]{} the elastic heterogeneities of vulcanized matters. To date, the prevalent strategy in studies of disordered systems is to assume a particular structure for the quenched disorder on phenomenological grounds (such as Gaussian, short ranged, etc.) and explore the consequences. Assumptions about disorder structure are usually based on symmetry arguments and also the preference for simplicity, but otherwise lack theoretical substantiation. It is one of the main advantage of vulcanization theory that it can [*predict*]{} some generic properties of the disordered structure in vulcanized matter, as is shown in the present work, which can be used to support and sharpen the assumptions underlying more phenomenological theories.
The classical theory of rubber elasticity, which was shown to be derivable from the saddle-point approximation of vulcanization theory, is known to fail to describe rubber elasticity in the intermediate and large deformation regimes [@Treloar1975]. While a recent study [@Xing2007] shows that long wave-length thermal elastic fluctuations account qualitatively for this failure, on general grounds, we expect that elastic heterogeneities should play an equally important role. It would therefore be interesting to explore how the elastic heterogeneities discovered in the present work modify the macroscopic elasticity of rubbery materials. Such a program is left for a future work.
The outline of this paper is as follows. In Section \[SEC:RLPM\], we analyze the semi-microscopic RLPM using the tools of vulcanization theory. Specifically, we use the replica method [@Mezard1987] to study a model network consisting of randomly linked particles, which exhibits a continuous phase transition from the liquid state to the random solid state, paying particular attention to the Goldstone fluctuations of the random solid state, which, as we have mentioned above, are related to the elastic shear deformations of the random solid state. In Section \[SEC:Phen\], we propose a nonlocal phenomenological model of a random elastic medium, and subsequently derive it from the RLPM, by identify that this phenomenological model is the low energy theory (i.e., it captures the Goldstone fluctuations) of the RLPM in the random solid state. Through this correspondence we learn information of the statistics of the quenched random nonlocal kernel. In Section \[SEC:relaxation\], we study the relaxation of the phenomenological model to a stable state for any fixed randomness (i.e., any realization of disorder), due to random stresses and attractive interactions. We re-expand the free energy about this relaxed state to obtain the true elastic theory. This relaxed reference state is, however, still randomly stressed [@Alexander1998]; nevertheless, the stress in this state—the so-called *residual stress*—does satisfy the condition of mechanical equilibrium, viz., $\partial_i \Stre_{ij}(x)=0$. In its local limit, the proposed phenomenological model reproduces a version of Lagrangian continuum elasticity theory that features random Lamé coefficients and residual stresses. In this section, we also use the phenomenological model to explore the related issue of elastic heterogeneity, viz., the nonaffine way in which the medium responds to external stress. In Section \[SEC:Heterogeneity\], we arrive at predictions for the statistics of the quenched random elastic parameters that feature in the phenomenological model in the relaxed state, along with the statistics of nonaffine deformations. Thus we provide a first principles account of the heterogeneous elasticity of soft random solids. We conclude, in Section \[SEC:ConcDisc\], with a brief summary and discussion of our results.
Semi-microscopic approach: The randomly linked particle model {#SEC:RLPM}
=============================================================
Randomly linked particle model {#SEC:RLPMBasics}
------------------------------
The Randomly Linked Particle Model (RLPM) consists of $\PartNumb$ particles in a volume $\Volu$ in $\Dime$ dimensions. In order to study elasticity, including bulk deformations, $\Volu$ is allowed to fluctuate under a given pressure $\Pres$. The positions of the particles in this fluctuating volume are denoted by $\{\PartPosi_j\}_{j=1}^{\PartNumb}$. The particles in the RLPM interact via two types of interactions: a repulsive interaction $\VExc$ between all pairs of particles (either direct or mediated via a solvent); and an attractive interaction $\VLink$ between the pairs of particles that are chosen at random to be linked. We take the latter to be a soft link (as opposed to the usual hard constraint of vulcanization theory). Thus, the Hamiltonian can be written as $$\begin{aligned}
\label{EQ:HRLPM}
H_{\RealDiso} =
\sum_{1\le i<j\le N}^{\PartNumb}
\VExc (\PartPosi_i-\PartPosi_j)+
\sum_{e=1}^{\LinkNumb}
\VLink \big(\vert \PartPosi_{i_e}-\PartPosi_{j_e}\vert\big) .\end{aligned}$$ The label $e$, which runs from $1$ to the total number of links $M$, indexes the links in a given realization of the quenched disorder, and specifies them via the quenched random variables $M$ and $\{i_{e},j_{e}\}_{e=1}^{M}$.
We take $\VExc$ to be a strong, short-ranged repulsion, which serves to penalize density fluctuations and thus render the system nearly incompressible, as is appropriate for regular or polymeric liquids. As we shall describe in Section \[SEC:FTMF\], we address the interactions in Eq. (\[EQ:HRLPM\]) by eliminating the particle coordinates in favor of collective fields, which have the form of joint densities of the replicas of the particles. This continuum approach enables us to focus on the physics of random particle localization, particularly at lengthscales that are relevant for such localization, which (except when the density of links is extremely high) are long compared with the ranges of $\VExc$ and $\VLink$. With a focus on these longer lengthscales in mind, we see that it is adequate to replace the repulsive interaction $\VExc(\PartPosi)$ by the model Dirac delta-function excluded-volume interaction $\ExclVolu\,\delta(\PartPosi)$, characterized by the strength $\ExclVolu$ [@Gennes1979; @Deam1976; @Doi1986]. This procedure amounts to making a gradient expansion in real space (or, equivalently, a wave vector expansion in Fourier space) of $\VExc$ and retaining only the zeroth-order term; it gives for the strength $\ExclVolu$ the value ${\int}d\PartPosi\,\VExc(\PartPosi)$. Terms of higher order in the gradient expansion would have a non-negligible impact on the suppression of density fluctuations only at lengthscales comparable to or shorter than the range of $\VExc$, and fluctuation modes at such lengthscales are not the ones driven via random linking to the instability associated with random localization (and thus are not modes in need of stabilization via $\VExc$). We remark that the approximate interaction $\ExclVolu\,\delta(\PartPosi)$ is not, in practice, singular, and is instead regularized via a high wave vector cut-off.
At our coarse-grained level of description, the particles of the RLPM can be identified with polymers or small molecules, and the soft links can be identified with molecular chains that bind the molecules to one another. The potential for the soft links can be modeled as Gaussian chains $$\begin{aligned}
\label{EQ:VGC}
\VLink^{(\textrm{GC})}(\vert r \vert) = \frac{\BoltCons T \vert r \vert^2}{2\LinkScal^2} ,\end{aligned}$$ i.e., a harmonic attraction, or a zero rest-length spring, of lengthscale $\LinkScal$ between the two particles. In making this coarse-graining one is assuming that microscopic details (e.g., the precise locations of the cross-links on the polymers, the internal conformational degrees of freedom of the polymers, and the effects of entanglement) do not play significant roles for the long-wavelength physics. In part, these assumptions are justified by studying more detailed models, in which the conformational degrees of freedom of the polymers are retained [@Goldbart1996]. However, we should point out that the precise form of $\VLink$ is not important for long-wavelength physics and, hence, for the elastic properties that we are aiming to investigate \[cf. the discussion at the end of Sec. \[SEC:FTMF\]\].
From the discussion above, the RLPM is a convenient minimal model of soft random solids, inasmuch as it adequately captures the necessary long-wavelength physics. It can be regarded as either a model of a chemical gel, or as a caricature of vulcanized rubber or other soft random solid. The RLPM can be viewed as a simplified version of vulcanization theory [@Castillo1994; @Goldbart2004], with microscopic details, such as polymer chain conformations, being ignored. Nevertheless, it is able to reproduce the same universality class as vulcanization theory at the liquid-to-random-solid transition. For the study of elasticity, we shall consider length-scales on which the system is a well-defined solid (i.e., scales longer than the localization length, as we shall see later in this paper). Both of these scales are much larger than the characteristic linear dimension of an individual polymer. The RLPM is a model very much in the spirit of lattice percolation, except that it naturally allows for particle motion as well as particle connectivity, and is therefore suitable for the study of continuum elasticity and other issues associated with the (thermal or deformational) motion of the constituent entities.
Equation (\[EQ:HRLPM\]) is a Hamiltonian for a given realization of quenched disorder $\RealDiso \equiv \{ i_e, j_e\}_{e=1}^{\LinkNumb}$, which describes the particular random instance of the linking of the particles. These links are the quenched disorder of the system, which are specified at synthesis and do not change with thermal fluctuations. This is because there is a wide separation between the timescale for the linked-particle system to reach thermal equilibrium and the much longer timescale required for the links themselves to break. Therefore, we treat the links as permanent. Later, we shall apply the replica technique [@Mezard1987] to average over these permanent random links.
Replica statistical mechanics of the RLPM {#SEC:RLPMReplica}
-----------------------------------------
For a given volume and a given realization of disorder $\RealDiso$ we can write the partition function $Z_{\RealDiso}$ for the RLPM as $$\begin{aligned}
\label{EQ:partition}
Z_{\RealDiso} (\Volu)
&\equiv& \int _{\Volu} \prod_{i=1}^{\PartNumb} d \PartPosi_i
\exp\Big(-\frac{H_{\RealDiso}}{{\BoltCons}T}\Big) \nonumber\\
&=& \PartitionLiquid(V) \Bigg\ltha \prod _{e=1}^{\LinkNumb}
\LinkPote\big( \vert \PartPosi_{i_e} - \PartPosi_{j_e} \vert \big)
\Bigg\rtha_{1}^{\HamiExcl},\end{aligned}$$ where $\HamiExcl \equiv \frac{\ExclVolu }{2}\sum_{i,j=1}^{\PartNumb} \delta (\PartPosi_i-\PartPosi_j)$ is the excluded-volume interaction part of the Hamiltonian, and $\PartitionLiquid(V) \equiv \int _{\Volu} \prod_{i=1}^{\PartNumb} d \PartPosi_i \exp\big(-\HamiExcl/{\BoltCons}T\big)$ is the partition function of the liquid in the absence of any links. The issue of the Gibbs factorial factor, which is normally introduced to compensate for the overcounting of identical configuration, is a genuinely subtle one in the context of random solids (for a discussion, see Ref. [@Goldbart1996]). However, our focus will be on observables such as order parameter rather than on free energies, and thus the omission of the Gibbs factor is of no consequence. The factor $$\begin{aligned}
\label{EQ:DeltReal}
\LinkPote\big( \vert \PartPosi_{i_e} - \PartPosi_{j_e} \vert \big)
\equiv e^{ -\frac{\vert \PartPosi_{i_e} - \PartPosi_{j_e} \vert^2}
{2 \LinkScal^2}}\end{aligned}$$ is associated with the link-induced attractive interaction term in the Hamiltonian. The average $\ltha \cdots \rtha_{1}^{\HamiExcl}$, taken with respect to a Boltzmann weight involving the excluded-volume interaction Hamiltonian $\HamiExcl$, is defined as $$\begin{aligned}
\ltha \cdots \rtha_{1}^{\HamiExcl} \equiv
\frac{1}{\PartitionLiquid(V)}\int _{\Volu} \prod_{i=1}^{\PartNumb} d \PartPosi_i \,
e^{-\frac{\HamiExcl}{{\BoltCons}T}}\ldots\, .\end{aligned}$$ The corresponding Helmholtz free energy is then given by $$\begin{aligned}
\HelmFreeEner _{\RealDiso} (\Volu) \equiv -\BoltCons T \ln Z_{\RealDiso} (\Volu).\end{aligned}$$
To perform the average of the free energy over the quenched disorder, we shall need to choose a probability distribution that assigns a sensible statistical weight $\DEDist (\{ i_e, j_e\}_{e=1}^{\LinkNumb})$ to each possible realization of the total number $\LinkNumb$ and location $\{i_e,j_e\}_{e=1}^{\LinkNumb}$ of the links. Following an elegant strategy due to Deam and Edwards [@Deam1976], we assume a version of the normalized link distribution as follows: $$\begin{aligned}
\DEDist(\RealDiso)
= \frac{
\Big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\Big)^{\LinkNumb}
Z_{\RealDiso}(\Volu_0)}
{\LinkNumb ! Z_1},\end{aligned}$$ where $\LinkDens$ is a parameter that controls the mean total number of links. We assume that the preparation state (i.e., the state in which the links are going to be introduced) is in a given volume $\Volu_0$. The $Z_{\RealDiso}(\Volu_0)$ factor is actually the partition function, as given in Eq. (\[EQ:partition\]), and can be regarded as probing the equilibrium correlations of the underlying unlinked liquid. The factor $\LinkPote_0=\big(2\pi \LinkScal^2 \big)^{d/2}$ is actually the $p=0$ value of the Fourier transform of the $\LinkPote$ function defined in Eq. (\[EQ:DeltReal\]), and we shall see later that these factors ensure that the (mean-field) critical point occurs at $\LinkDens_C =1$. The normalization factor $Z_1$ is defined to be $\sum_{\RealDiso} \big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\big)^{\LinkNumb} Z_{\RealDiso}(\Volu_0)/ \LinkNumb !$. The calculation for $Z_1$ is straightforward, and is given in Appendix \[APP:DisoAver\].
The Deam-Edwards distribution can be understood as arising from a realistic vulcanization process in which the links are introduced simultaneously and instantaneously into the liquid state in equilibrium. Specifically, it incorporates the notion that all pairs of particles that happen (at some particular instant) to be nearby are, with a certain probability controlled by the link density parameter $\LinkDens$, linked. Thus, the correlations of the link distribution reflect the correlations of the unlinked liquid, and it follows that realizations of links only acquire an appreciable statistical weight if they are compatible with some reasonably probable configuration of the unlinked liquid.
The factor $\big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\big)^{\LinkNumb} /\LinkNumb !$ in the Deam-Edwards distribution introduces a Poissonian character to the total number $\LinkNumb$ of links. These links are envisioned to be the product of a Poisson chemical linking process. The factor $Z_{\RealDiso}(\Volu_0)$ assures that the probability of having a given random realization of links is proportional to the statistical weight for, in the unlinked liquid state, finding the to-be-linked pairs to be co-located in the liquid state to within the shape function $\exp \big(- \vert \PartPosi_{i_e}-\PartPosi_{j_e}\vert ^2 / 2\LinkScal^2\big)$.
As a result of the Deam-Edwards distribution, the mean number of links per particle is given by $\lda \LinkNumb \rda/\PartNumb = \LinkDens/2$. Thus, $\LinkDens = 2 \lda \LinkNumb \rda/N$ is the *mean coordination number*, i.e., the average number of particles to which a certain particle is connected, [[the factor of $2$ results from the fact that each link is shared by two particles]{}]{}. For a detailed discussion of the Deam-Edwards distribution, see Ref. [@Deam1976; @Broderix2002]. By using this distribution of the quenched disorder, we can perform the disorder average of the Helmholtz free energy via the replica technique, thus obtaining $$\begin{aligned}
\label{EQ:HFReplica}
\lda \HelmFreeEner \rda
&\equiv& \sum_{\RealDiso} \DEDist(\RealDiso)
\HelmFreeEner_{\RealDiso} (\Volu) \nonumber\\
&=& -\BoltCons T \sum_{\RealDiso} \DEDist(\RealDiso) \ln Z_{\RealDiso} (\Volu)
\nonumber\\
&=& -\BoltCons T \lim _{n \to 0} \sum_{\RealDiso} \DEDist(\RealDiso)
\frac{Z_{\RealDiso}(\Volu)^{n}-1}{n}.\end{aligned}$$ We now insert the Deam-Edwards distribution to get $$\begin{aligned}
\label{EQ:HFReplicaDE}
\lda \HelmFreeEner \rda
&=& -\BoltCons T \lim _{n \to 0} \sum_{\RealDiso}
\frac{
\Big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\Big)^{\LinkNumb}
Z_{\RealDiso}(\Volu_0)}{\LinkNumb ! Z_1} \nonumber\\
&&\quad\times \frac{Z_{\RealDiso}(\Volu)^{n}-1}{n}.\end{aligned}$$ This disorder-averaged free energy differs from the form traditionally obtained via the replica technique, in that there is an extra replica $Z_{\RealDiso}(\Volu_0)$, which originates in the Deam-Edwards distribution. We shall call this extra replica the $0^{\textrm{th}}$ replica, and note that it represents the *preparation state* of the system. [^3] The summation over the realizations of the quenched disorder $\RealDiso$ can be performed, following the calculation in Appendix \[APP:DisoAver\]; thus we arrive at the form $$\begin{aligned}
\label{EQ:HFEDisoAver}
\lda \HelmFreeEner \rda
&=& -\BoltCons T \lim _{n \to 0} \frac{1}{n}\Big( \frac{Z_{1+n}}{Z_1}-1 \Big),\end{aligned}$$ which can also be expressed as $$\begin{aligned}
\label{EQ:HFEDisoAverDeri}
\lda \HelmFreeEner \rda = -\BoltCons T \lim _{n \to 0}
\frac{\partial}{\partial n} \ln Z_{1+n} \, ,\end{aligned}$$ where
$$\begin{aligned}
\label{EQ:ReplPart}
Z_{1+n} &\equiv& \sum_{\RealDiso}
\frac{\Big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\Big)^{\LinkNumb}}
{ \LinkNumb !} Z_{\RealDiso}(\Volu_0) Z_{\RealDiso}(\Volu)^n \nonumber\\
&=& \PartitionLiquid(\Volu_0)\PartitionLiquid(\Volu)^n \lthal \exp \Big(\frac{\LinkDens \Volu_0}
{2\PartNumb \LinkPote_0} \sum_{i\ne j}^{\PartNumb} \ReplProd
\LinkPote \big(
\vert \PartPosi_{i}\REPa - \PartPosi_{j}\REPa \vert
\big)\Big)\rthal_{1+n}^{\HamiExcl} .\end{aligned}$$
Notice that, here, the preparation state (i.e., $0^{\textrm{th}}$ replica) has a fixed volume $\Volu_0$ because, for convenience, we have assumed that the linking process was undertaken instantaneously in a liquid state of fixed volume and thus the pressure is fluctuating, whereas the measurement states (replicas $1$ through $n$) are put in a fixed-pressure $\Pres$ environment, the volume $\Volu$ of which is allowed to fluctuate. In the latter parts of the paper we shall set the pressure $\Pres$ to be the average pressure measured in the preparation state at volume $\Volu_0$. In particular, for a given volume of the liquid state in which the links are made, the average pressure is given by $$\begin{aligned}
\Pres = - \frac{\partial \FLiquid (\Volu_0)}{\partial \Volu_0}
\Big\vert_{T} \, ,\end{aligned}$$ where we have introduced the Helmholtz free energy of the unlinked liquid $\FLiquid (\Volu_0) \equiv - \BoltCons T \ln \PartitionLiquid (\Volu_0)$. We suppose that the excluded-volume interactions are so strong that the density fluctuations are suppressed, and the density of the unlinked liquid is just $\PartNumb/\Volu_0$. [^4] Thus, the mean-field value of Helmholtz free energy in the unlinked liquid state is $$\begin{aligned}
\label{EQ:FLiquid}
\FLiquid (\Volu_0) &=& -\PartNumb \BoltCons T \ln \Volu_0
+\frac{\ExclVolu \PartNumb^2}{2\Volu_0}.\end{aligned}$$ Therefore, the mean pressure in the unlinked liquid state is given by $$\begin{aligned}
\label{EQ:AverPres}
\Pres = \frac{\PartNumb \BoltCons T}{\Volu_0}
+\frac{\ExclVolu \PartNumb^2}{2\Volu_0^2},\end{aligned}$$ [[from which we can identify—by the standard way in which the second virial coefficient $B_2$ appears in the free energy of the equation of state for a fluid—that $B_2$ for the unlinked liquid is $\ExclVolu/2\BoltCons T$. Thus, without having actually performed a cluster expansion, we see that the Dirac delta-function interaction with coefficient $\ExclVolu$ indeed leads to a virial expansion with a suitable excluded volume, viz., $\ExclVolu/(\BoltCons T)$. ]{}]{} As mentioned above, we shall apply this pressure in the measurement states (described by $1^{\textrm{th}}$ through $n^{\textrm{th}}$ replicas), and let their volumes $\Volu$ fluctuate, in order to obtain an elastic free energy that can describe volume variations. In particular, by choosing the pressure $p$ to be exactly the mean pressure of the liquid state, we shall obtain an elastic free energy that takes the *state right after linking*, which has the same volume $\Volu_0$ as the liquid state, as the elastic reference state. This issue of the state right after linking and the elastic reference state will be discussed in detail in Section \[SEC:PhenFE\].
In light of this construction of the pressure ensemble, we have the capability of learning about the bulk modulus of the system, and to characterize volume changes caused by linking, a process that has the effect of eliminating translational degrees of freedoms.
To establish an appropriate statistical mechanics for the fixed-pressure ensemble, we shall make the following Legendre transformation of the Helmholtz free energy, which leads to the Gibbs free energy $\GibbFreeEner (\Pres,T)$:
\[EQ:LegeTran\] $$\begin{aligned}
\Pres &=& -\frac{\partial \HelmFreeEner(\Volu,T)}
{\partial \Volu}\big\vert_{T} \label{EQ:LegeTran1} \, , \\
\GibbFreeEner (\Pres,T) &=& \HelmFreeEner(\Volu,T) + \Pres \Volu \, . \label{EQ:LegeTran2}\end{aligned}$$
In Eq. (\[EQ:LegeTran2\]) the volume $\Volu$ takes the value (in terms of $\Pres$) that satisfies Eq. (\[EQ:LegeTran1\]), i.e., the volume that minimizes the Gibbs free energy at a given pressure $p$.
In the following sections, we shall first calculate the disorder-average of the Helmholtz free energy, and then make this Legendre transformation to obtain the disorder-averaged Gibbs free energy. This will allow us to explore the elasticity of the RLPM in detail.
Field-theoretic description of the RLPM {#SEC:FTMF}
---------------------------------------
We shall use field-theoretic methods to analyze the disorder-averaged free energy $\lda\HelmFreeEner\rda$ and, more specifically, the replicated partition function $Z_{1+n}$. To do this, we introduce a joint probability distribution for the particle density in the replicated space, i.e., the replicated density function $$\begin{aligned}
\label{EQ:DensReal}
\DensFunc(\REPX)\equiv \frac{1}{\PartNumb} \sum_{i=0}^{\PartNumb}
\ReplProd \delta^{(d)}(x\REPa-\PartPosi_i\REPa) ,\end{aligned}$$ where $\REPX \equiv (x^{0},x^{1},\ldots,x^{n})$ is a short-hand for the $(1+n)$-replicated position $\Dime$-vector. For convenience, we introduce a complete orthonormal basis set in replica space $\{\BASE\REPa\}_{\alpha=0}^{n}$, in terms of which a vector $\REPX$ can be expressed as $$\begin{aligned}
\REPX=\ReplSum x\REPa \BASE\REPa .\end{aligned}$$ Note that the components $x\REPa$ are themselves $\Dime$-vectors. With this notation, the density function of a single replica $\alpha$ is given by $\DensFunc_{p\REPa\BASE\REPa}$, which is the Fourier transform of the $\DensFunc(\REPX)$ field, $\DensFunc_{\REPP}$, with momentum nonzero only in replica $\alpha$, corresponding to integrating over the normalized densities in other replicas in real space.
The replicated partition function (\[EQ:ReplPart\]) can be written as a functional of the replicated density function $\DensFunc$ in momentum space as $$\begin{aligned}
\label{EQ:ZQ}
Z_{1+n} = \int_{\Volu_0} \prod_{i=1}^{\PartNumb} d\PartPosi_{i}^{0}
\int_{\Volu} \ReplProd \prod_{j=1}^{\PartNumb} d\PartPosi_{j}\REPa
\, e^{-\frac{H_{\DensFunc}\lbrack \DensFunc_{\REPP}\rbrack}{\BoltCons T}} ,\end{aligned}$$ with $$\begin{aligned}
\label{EQ:HQ}
H_{\DensFunc}\lbrack \DensFunc_{\REPP}\rbrack
&\equiv& -\frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP}\DensFunc_{\REPP}\DensFunc_{-\REPP}\LinkPoteRepl_{\REPP}\nonumber\\
&&+\frac{\ExclVolu \PartNumb^2}{2\Volu_0} \sum_{p}\DensFunc_{p\BASE^{0}}
\DensFunc_{-p\BASE^{0}} \nonumber\\
&&+\frac{\ExclVolu \PartNumb^2}{2\Volu} \sum_{p} \ReplSumOne
\DensFunc_{p\BASE\REPa} \DensFunc_{-p\BASE\REPa} ,\end{aligned}$$ where the factor $$\begin{aligned}
\LinkPoteRepl_{\REPP} = \big(\LinkPote_0\big)^{1+n}
e^{-\LinkScal^2 \vert \REPP \vert^2/2}\end{aligned}$$ is the replicated version of the Fourier transform of the function $\LinkPote(x)$, defined in Eq. (\[EQ:DeltReal\]). [[The summation $\sum_{\REPP}$ denotes a summation over all momentum $\Dime$-vectors $p\REPa$, one for each replica, with $\REPP$ taking the values $\REPP=\ReplSum p\REPa \BASE\REPa$. The cartesian components of the $p\REPa$ take the values $2\pi m/L$, where $L$ is the linear size of the system and $m$ is any integer. Similarly, summations $\sum_{p}$ over $\Dime$-vectors $p$ include components having the values $2\pi m/L$.]{}]{} The first term on the right hand side of Eq. (\[EQ:HQ\]) arises from the attractive, link-originating, interaction part \[see Eq. (\[EQ:ReplPart\])\]; the next two terms represent the excluded-volume interaction in $\HamiExcl$, for the $0^{\textrm{th}}$ replica and for replicas $1$ through $n$, respectively.
The excluded-volume interaction is taken to be very strong, and thus the density fluctuations in any single replica are heavily suppressed. This means that $\DensFunc_{p\REPa\BASE\REPa}$ are very small for all $p\REPa\ne 0$, and $\DensFunc_{p\REPa\BASE\REPa}\vert_{p\REPa=0}=1$, corresponding to a nearly homogeneous particle density. To manage this issue, we separate the replicated space into a Lower Replica Sector (LRS), in which the density fluctuations are suppressed, and a Higher Replica Sector (HRS), which captures the correlations between different replicas, and develops an instability at the liquid-to-random-solid transition. The definitions of the LRS and HRS are (in momentum space) as follows: if two or more components of a replicated momentum vector $\REPP \equiv (p^{0},p^{1},\ldots,p^{n})$ are non zero then $\REPP$ is an element of the HRS; on the other hand, if $\REPP$ has zero or only one component $p\REPa$ being nonzero, and all other $p\REPb=0$, then $\REPP$ is an element of the LRS. In addition, the LRS can be separated into a 1RS part, in which vectors $\REPP$ has exactly one component $p\REPa$ being nonzero, and a 0RS, which consists of only $\REPP=0$. With this separation we can rewrite the effective Hamiltonian as $$\begin{aligned}
\label{EQ:HQHL}
H_{\DensFunc}\lbrack \DensFunc_{\REPP}\rbrack
&=& -\frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\DensFunc_{\REPP}
\DensFunc_{-\REPP}\LinkPoteRepl_{\REPP}\nonumber\\
&&+\frac{ \PartNumb^2}{2\Volu_0} \sum_{p}\ExclVoluRN_0(p)\DensFunc_{p\BASE^{0}}
\DensFunc_{-p\BASE^{0}} \nonumber\\
&&+\frac{ \PartNumb^2}{2\Volu} \sum_{p}\ExclVoluRN(p) \ReplSumOne
\DensFunc_{p\BASE\REPa} \DensFunc_{-p\BASE\REPa} ,\end{aligned}$$ with the renormalized coefficients $$\begin{aligned}
\label{EQ:ExclVoluRN}
\frac{\ExclVoluRN_0(p) \PartNumb^2}{2\Volu_0}
&\equiv&\frac{\ExclVolu \PartNumb^2}{2\Volu_0}
-\frac{\PartNumb \LinkDens\BoltCons T\LinkPoteRepl_{\REPP}}{2\Volu^n\LinkPote_0} ,
\nonumber\\
\frac{\ExclVoluRN(p) \PartNumb^2}{2\Volu}
&\equiv&\frac{\ExclVolu \PartNumb^2}{2\Volu}
-\frac{\PartNumb \LinkDens\BoltCons T\LinkPoteRepl_{\REPP}}{2\Volu^n\LinkPote_0} .\end{aligned}$$ We suppose that $\frac{\ExclVolu N}{\BoltCons T \Volu}\gg \LinkDens$ (i.e., the excluded-volume repulsion is very strong, relative to the attractive effects of the links), so these coefficients $\frac{\ExclVoluRN(p) \PartNumb^2}{2\Volu}$ are always positive and large, relative to the energy-scale of the HRS that we are interested in.
The interactions in Eq. (\[EQ:HQHL\]) can be decoupled using a Hubbard-Stratonovich (HS) transformation (for details see Appendix \[APP:HSTransformation\]). Thus, we arrive at a field-theoretic formulation of the replicated partition function, in terms of the order parameter field $\VOP$: $$\begin{aligned}
\label{EQ:HS}
Z_{1+n} = \int \mathcal{D}\VOP_{\hat{p}}\ReplProd\mathcal{D}\VOP_{p\BASE\REPa}
e^{-\frac{H_{\VOP}\lbrack \VOP_{\hat{p}},\VOP_{p\BASE\REPa} \rbrack}{\BoltCons T}} ,\end{aligned}$$ where the effective Hamiltonian is given by
$$\begin{aligned}
\label{EQ:HEffcVOP}
\! H_{\VOP}\lbrack \VOP_{\hat{p}},\VOP_{p\BASE\REPa} \rbrack
&=& \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\!\VOP_{\REPP}
\VOP_{-\REPP}\LinkPoteRepl_{\REPP}
+\frac{ \PartNumb^2}{2\Volu_0} \sum_{p}\ExclVoluRN_0(p)\VOP_{p\BASE^{0}}
\VOP_{-p\BASE^{0}} \nonumber\\
&& +\frac{\PartNumb^2}{2\Volu} \sum_{p} \ExclVoluRN(p) \ReplSumOne
\VOP_{p\BASE\REPa} \VOP_{-p\BASE\REPa} \!-\! N\BoltCons T \ln \MyZzero \, .\end{aligned}$$
and $$\begin{aligned}
\MyZzero &=& \int_{\Volu_0} \! d\PartPosi^{0} \!
\int_{\Volu} \!\ReplProd \! d\PartPosi\REPa
\exp\!\Big\lbrack
\frac{ \LinkDens}{ \Volu^n \LinkPote_0}
\!\sum_{\REPP\in HRS}\!\!\VOP_{\REPP}\,\LinkPoteRepl_{\REPP}e^{i\REPP \cdot \hat{c}} \nonumber\\
&& \quad\quad + \frac{i \PartNumb}{\Volu_0\BoltCons T}
\sum_{p}\ExclVoluRN_0(p)\VOP_{p\BASE^{0}}\,e^{ip^{0}c^{0}}
+\frac{i \PartNumb}{\Volu\BoltCons T} \!\sum_{p} \ExclVoluRN(p)\! \ReplSumOne
\VOP_{p\BASE\REPa} \,e^{ip\REPa c\REPa} \! \Big\rbrack .\end{aligned}$$ The form of this HS transformation \[see Appendix \[APP:HSTransformation\], especially Eq. (\[EQ:HSAver\])\] ensures that the mean value of the order parameter field $\VOP$ is related to the mean value of the replicated density function field $\DensFunc$ as
$$\begin{aligned}
&&\textrm{HRS:} \quad\quad
\langle \DensFunc_{\REPP} \rangle_{H_{\DensFunc}} \label{EQ:HSrelationHRS}
= \langle \VOP_{\REPP} \rangle_{H_{\VOP}} , \\
&&\textrm{LRS:} \quad\quad
i\langle \DensFunc_{p\BASE\REPa} \rangle_{H_{\DensFunc}}
= \langle \VOP_{p\BASE\REPa} \rangle_{H_{\VOP}} ,\end{aligned}$$
where the averages on either sides are defined via
$$\begin{aligned}
\langle \, \cdots \rangle_{H_{\DensFunc}}
& \equiv \frac{1}{Z_{1+n}}
\int_{\Volu_0} \prod_{i=1}^{\PartNumb} d\PartPosi_{i}^{0}
\int_{\Volu} \ReplProdOne \prod_{j=1}^{\PartNumb} d\PartPosi_{j}\REPa \label{EQ:HQAver}
\,\, e^{-\frac{H_{\DensFunc}\lbrack \DensFunc_{\REPP}\rbrack}{\BoltCons T}} \cdots , \\
\langle \, \cdots \rangle_{H_{\VOP}}
& \equiv \frac{1}{Z_{1+n}}
\int \mathcal{D}\VOP_{\hat{p}}\ReplProd\mathcal{D}\VOP_{p}\REPa \label{EQ:HVOPAver}
\,\,e^{-\frac{H_{\VOP}\lbrack \VOP_{\hat{p}},\VOP_{p\BASE\REPa} \rbrack}{\BoltCons T}} \cdots . \end{aligned}$$
The leading-order terms in $H_{\VOP}\lbrack \VOP_{\hat{p}},\VOP_{p\BASE\REPa} \rbrack$ can be constructed by expanding the $\ln \MyZzero$ term in Eq. (\[EQ:HEffcVOP\]) in powers of the fields $\VOP_{\REPP}$ and $\VOP_{p\BASE\REPa}$, and thus we can obtain the leading-order terms in the Landau-Wilson effective Hamiltonian. To leading order this expansion gives $$\begin{aligned}
\label{EQ:HQExpansion}
H_{\VOP}\lbrack \VOP_{\hat{p}},\VOP_{p\BASE\REPa} \rbrack
= \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
&& \sum_{\REPP\in HRS}\VOP_{\REPP}\VOP_{-\REPP}\LinkPoteRepl_{\REPP}
\Bigg(1-\LinkDens \frac{\LinkPoteRepl_{\REPP}}{V^{n}\LinkPote_{0}}\Bigg)
+\frac{\ExclVoluRN_0 \PartNumb^2}{2\Volu_0} \sum_{p}\VOP_{p\BASE^{0}}
\VOP_{-p\BASE^{0}}
\Bigg(1+\frac{\ExclVoluRN_0 \PartNumb}{\Volu_0\BoltCons T}\Bigg)
\nonumber\\ \noalign{\medskip}
+\frac{\ExclVoluRN \PartNumb^2}{2\Volu} &&\sum_{p} \ReplSumOne
\VOP_{p\BASE\REPa} \VOP_{-p\BASE\REPa}
\Bigg(1+\frac{\ExclVoluRN \PartNumb}{\Volu_0\BoltCons T}\Bigg)
+O\big((\VOP_{\hat{p}})^3,(\VOP_{p\BASE\REPa})^3\big).\end{aligned}$$
For the LRS fields $\VOP_{p\BASE\REPa}$ we see the coefficients of the corresponding quadratic term are always positive (given that $\ExclVoluRN_0,\,\ExclVoluRN>0$, i.e., the excluded-volume repulsion is very strong), so this sector of the field theory does not undergo an instability. Furthermore, because these coefficients (the masses, in particle-physics language) are very large \[see Eq. (\[EQ:ExclVoluRN\])\], the fluctuations of these LRS fields $\VOP_{p\BASE\REPa}$ are heavily suppressed. For this reason, we ignore these fluctuations and, for all $\alpha =0,1,\ldots,n$, we take $$\begin{aligned}
\label{EQ:LRScons}
\VOP_{p\BASE\REPa}\vert_{p=0}&=& i , \nonumber\\
\VOP_{p\BASE\REPa}\vert_{p\ne 0}&=& 0 ,\end{aligned}$$ as a hard constraint.
Having implemented this constraint, we arrive at the HRS Hamiltonian (the full form, not just the leading-order expansion): $$\begin{aligned}
\label{EQ:HVOPHRS}
H_{\VOP}\lbrack\VOP_{\REPP}\rbrack
&=& \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP}
\VOP_{-\REPP}\LinkPoteRepl_{\REPP}\nonumber\\ \noalign{\medskip}
&&\quad - N\BoltCons T \ln \MyZzero ,\end{aligned}$$ where $$\begin{aligned}
\label{EQ:Z_zero}
\MyZzero&\equiv& \int_{\Volu_0} \! d\PartPosi^{0} \!\!
\int_{\Volu} \!\ReplProdOne d\PartPosi\REPa
\exp \Big\lbrack
\frac{ \LinkDens}{ \Volu^n \LinkPote_0} \nonumber\\ \noalign{\medskip}
&&\quad\quad\times \sum_{\REPP\in HRS}\VOP_{\REPP}\LinkPoteRepl_{\REPP}e^{i\REPP \cdot \hat{c}}
\Big\rbrack .\end{aligned}$$ The Landau theory of the vulcanization transition [@Peng1998] can be recovered by making the expansion of this HRS Hamiltonian that keeps only the leading-order terms in the order-parameter $\VOP$ and the momentum $\REPP$. Up to an additive constant and an appropriate rescaling of the order parameter, this expansion reads $$\begin{aligned}
\label{EQ:Landau}
H_{\VOP}\lbrack\VOP_{\REPP}\rbrack
&=& \frac{1}{2}\sum_{\REPP\in HRS} \big( r+ \vert \REPP \vert^2 \big)
\VOP_{\REPP} \VOP_{-\REPP} \nonumber\\
&& -\frac{v}{3!} \sum_{\REPP_1,\REPP_2 \in HRS} \VOP_{\REPP_1} \VOP_{\REPP_2} \VOP_{-\REPP_1-\REPP_2} \, ,\end{aligned}$$ where the potential of the links $\LinkPoteRepl_{\REPP}$ has been momentum-expanded. This is precisely the form of the Landau free energy that was constructed via *symmetry arguments* in Ref. [@Peng1998]. In the limit $n\to 0$, the coefficients become $$\begin{aligned}
r &\propto& \LinkDens(1-\LinkDens) , \nonumber\\
v &\propto& (\LinkDens)^3 .\end{aligned}$$
It is straightforward to see that the $r$ term leads to an *instability* for the link density parameter $\LinkDens$ larger than the critical value $\LinkDens_C=1$, and the lowest unstable modes are long-wavelength modes (i.e., $\REPP \to 0$). \[One should, however, keep in mind that the component $\REPP=0$ itself, which is the 0RS, is excluded from this HRS-only field theory; see Eq. (\[EQ:LRScons\]).\] This instability corresponds to the *liquid-to-soft-random-solid transition*, because the liquid state corresponds to the $\VOP_{\REPP}=0$ (in the HRS) state, and becomes unstable when the link density parameter $\LinkDens$ exceeds $1$.
[[ There are two points that we would like to discuss about this Hamiltonian. First, it can be seen from the expansion in Eq. (\[EQ:Landau\]) that near the critical point the exact form of interaction is irrelevant, because we have only kept terms to $\vert\REPP\vert^2$ in $\LinkPoteRepl_{\REPP}$, and this governs the long-distance physics. We have used the Gaussian-chain potential (\[EQ:VGC\]) in this calculation. It is clear that if we were to change to a different potential, such as a finite-rest-length spring $\VLink(\vert r \vert) = \frac{k}{2} \big(\vert r \vert -l \big)^2$, the long-distance (i.e., small-momentum) physics would be unchanged. Second, as we shall see in Sec. \[SEC:Heterogeneity\], neither do the statistics of the elastic modulus depend on the details of the interaction potential; instead, they only depend on the density of links. This results from the fact that the elasticity originates in the entropy of the *network*. The critical point for the vulcanization transition, and thus the entropic rigidity in the presence of thermal fluctuations, occurs at the same point as *connectivity percolation* does, rather than at the *rigidity percolation* critical point. The fact that we have a shear modulus scaling as $T$ results from the entropy of the network, and not from the factor of $T$ in the Gaussian-chain potential, is clear because if we were to change to another attractive potential, the same results would hold for the long-distance physics. ]{}]{}
Mean-field theory of the RLPM
-----------------------------
To understand the physics of the order parameter in vulcanization theory, and thus obtain the form of the stationary value of the order parameter, we need to recall the property of the HS transformation, Eq. (\[EQ:HSrelationHRS\]), which relates the average of the order parameter field $\VOP$ to the average of the replicated density function $\DensFunc$. According to this relation, we have $$\begin{aligned}
\langle \VOP_{\REPP} \rangle_{H_{\VOP}}
= \Big\langle \frac{1}{N}\sum_{j=1}^{\PartNumb}e^{i\REPP \cdot \hat{c}_j}\Big\rangle_{H_{\DensFunc}}
-\delta^{((1+n)d)}_{\REPP,0} .\end{aligned}$$ Here, the $\delta^{((1+n)d)}_{\REPP}$ removes the 0RS part of $\VOP_{\REPP}$. Equivalently, in real space we have $$\begin{aligned}
\langle \VOP(\REPX)\rangle_{H_{\VOP}} \!\!
=\!\Big\langle \frac{1}{N}\!\sum_{j=1}^{\PartNumb}\delta^{((1+n)d)}(\REPX-\hat{c}_j)\!\Big\rangle_{H_{\DensFunc}}
\!\!\!-\frac{1}{\Volu_0 \Volu^n} ,\end{aligned}$$ which can be interpreted as $$\begin{aligned}
\label{EQ:VOPinte}
\langle \VOP(\REPX) \rangle_{H_{\VOP}}
&=&\!\frac{1}{N}\!\sum_{j=1}^{\PartNumb} \big\lbrack
\langle \delta^{(d)}(x^0-c^{0}_{j})\rangle
\langle \delta^{(d)}(x^1-c^{1}_{j})\rangle \cdots \nonumber\\
&&\quad\times
\langle \delta^{(d)}(x^n-c^{n}_{j})\rangle
\big\rbrack - \frac{1}{\Volu_0 \Volu^n} \, .\end{aligned}$$ This average consists of the following two steps: One first constructs *independent thermal averages in each replica* (denoted by $\langle\cdots\rangle$) with a common given realization of disorder $\RealDiso$; one then forms the product over all replicas, and finally one *averages over all realizations of disorder* (an average denoted by $\big\lbrack\cdots\big\rbrack$). This interpretation can be understood from the definition of $H_{\DensFunc}$ via $Z_{1+n}$, as in Eq. (\[EQ:ZQ\]). Recall that $Z_{1+n}$, as defined in Eq. (\[EQ:ReplPart\]), contains thermal averages of the $(1+n)$ replicas, represented by the factor $Z_{\RealDiso}(\Volu_0)\, Z_{\RealDiso}(\Volu)^n$, together with an overall disorder average. This validates the interpretation given in Eq. (\[EQ:VOPinte\]). For a strict proof, see Ref. [@Goldbart1996].
The structure of Eq. (\[EQ:VOPinte\]) allows us to relate the value of the order parameter $\VOP$ to measurements on the system. In the liquid state, the single-particle densities $\ltha \delta^{(d)}(x\REPa-c\REPa_{j}) \rtha$ in each replica are simply $1/\Volu$ (or $1/\Volu_0$ for the $0^{\textrm{th}}$ replica), and thus the order parameter $\VOP$ vanishes. In the soft random solid state, it is hypothesized that *a finite fraction $\LocaPart$ of the particles become localized around random positions*. This happens when the density of links exceeds the *percolation threshold*, and the particles that constitute the infinite, percolating cluster become localized.
In the language of replica field theory, a localized particle remains near the same spatial position in each replica[^5]. This is because, as we discussed earlier, *each replica corresponds to a copy of the same disordered system but with independent thermal fluctuations*, and for a particle localized as a part of the percolating cluster, it fluctuates around its fixed mean position, which is common to all replicas. According to these considerations, it is reasonable to hypothesize the following form for the stationary value of the order parameter in real space: $$\begin{aligned}
\label{EQ:VOPAnsatzR}
\!\!\!\!\VOPSP(\REPX)\!\!\!\!&=&\!\!\!\LocaPart \int \frac{dz}{\Volu_0}\int d\ISLL \DistISLL(\ISLL)
\Big(\frac{\ISLL}{2\pi}\Big)^{\frac{(1+n)\Dime}{2}}
\nonumber\\
&& \!\!\times e^{-\frac{\ISLL}{2}
\{ \vert x^0-z\vert^2 +\ReplSumOne\vert x\REPa-\Contraction z \vert^2 \}} \!\!
-\!\frac{\LocaPart}{\Volu_0 \Volu^n} .\,\end{aligned}$$ To arrive at this form we have assumed that the fraction of particles that become localized is $\LocaPart$, and the localized single-particle density function is proportional to $e^{-\frac{\ISLL}{2}\vert x\REPa-\Contraction z \vert^2}$ in replica $\alpha$($=1,\ldots,n$), where $\Contraction z$ is the random position near to which the particle is localized, and $\Contraction$ corresponds to the uniform contraction of the entire volume in the measurement state with respect to the preparation state, due to linking \[see the discussion following Eq. (\[EQ:ContRLPM\])\]. Correspondingly, this particle was near the position $z$ in the preparation state, so, for replica $0$ the corresponding single-particle density function is proportional to $e^{-\frac{\ISLL}{2}\vert x^0-z \vert^2}$. For convenience of notation we define $\hat{z}\equiv (z,\Contraction z, \Contraction z,\ldots)$ as the mean position vector in the replicated space. The contraction $\Contraction$ is related to the change of volume as $$\begin{aligned}
\label{EQ:VContraction}
\frac{\Volu}{\Volu_0}=\Contraction^{d} .\end{aligned}$$ The localization of the particle is characterized by the localization length $\LocaLeng$, although for notational convenience we exchange this variable for the inverse square localization length $\ISLL\equiv 1/\LocaLeng^2$. Because the network is heterogeneous, the particles can have widely different localization lengths. This heterogeneity is characterized by the distribution $\DistISLL(\ISLL)$.
We can also write this stationarity-point order parameter in momentum space: $$\begin{aligned}
\label{EQ:VOPAnsatzM}
\!\!\!\!(\!\VOPSP\!)_{\REPP}\!\!\!&=&\!\!\!\LocaPart \int \frac{dz}{\Volu_0}\int d\ISLL \DistISLL(\ISLL)
\nonumber\\ \noalign{\medskip}
&& \!\!\times e^{-\frac{\vert\REPP\vert^2}{2\ISLL}
\!-\! ip^0 \cdot z-i\ReplSumOne p\REPa \cdot (\Contraction z) }
\!\!-\!\LocaPart\delta^{((1+n)d)}_{\REPP,0}\! .\,\,\,\end{aligned}$$ The parameters that characterize this order parameter, $\LocaPart$ and $\DistISLL(\ISLL)$, have been obtained by solving the stationarity condition for the Hamiltonian: $$\begin{aligned}
\frac{\delta H_{\VOP}}{\delta \VOP_{\REPP}}=0 .\end{aligned}$$ The form of the order parameter $(\VOPSP)_{\REPP}$ given in Eq. (\[EQ:VOPAnsatzM\]) exactly solves the above stationarity condition, thus we arrive at self-consistency equations of $\LocaPart$ and $\DistISLL(\ISLL)$. In particular, the equation for $\LocaPart$ is $$\begin{aligned}
\label{EQ:SPQ}
1-\LocaPart=e^{-\LinkDens \LocaPart}.\end{aligned}$$ For all values of $\LinkDens$, Eq. (\[EQ:SPQ\]) has a solution $\LocaPart=0$, corresponding to the liquid state. However, for $\LinkDens> 1$, an additional root appears, emerging continuously from $\LocaPart=0$ at $\LinkDens= 1$, and describing the equilibrium amorphous solid state. In Fig. \[FIG:ThetaGamma\] we show the dependence of the localized fraction $\LocaPart$ on the link density, which we characterize by $\LinkDens$. [[The critical point $\LinkDens_C=1$ corresponds to mean coordination number $z$ of $1$, and this agrees with the classic work on the statistical properties of random graphs by Erd[ő]{}s and R[é]{}nyi [@Erdos1960].]{}]{} For a detailed discussion and for the stationary-point distribution of inverse square localization lengths $\DistISLL(\ISLL)$, see Refs. [@Goldbart1996; @Castillo1994]. The contraction $\Contraction$, which is relevant to the elasticity of the random solid state, can be investigated by inserting the form (\[EQ:VOPAnsatzM\]) of the order parameter into the Hamiltonian $H_{\VOP}$, Eq. (\[EQ:HVOPHRS\]), which yields the dependence of the Hamiltonian on the parameters $\LocaPart$, $\DistISLL(\ISLL)$ and $\Contraction$. Through a lengthy derivation, and by keeping terms to $O(n)$, we arrive at the following Hamiltonian for the stationary point (cf. Appendix \[APP:HSP\]):
$$\begin{aligned}
\label{EQ:HSP}
\HVOPSP&=&\frac{\ExclVoluRN_0(0)\PartNumb^2}{2\Volu_0}
+\frac{n\ExclVoluRN(0)\PartNumb^2}{2\Volu}
-\PartNumb \BoltCons T \ln \Volu_0 -n\PartNumb \BoltCons T \ln \Volu
+ n\PartNumb \BoltCons T \Bigg\lbrace
\UnivPara \Big\lbrack
\frac{d}{2}\big(\ln(2\pi)+\Contraction^2\big)-\ln\Volu
\Big\rbrack \nonumber\\
&&- \frac{\LinkDens \LocaPart^2}{2}\cdot\frac{d}{2}\int_{\ISLL_1,\ISLL_2}
\ln\big( \frac{1}{\ISLL_1}+\frac{1}{\ISLL_2}+\LinkScal^2 \big)
-e^{-\LinkDens \LocaPart}\frac{d}{2}
\sum_{m=1}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}
\int_{\ISLL_1,\ldots,\ISLL_m} \ln
\Big( \frac{\tISLL_1\cdots\tISLL_m}{\tISLL_1 +\cdots +\tISLL_m} \Big)
\Bigg\rbrace .\end{aligned}$$
Here, the variable $\tISLL$ is defined as $\tISLL\equiv\big(\frac{1}{\ISLL}+\LinkScal^2\big)^{-1}$, where $\ISLL$ is the inverse square localization length $\ISLL\equiv 1/\LocaLeng^2$, and we have introduced the shorthand for the integrals $\int_{\ISLL}\equiv \int d\ISLL \DistISLL(\ISLL)$. The dimensionless factor $\UnivPara$ in Eq. (\[EQ:HSP\]) is given by $$\begin{aligned}
\label{EQ:theta}
\UnivPara\equiv -\frac{\LinkDens \LocaPart^2}{2}+\LinkDens \LocaPart
-1+e^{-\LinkDens \LocaPart}.\end{aligned}$$ We shall see in Section \[SEC:Heterogeneity\] that $\UnivPara$ also controls the mean value of the shear modulus, as well as the amplitude of the disorder correlators that involve the residual stress fields.
To obtain the disorder-averaged free energy, we shall make the stationary-point approximation: $$\begin{aligned}
Z_{1+n}\simeq e^{-\frac{\HVOPSP}{\BoltCons T}} .\end{aligned}$$ Thus, we can obtain the Helmholtz free energy using Eq. (\[EQ:HFEDisoAver\]), arriving at the result
$$\begin{aligned}
\lda \HelmFreeEner_{\textrm{SP}} \rda \!&=&\! -\BoltCons T \lim _{n \to 0} \frac{1}{n}
\Big( \frac{Z_{1+n}}{Z_1}-1 \Big) \nonumber\\
\!&=&\! -\PartNumb \BoltCons T \Big(1-\frac{\LinkDens}{2}+\UnivPara \Big) \ln\Volu
-\PartNumb \BoltCons T \frac{\UnivPara d}{2}\big(\ln(2\pi)+\Contraction^2 \big)
+\frac{\ExclVolu\PartNumb^2}{2\Volu}
-\frac{\LinkDens \PartNumb \BoltCons}{2} \ln \LinkPote_0 \nonumber\\
&&+ \PartNumb \BoltCons T \frac{\LinkDens\LocaPart^2}{2}\frac{d}{2}
\int_{\ISLL_1,\ISLL_2}
\!\ln\Big( \frac{1}{\ISLL_1}+\frac{1}{\ISLL_2}+\frac{\LinkScal^2}{\BoltCons T} \Big)\!
\!+\!\PartNumb \BoltCons T \frac{d}{2} e^{-\LinkDens \LocaPart}
\sum_{m=1}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}
\!\int_{\ISLL_1,\ldots,\ISLL_m}\! \!\!\ln
\Big( \frac{\tISLL_1\cdots\tISLL_m}{\tISLL_1 +\cdots +\tISLL_m} \Big) ,\end{aligned}$$
where we have used the mean-field value of $Z_1$, from Eq. (\[EQ:Z1SP\]), and we have also made an expansion for small $n$ of the renormalized excluded-volume parameter $\ExclVoluRN$, using $$\begin{aligned}
\LinkPoteRepl_{0}=(\LinkPote_{0})^{1+n}
=\LinkPote_{0}\big(1+n\ln \LinkPote_{0} +O(n^2)\big).\end{aligned}$$ In order to study elasticity, we shall need to know the disorder-averaged Gibbs free energy $\lda\GibbFreeEner\rda$, which is given by a Legendre transformation, Eq. (\[EQ:LegeTran\]): $$\begin{aligned}
\lda\GibbFreeEner_{\textrm{SP}}\rda=\lda\HelmFreeEner_{\textrm{SP}}\rda+\Pres\Volu .\end{aligned}$$ [[We can insert the pressure $p$, given by Eq. (\[EQ:AverPres\]), and drop the slowly-varying $\ln\Volu$ term given that $\UnivPara\simeq\LinkDens/2-1$ provided the system is not close to the critical point $\LinkDens_C=1$. In the limit $\frac{\ExclVolu N}{\BoltCons T \Volu}\gg \LinkDens$ we arrive at]{}]{} $$\begin{aligned}
\label{EQ:SPGIBBS}
\lda \GibbFreeEner_{\textrm{SP}} \rda
\!&\simeq&\! \frac{\ExclVolu\PartNumb^2}{2\Volu_0}
\Big\lbrack 2+\Big(\frac{\Volu}{\Volu_0}-1\Big)^2\Big\rbrack
-\PartNumb \BoltCons T \frac{\UnivPara d}{2}\big(\ln(2\pi)+\Contraction^2 \big)
-\frac{\LinkDens \PartNumb \BoltCons}{2} \ln \LinkPote_0 \nonumber\\
&&+ \PartNumb \BoltCons T \frac{\LinkDens\LocaPart^2}{2}\frac{d}{2}
\int_{\ISLL_1,\ISLL_2}\!\!
\ln\Big( \frac{1}{\ISLL_1}+\frac{1}{\ISLL_2}+\frac{\LinkScal^2}{\BoltCons T} \Big)
\!+\!\PartNumb \BoltCons T \frac{d}{2} e^{-\LinkDens \LocaPart}
\!\sum_{m=1}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}
\!\int_{\ISLL_1,\ldots,\ISLL_2} \!\!\!\ln
\Big( \frac{\tISLL_1\cdots\tISLL_m}{\tISLL_1 +\cdots +\tISLL_m} \Big).\,\,\end{aligned}$$
Using the relation (\[EQ:VContraction\]), we can obtain the stationary values of the contraction $\Contraction$ that minimizes the disorder-averaged Gibbs free energy by solving $$\begin{aligned}
\frac{\delta \lda \GibbFreeEner_{\textrm{SP}} \rda }{ \delta \Contraction}=0 .\end{aligned}$$ [[In the limit $\frac{\ExclVolu N}{\BoltCons T \Volu}\gg \LinkDens$, the solution is $$\begin{aligned}
\label{EQ:ContRLPM}
\Contraction \simeq 1-\frac{ \UnivPara\Volu_0 \BoltCons T}{\ExclVolu \PartNumb d}.\end{aligned}$$ The limit $\frac{\ExclVolu N}{\BoltCons T \Volu}\gg \LinkDens$ is the same as the limit taken below Eq. (\[EQ:ExclVoluRN\]), indicating that the excluded-volume repulsion is much stronger than the attractive effects of the links.]{}]{} This contraction of the volume due to the introduction of links at a given pressure is a result of both the reduction of the total number of translational degrees of freedom, i.e., the change of the osmotic pressure, and the attractive interactions induced by the links. We shall see later that this contraction is consistent with a particular phenomenological model of a disordered elastic medium that we shall introduce.
Goldstone fluctuations in the RLPM
----------------------------------
### Spontaneous symmetry breaking {#SEC:SSB}
To characterize the Goldstone modes of fluctuations associated with the random solid state, we shall first look at the pattern of symmetry breaking accompanying the transition to this state.
The Hamiltonian (\[EQ:HVOPHRS\]) for the liquid-to-soft-random-solid transition has the symmetry of independent translations and rotations of each replica. The translational invariance of the Hamiltonian can be readily verified by making the transformation $$\begin{aligned}
\hat{x} &\to& \hat{x}'=\hat{x}+\hat{a} , \nonumber\\
\VOP(\hat{x}) &\to& \VOP'(\hat{x}')=\VOP(\hat{x}) = \VOP(\hat{x}'-\hat{a}) ,\end{aligned}$$ where $\hat{a}\equiv(a^0,a^1,\ldots,a^n)$ represents a replicated translation. In momentum space this transformation reads $$\begin{aligned}
\VOP_{\hat{p}} &\to & \VOP'_{\hat{p}}=e^{i\hat{p}\cdot\hat{a}}\, \VOP_{\hat{p}}\, .\end{aligned}$$ It is easy to check that, by inserting this transformed order parameter back into the Hamiltonian (\[EQ:HVOPHRS\]) and making a change of variables, the same Hamiltonian but for the field $\VOP '$ is recovered. Similarly, one can verify invariance under independent rotations $\hat{\Tens{O}}\equiv(\Tens{O}^{0},\Tens{O}^{1},\ldots,\Tens{O}^{n})$ with $$\begin{aligned}
\VOP_{\hat{p}} \to \VOP'_{\hat{p}} = \VOP_{\hat{\Tens{O}}^{-1}\cdot\hat{p}} \, .\end{aligned}$$
The order parameter in the liquid state (i.e., $\VOP=0$) has the full symmetry of the Hamiltonian, and at the transition to the soft random solid state it is the symmetry of *relative* translations and rotations between different replicas that is spontaneously broken. However, the symmetry of *common* translations and rotations of all replicas are preserved, and this reflects the important notion that from the macroscopic perspective the system remains translationally and rotationally invariant, even in the random solid state. This entire pattern of symmetry breaking amounts to an unfamiliar but essentially conventional example of the Landau paradigm.
This broken symmetry of relative translations and rotations between different replicas can be understood as a result of particle localization. Because a delocalized liquid particle can explore the whole volume via its thermal fluctuations, and in thermal equilibrium its positions in different replicas are uncorrelated, the liquid state is invariant, under separate translation and rotation of individual replica. On the contrary, for a localized particle, its positions in the various replicas are strongly correlated, and therefore the symmetries of relative translations and rotations are broken.
It is straightforward to verify that the form of the random-solid-state order parameter, Eq. (\[EQ:VOPAnsatzR\]), correctly implements this pattern of symmetry breaking. To see this, we can use the complete orthonormal basis in replica space defined in Section \[SEC:FTMF\], and define an alternative basis involving a replica (almost) body-diagonal unit vector $$\begin{aligned}
\BASEl \equiv \frac{1}{\sqrt{1+n\Contraction^2}} \big(\BASE^{0}+\Contraction\ReplSumOne\BASE\REPa\big) .\end{aligned}$$ Relative to $\BASEl$, we may decompose a $(1+n)d$ dimensional vector $\hat{x}$ into its longitudinal ($\lambda$) and transverse ($\tau$) components: $$\begin{aligned}
\hat{x}=\hat{x}_{\lambda}+\hat{x}_{\tau}, \quad\,
\hat{x}_{\lambda} = (\hat{x}\cdot\BASEl)\BASEl, \quad\,
\hat{x}_{\tau} = \hat{x}-\hat{x}_{\lambda} \, .\end{aligned}$$ Note that $\hat{x}_{\lambda}$ and $\hat{x}_{\tau}$ are both $(1+n)d$-dimensional vectors, but $\hat{x}_{\lambda}$ has only $d$-degrees of freedom (given by $\hat{x}\cdot\BASEl$), and $\hat{x}_{\tau}$ has only $nd$-degrees of freedom.
By this decomposition, the vector $\hat{z}=(z,\Contraction z, \Contraction z, \ldots )$, which characterizes the mean positions of a particle in the stationary-point state, can be written as $$\begin{aligned}
\hat{z} = \sqrt{1+n\Contraction^2} \, z\, \BASEl ,\end{aligned}$$ which points purely in the $\BASEl$ direction. As a result, the stationary order parameter, Eq. (\[EQ:VOPAnsatzR\]), can be written as $$\begin{aligned}
\label{EQ:OPInv}
\VOPSP(\hat{x}) &=& \LocaPart \int \frac{dz}{\Volu_0}
\int d\ISLL \DistISLL(\ISLL)
\Big(\frac{\ISLL}{2\pi}\Big)^{\frac{(1+n)\Dime}{2}} \nonumber\\
&& \quad\times e^{-\frac{\ISLL}{2} \vert \hat{x}_{\lambda}-\hat{z}_{\lambda}\vert^2
-\frac{\ISLL}{2} \vert \hat{x}_{\tau}\vert^2} - \frac{\LocaPart}{\Volu_0\Volu^n} \nonumber\\
&=& \LocaPart \int d\ISLL \DistISLL(\ISLL)
\Big(\frac{\ISLL}{2\pi}\Big)^{\frac{(1+n)\Dime}{2}}
\Big(\frac{2\pi}{\ISLL(1+n\Contraction^2)}\Big)^{\frac{d}{2}} \nonumber\\
&& \quad\times e^{-\frac{\ISLL}{2} \vert \hat{x}_{\tau}\vert^2} - \frac{\LocaPart}{\Volu_0\Volu^n} ,\end{aligned}$$ where in the last line we have integrated out the $d$-dimensional vector $z$. It is evident that this value of order parameter does not depend on $\hat{x}_{\lambda}$, which means that it is *invariant under translations in the $\BASEl$ direction, corresponding to common translations and rotations of all replicas (albeit appropriately contracted by $\Contraction$ in replicas $1$ through $n$)*. This stationary order parameter is shown schematically in Fig. \[FIG:OPab\](a) for two replicas. The Gaussian-like form in the $\hat{x}_{\tau}$ direction indicates a condensation between different replicas. This is called a *molecular bound state* in Ref. [@Goldbart2004].
### Goldstone fluctuations {#SEC:GSFT}
With the pattern of continuous symmetry breaking just outlined, we can write down the form that the order parameter takes when it is subject to Goldstone fluctuations: $$\begin{aligned}
\label{EQ:VOPGSR}
\VOPGS(\REPX)&=&\LocaPart \int \frac{dz}{\Volu_0}\int d\ISLL\, \DistISLL(\ISLL)
\Big(\frac{\ISLL}{2\pi}\Big)^{\frac{(1+n)\Dime}{2}}
\nonumber\\ \noalign{\medskip}
&& \,\times e^{-\frac{\ISLL}{2}
\{ \ReplSum\vert x\REPa-\DefoPosi\REPa(z) \vert^2 \}}
-\frac{\LocaPart}{\Volu_0 \Volu^n} ,\end{aligned}$$ where $\DefoPosi^{0}(z)=z$. Therefore the Goldstone deformation of the order parameter is parameterized by the $n$ independent functions $\{\DefoPosi^{1}(z),\ldots,\DefoPosi^{n}(z)\}$. The stationary form of the order parameter, Eq. (\[EQ:VOPAnsatzR\]), describes a system in which the mean positions of the replicas of the thermally fluctuating particles are located at $(x^0,x^1,\ldots,x^n)=(z,\Contraction z,\ldots,\Contraction z)$. We shall refer to these positions as the *centers of the thermal cloud*. By comparing the undeformed order parameter (\[EQ:VOPAnsatzR\]) and the Goldstone-deformed one, Eq. (\[EQ:VOPGSR\]), we see that the Goldstone-deformed order parameter describes a system in which the mean positions of the replicas of the fluctuating particles are displaced from $(z,\Contraction z,\ldots,\Contraction z)$ to $(z, \DefoPosi^{1}(z),\ldots,\DefoPosi^{n}(z))$. Thus, $\DefoPosi\REPa(z)$ ($\alpha=1,2,\ldots,n$) represent the deformed mean positions in the measurement replicas.
![(a) Schematic plot of the value of the order parameter (brightness) at the stationary point, for the illustrative two replicas, labeled by $x^1$ and $x^2$. (b) Schematic plot of the value of the order parameter (brightness) for a Goldstone deformation of the stationary point for two replicas.[]{data-label="FIG:OPab"}](GoldstonePlot.eps){width=".48\textwidth"}
We require that the deformations $\Contraction z \to \DefoPosi\REPa(z)$ be *pure shear* deformations. This constraint can be expressed as $\textrm{det}\big(\partial \DefoPosi_i\REPa(z)/\partial (\Contraction z)_j\big)=1$; it guarantees that the Goldstone fluctuation does not excite the LRS (i.e., each replica still has homogeneous density), which would be extremely energetically costly, owing to the large excluded-volume interaction. The 0RS has already been removed from the theory, and one can easily check that it remains zero in this Goldstone-deformed order parameter. The vanishing of the order parameter in the 1RS can be verified by taking the momentum-space Goldstone deformation and making a change of variables:
$$\begin{aligned}
(\VOPGS)_{\hat{q}}\big\vert_{\hat{q}=p\BASE\REPa}
&=& \LocaPart \int \frac{dz}{\Volu_0}\int d\ISLL \DistISLL(\ISLL)
e^{-\frac{\vert p \vert^2}{2\ISLL}
-ip \, \DefoPosi\REPa(z)}
-\LocaPart\delta^{((1+n)d)}_{\REPP,0} \nonumber\\
&=&\LocaPart \int \frac{d\DefoPosi\REPa}{\Volu} \frac{\Volu}{\Volu_0}
\Big\vert\frac{\partial z}{\partial \DefoPosi\REPa}\Big\vert
\int d\ISLL \DistISLL(\ISLL)
e^{-\frac{\vert p \vert^2}{2\ISLL}
-ip \, \DefoPosi\REPa(z)}-\LocaPart\delta^{((1+n)d)}_{\REPP,0}=0.\end{aligned}$$
This result indicates that the deformed state has the same local density as the (undeformed) stationary-point state.
![Example of a Goldstone-deformed state. The system of replicas $0$ through $n$ are shown. The mean positions of the replicas of a thermally fluctuating particle are displaced to $(z, \DefoPosi^{1}(z),\ldots,\DefoPosi^{n}(z))$ in this Goldstone-deformed state, which characterizes an $n$-fold replicated deformation field. Here, for simplicity, we only show spatially homogeneous deformations, and it is worth noticing that the volumes of the measurement replicas, i.e., replicas $1$ through $n$, are contracted by a factor $\Contraction^{\Dime}$.[]{data-label="FIG:ReplGoldstone"}](ReplGold.eps){width=".45\textwidth"}
We ought to clarify the following point about this Goldstone deformation. As we have already mentioned, the symmetry that is broken at the transition is that of relative translations and rotations of the various replicas, the symmetry of common translations and rotations remaining intact. As a result, the Goldstone deformation should be constructed via $z$-dependent translations of the order parameter in the $\hat{x}_{\tau}$ direction, i.e., *the broken symmetry direction*. However, if we look at the deformation field defined by $\hat{\GSDF}\equiv\hat{\DefoPosi}-\hat{z}$, we find that $\hat{\GSDF}=(0,\DefoPosi^{1}(z)-\Contraction z, \DefoPosi^{2}(z)-\Contraction z , \ldots)$ is in fact *not* in the broken symmetry direction $\hat{x}_{\tau}$, because it has an $\hat{x}_{\lambda}$ component, viz., $$\begin{aligned}
\hat{\GSDF}_{\lambda}=(\hat{\GSDF}\cdot\BASEl )\,\, \BASEl
= \frac{\Contraction}{\sqrt{1+n\Contraction^2}} \ReplSumOne \GSDF\REPa(z) \,\, \BASEl .\end{aligned}$$ This $\hat{\GSDF}_{\lambda}$ component is actually redundant. This can be seen by decomposing the quadratic form $\vert \hat{x}-\hat{z}-\hat{\GSDF} \vert^2$ as $$\begin{aligned}
\vert \hat{x}-\hat{z}-\hat{\GSDF} \vert^2 = \vert \hat{x}_{\lambda}- \hat{z}_{\lambda}-\hat{\GSDF}_{\lambda} \vert^2
+ \vert \hat{x}_{\tau}-\hat{\GSDF}_{\tau} \vert^2 ,\end{aligned}$$ and noting that in the form of the order parameter (\[EQ:VOPGSR\]) one can change the integration variable (which is a $d$-dimensional vector) from $z$ to $$\begin{aligned}
y\equiv z+\frac{1}{\sqrt{1+n\Contraction^2}} \,\hat{\GSDF}\cdot\BASEl ,\end{aligned}$$ so the longitudinal component of the $(1+n)\Dime$-dimensional vector $\hat{y}\equiv(y,\Contraction y,\ldots,\Contraction y)$ is $\hat{y}_{\lambda}=\hat{z}_{\lambda}+\hat{\GSDF}_{\lambda}$. The Jacobian of this change of variables is unity, provided that each deformation $z\to \GSDF\REPa(z)$ is a pure shear deformation [^6]. With this change of variables the Goldstone-deformed order parameter attains the form
$$\begin{aligned}
\label{EQ:VOPGSRTWO}
\VOPGS(\hat{x}) &=& \LocaPart \int \frac{dy}{\Volu_0} \int d\ISLL \DistISLL(\ISLL)
\Big(\frac{\ISLL}{2\pi}\Big)^{\frac{(1+n)\Dime}{2}} e^{-\frac{\ISLL}{2} \vert \hat{x}_{\lambda} - \hat{y}_{\lambda}\vert^2
-\frac{\ISLL}{2} \vert \hat{x}_{\tau} - \hat{\GSDF}'_{\tau}(y)\vert^2}
- \frac{\LocaPart}{\Volu_0\Volu^n} \nonumber\\
&=& \LocaPart \int \frac{dy}{\Volu_0} \int d\ISLL \DistISLL(\ISLL)
\Big(\frac{\ISLL}{2\pi}\Big)^{\frac{(1+n)\Dime}{2}} e^{-\frac{\ISLL}{2} \vert \hat{x} - \hat{y}-\hat{\GSDF}'_{\tau}(y) \vert^2}
- \frac{\LocaPart}{\Volu_0\Volu^n} \nonumber\\
&\equiv& \VOPGS'(\hat{y}) ,\end{aligned}$$
where the transformed deformation field $\hat{\GSDF}'$ is defined via $\hat{\GSDF}'_{\tau}(y)=\hat{\GSDF}_{\tau}(z)$. With this change of variables, the order parameter field, $\VOPGS(\hat{x})$ of $\hat{x}$, is transformed to a new field, $\VOPGS'(\hat{y})$ of $\hat{y}$, which can be viewed as being the stationary point $\VOP_{SP}(\hat{y})$ but locally translated purely in the $\hat{y}_{\tau}$ directions, because the deformation is $\hat{\DefoPosi}'(y)=\hat{y}+\hat{\GSDF}'_{\tau}(y)$. Comparing with the deformation before the change of variables, $\hat{\DefoPosi}(z)=\hat{z}+\hat{\GSDF}(z)$, it is evident that the $\GSDF_{\lambda}$ component is actually removed, and the deformation field $\hat{\GSDF}$ only affects the $\hat{x}_{\tau}$ direction. Therefore, $\GSDF_{\lambda}$ is a redundant component in this field-theoretic description of the Goldstone fluctuation. Note that in these two representations of the Goldstone fluctuation, Eqs. (\[EQ:VOPGSR\]) and (\[EQ:VOPGSRTWO\]), the number of degrees of freedom of the deformation field, $\GSDF(z)$ or $\hat{\GSDF}'_{\tau}(y)$, is $nd$, because in $\GSDF(z)$ one has the constraint $\GSDF^{0}=0$.
The reason we choose to adopt the form of Goldstone fluctuation given in Eq. (\[EQ:VOPGSR\]) is that in the true physical system that we are intending to describe, the preparation state (replica $0$) is not deformed. Although in the field theory the $1+n$ replicas feature symmetrically (apart from the contraction $\Contraction$), physically, one should only have Goldstone fluctuations that deform replicas $1$ through $n$, as these are the replicas associated with the measurement states, on which deformations are actually performed. Therefore, although Eqs. (\[EQ:VOPGSR\]) and (\[EQ:VOPGSRTWO\]) are mathematically equivalent, Eq. (\[EQ:VOPGSR\]) provides a better physical description of the Goldstone fluctuations [^7].
### Energetics of Goldstone deformations {#SEC:EnerGoldSton}
To obtain the energy of a Goldstone deformed state, we take the momentum-space version of the Goldstone-deformed order parameter, Eq. (\[EQ:VOPGSR\]), $$\begin{aligned}
\label{EQ:VOPGSM}
(\VOPGS)_{\REPP}&=&\LocaPart \int \frac{dz}{\Volu_0}\int d\ISLL \DistISLL(\ISLL)
\nonumber\\
&& \quad\times e^{-\frac{\vert\REPP\vert^2}{2\ISLL}
-i\REPP\cdot\hat{\DefoPosi}(z)}
-\LocaPart \delta^{(1+n)d}_{\REPP,0} ,\end{aligned}$$ and insert it into the Hamiltonian (\[EQ:HVOPHRS\]). After a lengthy calculation (see Appendix \[APP:HGS\]), similar to the one for the stationary-point free energy, we arrive at the energy of the Goldstone deformed state: $$\begin{aligned}
\label{EQ:HGS}
\HVOPGF= \HVOPSP+H_{\VOP}^{\DefoScalPsi}.\end{aligned}$$ Here, we use the short-hand $$\begin{aligned}
\!\!\!\!\DefoScalPsi(z_1,z_2)\!\!&\equiv&\!\!(\hat{\DefoPosi}(z_1)\!-\!\hat{\DefoPosi}(z_2))^2\!-\!(1+n)(z_1\!-\!z_2)^2 \nonumber\\
\!\!&=&\!\!\!\! \ReplSumOne\! \big((\DefoPosi\REPa(z_1)\!\!\!-\!\!\DefoPosi\REPa(z_2))^2\!-\!(z_1\!-\!z_2)^2\!\big)\end{aligned}$$ to denote the deformation. The term $\HVOPSP$ is the Hamiltonian at the stationary point, as given in Eq. (\[EQ:HSP\]), and the term $H_{\VOP}^{\DefoScalPsi}$ accounts for the increase in the energy due to Goldstone deformation, and is given by $$\begin{aligned}
\label{EQ:GibbEnerGoldSton}
\!\!\!\! H_{\VOP}^{\DefoScalPsi}
\!\!\!\!&=&\!\!\frac{1}{2}\int dz_1 dz_2 \KernOne(z_1,z_2)\DefoScalPsi(z_1,z_2)
\nonumber\\
&&\!\!-\frac{1}{8\BoltCons T}\int dz_1 dz_2 dz_3 dz_4 \KernTwo(z_1,z_2,z_3,z_4)
\nonumber\\
&& \! \times \DefoScalPsi(z_1,z_2)\DefoScalPsi(z_3,z_4)
\!-\!\!n\PartNumb \BoltCons T \frac{\UnivPara d}{2}(\Contraction^2\!-\!1\!) , \,\end{aligned}$$ [[where we have kept terms to linear order in $n$ and quadratic order in $\DefoScalPsi^2$ in this expansion.]{}]{} The functions $\KernOne(z_1,z_2)$ and $\KernTwo(z_1,z_2,z_3,z_4)$ are bell-shaped functions of the distances $(z_1-z_2)$ and $(z_1-z_2)$, $(z_2-z_3)$, $(z_3-z_4)$, as shown in Fig. \[FIG:KK\]. They are independent of the centers of mass, $(z_1+z_2)/2$ for $\KernOne(z_1,z_2)$, and $(z_1+z_2+z_3+z_4)/4$ for $\KernTwo(z_1,z_2,z_3,z_4)$. The forms of $\KernOne(z_1,z_2)$ and $\KernTwo(z_1,z_2,z_3,z_4)$ are given in Appendix \[APP:HGS\]. In specific, $\KernOne(z_1,z_2)$ has the following schematic functional form $$\begin{aligned}
\KernOne(z_1,z_2) = \int dr \, C(r) e^{-\frac{\vert z_1-z_2\vert^2}{2r^2}} ,\end{aligned}$$ which is a superposition of Gaussian distributions on the scale of $r$, where $r$ is a certain combination of localization lengths $\LocaLeng$, weighted by the distribution of those localization lengths. The kernel $\KernTwo(z_1,z_2,z_3,z_4)$ has a similar structure, except that it also contains delta-function factors such as $\delta(z_1-z_3)$, which should be associated with the scale of the short-distance cutoff of the theory, which is also on the scale of typical localization length.
![Diagrams for the functions $\KernOne(z_1,z_2)$ and $\KernTwo(z_1,z_2,z_3,z_4)$. In these diagrams, the straight lines represent delta functions, and the wavy lines represent Gaussian potentials between pairs of points, averaged over the distribution of localization lengths, as indicated schematically by the summation signs, which also indicate symmetrization over the arguments. The typical localization length provides the characteristic lengthscale for $\KernOne$ and $\KernTwo$. The full expression is listed in Appendix \[APP:HGS\].[]{data-label="FIG:KK"}](K1K2.eps){width=".4\textwidth"}
The form of the energy increase due to Goldstone fluctuations, Eq. (\[EQ:GibbEnerGoldSton\]), can be understood intuitively as follows. This energy actually describes the elastic energy of replicated shear deformations of the system, as we have explained in Section \[SEC:GSFT\] (i.e., the Goldstone fluctuations in the RLPM are replicated shear deformations). Thus, it is evident that the first term in Eq. (\[EQ:GibbEnerGoldSton\]) represents a coupling of the deformations $\hat{\DefoPosi}(z_1)$ and $\hat{\DefoPosi}(z_2)$ at the points $z_1$ and $z_2$. The coupling function, $\KernOne(z_1,z_2)$, given in Eq. (\[EQ:KoneApp\]), has a magnitude controlled by the probability $\LocaPart$ for a particle to be localized, and a lengthscale controlled by the typical localization length; these two quantities are, in turn, determined by the link density parameter $\LinkDens$, which is the control parameter for the RLPM. In particular, the first term in $\KernOne(z_1,z_2)$ carries a factor $\LocaPart^2$, which is the probability for both of the two mass-points, $z_1$ and $z_2$, to be in the infinite cluster, and therefore to have an elastic interaction between their deformations. The second term in $\KernOne(z_1,z_2)$, which involves a summation over $m$ from $2$ to $\infty$, takes into account the interactions between these two points that are mediated via other mass points. The four-point term, with coupling function $\KernTwo(z_1,z_2,z_3,z_4)$, is a bit more complicated. It couples replicas of the shear deformation to one another and to themselves, also on the scale of the localization length. These two terms in the elastic energy embody the statistics of the elastic free energy of the soft random solid in the language of replicas, and thus encode information about the statistics of the quenched random elastic parameters. In Section \[SEC:Phen\] we shall uncover this statistical content via the introduction of a phenomenological model inspired by this elastic energy.
In order to compare with the phenomenological elastic free energy that we shall discuss in Section \[SEC:Phen\], it is useful to make an alternative decomposition of the Hamiltonian to the one made in Eq. (\[EQ:HGS\]). That decomposition (\[EQ:HGS\]) was in terms of the stationary-point part and the part due to fluctuations. The alternative decomposition reads $$\begin{aligned}
\label{EQ:HSepa}
\HVOPGF= H_{\VOP}^{(0)}+H_{\VOP}^{(\DefoPosi)}.\end{aligned}$$ The relation between the two decompositions is given by $$\begin{aligned}
H_{\VOP}^{(0)} &=& \HVOPSP-\ContEner(\Contraction) , \nonumber\\
H_{\VOP}^{(\DefoPosi)} &=& H_{\VOP}^{\DefoScalPsi} \,\,\,\, + \ContEner(\Contraction) ,\end{aligned}$$ where $\ContEner(\Contraction)$ accounts for the energy of the stationary point, measured with respect to the state right after linking, which is actually the elastic energy of the contraction $\Contraction$, i.e., $$\begin{aligned}
\label{EQ:CE}
\ContEner(\Contraction)
= \frac{\ExclVolu \PartNumb^2}{2\Volu_0} \Big( \frac{\Volu}{\Volu_0}-1 \Big)^2
+\PartNumb \BoltCons T \frac{\UnivPara d}{2}(\Contraction^2-1).\end{aligned}$$ In Eq. (\[EQ:HSepa\]), we are separating the Hamiltonian into two parts. $H_{\VOP}^{(0)}$ gives energy of the state right after linking, which is a state that has not been allowed to contract after the links were made and thus has the same volume and shape as the liquid state. In the state right after linking, the mean positions of the replicas of the thermally fluctuating particle (i.e., the centers of the thermal cloud) are located at the positions $(x^0,x^1,\ldots,x^n)=(z,z,\ldots,z)$. The other part of the Hamiltonian is the elastic energy of the deformation away from this state, $H_{\VOP}^{(\DefoPosi)}$. This is different from the separation made in Eq. (\[EQ:HGS\]), in which one has the stationary point energy $\HVOPSP$ and the energy-increase due to Goldstone deformations $H_{\VOP}^{\DefoScalPsi}$.
Phenomenological approach to the elasticity of soft random solids {#SEC:Phen}
=================================================================
Phenomenological nonlocal elastic free energy {#SEC:PhenFE}
---------------------------------------------
As discussed in Section \[SEC:Intr\], in the classical theory of rubber elasticity [@Treloar1975], rubbery materials are modeled as incompressible networks of entropic Gaussian chains, and the resulting elastic free energy density is given by $$\begin{aligned}
f = \frac{\SheaModu}{2} \, \textrm{Tr} \,\, \DefoGrad^{\textrm{T}} \DefoGrad\end{aligned}$$ for spatially uniform deformations $r\to \DefoGrad \cdot r$. Incompressibility is incorporated via the constraint $\textrm{det} \DefoGrad = 1$. For the shear modulus $\SheaModu$, the classical theory gives the result $n_c \, \BoltCons \, T$, where $n_c$ is the density of effective chains in the network.
The phenomenological model that we now discuss is in the same spirit as the classical theory of rubber elasticity. However, to account for the heterogeneity of the medium we need to introduce the additional feature of quenched randomness into the model, and thus the entropic Gaussian chains are allowed to be of heterogeneous length and density. Furthermore, the classical theory is a local elasticity theory, which is valid at length scales that are much longer than the effective chain length of the polymers. By contrast, our phenomenological model is a nonlocal theory, which explicitly takes into account the finite length of polymer chains, as well as their variations.
Inspired by the form of the energy of Goldstone fluctuations determined from the RLPM in Section \[SEC:EnerGoldSton\], we choose the following elastic free energy $\FreeEnerPhen_{\NonLocaKern}$, associated with a deformation of the soft random solid state that maps the mass point at $z$ to the new location $\DefoPosi(z)$: $$\begin{aligned}
\label{EQ:phenom_model}
\FreeEnerPhen_{\NonLocaKern} \!\!&=& \! \frac{1}{2} \!\int\! dz_1\, dz_2\, \NonLocaKern (z_1,z_2)
\big(\vert \DefoPosi(z_1)-\DefoPosi(z_2) \vert^2 \! -\vert z_1-z_2 \vert^2\big) \nonumber\\
\noalign{\medskip}
&& +\frac{\BulkModuZero}{2}\int dz
\Big\lbrace\textrm{det}\Big(\frac{\partial\DefoPosi_i(z)}{\partial z_j}\Big)-1 \Big\rbrace^2 ,\end{aligned}$$ where $\NonLocaKern (z_1,z_2)$ is a *nonlocal harmonic attraction* that serves to pull the two mass points (i.e., coarse-grained volume-elements) at $z_1$ and $z_2$ towards one another. The kernel $\NonLocaKern (z_1,z_2)$ originates in the entropy of the molecular chains of the heterogeneous network, and we model it as zero rest-length springs having random spring coefficient. Notice that $\NonLocaKern (z_1,z_2)$ is a *coarse grained* consequence of many molecular chains and, more importantly, is an *entropic* effect and does not depend on the choice of precise form of microscopic attractive interactions.
We take $\NonLocaKern (z_1,z_2)$ to be a quenched random function of the two positions, $z_1$ and $z_2$, symmetric under $z_1 \leftrightarrow z_2$. We assume that the disorder average of $\NonLocaKern (z_1,z_2)$ is $\NonLocaKernZero(z_{1}-z_{2})\equiv [\NonLocaKern(z_{1},z_{2})]$, i.e., is translationally invariant. Furthermore, we define the fluctuation part of $\NonLocaKern (z_1,z_2)$ to be $\NonLocaKernOne (z_1,z_2)\equiv \NonLocaKern (z_1,z_2) - \NonLocaKernZero(z_{1}-z_{2})$. In the following analysis, we assume that $\NonLocaKernOne \ll \NonLocaKernZero$ in order to make a necessary perturbative expansion.
In the second term in Eq. (\[EQ:phenom\_model\]), the determinant of the deformation gradient tensor $\DefoGrad_{ij} (z)$\[$\equiv \partial \DefoPosi_i/\partial z_j$\] captures the change of the volume and, correspondingly, the parameter $\BulkModuZero$, which we take to be large, heavily penalizes density variations. This large $\BulkModuZero$ results from a competition between (i) repulsions (either direct or mediated via a solvent, e.g., excluded-volume), and (ii) intermolecular attractions and external pressure.
In discussion of elasticity that follows, we exploit the notions of a reference space and a target space for any deformation $\DefoPosi(z)$. The reference space, labeled by the $\Dime$-dimensional vector $z$, is the space *before* the deformation, whereas the target space, labeled by the $\Dime$-dimensional vector $\DefoPosi(z)$, is the space *after* the deformation.
Disorder average of the phenomenological model via the replica method
---------------------------------------------------------------------
To make a comparison with the RLPM, and thus to obtain information about the statistics of the nonlocal kernel $\NonLocaKern$ that characterizes the disorder present in the phenomenological model, we shall use the replica method to average the elastic free energy (\[EQ:phenom\_model\]) over the quenched disorder, whose statistics will be specified below.
We follow a recipe similar to the one used in Section \[SEC:RLPMReplica\] \[see Eq. (\[EQ:HFReplica\])\]. The elastic free energy, Eq. (\[EQ:phenom\_model\]), contains the random ingredient $\NonLocaKern$. As with the RLPM, the physical quantity to be disorder-averaged is the free energy at a given pressure, but now with the deformations $\DefoPosi(z)$ as thermally fluctuating field. Therefore, we need to take $\FreeEnerPhen$, defined in Eq. (\[EQ:phenom\_model\]), as the effective Hamiltonian, because it is the elastic energy for a given deformation field specified by $\DefoPosi(z)$, and then calculate the free energy at a given temperature, via the partition function $$\begin{aligned}
Z_{\NonLocaKern}
=\int \mathcal{D} \DefoPosi
e^{-\FreeEnerPhen_{\NonLocaKern}(\DefoPosi(z))/\BoltCons T} ,\end{aligned}$$ with $\FreeEnerPhen$ depending on the quenched randomness through its kernel $\NonLocaKern$. The Gibbs free energy is related to this partition function via $\GibbFreeEner = -\BoltCons T \ln Z$, and $\GibbFreeEner$ is the quantity that should be averaged over the quenched disorder. Note that it is the Gibbs free energy, instead of the Helmholtz free energy, that is related to this partition function $Z$, because in the elastic energy $\FreeEnerPhen_{\NonLocaKern}$ one has a fixed pressure, which is accounted for by the $\BulkModuZero$ term, the volume being allowed to fluctuated. The disorder average of the Gibbs free energy can be computed using the replica technique: $$\begin{aligned}
\lda \GibbFreeEner \rda
&=& -\BoltCons T \int \mathcal{D} \NonLocaKern\, \ProbG
\ln Z_{\NonLocaKern} \nonumber\\
&=& -\BoltCons T\int \mathcal{D} \NonLocaKern\, \ProbG \lim_{n\to 0}
\frac{Z_{\NonLocaKern}^n-1}{n} \nonumber\\
&=& -\BoltCons T \lim_{n\to 0} \frac{\partial}{\partial n}\Big\vert_{n \to 0}
\lda Z_{\NonLocaKern}^n \rda ,\end{aligned}$$ where we have used $\lda\ldots\rda\equiv
\int \mathcal{D} \NonLocaKern \, \ProbG \ldots$ once again to denote a disorder average, but this time over the values of the quenched random kernel $\NonLocaKern$, weighted by an as-yet unknown distribution $\ProbG$. In the present setting, we do not have a zeroth replica, as such a replica arises from the Deam-Edwards distribution of the links, and this is not the type of quenched disorder that we have in mind. Rather, in the present setting we regard the distribution of disorder $\ProbG$ as a physical quantity that is unknown but is to be determined through a comparison with the analysis of the Goldstone fluctuations of the RLPM. The replica partition function is then given by $$\begin{aligned}
\label{EG:PhenZn}
\mathbb{Z}_{n}&\equiv& \lda Z_{\NonLocaKern}^n \rda
= \int \mathcal{D} \NonLocaKern \,\ProbG\, Z_{\NonLocaKern}^n \nonumber\\
&=& \int \mathcal{D} \NonLocaKern \,\ProbG
\int \ReplProdOne \mathcal{D} \DefoPosi\REPa
e^{-\ReplSumOne\FreeEnerPhen_{\NonLocaKern}(\DefoPosi\REPa(z))/\BoltCons T}
\nonumber\\
&=& \int \ReplProdOne \mathcal{D} \DefoPosi\REPa\,
e^{-\FreeEnerPhen_n/\BoltCons T} ,\end{aligned}$$ in which we functionally integrate over the configurations of the $n$-fold replicated displacement fields $\DefoPosi\REPa$. We have also introduced the effective pure Hamiltonian $\FreeEnerPhen_n$ that governs the replicated deformation fields: $$\begin{aligned}
\label{EQ:PhenReplPartFunc}
\FreeEnerPhen_n &\equiv& -\BoltCons T \ln \big\lda
e^{-\ReplSumOne\FreeEnerPhen_{\NonLocaKern}(\DefoPosi\REPa(z))/\BoltCons T}
\big\rda .\end{aligned}$$ The exponential and the logarithm in Eq. (\[EQ:PhenReplPartFunc\]) can jointly be expanded in terms of cumulants, and thus we arrive at the form
$$\begin{aligned}
\FreeEnerPhen_n &=& -\BoltCons T \Big\lbrace
-\Big\lda \ReplSumOne\FreeEnerPhen_{\NonLocaKern}(\DefoPosi\REPa(z))/\BoltCons T \Big\rda_c
+\frac{1}{2} \Big\lda
\sum_{\alpha,\beta=1}^{n}
\FreeEnerPhen_{\NonLocaKern}(\DefoPosi\REPa(z))
\FreeEnerPhen_{\NonLocaKern}(\DefoPosi\REPb(z))/\BoltCons T
\Big\rda_c
- \cdots
\Big\rbrace ,\end{aligned}$$
where $\lda \ldots \rda_c$ are connected statistical moments (i.e., cumulants) associated with the probability distribution of the disorder $\ProbG$, and the omitted terms are $O[(\FreeEnerPhen/\BoltCons T)^3]$. The elastic energy $\FreeEnerPhen_{\NonLocaKern}$ for a given realization of disorder $\NonLocaKern$ and a given deformation field $\DefoPosi(z)$ is given in Eq. (\[EQ:phenom\_model\]); inserting this form for $\FreeEnerPhen_{\NonLocaKern}$ we obtain $$\begin{aligned}
\label{EQ:FEPn}
\FreeEnerPhen_n &=& \frac{\BulkModuZero}{2} \int _{z} \ReplSumOne
\big(\vert\partial \DefoPosi\REPa \vert-1\big)^2
+\frac{1}{2}\int_{z_1,z_2} \lda \NonLocaKern(z_1,z_2) \rda_c \DefoScalPsi (z_1,z_2)
\nonumber\\
\noalign{\medskip}
&& -\frac{1}{8\BoltCons T} \int_{z_1,z_2,z_3,z_4}
\lda \NonLocaKern(z_1,z_2) \NonLocaKern(z_3,z_4) \rda_c
\DefoScalPsi (z_1,z_2)\DefoScalPsi (z_3,z_4)
+ O(\DefoScalPsi^3) ,\end{aligned}$$
and we remind the reader of the definition of $\DefoScalPsi$, first given in Section \[SEC:EnerGoldSton\]: $$\begin{aligned}
\DefoScalPsi\equiv\big(\hat{\DefoPosi}(z_1)-\hat{\DefoPosi}(z_2)\big)^2-(1+n)(z_1-z_2)^2 .\end{aligned}$$ Up to quadratic order in $\DefoScalPsi$, the effective pure Hamiltonian $\FreeEnerPhen_n$ of Eq. (\[EQ:FEPn\]) has *precisely the same form* as the energy of the Goldstone fluctuations (\[EQ:GibbEnerGoldSton\]), derived microscopically from the RLPM. Thus, the RLPM actually provides a *derivation* of the phenomenological model we proposed in Section \[SEC:PhenFE\], and *justifies*, from a microscopic perspective, the phenomenological elastic free energy (\[EQ:phenom\_model\]) with its quenched randomness. Therefore, the probability distribution $\ProbG$ of the quenched randomness in Eq. (\[EQ:phenom\_model\]) is contained in the RLPM. By comparing the two schemes, i.e., Eqs. (\[EQ:GibbEnerGoldSton\]) and (\[EQ:FEPn\]), we arrive at a statistical description of the quenched random kernel $\NonLocaKern$ in the phenomenological model (\[EQ:phenom\_model\]), as we shall now show.
Comparing the Gibbs free energies of the RLPM and the phenomenological model
----------------------------------------------------------------------------
The RLPM is a semi-microscopic random network model, our analysis of which led to the disorder-averaged Gibbs free energy (\[EQ:HFEDisoAverDeri\],\[EQ:LegeTran\]): $$\begin{aligned}
\label{EQ:GRLPM}
\lda \GibbFreeEner \rda = -\BoltCons T \lim _{n \to 0}
\frac{\partial}{\partial n} \ln Z_{1+n} + \Pres\Volu ,\end{aligned}$$ with $$\begin{aligned}
\label{EQ:ZRLPMVOP}
Z_{1+n}=\int \mathcal{D} \VOP \, e^{-H_{\VOP}/\BoltCons T} .\end{aligned}$$ [[ The functional integration here is over all possible configurations of the order-parameter field $\VOP$. By contrast, in the phenomenological model the replicated partition function $\mathbb{Z}_{n}$ involves a functional integration over the $n$-fold replicated deformation field $\DefoPosi\REPa$, as in Eq. (\[EG:PhenZn\]). To obtain the equivalence between the RLPM and the phenomenological model it is useful to proceed in two steps. First, we note that in the random solid state the Boltzmann weight in Eq. (\[EQ:ZRLPMVOP\]) \[i.e., $\exp\left(-H_{\VOP}/\BoltCons T\right)$\] is heavily concentrated near the stationary point $\VOP_{SP}$ and the Goldstone-deformed states $\VOPGS$, and decreases steeply for other sectors of fluctuations. This suggests that we parametrize the fluctuating field $\VOP$ in terms of an amplitude, which we take to be $\VOP_{SP}$, plus radial fluctuations around it, together with an appropriate set of generalized angular [Goldstone]{} variables, which are the $n$ independent $\Dime$-vector fields in $\hat{\DefoPosi}$, as, e.g., in Eq. (\[EQ:VOPGSM\]). We then make an exact change of functional integration variables, from $\VOP$ to these radial and angular variables, and this introduces a corresponding Jacobian factor. The second step is to recognize that the radial fluctuations are massive (i.e., they have restoring forces, in contrast with the angular fluctuations, which are massless). Thus, if we were to integrate these radial fluctuations we would obtain small corrections to the terms of the remaining, effective, angular-variable theory. We therefore elect to treat the radial variables at the classical level, which amounts to neglecting the radial fluctuations. Under this condition, the aforementioned Jacobian factor does not depend on the angular variables, and therefore it contributes only a constant multiplicative factor to the functional integral, which can be safely omitted. This procedure enables us to arrive at the following approximate form for the replica partition function of the RLPM, Eq. (\[EQ:ZRLPMVOP\]): ]{}]{} $$\begin{aligned}
Z_{1+n}\approx e^{-\HVOPSP/\BoltCons T}
\int \ReplProdOne \mathcal{D} \DefoPosi\REPa\, e^{-H_{\VOP}^{\DefoScalPsi}/\BoltCons T},\end{aligned}$$ with $\HVOPSP$ and $H_{\VOP}^{\DefoScalPsi}$ given in Eqs. (\[EQ:HSP\]) and (\[EQ:GibbEnerGoldSton\]). In our phenomenological model, introduced in Section \[SEC:PhenFE\], the disorder-averaged Gibbs free energy is given by $$\begin{aligned}
\label{EQ:GPhen}
\lda \GibbFreeEner \rda =
-\BoltCons T \lim_{n\to 0} \frac{\partial}{\partial n}\Big\vert_{n \to 0}
\mathbb{Z}_{n}\, ,\end{aligned}$$ with $$\begin{aligned}
\mathbb{Z}_{n}=\int \ReplProdOne \mathcal{D} \DefoPosi\REPa\, e^{-\FreeEnerPhen_n/\BoltCons T},\end{aligned}$$ where $\FreeEnerPhen_n$ is given in Eq. (\[EQ:FEPn\]). The Gibbs free energies—for the RLPM and for the phenomenological model—are supposed to be equal, up to an additive constant, because they both capture the Gibbs free energy of a soft random solid system having elastic deformations. It is this equality that we shall now exploit to characterize, via the RLPM, the distribution of quenched disorder $\ProbG$ in the phenomenological model. Actually, we can directly identify the Hamiltonian $\FreeEnerPhen_n$ with $H_{\VOP}^{(\DefoPosi)}$, because the functional integration over the replicated deformation field $\DefoPosi\REPa$ is common to both the RLPM and the phenomenological model, in the sense that the deformation $z\to\DefoPosi\REPa(z)$, in both the RLPM and the phenomenological model, takes the *state right after linking* as the reference state. Therefore, we have the relation $$\begin{aligned}
\label{EQ:Comp}
\FreeEnerPhen_n = H_{\VOP}^{(\DefoPosi)} .\end{aligned}$$ Notice that, here, the RLPM Hamiltonian is $H_{\VOP}^{(\DefoPosi)}$, not $H_{\VOP}^{\DefoScalPsi}$, because it is $H_{\VOP}^{(\DefoPosi)}$ that is the energy measured from the *state right after linking*, which matches the definition of reference state in the phenomenological theory, whereas $H_{\VOP}^{\DefoScalPsi}$ is the energy measured from the *stationary point*, which differs from the *state right after linking* by the energy associated with the contraction $\ContEner(\Contraction)$, given in Eq. (\[EQ:CE\]).
By the comparison stated Eq. (\[EQ:Comp\]), we arrive at the following determination of the quenched-disorder characteristics of the phenomenological model (LHS) in terms of the elastic properties of the RLPM (RHS):
\[EQ:CompDA\] $$\begin{aligned}
\lda \NonLocaKern(z_1,z_2) \rda_c &=& \KernOne(z_1,z_2),
\label{EQ:KOne}\\
\lda \NonLocaKern(z_1,z_2) \, \NonLocaKern(z_3,z_4) \rda_c &=& \KernTwo(z_1,z_2,z_3,z_4),
\label{EQ:KTwo}\\
\BulkModuZero &=& \ExclVolu \PartDens^2 , \label{EQ:BMZ}\end{aligned}$$
where $\PartDens \equiv \PartNumb / \Volu_0$ is the number-density of the particles in the preparation state. The functions $\KernOne(z_1,z_2)$ and $\KernTwo(z_1,z_2,z_3,z_4)$ are defined in Appendix \[APP:HGS\], and have been discussed in Section \[SEC:EnerGoldSton\]
Phenomenological model at fixed disorder: Relaxation, excitation, and deformation {#SEC:relaxation}
=================================================================================
Relaxation to the stable state at fixed disorder {#SEC:subrelaxation}
------------------------------------------------
The free energy $\FreeEnerPhen$ provides a natural description of the heterogeneous elasticity of soft random solids. However, its stable state is not $\DefoPosi(z)=z$ (i.e., the state $\DefoPosi(z)=z$ does not satisfy the stationarity condition $\delta \FreeEnerPhen/\delta \DefoPosi(z)=0$). There are two reasons for this instability. First, the attraction $\NonLocaKern$ causes a small, spatially uniform, contraction \[the fractional volume change being $O(1/\BulkModuZero)$\]. Second, the randomness of $\NonLocaKern$ additionally destabilizes this contracted state, causing the adoption of a randomly deformed stable state. We denote this relaxation as $$\begin{aligned}
\label{EQ:Defiv}
z\to\tz\equiv\Contraction z + \RelaRand(z) ,\end{aligned}$$ in which $\Contraction$ describes the uniform contraction and $\RelaRand(z)$ describes the random local deformation. This relaxation process can be understood in the setting of the preparation of a sample of rubber via a hypothetic instantaneous cross-linking: cross-linking not only drives the liquid-to-random-solid transition but it also generates a uniform inward pressure, as well as introducing random stresses, as shown in Fig. \[FIG:Rela\] As a result, immediately after cross-linking the state is not stable, but relaxes to a new stable state, determined by the particular realization of randomness created by the cross-linking. In the following discussion, we shall use the following nomenclature:
$$\begin{aligned}
\DefoPosi(z)\!\!&=&\!\!z \!\iff\! \textrm{the state right after linking,} \\
\DefoPosi(z)\!\!&=&\!\!\tz\equiv\Contraction z + \RelaRand(z) \!\iff\! \textrm{the relaxed state.}\,\end{aligned}$$
The state right after linking, here, is the same as the one just defined, following Eq. (\[EQ:HSepa\]), which has energy $H_{\VOP}^{(0)}$ in the RLPM, because they both describe the state that has undergone no deformation since being linked.
![Schematic plot of the relaxation process under a fixed pressure. (a) The liquid state with no linking. (b) The *state right after linking*. cross-links are added to the system, and an infinite cluster is formed. This state is not stable, because of the inward pressure and local stresses. (c) The *relaxed state*. The system undergoes a uniform contraction and random local deformations that release the unbalanced stress introduced by cross-linking.[]{data-label="FIG:Rela"}](relaxation.eps){width=".45\textwidth"}
By writing the relaxation as $z\to\tz \,[\,\equiv\Contraction z + \RelaRand(z)]$ we are making the approximations that the contraction $\Contraction$ is homogeneous and that the random deformations $\RelaRand(z)$ are pure shear, which means that any randomness in the *bulk* deformation is ignored. This can be understood by looking at the orders of magnitude of the deformations. The uniform contraction is of order $O(\NonLocaKernZero/\BulkModuZero)$, and the random local shear deformations are of order $O(\NonLocaKernOne/\NonLocaKernZero)$. The random local bulk deformations is, however, of order $O(\NonLocaKernOne/\BulkModuZero)$, and is thus much smaller than the other two deformations, given the assumptions that (i) the fluctuations of the shear modulus are much smaller than the mean value (the shear modulus corresponds to $\NonLocaKern$, as we shall see later), and (ii) the shear modulus is much smaller than the bulk modulus.
With these assumptions, we can insert the form $\DefoPosi(z)=\tz(z)$ into the stationarity condition, and solve for the relaxed state, which is characterized by $\Contraction$ and $\RelaRand(z)$. As we have just discussed, for the contraction, only the homogeneous part is admitted, so the variational equation for $\Contraction$ assumes $\NonLocaKernOne=0$ and thus $\RelaRand(z)=0$ and, stationarity requires $$\begin{aligned}
\label{EQ:StabGlob}
\frac{\partial \FreeEnerPhen}{\partial \Contraction} = 0 .\end{aligned}$$ Thus, for the present model, Eq. (\[EQ:phenom\_model\]), we have $$\begin{aligned}
\label{EQ:StabCont}
\!\!0\!\!&=&\!\!\frac{\partial }{\partial \Contraction}
\Big\lbrace
\frac{1}{2} \int dz_1 \, dz_2 \, \NonLocaKernZero (z_1,z_2) \,
(\Contraction^2-1)\, \vert z_1-z_2 \vert ^2 \nonumber\\
&&\quad\quad +\frac{\BulkModuZero}{2}\int dz \big(\Contraction^{\Dime}-1 \big)^2
\Big\rbrace\nonumber\\ \noalign{\medskip}
&=& \Volu_0 \big(
\Dime \, \MeanSheaModu \, \Contraction
+ \BulkModuZero \, \big(\Contraction^{\Dime}-1 \big) \Dime \Contraction^{\Dime-1}
\big) .\end{aligned}$$ By solving this equation to leading order in $\MeanSheaModu/\BulkModuZero$, we obtain $$\begin{aligned}
\label{EQ:SoluCont}
\Contraction \approx 1-({\MeanSheaModu}/{d\BulkModuZero}) ,\end{aligned}$$ where $$\label{EQ:MSheaModu}
\MeanSheaModu\equiv
\frac{1}{\Dime}
\int dz_2\,(z_1-z_2)^2\,\NonLocaKernZero(z_1-z_2).$$ As we shall see below, $\MeanSheaModu$ is actually the mean shear modulus.
The stationarity condition for the random local deformation $\RelaRand(z)$ reads $$\begin{aligned}
\label{EQ:StabLoca}
\frac{\delta \FreeEnerPhen}{\delta \RelaRand_a(z)} = 0 ,\end{aligned}$$ and for the present model, Eq. (\[EQ:phenom\_model\]), this condition becomes $$\begin{aligned}
\label{EQ:RelaV}
&& 2(\Contraction z_a + \RelaRand_a(z))\int dz_2 \NonLocaKern(z,z_2) \nonumber\\
&& -2\!\int dz_2\NonLocaKern(z,z_2)\big(\Contraction z_{2,a} + \RelaRand_a(z_2)\big) \nonumber\\
&& -\BulkModuZeroP \partial_{a} \big(\partial_i \RelaRand_i(z)\big) =0 .\end{aligned}$$ Here, the last term, $\BulkModuZeroP \partial_{a} \big(\partial_i \RelaRand_i(z)\big)$, is associated with density variations, and arises from the variation of the second term in the elastic free energy (\[EQ:phenom\_model\]), which is $\frac{\BulkModuZero}{2}\int dz
\big\lbrack\textrm{det}\big(\partial\DefoPosi_i(z)/\partial z_j\big)-1 \big\rbrack^2$. In the following discussion we shall call this the bulk term in the elastic free energy, and we have made the definition $\BulkModuZeroP\equiv\Contraction^{2d-2}\BulkModuZero$; see Appendix \[APP:Relaxation\] for the expansion.
The stationarity equation (\[EQ:RelaV\]) for $\RelaRand(z)$ can be solved perturbatively, by assuming that $\NonLocaKernZero$ is of zeroth order and that $\NonLocaKernOne$ and $\RelaRand(z)$ are of first order; see Appendix \[APP:Relaxation\] for the explicit calculation. In momentum space, the result is $$\begin{aligned}
\label{EQ:SoluV}
\vec{\RelaRand}_{p}
&=& \frac{\PPerpT \cdot \vec{\RandForc}_{p}}
{2\DiffG_p}
+\frac{\PLongT \cdot \vec{\RandForc}_{p}}
{\BulkModuZeroP+ 2\DiffG_p} ,\end{aligned}$$ where $p$ is a $\Dime$-dimensional momentum vector. The notation $\vec{\RandForc}_{p}$ and $\DiffG_p$ are defined as $$\begin{aligned}
\RandForc_{a,p}&\equiv& -2\Contraction \Big(
i\frac{\partial}{\partial p_{a}} \NonLocaKernOne_{p,0}
-i\frac{\partial}{\partial p'_{a}}\Big\vert_{p'=0} \NonLocaKernOne_{p,p'}
\Big), \nonumber\\
\DiffG_p &\equiv& \NonLocaKernZero_{0}-\NonLocaKernZero_{p} .\end{aligned}$$ Notice that $\RandForc_{a,p}$ is actually the random force in the state that is contracted but not yet equilibrated for randomness. The definitions of the projection operators $\PLongT$ and $\PPerpT$ are respectively, $$\begin{aligned}
\label{EQ:DefiProjText}
\PLong_{ij} &\equiv& p_i p_i /p^2 , \nonumber\\
\PPerp_{ij} &\equiv& \delta_{ij}-p_i p_i /p^2 .\end{aligned}$$ We use bold letters to denote $\Dime$-dimensional rank-$2$ tensors, and add an overhead arrow \[such as $\vec{\RelaRand}$\] to denote vectors, when needed.
[[In the solution (\[EQ:SoluV\]), the second term is much smaller than the first term, due to the large bulk modulus $\BulkModuZeroP$. In the incompressible limit (i.e., $\BulkModuZero \to \infty$), we have $$\begin{aligned}
\vec{\RelaRand}_{p}=\frac{\PPerpT \cdot \vec{\RandForc}_{p}}
{2\DiffG_p} ,\end{aligned}$$ which is a purely transverse field, meaning that it satisfies $p_i\,\RelaRand_{i,p}=0$ or, equivalently, that $\partial_i\,\RelaRand_{i}(x)=0$, which is the only type of deformation that can occur in an incompressible medium.]{}]{}
Excitation around the relaxed state at fixed disorder {#SEC:EFERS}
-----------------------------------------------------
In order to obtain a description of the elasticity of the relaxed state $z\to\tz$, which is a stable state and thus relevant for experimental observations, we re-expand the phenomenological elastic free energy (\[EQ:phenom\_model\]) around the relaxed state. This amounts to taking the relaxed state $\tz(z)\,[\,\equiv\Contraction z+\RelaRand(z)]$ as the new reference state, and deriving the elastic free energy for deformations (i.e., excitations) relative to this state.
To do this, we study the free energy for the following elastic deformation: $$\begin{aligned}
z\to \DefoPosi(z)=\tz(z) + \deformation(z) ,\end{aligned}$$ where $\deformation(z)$ is a deformation away from the relaxed reference state. It will be convenient to make the following change of the independent variables: $$\begin{aligned}
z\to\tz(z)\equiv\Contraction z+\RelaRand(z) ,\end{aligned}$$ which has the Jacobian determinant $$\begin{aligned}
\label{EQ:zJaco}
\Jaco (z) \equiv \Big\vert \frac{\partial \tz_i}{\partial z_j}\Big\vert
\approx \Contraction^{\Dime} \big(1+\Contraction^{-1}\partial_j\RelaRand_j(z)\big).\end{aligned}$$ With this change of variables, the phenomenological elastic free energy is expressed as
$$\begin{aligned}
\label{EQ:REExpaEner}
\FreeEnerPhen &=& \FreeEnerPhen_0 +
\frac{1}{2} \int d\tz_1\, d\tz_2\, \Jaco(z_1)^{-1}\,\Jaco(z_2)^{-1}
\tNonLocaKern (\tz_1,\tz_2)\Big(
\big\vert \tz_1+\tu(\tz_1)-\tz_2-\tu(\tz_2)\big\vert^2
- \big\vert \tz_1-\tz_2\big\vert^2
\Big) \nonumber\\
\noalign{\medskip}
&&+ \frac{\BulkModuZero}{2} \int d\tz\, \Jaco(z)^{-1}
\Big\lbrace
\Jaco(z) \textrm{det} \Big( \frac{\partial \tz_i+\tu_i(\tz)}{\partial \tz_j}
\Big)-1
\Big\rbrace^2 ,\end{aligned}$$
where we have made the definitions
$$\begin{aligned}
\label{EQ:tGtu}
\tNonLocaKern(\tz_1,\tz_2)&\equiv&\NonLocaKern(z(\tz_1),z(\tz_2)), \\
\tu(\tz)&\equiv&\deformation(z(\tz)), \\
\tDefoPosi(\tz)&\equiv& \tz+\tu(\tz),\end{aligned}$$
with $z(\tz)$ denoting the mapping of the mass point $\tz$ in relaxed state back to the mass point $z$ in the state right after linking, i.e., the inverse of the $\tz(z)$ mapping. The change of the free energy due to choosing a different reference state is defined as $$\begin{aligned}
\!\!\FreeEnerPhen_0\!\equiv\frac{1}{2} \!\int\!\! dz_1\, dz_2 \,
\NonLocaKern (z_1,z_2)\big(\vert \tz_1-\tz_2\vert^2 \!\!-\! \vert z_1-z_2\vert^2 \big) ,\end{aligned}$$ which is a constant for any given realization of the randomness.
In order to obtain a direct description of the elastic energy relative to the relaxed state, we expand the quenched random nonlocal kernel in the relaxed state, $\tNonLocaKern(\tz_1,\tz_2)$, defined in Eq. (\[EQ:tGtu\]) as $$\begin{aligned}
\label{EQ:tNonLocaKernText}
\!\!\!\! \tNonLocaKern_{\tp_1,\tp_2} \!\!
&\approx& \!\! \NonLocaKernZero_{\tp_1,\tp_2}+\NonLocaKernOne_{\tp_1,\tp_2}
\nonumber\\
&&\!\! -i \big(
\tp_1\cdot\!\vec{\RelaRand}_{(\tp_1+\tp_2)} \NonLocaKernZero_{\tp_2}
\!+\!\tp_2\cdot\!\vec{\RelaRand}_{(\tp_1+\tp_2)} \NonLocaKernZero_{\tp_1}
\big),\,\end{aligned}$$ where $\vec{\RelaRand}$ is the random local deformation field defined in Eq. (\[EQ:Defiv\]). We then expand the elastic free energy (\[EQ:REExpaEner\]) as a power series in the small deformation $\tu(\tz)$ away from the relaxed state $\tz$. The computation of this expansion is given in Appendix \[APP:ReExpandFreeEner\].
As we shall show in Section \[SEC:Heterogeneity\], the statistics of the quenched randomness present in this phenomenological theory can be determined via a comparison with the RLPM. Through this comparison, we find that, the lengthscale of the nonlocal kernel $\NonLocaKern$ is actually the typical localization length, which is small compared to the lengthscales on which our theory of elasticity applies, because the deformations in this theory are associated with Goldstone fluctuations in the RLPM, which feature lengthscales larger than the typical localization length.
Thus, it is reasonable to make a *local expansion* of the elastic energy (\[EQ:REExpaEner\]) relative to the relaxed state in terms of the strain tensor $\StraTensT$. The resulting form, which we shall call the local form of the elastic energy relative to the relaxed state,is in the form of Lagrangian elasticity. As will be seen in Section \[SEC:Corr\], the advantage of this local form of the elastic energy is that one can extract from it *large-distance behavior* of the disorder correlators of the elastic parameters, which turn out to be *universal*. The calculation for this local expansion is given in Appendix \[APP:ReExpandFreeEner\]. The resulting local form of elastic energy is $$\begin{aligned}
\label{EQ:FELagr}
\FreeEnerPhen
&=& \int d\tz \big\lbrace
\textrm{Tr}(\StreT(\tz)\cdot\tStraTensT(\tz)) \nonumber\\
&& + \SheaModu(\tz) \textrm{Tr} \tStraTensT (\tz)^2
+ \frac{\BulkModu(\tz)}{2} (\textrm{Tr} \tStraTensT (\tz))^2
\big\rbrace ,\end{aligned}$$ where the strain tensor relative to the relaxed state is defined as $$\begin{aligned}
\tStraTens_{ij} (\tz) \equiv \frac{1}{2}\Big(\frac{\partial \tu_j}{\partial \tz_i}+\frac{\partial \tu_i}{\partial \tz_j} +\frac{\partial \tu_l}{\partial \tz_i}\frac{\partial \tu_l}{\partial \tz_j}\Big) ,\end{aligned}$$ and the heterogeneous elastic parameters, viz., the residual stress $\StreT$, the shear modulus $\SheaModu$, and the bulk modulus $\BulkModu$, are given in momentum space, by
\[EQ:RelaStre\] $$\begin{aligned}
\Stre_{ij,\tp}
&=& -\frac{\partial^2}{\partial \tq_i \partial \tq_j}
\Big\vert_{q=0} \NonLocaKernOne_{\tp-\tq,\tp}
+ i \delta_{ij} \frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\vert \tp \vert ^2}\label{EQ:Stress}\\
&& \quad -\frac{\RandForc_{a,\tp}}{\vert \tp \vert ^2}\big(
\tp_{i} \PPerp_{ja,\tp} + \tp_{j} \PPerp_{ia,\tp}
\big) , \nonumber\\
\SheaModu_{\tp}
&=& \MeanSheaModu \Volu_0 \delta_{\tp} -
\frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\vert \tp \vert ^2} , \label{EQ:Shear}\\
\BulkModu_{\tp}
&=& \BulkModuZero \Volu_0 \delta_{\tp} + 2\Big\lbrace
\frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\vert \tp \vert ^2}
-\MeanSheaModu \Volu_0 \delta_{\tp}
\Big\rbrace . \label{EQ:Bulk}\end{aligned}$$ \[Eq:ThreeResults\]
In the expression for $\Stre_{ij,\tp}$ we have kept terms only to leading order in the momentum $\tp$ (see Appendix \[APP:ReExpandFreeEner\] for the derivation) [^8]. It is worth mentioning that, to leading order in the momentum $\tp$, this residual stress satisfies the stability condition $\tp_i\,\Stre_{ij,\tp}=0$, because the reference state of this elastic free energy, the relaxed state, is a stable state. This will also be shown more directly in the final results in Section \[SEC:Corr\].
Nonaffine deformation at fixed disorder {#SEC:NAD}
---------------------------------------
Because of the quenched disorder present in the elastic parameter $\NonLocaKern$ of our phenomenological model, Eq. (\[EQ:phenom\_model\]), upon the application of external stress, the system will respond by adopting a strain field that is nonaffine. This means that the strain tensor will be spatially inhomogeneous even though the applied stress is homogeneous. Such nonaffine deformations reflect the quenched randomness of the elasticity, and can be derived for a given realization of the disorder and a given macroscopic deformation by external stress. Because the deformation is the quantity that is directly measurable in experiments, it is useful to derive the relationship between the nonaffine deformation and the quenched randomness in the elastic parameters. Then, by comparing with the RLPM, we shall obtain a statistical description of the nonaffine deformations, as we shall discuss in Section \[SEC:SNAD\].
To study nonaffine deformations, it is convenient to take the state right after linking \[i.e., the state $\DefoPosi(z)=z$\] as the reference state, and to re-derive the relaxation in the presence of *a given deformation* $\DefoGrad$. This is equivalent to applying the deformation $\DefoGrad$ to the relaxed state, and then letting the system further relax for this given deformation, as shown in Fig. \[FIG:RelaDefo\]. The relaxed state for this given deformation $\DefoGrad$, which we term the relaxed deformed state, is described by the deformation $z\to\tzL (z) $. We suppose that $$\begin{aligned}
\label{EQ:tzl}
\tzL (z) = \Contraction \DefoGrad \cdot z + \RelaRandL(z) .\end{aligned}$$ For simplicity, we assume that the deformation $\DefoGrad$ is pure shear (i.e., $\textrm{det}\,\DefoGrad=1$).
![Illustration of the *relaxed deformed state*. Theoretically, the relaxed deformed state can be reached via two routes: in the first, the upper route in this plot, one applies the deformation $\DefoGrad$ on the already-relaxed state, and lets the system re-relax while keeping the external deformation $\DefoGrad$ to the relaxed deformed state; the second route, the lower route in this plot, one deforms the system before relaxation is allowed, and then lets the system relax while maintaining the deformation $\DefoGrad$. These two routes reach the same final state, the relaxed deformed state, which characterizes the *nonaffine deformations* that the system undergoes under external deformation. For convenience of calculation, we use the lower route to determine the nonaffine deformations.[]{data-label="FIG:RelaDefo"}](deformation.eps){width=".45\textwidth"}
Next, we use the two stationarity conditions, Eqs. (\[EQ:StabGlob\]) and (\[EQ:StabLoca\]) to solve for the relaxed deformed state. The condition (\[EQ:StabCont\]) for the homogeneous contraction $\Contraction$ is unchanged, so we still obtain $\Contraction \approx 1-({\MeanSheaModu}/{d\BulkModuZero})$. For the stability condition for the random local deformations $\RelaRandL$ we follow a similar expansion to the one given in Eqs. (\[EQ:DetPPExp\]) and (\[EQ:BulkTermExp\]), arriving at $$\begin{aligned}
&& 2(\Contraction \DefoGrad_{ai}z_i+(\RelaRandL)_{a}(z)) \int dz_2\, \NonLocaKern (z,z_2) \nonumber\\
&& \quad -2 \int dz_2\, \NonLocaKern (z,z_2)
\,(\Contraction \DefoGrad_{ai}z_{2,i}+(\RelaRandL)_{a}(z_2)\,) \nonumber\\
&& \quad - \BulkModuZeroP \,
\DefoGrad^{-1}_{ia}\DefoGrad^{-1}_{jb}\,\partial_i \partial_j (\RelaRandL)_{b}(z) =0 .\end{aligned}$$ As with the derivation given in Section \[SEC:relaxation\], we can solve this equation perturbatively, to leading order in $\NonLocaKernOne$ and $\RelaRandL$; see Appendix \[APP:DefoRela\] for details. The result is $$\begin{aligned}
(\vec{\RelaRandL})_p = \Bigg\lbrace \frac{\PPerpL}{2 D_p} + \frac{\PLongL}{\BulkModuZeroP \trOne \vert p \vert ^2+2 D_p}
\Bigg\rbrace \cdot (\vec{\RandForcL})_p \, ,\end{aligned}$$ where $\PPerpL$ and $\PLongL$, defined in Appendix \[APP:DefoRela\], are the *deformed* versions of the projection operators, and $\trOne$ is also defined in Appendix \[APP:DefoRela\].
In the literature the nonaffine deformations are often characterized by the nonaffine deformation field $\NADF$, which is defined in momentum space as $$\begin{aligned}
\label{EQ:DefNADF}
(\NADF) _p &\equiv& \DefoGrad^{-1} \cdot (\tzL)_p - \tz_p \, , \nonumber\\
&=& \DefoGrad^{-1} \cdot (\RelaRandL)_p - \RelaRand_p ,\end{aligned}$$ where $\tz$ denotes the relaxed state of the undeformed system (as discussed in Section \[SEC:relaxation\]), and $\tzL$ is the relaxed deformed state.
Inserting the solution for $\RelaRandL$ into the expression for nonaffine deformation field, Eq. (\[EQ:DefNADF\]), we have $$\begin{aligned}
\label{EQ:NADFp}
(\NADF) _p &=& 2i\Contraction \Bigg\lbrace
\frac{\BulkModuZeroP \vert p \vert ^2 }{(\BulkModuZeroP \vert p \vert ^2 \trOne + 2D_p)2D_p} \MetrTens^{-1}
\nonumber\\
&& \quad - \frac{\BulkModuZeroP \vert p \vert ^2 }{(\BulkModuZeroP \vert p \vert ^2 + 2D_p)2D_p} \IdenT
\Bigg\rbrace
\cdot \PLongT \cdot S_p \, ,\end{aligned}$$ where $$\begin{aligned}
S_p \equiv \frac{\partial}{\partial p_{1,a}} \NonLocaKernOne_{p_1,0}
-\frac{\partial}{\partial p_{2,a}}\Big\vert_{p_2=0} \NonLocaKernOne_{p_1,p_2} ,\end{aligned}$$ and $$\begin{aligned}
\MetrTens \equiv \DefoGrad\Tran\DefoGrad\end{aligned}$$ In the incompressible limit, we have $$\begin{aligned}
\label{EQ:NADFpInco}
(\NADF) _p &\approx& 2i\Big\lbrace
\frac{1 }{2D_p \trOne} \MetrTens^{-1}
- \frac{1 }{2D_p} \IdenT
\Big\rbrace
\cdot \PLongT \cdot S_p \,\, .\end{aligned}$$ In Section \[SEC:SNAD\] we shall compute the mean value and disorder correlator of this nonaffine deformation field.
Characterizing the physical elastic quenched disorder {#SEC:Heterogeneity}
=====================================================
In Section \[SEC:EFERS\] we addressed three elastic parameter fields—the residual stress $\StreT$ and the Lamé coefficients $\SheaModu_{\tp}$ and $\BulkModu_{\tp}$—which characterize the elastic energy relative to the *relaxed state* and are therefore the physically relevant parameters for describing spatially heterogeneous elasticity. There, we showed how these parameters are determined, for a given realization of the quenched disorder, by $\NonLocaKern$, i.e., the random nonlocal kernel of the phenomenological model, giving the connection in Eqs. (\[EQ:RelaStre\]). In Section \[SEC:Phen\] we obtained a statistical characterization of $\NonLocaKern$ via a comparison with the semi-microscopic RLPM, giving the results in Eqs. (\[EQ:CompDA\]). Thus, we have the ingredients for constructing a statistical characterization of the physical position-dependent elastic parameters, as we do in the present section.
Disorder averages of the elastic parameters {#SEC:DAEP}
-------------------------------------------
Our first step is to determine the disorder average of the nonlocal kernel in the relaxed state $\tNonLocaKern$. To do this, we note that $\tNonLocaKern$ is related to $\NonLocaKern$ via Eq. (\[EQ:tGtu\]); the leading-order expansion of this relationship is given in Eq. (\[EQ:tNonLocaKernText\]). By taking the disorder average on both sides of Eq. (\[EQ:tNonLocaKernText\]), we find that only the first term on the RHS survives, because all other terms are linear in the fluctuation part of $\NonLocaKern$ and this vanishes upon disorder–averaging. Thus, we find that the disorder average of $\tNonLocaKern$ is given by $$\begin{aligned}
\lda \tNonLocaKern(z_1,z_2) \rda = \lda \NonLocaKern(z_1,z_2) \rda = \KernOne(z_1,z_2) ,\end{aligned}$$ which means that the disorder average of $\tNonLocaKern$ is the same as the disorder average of $\NonLocaKern$, where we have dropped the tilde on $\tz$ because now we discuss elastic parameters in the relaxed reference state only. It is worth noting that, as expected, because $\KernOne(z_1,z_2)$ is independent of the center of mass coordinate $(z_1+z_2)/2$, the disorder average of the nonlocal kernel $\lda \tNonLocaKern(z_1,z_2) \rda$ is translationally and rotationally invariant, depending only on $\vert z_1-z_2\vert$. This is a consequence of the macroscopic translational and rotational invariance of the random solid state discussed in Section \[SEC:SSB\].
Second, we determine the disorder averages of the position-dependent elastic parameters in the local form of the elastic energy relative to the relaxed state, including the residual stress $\StreT$, the shear modulus $\SheaModu$, and the bulk modulus $\BulkModu$. For any given realization of the disorder, these elastic parameters are related to $\NonLocaKern$ and $\BulkModuZero$ via Eqs. (\[EQ:RelaStre\]). Thus, as with the nonlocal kernel, we obtain the disorder averages of these elastic parameters via the statistics of $\NonLocaKern$.
The disorder average of the residual stress $\StreT$ is straightforwardly seen to vanish: $$\begin{aligned}
\lda \Stre_{ij}(z) \rda =0 .\end{aligned}$$ Thus, the residual stress is a quenched random field with zero mean. As for the shear modulus $\SheaModu$, its disorder average is given by $$\begin{aligned}
\lda \SheaModu(z) \rda &=& \MeanSheaModu
= -\frac{1}{\Dime} \int dz_2 \NonLocaKernZero(z-z_2) \vert z-z_2\vert^2 \nonumber\\
&=& \PartDens \BoltCons T \, \UnivPara ,\end{aligned}$$ with $\UnivPara$ given in Eq. (\[EQ:theta\]). This has been obtained in Ref. [@Ulrich2006]. This mean shear modulus is linear in temperature $T$, reflecting its entropic nature, a result that confirms this aspect of the classical theory of rubber elasticity. As for the disorder average of bulk modulus $\BulkModu$, it is obtained via Eq. (\[EQ:BMZ\]), which gives $$\begin{aligned}
\lda \BulkModu(z) \rda = \ExclVolu \PartDens^2 \, .\end{aligned}$$ As one might expect, the mean bulk modulus depends on the particle number density $\PartDens$ and the strength of the excluded-volume interaction $\ExclVolu$. The disorder average of the these three elastic parameters of the local form of the elastic energy relative to the relaxed state (viz., $\lda \Stre_{ij}(z) \rda$, $\lda \SheaModu(z) \rda$ and $\lda \BulkModu(z) \rda$) are all spatially homogeneous and isotropic; this is also a consequence of the macroscopic translational and rotational invariance of the random solid state discussed in Section \[SEC:SSB\].
Disorder correlators of the elastic parameters {#SEC:Corr}
----------------------------------------------
### Disorder correlator of the nonlocal kernel {#SEC:CorrNK}
The nonlocal kernel $\tNonLocaKern$ characterizes the quenched random nonlocal interactions in the relaxed state. Its statistics can be described via its moments. In Section \[SEC:DAEP\], we already determined the disorder average of $\tNonLocaKern$; in the present section we determine the disorder correlator of $\tNonLocaKern$.
To do this, we use Eq. (\[EQ:tNonLocaKernText\]), which relates $\tNonLocaKern$ to any given random configuration of $\NonLocaKern$. Using the disorder correlator of $\NonLocaKern$, Eq. (\[EQ:KTwo\]), we then arrive at the disorder correlator of $\tNonLocaKern$, viz., $\lda\tNonLocaKern(z_1,z_2)\tNonLocaKern(z_3,z_4)\rda$. This is a combination of Gaussian and delta-function factors in the separations of the six pairs formed by the four points $\{z_1,z_2,z_3,z_4\}$. The derivation and the momentum-space expression of the result for this disorder correlator is given in Appendix \[APP:MM\].
To reveal the universal characteristics of the disorder correlator $\lda\tNonLocaKern(z_1,z_2)\tNonLocaKern(z_3,z_4)\rda$, we investigate its large-distance behavior. The nonlocal kernel itself describes a short-distance attractive interaction, because the disorder average $\lda\tNonLocaKern(z_1,z_2)\rda$ is a short-ranged function in $\vert z_1-z_2\vert$ characterized by the typical localization length. Thus, to extract the long-distance behavior of $\lda\tNonLocaKern(z_1,z_2)\tNonLocaKern(z_3,z_4)\rda$, we take the limit that the two pairs $\{z_1,z_2\}$ and $\{z_3,z_4\}$ are far apart from one another, but $z_1$ is near $z_2$, and $z_3$ is near $z_4$. This is precisely the construction of the local description of elasticity of the relaxed state introduced in Section \[SEC:EFERS\], featuring the quenched random residual stress and the Lamé coefficients fields. We shall now discuss the disorder correlators of these elastic parameters in the local description of elasticity.
### Disorder correlators of the elastic parameters in the local form of the elastic energy {#SEC:Local}
The elastic parameters in the local form of the elastic energy, including the residual stress $\StreT$, the shear modulus $\SheaModu$, and the bulk modulus $\BulkModu$, are related to any given configuration $\NonLocaKern$ via Eq. (\[EQ:RelaStre\]). Using these relations and the disorder correlator of $\NonLocaKern$, Eq. (\[EQ:KTwo\]), we arrive at the disorder correlators of the elastic parameters. The details of this calculation are given in Appendix \[APP:CFLL\]; we summarize the results in Table \[TABLE:CorrTable\].
[|r|c c c|]{} &$\myshift\Stre_{kl,p^{\prime}}\myshift$
------------------------------------------------------------------------
&$\myshift\SheaModu_{p^{\prime}}\myshift$& $\myshift\BulkModu_{p^{\prime}}\myshift$\
$\myshift \Stre_{ij,p}\myshift$
------------------------------------------------------------------------
&$\myshift \UnivPara A_{ijkl}\myshift$ &$\myshift -2\UnivPara\PPerp_{ij}\myshift$ &$\myshift\phantom{-}4\UnivPara\PPerp_{ij}\myshift$\
$\myshift \SheaModu_{p}\myshift$
------------------------------------------------------------------------
&$\myshift -2\UnivPara\PPerp_{kl}\myshift$ &$\myshift\phantom{-2} \Corr\myshift$ &$\myshift -2 \Corr\myshift$\
$\myshift \BulkModu_{p}\myshift$
------------------------------------------------------------------------
&$\myshift\phantom{-}4\UnivPara\PPerp_{kl}\myshift$ &$\myshift -2 \Corr\myshift$ &$\myshift\phantom{-}4 \Corr\myshift\myshift$\
The correlation function $\lda\StreT\StreT\rda$ features the tensor $A_{ijkl}$, which is defined as $$\begin{aligned}
A_{ijkl} \equiv 2 \PPerp_{ij} \PPerp_{kl} + \PPerp_{ik} \PPerp_{jl} + \PPerp_{il} \PPerp_{jk}\, .\end{aligned}$$ where the projection operator $\PPerpT$ is defined in Section \[SEC:relaxation\]. The stability condition on the residual stress field $\StreT$ requires that its Fourier transform vanishes when contracted with the momentum $p$. It is straightforward to see that this feature is obeyed by the correlation function $\lda\StreT\StreT\rda$ given in Table \[TABLE:CorrTable\], owing to the structure of $A$.
The parameters $\UnivPara$ and $\Corr$, on which the correlators in Table \[TABLE:CorrTable\] depend, are given by
$$\begin{aligned}
\UnivPara \!&\equiv&\! -\frac{\LinkDens \LocaPart^2}{2}+\LinkDens \LocaPart
-1+e^{-\LinkDens \LocaPart} , \\
\Corr \!&\equiv&\!
-\frac{3}{2}\LinkDens \LocaPart^2 \!+(\LinkDens \LocaPart)^2 +\LinkDens \LocaPart -1 + e^{\LinkDens \LocaPart} .\end{aligned}$$
![Plot of $\LocaPart$, $\UnivPara$ and $\Corr$ as functions of the links density parameter $\LinkDens$.[]{data-label="FIG:ThetaGamma"}](theta.eps){width=".45\textwidth"}
The dependence of $\UnivPara$ and $\Corr$ on the density of link $\LinkDens$ is shown in Fig. \[FIG:ThetaGamma\]. The asymptotic behaviors of $\UnivPara$ and $\Corr$ are as follows: $$\begin{aligned}
\UnivPara =
\left\{
\begin{array}{ll}
\frac{2}{3} (\LinkDens-1)^3 , & \textrm{for} \quad\LinkDens\gtrsim 1 ; \\
\LinkDens/2 , & \textrm{for} \quad\LinkDens\gg 1 ;
\end{array}
\right.\end{aligned}$$ $$\begin{aligned}
\Corr =
\left\{
\begin{array}{ll}
\frac{14}{3} (\LinkDens-1)^3 , & \textrm{for} \quad\LinkDens\gtrsim 1 ; \\
\eta^4 , & \textrm{for} \quad\LinkDens\gg 1 ,
\end{array}
\right.\end{aligned}$$ where $\LinkDens$ is equal to the mean coordination number of the particles. Although the connected disorder correlators of the elastic parameters increase with the density of links $\LinkDens$, it is worth noting that the *relative fluctuations* are decreasing functions of $\LinkDens$. For example, the relative fluctuation in the shear modulus, defined as $\lda \SheaModu \SheaModu\rda_c / \lda \SheaModu \rda^2$, scales as $$\begin{aligned}
\frac{\lda \SheaModu \SheaModu\rda_c}{ (\lda \SheaModu \rda)^2}
\sim \frac{\Corr}{\UnivPara^2} ,\end{aligned}$$ which is shown in Fig. \[FIG:RelaFluc\]. This is a decreasing function of $\LinkDens$, which means that the relative fluctuations of the shear modulus actually decrease as links are added.
![Plot of the function $\phi(\LinkDens)=\frac{\Corr}{\UnivPara^2}$, which characterizes the relative fluctuations of both the shear and bulk moduli, and also the connected correlator of the nonaffine deformations, as will be shown in Section \[SEC:SNAD\].[]{data-label="FIG:RelaFluc"}](phi.eps){width=".45\textwidth"}
It is interesting to look at the real-space behavior of disorder correlators of the elastic parameters in the local form of elastic energy. First, it is easy to see that the disorder correlators $\lda \SheaModu(0)\,\SheaModu(r) \rda$, $\lda \BulkModu(0)\,\BulkModu(r) \rda$ and $\lda \SheaModu(0)\,\BulkModu(r) \rda$ are short-ranged in real space: more precisely, they are proportional to $\delta_s(r)$, i.e., to a Dirac delta-function that has been smoothed on the scale of the short-distance cutoff. This cutoff should be taken to be the typical localization length, in order to validate the Goldstone-fluctuation framework for elastic deformations, because the Goldstone fluctuations in the RLPM are long-wavelength, low-energy excitations of the random solid state, and these do not touch lengthscales shorter than the typical localization length.
By contrast, entities involving the residual stress have a more interesting spatial correlations: in three dimensions and at large lengthscales we find that
$$\begin{aligned}
\!\!\!\!\!\!\lda \Stre_{ij}(0) \, \Stre_{kl}(\vec{r})\rda_c
\!\!\!&=&\!\! \frac{(\BoltCons T)^2 \PartDens \UnivPara}{\pi \vert \vec{r}\vert^3} B_{ijkl}\, , \\
\!\!\!\!\!\!\lda \Stre_{ij}(0) \, \SheaModu(\vec{r})\rda_c
\!\!\!&=&\!\!\! - \frac{(\BoltCons T)^2 \PartDens \UnivPara}{\pi \vert \vec{r}\vert^3}
\Big(\!\PLong_{ij}(\vec{r})\!-\!\PPerp_{ij}(\vec{r})\!\Big), \,\, \\
\!\!\!\!\!\!\lda \Stre_{ij}(0) \, \BulkModu(\vec{r})\rda_c
\!\!\!&=&\!\! \!\frac{2(\BoltCons T)^2 \PartDens \UnivPara}{\pi \vert \vec{r}\vert^3}
\Big(\!\PLong_{ij}(\vec{r})\!-\!\PPerp_{ij}(\vec{r})\!\Big) , \,\end{aligned}$$
where $\PLong_{ij}(\vec{r})$ and $\PPerp_{ij}(\vec{r})$ are, respectively, longitudinal and transverse projection operators in real space, which are given by $$\begin{aligned}
\PLong_{ij}(\vec{r})\equiv \frac{r_i r_j}{\vert \vec{r}\vert^2}, \quad\quad
\PPerp_{ij}(\vec{r})\equiv \delta_{ij}-\PLong_{ij}(\vec{r}) .\end{aligned}$$ The tensor $B_{ijkl}$ has a complicated structure comprising terms built from projection operators of $\vec{r}$, together with various index combinations, and also depends on the large-momentum cutoff, which can be identified with the inverse of the typical localization length [^9] .
[[ As the above results show, the mean shear modulus and the long-distance behavior of the disorder correlators depend only on the link density $\LinkDens$, the temperature $T$, and the particle density $\PartDens$, and do not depend on the details of the link potential. This verifies the argument that the shear rigidity of the RLPM is a result of the network entropy.]{}]{}
Statistics of nonaffine deformations {#SEC:SNAD}
------------------------------------
In this section we develop a statistical characterization of the nonaffine deformations of the soft random solid state. In Section \[SEC:NAD\] we discussed why soft random solids undergo nonaffine deformations in presence of a given shear deformation $\DefoGrad$, and explained how to characterize these deformations in terms of the nonaffine deformation field $\NADF$.
The nonaffine deformation field is related to any given nonlocal random kernel $\NonLocaKern$ via Eq. (\[EQ:NADFp\]). It is straightforward to see that the disorder average of the nonaffine deformation field vanishes, i.e., $\lda \NADF \rda =0$: it is proportional to the fluctuation part of the quenched random nonlocal kernel $\NonLocaKernOne$.
Next, we calculate the disorder correlator of the nonaffine deformations. For convenience, we take the incompressible limit, i.e., $\BulkModuZero\to\infty$, in which limit the nonaffine deformation field $\NADF$ is given by Eq. (\[EQ:NADFpInco\]). Using Eq. (\[EQ:NADFpInco\]), as well as the disorder correlator of the nonlocal kernel $\lda\NonLocaKern\NonLocaKern\rda_c$, the disorder correlator of the nonaffine deformation field $\NADF$ is found to be $$\begin{aligned}
\label{EQ:CorrNADF}
\lda (\NADF)_p \cdot (\NADF)_{-p} \rda_c
&=& \frac{1}{\vert \vec{p} \vert^2} \frac{1}{\PartDens} \frac{\Corr}{\UnivPara^2}
\Big(\frac{\trTwo}{\trOne^2}-1\Big) ,\end{aligned}$$ where
$$\begin{aligned}
\trOne &\equiv& \textrm{Tr}(\PLongT \MetrTens^{-1} ) , \\
\trTwo &\equiv& \textrm{Tr}(\PLongT \MetrTens^{-1} \MetrTens^{-1}) , \\
\MetrTens &\equiv& \DefoGrad\Tran\DefoGrad .\end{aligned}$$
The dependence of the connected disorder-correlator of the nonaffine deformation field $\lda \NADF \NADF \rda_c$ on the density of links comes through the factor $\phi\equiv \Corr/(\UnivPara^2)$, which is shown in Fig. \[FIG:RelaFluc\]. It is evident that, as the density of links increases, the system has smaller relative fluctuations of its elasticity, i.e., the relative fluctuations in the elastic moduli decrease, and thus the nonaffine deformations also decrease, corresponding to the system becoming *less heterogeneous*.
The disorder correlator of the nonaffine deformation field, Eq. (\[EQ:CorrNADF\]), is consistent with the disorder correlator of the nonaffine deformations given in Ref. [@DiDonna2005]. In Eq. (3.22) of Ref. [@DiDonna2005], the disorder correlator of the nonaffine deformation $u^{\prime}$ (which corresponds to $\RelaRandL$ in our notation) was found to depend on the random local elastic modulus $K_{ijkl}$ in momentum-space as $$\begin{aligned}
\lda u^{\prime}(q) \, u^{\prime}(-q)\rda \propto \frac{\gamma^2}{q^2} \frac{\Delta^{K}(q)}{K^2} ,\end{aligned}$$ where $\gamma$ represents the appropriate components of the tensorial externally applied deformation (i.e., $\DefoGrad$ in our notation), $\Delta^{K}$ represents components of the variance of the elastic-modulus tensor, and $K$ represents components of the average of the elastic-modulus tensor. Consistency with Ref. [@DiDonna2005] is revealed by taking Eq. (\[EQ:CorrNADF\]), and recalling that $\lda\SheaModu\rda\sim\UnivPara$ and $\lda\SheaModu\SheaModu\rda_c\sim\Corr$, and therefore that our disorder correlator of the nonaffine deformation field can be written as $$\begin{aligned}
\lda (\NADF)_p (\NADF)_{-p} \rda_c \propto \frac{1}{\vert p \vert^2}
\frac{\lda\SheaModu\SheaModu\rda_c}{\lda\SheaModu\rda^2} ,\end{aligned}$$ which exhibits the same dependence on the mean and variance of the quenched random elastic modulus as Eq. (3.22) of Ref. [@DiDonna2005] does.
By transforming the disorder correlator of the nonaffine deformation field (\[EQ:CorrNADF\]) back to real space, we find that the large-distance behavior of the disorder correlator of the nonaffine deformation field $\lda \NADF(0) \cdot \NADF(r) \rda_c$ is proportional to $\vert r\vert^{-1}$ in three dimensions, which is also long-ranged.
Concluding remarks {#SEC:ConcDisc}
==================
The heterogeneous elasticity of soft random solids has been investigated via a semi-microscopic approach. By starting with the Randomly Linked Particle Model (RLPM), which describes networks of particles randomly connected by soft links, and applying the concepts and techniques of vulcanization theory, we have established a field-theoretic description of the liquid-to-random-solid transition, and have analyzed the corresponding pattern of spontaneous symmetry breaking and the structure of the associated Goldstone fluctuations. We have identified these Goldstone fluctuations as being related to shear deformations of the random solid state and, via this identification, we have obtained a statistical characterization of the quenched randomness exhibited by the heterogeneous elasticity of soft random solids, which features a random nonlocal kernel describing attractive interactions between mass-points.
The heterogeneous elasticity studied via the Goldstone fluctuations in the RLPM is a description of the elastic properties of the state right after linking (i.e., an elastic free energy that takes the state right after linking as its elastic reference state). We have shown that, after linking, the system relaxes to a stable state for any given realization of disorder (i.e., for any given heterogeneous configuration of the elastic parameters in the state right after linking), and this relaxed state, which is a state of mechanical equilibrium, is actually the state of experimental relevance. By solving for the relaxed state for any given realization of disorder, and expanding the elastic free energy for deformations relative to this relaxed state, we have obtained an elastic free energy relative to the relaxed state (i.e., taking the relaxed state as the new elastic reference state). The statistics of the quenched randomness in this elastic free energy is subsequently determined.
The first statistical moments of the quenched random elastic parameters (i.e., the disorder averages of the elastic parameters), unveil the basic homogeneous macroscopic properties of the heterogeneous elastic medium. We have found that the disorder average of the nonlocal kernel of attractive interactions is characterized by the typical localization lengthscale of the RLPM, which is a scale smaller than the lengthscale of the elastic deformations that we are considering. Thus, it is reasonable to make a local expansion of the elastic energy, relative to the relaxed state. The resulting local form of the elastic energy is a version of Lagrangian elasticity, featuring heterogeneous (i.e., spatially randomly varying) residual stress and Lamé coefficients. The disorder average of the residual stress vanishes. The disorder average of the shear modulus is found to be proportional to temperature, reflecting the entropic nature of the shear rigidity of soft random solids. The disorder average of the bulk modulus depends on the particle number-density and the strength of the excluded-volume interaction. In particular, the disorder averages of these elastic parameters of the relaxed state are all translationally and rotationally invariant, reflecting the macroscopic translational and rotational invariance of the soft random solid state.
The second statistical moments of the quenched random elastic parameters (i.e., the spatial correlations of these elastic parameters) characterize the fluctuations of the quenched randomness in the elastic properties. The disorder correlators of the elastic parameters that appear in the local form of the elastic energy (relative to the relaxed state) exhibit interesting universal behaviors. In particular, the disorder-correlators involving the residual stress are found to be long-ranged and governed by a universal parameter that also determines the mean shear modulus, but the disorder-correlators of the shear and bulk moduli are found to be short ranged.
Because of the heterogeneity present in the elasticity of soft random solids, upon the application external stress, the system responds by adopting a strain field that is nonaffine (i.e., a strain field that is characterized by an inhomogeneous deformation gradient). We have also obtained a statistical description of these nonaffine deformations. The disorder average of the nonaffine deformations vanishes, and their disorder correlator is also found to be long ranged.
So far, we have studied the first two statistical moments of the quenched random elastic parameters of soft random solids. The entire probability distribution of the quenched random elastic parameters can also be explored using the formalism presented here, via the RLPM, and one can also progress beyond the local limit of the elasticity theory.
This approach to the heterogeneous elasticity of soft random solids can also be applied to the setting of liquid crystal elastomers, in which the constituent polymers of the random network possesses (or are capable of exhibiting) liquid-crystalline order [@Warner2003; @Lubensky2002; @Xing2008; @Clarke1998]. In liquid crystal elastomers, the strain field is coupled to the liquid-crystalline order, and this produces a rich collection of interesting phenomena, such as spontaneous sample-shape deformation upon changes of temperature, anomalously soft modes in the elasticity. The interplay of the heterogeneity of the random network and the liquid-crystalline order has interesting consequences: e.g., it can give rise to a polydomain structure in the liquid crystalline order. [These interesting topics shall be reserved for future studies. ]{}
We thank Tom Lubensky for stimulating discussions. This work was supported by National Science Foundation Grant No. DMR 06-05816 (X.M. and P.M.G.), American Chemical Society Grant No. PRF 44689-G7 (X.X.), Deutsche Forschungsgemeinschaft through SFB 602 (standard) (A.Z.), and the National Science Foundation Grant No. DMR 0804900 (X.M.).
Disorder average with the Deam-Edwards distribution {#APP:DisoAver}
===================================================
In this Appendix we calculate disorder averages weighted by the Deam-Edwards distribution, in particular, $Z_1$ and $Z_{1+n}$ of Section \[SEC:RLPMReplica\].
First, we calculate the factor $Z_1$, which is defined as $$\begin{aligned}
\label{EQ:AZOne}
Z_1 &\equiv& \sum_{\RealDiso}
\frac{\left(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\right)^{\LinkNumb}
Z_{\RealDiso}(\Volu_0)}{ \LinkNumb !}.\end{aligned}$$ The summation over the quenched disorder $\sum_{\RealDiso}$ includes two steps: a summation over the number of links $\LinkNumb$, and a summation over all possible ways of making these $\LinkNumb$ links, i.e., of assigning the $\LinkNumb$ links to different collections of pairs. Thus Eq. (\[EQ:AZOne\]) can be written as
$$\begin{aligned}
Z_1 &=& \sum_{\LinkNumb}
\frac{\Big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\Big)^{\LinkNumb}
Z_{\RealDiso}(\Volu_0)}{ \LinkNumb !} \nonumber\\
&=& \sum_{\LinkNumb=0}^{\infty} \sum_{i_1\ne j_1}^{\PartNumb}
\sum_{i_2\ne j_2}^{\PartNumb} \cdots
\sum_{i_{\LinkNumb}\ne j_{\LinkNumb}}^{\PartNumb}
\frac{\Big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\Big)^{\LinkNumb}}
{ \LinkNumb !} \PartitionLiquid(\Volu)
\left\langle
\prod_{e=1}^{\LinkNumb} \LinkPote \big(
\vert \PartPosi_{i_e} - \PartPosi_{j_e} \vert
\big) \right\rangle_{1}^{\HamiExcl} \nonumber\\
&=& \PartitionLiquid(\Volu)\lthal
\sum_{\LinkNumb=0}^{\infty} \frac{\big(\frac{\LinkDens \Volu_0}
{2\PartNumb \LinkPote_0}\big)^{\LinkNumb}}{ \LinkNumb !}
\Big(\sum_{i\ne j}^{\PartNumb}
\LinkPote \big(
\vert \PartPosi_{i} - \PartPosi_{j} \vert
\big)\Big)^{\LinkNumb}\rthal_{1}^{\HamiExcl} \nonumber\\
\noalign{\smallskip}
&=& \PartitionLiquid(\Volu)\left\langle
\exp\Big(
\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}
\sum_{i\ne j}^{\PartNumb}
\LinkPote \big(
\vert \PartPosi_{i} - \PartPosi_{j} \vert
\big)
\Big)\right\rangle_{1}^{\HamiExcl}.\end{aligned}$$
The mean-field approximation for $Z_1$ amounts to taking the number density of the unlinked liquid to be $\PartNumb/\Volu_0$, which is similar to the calculation that yields Eq. (\[EQ:FLiquid\]), and hence we arrive at $$\begin{aligned}
\label{EQ:Z1SP}
Z_1=\exp\big(\PartNumb\ln\Volu_0
-\frac{\ExclVolu\PartNumb^2}{2\Volu_0 \BoltCons T}+\frac{\PartNumb \LinkDens}{2} \big).\end{aligned}$$
Second, we calculate $Z_{1+n}$, which is defined as $$\begin{aligned}
Z_{1+n}\equiv\sum_{\RealDiso}
\frac{\big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\big)^{\LinkNumb}}
{ \LinkNumb !} Z_{\RealDiso}(\Volu_0) Z_{\RealDiso}(\Volu)^n .\end{aligned}$$ The factor $Z_{\RealDiso}(\Volu)^n$ can be written in terms of replicas as $$\begin{aligned}
Z_{\RealDiso}(\Volu)^n
&=& \int_{\Volu} \ReplProd \prod_{j=1}^{\PartNumb} d\PartPosi_{j}\REPa
\, e^{-\ReplSum \HamiExcl \REPa/\BoltCons T} \nonumber\\
&&\quad\quad\times
\ReplProd \prod_{e=1}^{\LinkNumb}
\LinkPote \big(
\vert \PartPosi_{i_e}\REPa - \PartPosi_{j_e}\REPb \vert \big) ,\end{aligned}$$ where $\HamiExcl \REPa$ is the part of the Hamiltonian $\HamiExcl$ (i.e., the excluded-volume interaction) for replica $\alpha$, as defined in Eq. (\[EQ:partition\]).
We define the $\HamiExcl$ average for $1+n$ replicas as $$\begin{aligned}
\ltha \cdots \rtha_{1+n}^{\HamiExcl} &\equiv&
\frac{1}{\PartitionLiquid(V_0) \PartitionLiquid(V)^n}\nonumber\\
&&\quad\quad\times
\int _{\Volu_0} \prod_{i=1}^{\PartNumb} d \PartPosi_i \REP0
\int _{\Volu} \ReplProdOne \prod_{i=1}^{\PartNumb}
d \PartPosi_i \REPa \nonumber\\
&&\quad\quad\times
e^{-\frac{\HamiExcl\REP0}{{\BoltCons}T}
-\frac{\ReplSumOne \HamiExcl\REPa}{{\BoltCons}T}}\cdots .\end{aligned}$$ Using this notation we arrive at
$$\begin{aligned}
Z_{1+n}&=&\sum_{\LinkNumb=0}^{\infty} \sum_{i_1\ne j_1}^{\PartNumb}
\sum_{i_2\ne j_2}^{\PartNumb} \cdots
\sum_{i_{\LinkNumb}\ne j_{\LinkNumb}}^{\PartNumb}
\frac{\big(\frac{\LinkDens \Volu_0}{2\PartNumb \LinkPote_0}\big)^{\LinkNumb}}
{ \LinkNumb !} \PartitionLiquid(\Volu_0)\PartitionLiquid(\Volu)^n
\ltha
\ReplProd \prod_{e=1}^{\LinkNumb} \LinkPote \big(
\vert \PartPosi_{i_e}\REPa - \PartPosi_{j_e}\REPa \vert
\big) \rtha_{1+n}^{\HamiExcl} \nonumber\\
&=& \PartitionLiquid(\Volu_0)\PartitionLiquid(\Volu)^n \lthal
\sum_{\LinkNumb=0}^{\infty} \frac{\big(\frac{\LinkDens \Volu_0}
{2\PartNumb \LinkPote_0}\big)^{\LinkNumb}}{ \LinkNumb !}
\Big(\sum_{i\ne j}^{\PartNumb} \ReplProd
\LinkPote \big(
\vert \PartPosi_{i}\REPa - \PartPosi_{j}\REPa \vert
\big)\Big)^{\LinkNumb}\rthal_{1+n}^{\HamiExcl} \nonumber\\
&=& \PartitionLiquid(\Volu_0)\PartitionLiquid(\Volu)^n
\lthal \exp \Big(\frac{\LinkDens \Volu_0}
{2\PartNumb \LinkPote_0}
\sum_{i\ne j}^{\PartNumb} \ReplProd
\LinkPote \big(
\vert \PartPosi_{i}\REPa - \PartPosi_{j}\REPa \vert
\big)\Big)\rthal_{1+n}^{\HamiExcl}\end{aligned}$$
Hubbard-Stratonovich Transformation {#APP:HSTransformation}
===================================
The effective Hamiltonian, Eq. (\[EQ:HQHL\]), can be analyzed via a Hubbard-Stratonovich (HS) transformation—a field-theoretic tool that is often applied to strongly coupled models to decouple interactions and develop a convenient representation in terms of functional integrals [@Hubbard1959; @Stratonovich1957].
The version of the HS transformation that we use for the RLPM can be illustrated via the following simple example. Consider a statistical-mechanical system having the following partition function: $$\begin{aligned}
Z(h) =\int dq\, e^{-H_0(q)}e^{Jq^2+hq} = Z_0\,\langle e^{Jq^2+hq} \rangle_{H_0(q)} ,\end{aligned}$$ where $H(q)\equiv H_0(q)-Jq^2-hq$ is the total Hamiltonian for the variable $q$, with $H_0(q)$ being the leading-order term and $Jq^2$ being considered as a perturbation. (Although it is just a simple quadratic term, we use it to illustrate the method.) The factor $Z_0\equiv \int dq \, e^{-H_0(q)}$. The term $hq$ denotes the coupling to an external field, which generates the statistical moments of $q$ via $$\begin{aligned}
\langle q \rangle_{H(q)}=\frac{\partial }{\partial h}\Big\vert_{h=0} \ln Z(h) .\end{aligned}$$ The $Jq^2$ term in the exponent can be decoupled using the following version of the HS transformation: $$\begin{aligned}
Z&=&\Big(\frac{J}{\pi}\Big)^{1/2} Z_0 e^{-\frac{h^2}{4J}}
\int d \omega \, e^{-J\omega^2+h \omega}
\langle e^{2J\omega q} \rangle_{H_0(q)} \nonumber\\
&=&\Big(\frac{J}{\pi}\Big)^{1/2} \int d \omega\, e^{-\mathcal{H}(\omega)} ,\end{aligned}$$ with $$\begin{aligned}
\label{EQ:Homega}
\mathcal{H}(\omega)\equiv J\omega^2-\ln Z_0\,\langle e^{2J\omega q} \rangle_{H_0(q)}-h\omega +\frac{h^2}{4J} .\end{aligned}$$ In this form, the partition function is expressed as an integral over the variable $\omega$, and the quadratic term in the original variable $q$ is now decoupled. If fluctuations with large $q$ only appear with very small probabilities, as governed by $H_0(q)$, one can expand the $\ln\langle e^{2J\omega q} \rangle_{H_0(q)}$ term as a power series in $q$. Thus, one can obtain an effective Hamiltonian $\mathcal{H}(\omega)$ via the low-order terms in $\omega$, which has the form of a Landau free energy, and is convenient to analyze.
It is evident that the average of $\omega$, taken with the statistical weight defined by $\mathcal{H}(\omega)$, equals the average of $q$, taken with the statistical weight defined by $H(q)$: $$\begin{aligned}
\label{EQ:HSAver}
\langle \omega \rangle_{\mathcal{H}(\omega)}
=\frac{\partial }{\partial h}\Big\vert_{h=0} \ln Z(h)
=\langle q \rangle_{H(q)}.\end{aligned}$$ Thus, the statistical mechanics of $q$ can be examined by studying the statistical mechanics of $\omega$. In the Hamiltonian $\mathcal{H}(\omega)$, $q$ appears linearly; therefore, in cases in which $q$ is a variable that involves a summation over many particles, this method will allow us to decouple the problem into a single-particle one, as will be seen in the following application of the HS transformation to vulcanization theory.
In the RLPM, the partition function we are going to decouple is Eq. (\[EQ:ZQ\]): $$\begin{aligned}
Z_{1+n} = \int_{\Volu_0} \prod_{i=1}^{\PartNumb} d\PartPosi_{i}^{0}
\int_{\Volu} \ReplProdOne \prod_{j=1}^{\PartNumb} d\PartPosi_{j}\REPa
\, e^{-\frac{H_{\DensFunc}\lbrack \DensFunc_{\REPP}\rbrack}{\BoltCons T}} ,\end{aligned}$$ with
$$\begin{aligned}
H_{\DensFunc}\lbrack \DensFunc_{\REPP}\rbrack
= -\frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\DensFunc_{\REPP}
\DensFunc_{-\REPP}\LinkPoteRepl_{\REPP}
+\frac{\ExclVoluRN_0 \PartNumb^2}{2\Volu_0} \sum_{p}\DensFunc_{p\BASE^{0}}
\DensFunc_{-p\BASE^{0}}
+\frac{\ExclVoluRN \PartNumb^2}{2\Volu} \sum_{p} \ReplSumOne
\DensFunc_{p\BASE\REPa} \DensFunc_{-p\BASE\REPa} .\end{aligned}$$
The field $\DensFunc_{\REPP}=(1/N)\sum_{j=1}^{\PartNumb}e^{-i\REPP \cdot \hat{c}_j}$ is a complex field, so we apply the following equalities for the complex variables $q$ and $\omega$:
\[EQ:HSComplex\] $$\begin{aligned}
\!\!e^{-J\vert q \vert^2}\!\!
&=&\!\! \frac{J}{\pi}\int\!\! d(\RealPart\, \omega)d(\ImagPart\, \omega) \label{EQ:HS1}
e^{-J\vert \omega \vert^2+2iJ \RealPart\, q\omega^{*}} , \\
\!\!e^{+J\vert q \vert^2}\!\!
&=&\!\! \frac{J}{\pi}\int \!\!d(\RealPart\, \omega)d(\ImagPart\, \omega)
e^{-J\vert \omega \vert^2+2J \RealPart\, q\omega^{*}} , \label{EQ:HS2}\end{aligned}$$
noticing that the product $\RealPart \, (q\omega^{*}) = (\RealPart\, q)(\RealPart\,\omega)+ (\ImagPart\, q)(\ImagPart\,\omega)$. We use Eq. (\[EQ:HS1\]) for the HS transformation for the LRS fields, and Eq. (\[EQ:HS2\]) for the HS transformation for the HRS field, and hence arrive at the form (For more details, see Section V in [@Goldbart1996].) $$\begin{aligned}
\label{EQ:HSPartition}
Z_{1+n} = \int \mathcal{D}\VOP_{\hat{p}}\ReplProd\mathcal{D}\VOP_{p}\REPa
e^{-\frac{H_{\VOP}\lbrack\VOP_{p}\REPa,\VOP_{\hat{p}}\rbrack}{\BoltCons T}} ,\end{aligned}$$ with $$\begin{aligned}
H_{\VOP}\lbrack\VOP_{p}\REPa,\VOP_{\hat{p}}\rbrack
= \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\!\sum_{\REPP\in HRS}\!\!\VOP_{\REPP}
\VOP_{-\REPP}\LinkPoteRepl_{\REPP}
+\!\frac{\ExclVoluRN_0 \PartNumb^2}{2\Volu_0}\!\sum_{p}\VOP_{p\BASE^{0}}
\VOP_{-p\BASE^{0}}
+\!\frac{\ExclVoluRN \PartNumb^2}{2\Volu} \sum_{p}\!\ReplSumOne
\VOP_{p\BASE\REPa} \VOP_{-p\BASE\REPa} \!
- \! N\BoltCons T \ln \MyZzero,\end{aligned}$$ where the $N\BoltCons T \ln \MyZzero$ term is analog of the $\ln \langle e^{2J\omega q} \rangle_{H_0(q)}$ term in Eq. (\[EQ:Homega\]) and, using $\DensFunc_{-\REPP}=(1/\PartNumb)\sum_{j=1}^{\PartNumb} e^{i\REPP\cdot\hat{c}_j}$ and $\DensFunc_{-p}\REPa=(1/\PartNumb)\sum_{j=1}^{\PartNumb} e^{ip\cdot c_j\REPa}$, we have $$\begin{aligned}
\PartNumb \ln \MyZzero
&\equiv&
\ln \Big\lbrace \int_{\Volu_0} \prod_{i=1}^{\PartNumb} d\PartPosi_{i}^{0}
\int_{\Volu} \ReplProdOne \prod_{j=1}^{\PartNumb} d\PartPosi_{j}\REPa
\exp\big\lbrack
\frac{\PartNumb \LinkDens }{ \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP}
\frac{1}{\PartNumb}\sum_{j=1}^{\PartNumb}
e^{i\REPP\cdot\hat{c}_j}\LinkPoteRepl_{\REPP} \nonumber\\
&&\quad\quad + \frac{i\ExclVoluRN_0 \PartNumb^2}{\Volu_0\BoltCons T}
\sum_{p}\VOP_{p\BASE^{0}}
\frac{1}{\PartNumb}\sum_{j=1}^{\PartNumb} e^{ip\cdot c_j^{0}}
+\frac{i\ExclVoluRN \PartNumb^2}{\Volu\BoltCons T} \sum_{p} \ReplSumOne
\VOP_{p\BASE\REPa} \frac{1}{\PartNumb}\sum_{j=1}^{\PartNumb} e^{ip\cdot c_j^{0}}
\big\rbrack\Big\rbrace \nonumber\\
&=&\PartNumb \ln \Bigg\lbrace \int_{\Volu_0} \! d\PartPosi^{0} \!\!
\int_{\Volu} \!\ReplProd d\PartPosi\REPa
\exp\Big\lbrack
\frac{ \LinkDens}{ \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP}\LinkPoteRepl_{\REPP}e^{i\REPP \cdot \hat{c}}
\nonumber\\
&& \quad + \frac{i\ExclVoluRN_0 \PartNumb}{\Volu_0\BoltCons T}
\sum_{p}\VOP_{p\BASE^{0}}e^{ip^{0}c^{0}}
\!\!+\frac{i\ExclVoluRN \PartNumb}{\Volu\BoltCons T} \sum_{p} \ReplSumOne
\VOP_{p\BASE\REPa} e^{ip\REPa c\REPa} \Big\rbrack\Bigg\rbrace . \end{aligned}$$
In this form it is evident that the $\PartNumb$ particles are actually *decoupled*. Notice that in Eq. (\[EQ:HSPartition\]) the functional integrals $\int \mathcal{D}\,\VOP_{\hat{p}}\ReplProd\mathcal{D}\VOP_{p}\REPa $ have carefully chosen prefactors \[as in Eq. (\[EQ:HSComplex\])\], to ensure that the integration is properly normalized.
Hamiltonian of the stationary point {#APP:HSP}
===================================
In this Appendix we calculate the value of the Hamiltonian at the stationary point by inserting the stationary point order parameter (\[EQ:VOPAnsatzM\]) into the Hamiltonian (\[EQ:HVOPHRS\]).
The first term in the Hamiltonian is
$$\begin{aligned}
&& \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP} \VOP_{-\REPP}\LinkPoteRepl_{\REPP} \nonumber\\
&=&\frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP} (\LinkPote)^{1+n}
e^{-\frac{\LinkScal^2\vert\REPP\vert^2}{2\BoltCons T}}
\big\lbrace \LocaPart \int \frac{dz_1}{\Volu_0}\int _{\ISLL_1}
e^{-\frac{\vert\REPP\vert^2}{2\ISLL_1}
-i \REPP \cdot \hat{z}_{\Contraction,1} }
-\LocaPart\delta^{((1+n)d)}_{\REPP}\big\rbrace \nonumber\\
&&\quad\times \big\lbrace \LocaPart \int \frac{dz_2}{\Volu_0}\int _{\ISLL_2}
e^{-\frac{\vert\REPP\vert^2}{2\ISLL_2}
+i \REPP \cdot \hat{z}_{\Contraction,2} }
-\LocaPart\delta^{((1+n)d)}_{\REPP}\big\rbrace \nonumber\\
&=&\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2}{2 \Volu^n \LinkPote_0}
(\LinkPote)^{1+n}\nonumber\\
&& +\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2}{2 \Volu^n \LinkPote_0}
(\LinkPote)^{1+n}
\sum_{\REPP}
\int\frac{dz_1 dz_2}{\Volu_0^2}\int _{\ISLL_1,\ISLL_2}
e^{\big(\frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)\vert\REPP\vert^2
-i\REPP\cdot (\hat{z}_{\Contraction,1}-\hat{z}_{\Contraction,2})} \nonumber\\
&=&\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2(\LinkPote)^{n}}{2 \Volu^n}
+\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2(\LinkPote)^{n}}{2 \Volu^n }
(1+n\Contraction^2)^{-d/2}\int _{\ISLL_1,\ISLL_2}
\Big\lbrace 2\pi \Big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\Big)
\Big\rbrace^{-nd/2} ,\end{aligned}$$
where $\hat{z}_{\Contraction}\equiv \{z,\Contraction z,\ldots,\Contraction z\}$. The sum $\sum_{\REPP\in HRS}$ can be changed into the $\sum_{\REPP}$ because the order parameter we have inserted vanishes for $\REPP\in\,$LRS. We have also changed momentum summation into an integral by using $\frac{1}{\Volu_0\Volu^n}\sum_{\REPP}=\int\frac{d^{(1+n)d}\REPP}{(2\pi)^{(1+n)d}}$.
The free energy of the system is related to the $O(n)$ term of this Hamiltonian, as given by Eqs. (\[EQ:HFEDisoAver\], \[EQ:HFEDisoAverDeri\]). Thus, we make the small-$n$ expansion. It is straightforward to see that the $O(1)$ terms cancel, and that the leading-order term is given by $O(n)$ $$\begin{aligned}
&& \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP} \VOP_{-\REPP}\LinkPoteRepl_{\REPP} \nonumber\\
&=& n \frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2}{2 \Volu^n}
\Big\lbrace
\ln \Volu -\frac{d}{2}\big(\ln(2\pi)+\Contraction^2\big)
-\frac{d}{2} \int _{\ISLL_1,\ISLL_2}
\Big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\Big)
\Big\rbrace .\end{aligned}$$
Similarly, we can calculate the $\ln \MyZzero$ term. By inserting the saddle-point value of the order parameter into $\MyZzero$, and summing (or, more precisely, integrating) over momentum $\REPP$, we have $$\begin{aligned}
\MyZzero\!&=&\!\int_{\Volu_0} \! d\PartPosi^{0} \!\!
\int_{\Volu} \!\ReplProd d\PartPosi\REPa
\exp\Big\lbrace
\frac{ \LinkDens}{ \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP}\LinkPoteRepl_{\REPP}e^{i\REPP \cdot \hat{c}}
\Big\rbrace \nonumber\\
&=& e^{-\LinkDens \LocaPart (\LinkPote_0/\Volu)^n}
\int d\hat{c} \exp \big\lbrace
\LinkDens \LocaPart (\LinkPote_0)^n
\int dz \int_{\ISLL} \big(\frac{\tISLL}{2\pi}\big)^{\frac{(1+n)d}{2}}
e^{-\frac{\tISLL}{2}(\hat{z}_{\Contraction}-\hat{c})^2}
\big\rbrace ,\end{aligned}$$ where $\tISLL\equiv\big(\frac{1}{\ISLL}+\LinkScal^2\big)^{-1}$. We then Taylor-expand the exponential (keeping all orders) and integrate out $\hat{c}$ to get $$\begin{aligned}
\label{EQ:Z0intec}
\MyZzero &=& e^{-\LinkDens \LocaPart (\LinkPote_0/\Volu)^n}\Volu_0\Volu^n
\Big\lbrace
1+\LinkDens \LocaPart (\LinkPote_0/\Volu)^n
+\frac{1}{\Volu_0\Volu^n} \sum_{m=2}^{\infty}
\frac{\big(\LinkDens \LocaPart (\LinkPote_0)^n\big)^m}{m!}
\int dz_1 \cdots dz_m
\nonumber\\
&&\quad \times \int_{\ISLL_1,\ldots,\ISLL_m}
\prod _{j=1}^{m} \Big(\frac{\tISLL_j}{2\pi}\Big)^{\frac{(1+n)d}{2}}
\Big(\frac{2\pi}{\tISLL_1+\tISLL_m}\Big)^{\frac{(1+n)d}{2}}
e^{-\frac{\tISLL_1\tISLL_2
(\hat{z}_{\Contraction,1}-\hat{z}_{\Contraction,2})^2+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}}
\Big\rbrace ,\end{aligned}$$ where in the exponential the terms following $(\hat{z}_{\Contraction,1}-\hat{z}_{\Contraction,2})^2$ include all pairs of the $m$ variables \[there are $m(m-1)/2$ such terms\]. By using $(\hat{z}_{\Contraction,1}-\hat{z}_{\Contraction,2})^2
=(1+n\Contraction^2)(z_1-z_2)^2$ \[recall that $\hat{z}_{\Contraction,1}$ is a $(1+n)d$-dimensional vector, and that $z_1$ is a $d$-dimensional vector\], the integration $\int dz_1 \cdots dz_m $ can be readily performed, and we thus obtain $$\begin{aligned}
\label{EQ:Z0intez}
\MyZzero&=& e^{-\LinkDens \LocaPart (\LinkPote_0/\Volu)^n}\Volu_0\Volu^n
\Big\lbrace
1+\LinkDens \LocaPart (\LinkPote_0/\Volu)^n \nonumber\\
&& \quad +\frac{1}{\Volu^n} \sum_{m=2}^{\infty}
\frac{\big(\LinkDens \LocaPart (\LinkPote_0)^n\big)^m}{m!}
\int_{\ISLL_1,\ldots,\ISLL_m}
\prod _{j=1}^{m} \Big(\frac{\tISLL_j}{2\pi}\Big)^{\frac{nd}{2}}
\Big(\frac{2\pi}{\tISLL_1+\tISLL_m}\Big)^{\frac{nd}{2}}
(1+n\Contraction^2)^{\frac{(1-m)d}{2}}
\Big\rbrace .\end{aligned}$$ From this, $-\ln \MyZzero$ can obtained by making the small-$n$ expansion, using the following equality: $$\begin{aligned}
\ln(x+ny+O(n^2))=\ln(x(1+n(y/x)+O(n^2)))=\ln x + n (y/x) +O(n^2),\end{aligned}$$ so that we have $$\begin{aligned}
-\ln \MyZzero
&=& -\ln\Volu_0 +n \Big\lbrace
-\ln \Volu +(\LinkDens \LocaPart + e^{-\LinkDens \LocaPart}-1)
\Big(\frac{d}{2}\big(\ln(2\pi)+\Contraction^2\big)-\ln \Volu \Big) \nonumber\\
&&\quad -e^{-\LinkDens \LocaPart} \frac{d}{2} \sum_{m=1}^{\infty}
\frac{\big(\LinkDens \LocaPart \big)^m}{m!}
\ln \Big(\frac{\tISLL_1\cdots\tISLL_m}{\tISLL_1+\cdots+\tISLL_m}\Big)
\Big\rbrace .\end{aligned}$$
Therefore, the small-$n$ expansion of the stationary-point Hamiltonian $\HVOPSP$ to $O(n)$ is given by $$\begin{aligned}
\HVOPSP&=&\frac{\ExclVoluRN_0(0)\PartNumb^2}{2\Volu_0}
+\frac{n\ExclVoluRN(0)\PartNumb^2}{2\Volu}
-\PartNumb \BoltCons T \ln \Volu_0 -n\PartNumb \BoltCons T \ln \Volu
+n\PartNumb \BoltCons T \Bigg\lbrace
\UnivPara \big\lbrack
\frac{d}{2}\big(\ln(2\pi)+\Contraction^2\big)-\ln\Volu
\big\rbrack \nonumber\\
&&- \frac{\LinkDens \LocaPart^2}{2}\cdot\frac{d}{2}\int_{\ISLL_1,\ISLL_2}
\ln\big( \frac{1}{\ISLL_1}+\frac{1}{\ISLL_2}+\frac{\LinkScal^2}{\BoltCons T} \big)
-e^{-\LinkDens \LocaPart}\frac{d}{2}
\sum_{m=1}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}
\int_{\ISLL_1,\ldots,\ISLL_2} \ln
\Big( \frac{\tISLL_1\cdots\tISLL_m}{\tISLL_1 +\cdots +\tISLL_m} \Big)
\Bigg\rbrace ,\end{aligned}$$
where the parameter $\UnivPara$ is defined via $$\begin{aligned}
\label{EQ:UnivParaDef}
\UnivPara\equiv -\frac{\LinkDens \LocaPart^2}{2}+\LinkDens \LocaPart
+e^{-\LinkDens \LocaPart}-1.\end{aligned}$$
Hamiltonian of the Goldstone deformed order parameter {#APP:HGS}
=====================================================
In this Appendix we calculate the value of the Hamiltonian for the Goldstone-deformed order parameter by inserting the Goldstone-deformed order parameter (\[EQ:VOPGSR\]) into the Hamiltonian (\[EQ:HVOPHRS\]), following a calculation similar to that in Appendix \[APP:HSP\]. To obtain a description of the elasticity, we shall expand the Hamiltonian for small deformations, specifically in a series in the small, scalar variable that characterizes the replicated deformation field, viz., $\DefoScalPsi(z_1,z_2)\equiv
(\hat{\DefoPosi}(z_1)-\hat{\DefoPosi}(z_2))^2-(1+n)(z_1-z_2)^2$.
The quadratic term in the Hamiltonian is given by
$$\begin{aligned}
&& \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP} \VOP_{-\REPP}\LinkPoteRepl_{\REPP} \nonumber\\
&=&\frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP} (\LinkPote)^{1+n}
e^{-\frac{\LinkScal^2\vert\REPP\vert^2}{2\BoltCons T}}
\big\lbrace \LocaPart \int \frac{dz_1}{\Volu_0}\int _{\ISLL_1}
e^{-\frac{\vert\REPP\vert^2}{2\ISLL_1}
-i \REPP \cdot \hat{\DefoPosi}(z_1) }
-\LocaPart\delta^{((1+n)d)}_{\REPP}\big\rbrace \nonumber\\
&& \quad\times \big\lbrace \LocaPart \int \frac{dz_2}{\Volu_0}\int _{\ISLL_2}
e^{-\frac{\vert\REPP\vert^2}{2\ISLL_2}
+i \REPP \cdot \hat{\DefoPosi}(z_2) }
-\LocaPart\delta^{((1+n)d)}_{\REPP}\big\rbrace \nonumber\\
&=&\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2(\LinkPote)^{n}}{2 \Volu^n} \nonumber\\
&& +\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2(\LinkPote)^{n}}{2 \Volu^n }
\Volu_0 \Volu^n \int\frac{dz_1 dz_2}{\Volu_0^2}\int _{\ISLL_1,\ISLL_2}
\Big\lbrace 2\pi \Big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\Big)
\Big\rbrace^{-\frac{(1+n)d}{2}}
e^{-\frac{(\hat{\DefoPosi}(z_1)-\hat{\DefoPosi}(z_2))^2}
{2\big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)}} .\end{aligned}$$
Next, we expand for small $\DefoScalPsi$, adopting the notation $\DefoScalPsi(z_1,z_2)\equiv
(\hat{\DefoPosi}(z_1)-\hat{\DefoPosi}(z_2))^2-(1+n)(z_1-z_2)^2$. Note that $\DefoScalPsi$ is *not* related to the deformation relative to the stationary point, this stationary point being characterized by the mean positions of the replicas of the particle $\hat{z}_{\Contraction}=\{z,\Contraction z,\ldots,\Contraction z\}$. Instead, $\DefoScalPsi$ describes deformations relative to the state right after linking (i.e., prior to relaxation), this state being characterized by the mean positions of the replicas of the particle $\hat{z}=\{z,z,\ldots,z\}$, as discussed in Sections \[SEC:EnerGoldSton\] and \[SEC:relaxation\]. The expansion of the quadratic term for small $\DefoScalPsi$ is given by $$\begin{aligned}
&& \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP} \VOP_{-\REPP}\LinkPoteRepl_{\REPP} \nonumber\\
&=&\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2(\LinkPote)^{n}}{2 \Volu^n}
\nonumber\\
&& +\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2(\LinkPote)^{n}}{2 }
\int\frac{dz_1 dz_2}{\Volu_0}\int _{\ISLL_1,\ISLL_2}
\Big\lbrace 2\pi \Big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\Big)
\Big\rbrace^{-(1+n)d/2}
e^{-\frac{(1+n)(z_1-z_2)^2}
{2\big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)}} \nonumber\\
&&\times \Bigg\lbrace
1-\frac{\DefoScalPsi(z_1,z_2)}{2\big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)}
+\frac{1}{2}\Big(\frac{\DefoScalPsi(z_1,z_2)}
{2\big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)}\Big)^2
+O(\DefoScalPsi(z_1,z_2)^3)
\Bigg\rbrace .\end{aligned}$$ The small-$n$ expansion on this quadratic term is then given by $$\begin{aligned}
&& \frac{\PartNumb \LinkDens \BoltCons T}{2 \Volu^n \LinkPote_0}
\sum_{\REPP\in HRS}\VOP_{\REPP} \VOP_{-\REPP}\LinkPoteRepl_{\REPP} \nonumber\\
&=&\frac{\PartNumb \LinkDens \BoltCons T\LocaPart^2}{2} \Bigg\lbrace
n \Big\lbrace \ln \Volu -\frac{d}{2}\big(\ln(2\pi)+1\big)
-\frac{d}{2}\int _{\ISLL_1,\ISLL_2}
\ln \Big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\Big) \Big\rbrace \nonumber\\
&&\quad +\int\frac{dz_1 dz_2}{\Volu_0}\int _{\ISLL_1,\ISLL_2}
\Big\lbrace 2\pi \Big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\Big)
\Big\rbrace^{-d/2}
e^{-\frac{(1+n)(z_1-z_2)^2}
{2\big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)}} \nonumber\\
&&\quad\quad\times \Big\lbrace
-\frac{\DefoScalPsi(z_1,z_2)}{2\big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)}
+\frac{1}{2}\Big(\frac{\DefoScalPsi(z_1,z_2)}
{2\big( \frac{1}{2\ISLL_1}+\frac{1}{2\ISLL_2}+
\frac{\LinkScal^2}{2\BoltCons T}\big)}\Big)^2
+O(\DefoScalPsi(z_1,z_2)^3)
\Big\rbrace
\Bigg\rbrace .\end{aligned}$$ The calculation for the $\ln \MyZzero$ term is similar to the above calculation of the quadratic term. The expansion in small quantity $\DefoScalPsi$ reads $$\begin{aligned}
\MyZzero
&=& e^{-\LinkDens \LocaPart (\LinkPote_0/\Volu)^n}\Volu_0\Volu^n
\Big\lbrace
1+\LinkDens \LocaPart (\LinkPote_0/\Volu)^n \nonumber\\
&&\quad
+\frac{1}{\Volu_0\Volu^n} \sum_{m=2}^{\infty}
\frac{\big(\LinkDens \LocaPart (\LinkPote_0)^n\big)^m}{m!}
\int dz_1 \cdots dz_m \int_{\ISLL_1,\ldots,\ISLL_m}
\prod _{j=1}^{m} \Big(\frac{\tISLL_j}{2\pi}\Big)^{\frac{(1+n)d}{2}}
\Big(\frac{2\pi}{\tISLL_1+\cdots+\tISLL_m}\Big)^{\frac{(1+n)d}{2}}
e^{-\frac{\tISLL_1\tISLL_2(z_1-z_2)^2+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}} \nonumber\\
&&\quad\times\big\lbrace
1-\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}
+\frac{1}{2}\big(\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}
\big)^2
\big\rbrace
\Big\rbrace ,\end{aligned}$$ where the summations that we have abbreviated with $\cdots$ include all pairs formed by $\{z_1,\ldots,z_m\}$. Next, we expand for small $n$, keep terms to $O(n)$ in the $\ln \MyZzero$ term, [assuming that both $\DefoScalPsi$ and $\DefoScalPsi^2$ contain $O(n)$ terms. ]{} After a tedious calculation we have $$\begin{aligned}
\label{EQ:AppLnz}
-\ln \MyZzero
&=& -\LinkDens \LocaPart (\frac{\LinkPote}{\Volu})^n
-\ln \Volu_0 -n\ln \Volu - \LinkDens \LocaPart \nonumber\\
&& -n e^{-\LinkDens \LocaPart} \Big\lbrace
(1-e^{\LinkDens \LocaPart})\ln \Volu
+\LinkDens\LocaPart e^{\LinkDens \LocaPart} \ln \Volu_0
+\big(e^{\LinkDens \LocaPart}-1
-\LinkDens\LocaPart e^{\LinkDens \LocaPart}\big)\frac{d}{2}
\big( \ln(2\pi)+1 \big) \nonumber\\
&&\quad\quad +\frac{d}{2}\sum_{m=1}^{\infty}\int_{\ISLL _1,\ldots\ISLL _m}
\ln\Big(\frac{\tISLL_1\cdots\tISLL_m}{\tISLL_1+\cdots+\tISLL_m}\Big)
\Big\rbrace \nonumber\\
&& -e^{-\LinkDens\LocaPart} \frac{1}{\Volu_0}
\sum_{m=2}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}\int dz_1\cdots dz_m
\int_{\ISLL_1,\ldots,\ISLL_m} \prod_{j=1}^{m}
\Big(\frac{\tISLL_j}{2\pi}\Big)^{\frac{d}{2}}
\Big(\frac{2\pi}{\tISLL_1+\cdots+\tISLL_m}\Big)^{\frac{d}{2}}
e^{-\frac{\tISLL_1\tISLL_2(z_1-z_2)^2+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}} \nonumber\\
&&\quad\quad\times \Bigg\lbrace
-\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}
+\frac{1}{2}\Big(\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}
\Big)^2
\Bigg\rbrace .\end{aligned}$$ To further simplify the expression, first consider the $O(\DefoScalPsi)$ terms in the expansion, $-\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}{2(\tISLL_1+\cdots+\tISLL_m)}$. The first term has a factor of $\DefoScalPsi(z_1,z_2)\equiv
(\hat{\DefoPosi}(z_1)-\hat{\DefoPosi}(z_2))^2-(1+n)(z_1-z_2)^2$, which only involves two variables $z_1$ and $z_2$, so we can integrate out the other $(m-2)$ variables, i.e., $z_3,\ldots,z_m$. (Of course, for $m=2$, no integrals are needed.)In total, there are $\frac{m(m-1)}{2}$ such terms (i.e., the number of pairs among $m$ variables). Thus, the $O(\DefoScalPsi)$ term in $-\ln \MyZzero$ is given by $$\begin{aligned}
\label{EQ:FirsOrdeInte}
&& -e^{-\LinkDens\LocaPart} \frac{1}{\Volu_0}
\sum_{m=2}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}\int dz_1\cdots dz_m
\int_{\ISLL_1,\ldots,\ISLL_m} \prod_{j=1}^{m}
\Big(\frac{\tISLL_j}{2\pi}\Big)^{\frac{d}{2}}
\Big(\frac{2\pi}{\tISLL_1+\cdots+\tISLL_m}\Big)^{\frac{d}{2}}
\nonumber\\
&& \quad\times e^{-\frac{\tISLL_1\tISLL_2(z_1-z_2)^2+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}}
\Bigg\lbrace
-\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}
\Bigg\rbrace \nonumber\\
&=& -e^{-\LinkDens\LocaPart} \frac{1}{\Volu_0}
\sum_{m=2}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}\int dz_1\cdots dz_m
\int_{\ISLL_1,\ldots,\ISLL_m} \prod_{j=1}^{m}
\Big(\frac{\tISLL_j}{2\pi}\Big)^{\frac{d}{2}}
\int dc
\nonumber\\
&& \quad\times e^{-\frac{\tISLL_1}{2}(z_1-c)^2-\frac{\tISLL_2}{2}(z_2-c)^2-\cdots}
\Bigg\lbrace
-\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}
\Bigg\rbrace \nonumber\\
&=& -e^{-\LinkDens\LocaPart} \frac{1}{\Volu_0}
\sum_{m=2}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!} \frac{m(m-1)}{2}
\int dz_1 dz_2 \int_{\ISLL_1,\ldots,\ISLL_m}
\Big(\frac{\tISLL_1 \tISLL_2}{2\pi(\tISLL_1+\tISLL_2)}\Big)^{\frac{d}{2}}
\nonumber\\
&&\quad\times e^{-\frac{\tISLL_1 \tISLL_2 (z_1-z_2)^2}{2(\tISLL_1+\tISLL_2)}}
\Big\lbrace
-\frac{\tISLL_1\tISLL_2 }
{2(\tISLL_1+\cdots+\tISLL_m)}
\Big\rbrace \DefoScalPsi(z_1,z_2) ,\end{aligned}$$ where in the last line here we have used the fact that $\{z_1,z_2,\ldots,z_m\}$ appear symmetrically, so that the $\frac{m(m-1)}{2}$ terms are identical.
Similarly, for the $O(\DefoScalPsi^2)$ terms in the expansion in Eq. (\[EQ:AppLnz\]) there are terms involving two points, such as $\DefoScalPsi(z_1,z_2)^2$, three points, such as $\DefoScalPsi(z_1,z_2)\DefoScalPsi(z_1,z_3)$, and four points, such as $\DefoScalPsi(z_1,z_2)\DefoScalPsi(z_3,z_4)$. (Of course, for $m=3$ there are no four-point terms, and for $m=2$ there are no three or four points terms.)Thus, the $O(\DefoScalPsi^2)$ terms can be written as $$\begin{aligned}
\label{EQ:QuadExpa}
&& \Big\lbrace\frac{\tISLL_1\tISLL_2 \DefoScalPsi(z_1,z_2)+\cdots}
{2(\tISLL_1+\cdots+\tISLL_m)}\Big\rbrace ^2 \nonumber\\
&\to& \frac{1} {4(\tISLL_1+\cdots+\tISLL_m)^2}
\Big\lbrace \frac{m(m-1)}{2}\tISLL_1^2\tISLL_2^2 \DefoScalPsi(z_1,z_2)^2 \nonumber\\
&&+ m(m-1)(m-2)\tISLL_1\tISLL_2^2\tISLL_3 \DefoScalPsi(z_1,z_2)\DefoScalPsi(z_2,z_3) \nonumber\\
&&+ \frac{m(m-1)(m-2)(m-3)}{4}\tISLL_1\tISLL_2\tISLL_3\tISLL_4
\DefoScalPsi(z_1,z_2)\DefoScalPsi(z_3,z_4)
\Big\rbrace .\end{aligned}$$ Following a calculation similar to that in Eq. (\[EQ:FirsOrdeInte\]), we can integrate out the integration variables that are not present in $\DefoScalPsi$, and thus obtain the $O(\DefoScalPsi^2)$ term in $\ln \MyZzero$.
Summing up the contributions from the quadratic term and the $\ln \MyZzero$ term, we arrive at the Hamiltonian of the Goldstone deformed state: $$\begin{aligned}
H_{\VOP}^{(G)}= \HVOPSP+H_{\VOP}^{\DefoScalPsi},\end{aligned}$$ with $\HVOPSP$ being the Hamiltonian of the stationary point, and the increase of the Hamiltonian due to Goldstone deformation is given by $$\begin{aligned}
\label{EQ:HPsi}
H_{\VOP}^{\DefoScalPsi}
&=& -n\PartNumb \BoltCons T \frac{\UnivPara d}{2}(\Contraction^2-1)
+ \frac{1}{2}\int dz_1 dz_2
\KernOne(z_1,z_2)\DefoScalPsi(z_1,z_2) \nonumber\\
&& -\frac{1}{8\BoltCons T}\int dz_1 dz_2 dz_3 dz_4 \KernTwo(z_1,z_2,z_3,z_4)
\DefoScalPsi(z_1,z_2)\DefoScalPsi(z_3,z_4) .\end{aligned}$$ The first term here, i.e., $-\PartNumb \BoltCons T \frac{\UnivPara d}{2}(\Contraction^2-1)$, is present due to the fact that the expansion variable $\DefoScalPsi(z_1,z_2)$ measures departures from the *state right after linking*, not the stationary point, as we have previously discussed. Note that $H_{\VOP}^{\DefoScalPsi}$ involves only the energy of shear deformation, because the Goldstone modes contain only pure shear deformation. The energy of volume variations is in the stationary-point Hamiltonian part, which contains the variable contraction parameter $\Contraction$.
The kernels in Eq. (\[EQ:HPsi\]) are given by $$\begin{aligned}
\label{EQ:KoneApp}
\frac{1}{2}\KernOne(z_1,z_2)
&=& \frac{\PartNumb \LinkDens \BoltCons T \LocaPart^2}{4 \Volu_0} \int_{\tISLL_1,\tISLL_2}
\Big(2\pi\Big(\frac{1}{\tISLL_1}+\frac{1}{\tISLL_2}+\frac{\LinkScal^2}{\BoltCons T}\Big)\Big)^{-d/2}
\Big(\frac{1}{\tISLL_1}+\frac{1}{\tISLL_2}+\frac{\LinkScal^2}{\BoltCons T}\Big)^{-1}
e^{-\frac{(z_1-z_2)^2}{2\big(\frac{1}{\tISLL_1}+\frac{1}{\tISLL_2}+\frac{\LinkScal^2}{\BoltCons T}\big)}}
\nonumber\\
&& +\frac{\PartNumb \BoltCons T}{2 \Volu_0} e^{-\LinkDens \LocaPart}
\sum_{m=2}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}\frac{m(m-1)}{2}\int_{\tISLL_1,\ldots,\tISLL_m}
\Big(\frac{\tISLL_1\tISLL_2}{2\pi(\tISLL_1+\tISLL_2)}\Big)^{d/2} \nonumber\\
&& \quad\quad\times
e^{-\frac{\tISLL_1\tISLL_2(z_1-z_2)^2}{2(\tISLL_1+\tISLL_2)}}
\frac{\tISLL_1\tISLL_2}{\tISLL_1+\cdots+\tISLL_m} ,\end{aligned}$$ and $$\begin{aligned}
\label{EQ:KtwoApp}
&& -\frac{1}{8\BoltCons T}\KernTwo(z_1,z_2,z_3,z_4) \nonumber\\
&=& \frac{\PartNumb \LinkDens \BoltCons T \LocaPart^2}{16 \Volu_0} \int_{\tISLL_1,\tISLL_2}
\Big(2\pi\Big(\frac{1}{\tISLL_1}+\frac{1}{\tISLL_2}+\frac{\LinkScal^2}{\BoltCons T}\Big)\Big)^{-d/2}
\Big(\frac{1}{\tISLL_1}+\frac{1}{\tISLL_2}+\frac{\LinkScal^2}{\BoltCons T}\Big)^{-1} \nonumber\\
&& \quad\quad\times
e^{-\frac{(z_1-z_2)^2}{2\big(\frac{1}{\tISLL_1}+\frac{1}{\tISLL_2}+\frac{\LinkScal^2}{\BoltCons T}\big)}}
\delta^{(d)}(z_1-z_3) \delta^{(d)}(z_2-z_4) \nonumber\\
&& -\frac{\PartNumb \BoltCons T}{8 \Volu_0} e^{-\LinkDens \LocaPart}
\sum_{m=2}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}\frac{m(m-1)}{2}\int_{\tISLL_1,\ldots,\tISLL_m}
\Big(\frac{\tISLL_1\tISLL_2}{2\pi(\tISLL_1+\tISLL_2)}\Big)^{d/2} \nonumber\\
&& \quad\quad\times
e^{-\frac{\tISLL_1\tISLL_2(z_1-z_2)^2}{2(\tISLL_1+\tISLL_2)}}
\frac{\tISLL_1^2\tISLL_2^2}{(\tISLL_1+\cdots+\tISLL_m)^2}\delta^{(d)}(z_1-z_3) \delta^{(d)}(z_2-z_4)
\nonumber\\
&& -\frac{\PartNumb \BoltCons T}{8 \Volu_0} e^{-\LinkDens \LocaPart}
\sum_{m=3}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}m(m-1)(m-2)\int_{\tISLL_1,\ldots,\tISLL_m}
\Big(\frac{\tISLL_1\tISLL_2\tISLL_3}{4\pi^2(\tISLL_1+\tISLL_2+\tISLL_3)}\Big)^{d/2} \nonumber\\
&& \quad\quad\times
e^{-\frac{\tISLL_1\tISLL_2(z_1-z_2)^2+\tISLL_2\tISLL_3(z_2-z_3)^2+\tISLL_3\tISLL_1(z_3-z_1)^2}
{2(\tISLL_1+\tISLL_2+\tISLL_3)}}
\frac{\tISLL_1\tISLL_2^2\tISLL_3}{(\tISLL_1+\cdots+\tISLL_m)^2}\delta^{(d)}(z_2-z_4)
\nonumber\\
&& -\frac{\PartNumb \BoltCons T}{8 \Volu_0} e^{-\LinkDens \LocaPart}
\sum_{m=3}^{\infty}\frac{(\LinkDens \LocaPart)^m}{m!}\frac{m(m-1)(m-2)(m-3)}{4}\int_{\tISLL_1,\ldots,\tISLL_m}
\Big(\frac{\tISLL_1\tISLL_2\tISLL_3\tISLL_4}{8\pi^3(\tISLL_1+\tISLL_2+\tISLL_3+\tISLL_4)}\Big)^{d/2} \nonumber\\
&& \quad\quad\times
e^{-\frac{\tISLL_1\tISLL_2(z_1-z_2)^2+\tISLL_1\tISLL_3(z_1-z_3)^2+\tISLL_1\tISLL_4(z_1-z_4)^2
+\tISLL_2\tISLL_3(z_2-z_3)^2+\tISLL_2\tISLL_4(z_2-z_4)^2+\tISLL_3\tISLL_4(z_3-z_4)^2}
{2(\tISLL_1+\tISLL_2+\tISLL_3+\tISLL_4)}}
\frac{\tISLL_1\tISLL_2\tISLL_3\tISLL_4}{(\tISLL_1+\cdots+\tISLL_m)^2} .\end{aligned}$$ Strictly speaking, the kernel $\KernTwo$ should be symmetric under the exchanges of variables $z_1 \leftrightarrow z_2$ or $z_3 \leftrightarrow z_4$. Here, to save space, we have written the above non-symmetric form. The true (i.e., symmetric) form can be recovered by averaging: $$\begin{aligned}
\label{EQ:KSym}
\KernTwo(z_1,z_2,z_3,z_4) \to
\frac{1}{4} \big(
\KernTwo(z_1,z_2,z_3,z_4)+
\KernTwo(z_1,z_2,z_4,z_3)+
\KernTwo(z_2,z_1,z_3,z_4)+
\KernTwo(z_2,z_1,z_4,z_3)
\big) .\end{aligned}$$
Relaxation of the phenomenological elastic free energy for a given realization of disorder {#APP:Relaxation}
==========================================================================================
In this Appendix we solve the stationarity condition for the random local deformations $\RelaRand$. First, we need to calculate the variation of the bulk term, which can be expanded, to leading order in small $\RelaRand$, as [^10] $$\begin{aligned}
\label{EQ:DetPPExp}
\textrm{det}\Big(\frac{\partial\DefoPosi_i(z)}{\partial z_j}\Big)
&=& \textrm{det} \Big( \Contraction \delta_{ij} + \partial_j \RelaRand_i(z) \Big) \nonumber\\
&=& \Contraction^{d} \textrm{det} \Big( \delta_{ij} + \Contraction^{-1}\partial_j \RelaRand_i(z) \Big) \nonumber\\
&\simeq& \Contraction^{d} \big(1+\Contraction^{-1} \partial_i \RelaRand_i(z)\big).\end{aligned}$$ Using this expansion, we have $$\begin{aligned}
\label{EQ:BulkTermExp}
&& \Big\lbrace \textrm{det}\Big(\frac{\partial\DefoPosi_i(z)}{\partial z_j}\Big)-1 \Big\rbrace^2 \nonumber\\
&=& (\Contraction^{d}-1)^2 + 2(\Contraction^d-1)\Contraction^{d-1}\partial_i \RelaRand_i(z) \nonumber\\
&& + \Contraction^{2d-2}\partial_i \RelaRand_i(z)\partial_j \RelaRand_j(z) .\end{aligned}$$ Thus, the stationarity equation reads
$$\begin{aligned}
0 &=& 2(\Contraction z_a + \RelaRand_a(z))\int dz_2 \NonLocaKern(z,z_2)
-2\!\int dz_2\NonLocaKern(z,z_2)(\Contraction z_{2,a} + \RelaRand_a(z_2))
-\BulkModuZeroP \partial_{a} (\partial_i \RelaRand_i(z)) ,\end{aligned}$$
where $$\begin{aligned}
\BulkModuZeroP \equiv \BulkModuZero \Contraction^{2d-2} .\end{aligned}$$ We take the disorder average of the nonlocal kernel $\NonLocaKernZero$ to be a zeroth-order quantity, and the fluctuation part $\NonLocaKernOne$ to be a first-order quantity and, thus, $\RelaRand(z)$ is also of first order. The he zeroth-order equation is then $$\begin{aligned}
0=2z_a\int dz_2 \NonLocaKernZero(z-z_2)-2\int dz_2 \NonLocaKernZero(z-z_2)z_{2,a}\, ,\end{aligned}$$ which is automatically satisfied, given that $\NonLocaKernZero(z-z_2)$ is even in $(z-z_2)$.
The first order equation reads $$\begin{aligned}
\! 0\! = \Contraction z_a \! \int\! dz_2 \NonLocaKernZero(z,z_2)
\!+ \RelaRand_a(z)\!\! \int\! dz_2 \NonLocaKernZero(z-z_2)
\!- \!\!\int \!\! dz_2 \NonLocaKernOne(z,z_2)\Contraction z_{2,a}
\!- \!\int \!\! dz_2 \NonLocaKernZero(z-z_2) \RelaRand_a(z)
\!- \frac{\BulkModuZeroP}{2} \partial_{a} (\partial_i \RelaRand_i(z)) . \,\end{aligned}$$ We address this equation in momentum space. We define the following Fourier transforms (on a specific finite volume—the volume of the state right after linking, viz., $\Volu_0$): $$\begin{aligned}
\NonLocaKernZero_{p} &=& \int dx e^{-ipx} \NonLocaKernZero (x) , \nonumber\\
\NonLocaKernOne_{p_1,p_2} &=& \int dx e^{-ip_1 x_1-ip_2 x_2}
\NonLocaKernZero (x_1,x_2) ,\end{aligned}$$ so that the momentum-space stationarity equation becomes $$\begin{aligned}
\label{mom-space-stat}
0 &=& i \Contraction \frac{\partial}{\partial p_{1,a}} \NonLocaKernOne_{p_1,0}
-i \Contraction \frac{\partial}{\partial p_{2,a}} \Big\vert_{p_2=0}
\NonLocaKernOne_{p_1,p_2}
+ (\NonLocaKernZero_{0}-\NonLocaKernZero_{p_1})\RelaRand_{a,p_1}
+ \frac{\BulkModuZeroP}{2}p_{1,a} p_{1,b}\RelaRand_{b,p_1} .\end{aligned}$$
Strictly speaking, the derivatives here should instead be understood as difference quotients because we are using a finite-volume version of the Fourier transform; but for convenience we write it a derivatives.
Equation (\[mom-space-stat\]) can be written in the tensorial form $$\begin{aligned}
\label{EQ:OrdeOneMomeTens}
\Big\lbrace
2\Big( \frac{\DiffG_p}{p^2}\Big) \IdenT
+\BulkModuZeroP \PLongT
\Big\rbrace \cdot \vert p \vert ^2 \vec{\RelaRand}(p)
= \vec{\RandForc}(p) ,\end{aligned}$$ where $$\begin{aligned}
\label{EQ:VectRelaRand}
\RandForc_{a,p_1}\equiv -2\Contraction \Big(
i\frac{\partial}{\partial p_{1,a}} \NonLocaKernOne_{p_1,0}
-i\frac{\partial}{\partial p_{2,a}}\Big\vert_{p_2=0} \NonLocaKernOne_{p_1,p_2}
\Big).\end{aligned}$$ This quantity $\RandForc_{a,p_1}$ is actually the random force in the state that is contracted but has not yet been equilibrated for the randomness, and $$\begin{aligned}
\DiffG_p &\equiv& \NonLocaKernZero_{0}-\NonLocaKernZero_{p}.\end{aligned}$$ Furthermore, $\IdenT$ is the $\Dime$-dimensional identity matrix, and the projection operators in momentum space, $\PLong_{ij}$ and $\PPerp_{ij}$, are defined as $$\begin{aligned}
\label{EQ:DefiProj}
\PLong_{ij} &\equiv& (p_i\,p_j)/p^2 , \nonumber\\
\PPerp_{ij} &\equiv& \delta_{i,j}-(p_i\,p_j)/p^2;\end{aligned}$$ They satisfy the following relations: $$\begin{aligned}
(\PLongT)^2 = \PLongT, \quad
(\PPerpT)^2 = \PPerpT, \quad
\PLongT \cdot \PPerpT =0 .\end{aligned}$$ In the following we shall use bold-face letters to denote rank-2 tensors, and letters with an overhead arrow (such as $\vec{\RelaRand}(p)$) to denote a vector, when needed.
By this decomposition we arrive at the solution to Eq. (\[EQ:OrdeOneMomeTens\]): $$\begin{aligned}
\label{EQ:SoluRelaRand}
\vec{\RelaRand}_{p}
&=& \frac{\PPerpT \cdot \vec{\RandForc}_{p}}
{2\DiffG_p}
+\frac{\PLongT \cdot \vec{\RandForc}_{p}}
{\BulkModuZeroP + 2\DiffG_p} .\end{aligned}$$ Notice that the second term is much smaller than the first term, due to the large bulk modulus $\BulkModuZeroP$. In the incompressible limit (i.e., $\BulkModuZero \to \infty$), we have $$\begin{aligned}
\vec{\RelaRand}_{p}=\frac{\PPerpT \cdot \vec{\RandForc}_{p}}
{2\DiffG_p} ,\end{aligned}$$ which is a purely transverse field, meaning that it satisfies $p_i\,\RelaRand_{i,p}=0$ or, equivalently, $\partial_i\,\RelaRand_{i}(x)=0$, which is the only deformation allowed in an incompressible medium.
Re-expanding the elastic energy around the equilibrium reference state {#APP:ReExpandFreeEner}
======================================================================
In this Appendix we re-expand the elastic energy for deformations relative to the relaxed state, $\tz=\Contraction z + \RelaRand(z)$, as discussed in Section \[SEC:EFERS\]. The small variable in this expansion is the deformation field $\tu(\tz)$ relative to the relaxed state. Furthermore, to obtain a continuum description of the elasticity, we adopt a notation involving the strain tensor $\StraTens_{ij}(x)$: $$\begin{aligned}
\!\StraTens_{ij}(x)\!&\equiv&\! \frac{1}{2}(\DefoGrad_{ij}(x)\DefoGrad_{ij}(x)-\delta_{ij})
\nonumber\\
\!&=&\! \frac{1}{2}(\partial_i \deformation_j(x)\!+\!\partial_j \deformation_i(x)
\!+\!\partial_i \deformation_l(x)\partial_j \deformation_l(x) ).\end{aligned}$$ where $\DefoGrad_{ij}(x)\equiv \partial \DefoPosi_i(x)/\partial x_j$ is the deformation gradient tensor. This strain tensor transforms as a tensor in the reference space labeled by $x$, and as a scalar in the target space labeled by $\DefoPosi$.
Expanding the nonlocal kernel $\tNonLocaKern$ in the relaxed state {#SEC:NLKRR}
------------------------------------------------------------------
The definition of $\tNonLocaKern$, given in Section \[SEC:EFERS\], is $$\begin{aligned}
\tNonLocaKern(\tz_1,\tz_2)&\equiv&\NonLocaKern(z(\tz_1),z(\tz_2)) .\end{aligned}$$ It can be expanded for small $\RelaRand$ to yield a direct expression for $\tNonLocaKern$. In momentum-space this reads
$$\begin{aligned}
\tNonLocaKern_{\tp_1,\tp_2}
=& \int d\tz_1 d\tz_2 e^{-i\tp_1 \tz_1-i\tp_2 \tz_2}
\tNonLocaKern(\tz_1,\tz_2)
= \!\int d\tz_1 d\tz_2 e^{-i\tp_1 \tz_1-i\tp_2 \tz_2}
\NonLocaKern(z(\tz_1),z(\tz_2)) \nonumber\\
=& \int dz_1 dz_2 \Jaco(z_1) \Jaco(z_2)
e^{-i\tp_1 (\Contraction z_1+\RelaRand(z_1))
-i\tp_2 (\Contraction z_2+\RelaRand(z_2))}
\NonLocaKern(z_1,z_2) ,\end{aligned}$$
where, in the first line, $z(\tz_1)$ is the mapping of a mass point $\tz_1$ in the relaxed state back to the position $z(\tz_1)$, at which it was located in the state right after linking. Inserting in the expressions for $\Contraction$ and $\RelaRand(z)$, given in Eqs. (\[EQ:SoluCont\], \[EQ:SoluV\]), and keeping terms to $O((1/\BulkModuZero)^0)$ \[which gives $\Contraction \simeq 1$ and $\Jaco(z)\simeq 1$\], we can expand $\RelaRand$ down from the exponent, and keep terms to first order in $\NonLocaKernOne$ (noting that $\RelaRand$ is the same order as $\NonLocaKernOne$), and thus arrive at $$\begin{aligned}
\label{EQ:tNonLocaKern}
\tNonLocaKern_{\tp_1,\tp_2}
\simeq& \int dz_1 dz_2
\big(1-i\tp_1 \RelaRand(z_1)
-i\tp_2 \RelaRand(z_2)\big) e^{-i\tp_1 z_1-i\tp_2 z_2}
(\NonLocaKernZero(z_1,z_2)+\NonLocaKernOne(z_1,z_2)) \nonumber\\
\simeq& \,\,\NonLocaKernZero_{\tp_1,\tp_2}+\NonLocaKernOne_{\tp_1,\tp_2}
-i \int dz_1 dz_2 \big(\tp_1 \RelaRand(z_1)
+\tp_2 \RelaRand(z_2)\big) e^{-i\tp_1 z_1-i\tp_2 z_2}
\NonLocaKernZero(z_1,z_2) \nonumber\\
=& \,\, \NonLocaKernZero_{\tp_1,\tp_2}+\NonLocaKernOne_{\tp_1,\tp_2}
-i \big(
\tp_1\cdot\RelaRand_{(\tp_1+\tp_2)} \NonLocaKernZero_{\tp_2}
+\tp_2\cdot\RelaRand_{(\tp_1+\tp_2)} \NonLocaKernZero_{\tp_1}
\big) .\end{aligned}$$
Local expansion of the harmonic attraction {#APP:REHT}
------------------------------------------
In this section we make a local expansion of the nonlocal term in the elastic free energy of the equilibrium reference state, Eq. (\[EQ:REExpaEner\]), i.e., the term $$\begin{aligned}
\label{EQ:FNL}
\FreeEnerPhen_{\textrm{nonlocal}} = & \,
\frac{1}{2} \!\int\! d\tz_1 d\tz_2 \Jaco(z_1)^{-1}\Jaco(z_2)^{-1}
\tNonLocaKern (\tz_1,\tz_2)
\Big(
\big\vert \tDefoPosi(\tz_1)-\tDefoPosi(\tz_2)\big\vert^2
- \big\vert \tz_1-\tz_2\big\vert^2
\Big) .\end{aligned}$$ For convenience of notation, we define the following change of variables: $$\begin{aligned}
z&=&z_1 ,\nonumber\\
y&=&z_2-z_1 ,\nonumber\\
\NLM(z,y)&\equiv&\tNonLocaKern(z_1,z_2),\end{aligned}$$ so that the nonlocal kernel in the relaxed state, Eq. (\[EQ:tNonLocaKernText\]), can be written (in momentum space) as $$\begin{aligned}
\label{EQ:tMinM}
\tNLM_{\tp,\tq}&\simeq&\NLMZero_{\tp,\tq}+\NLMOne_{\tp,\tq}
+ \tg_{\tp,\tq}\big( \RandForc_{\tp} \cdot \PPerpT \cdot \tq \big) ,\end{aligned}$$ with the definition of $\tg_{\tp,\tq}$, and then its leading-order expansion in momentum, being given by $$\begin{aligned}
\tg_{\tp,\tq}&\equiv&\frac{i(\NonLocaKernZero_{\tq}-\NonLocaKernZero_{\tp-\tq})}
{2(\NonLocaKernZero_{0}-\NonLocaKernZero_{\tp})} \simeq \frac{i(\tp^2-2\tp\cdot\tq)}{2\tp^2} .\end{aligned}$$ The local expansion of Eq. (\[EQ:FNL\]) then becomes $$\begin{aligned}
\FreeEnerPhen_{\textrm{nonlocal}}
&=&
\frac{1}{2} \int d\tz d\ty \, \tNLM (\tz,\ty) \Big(
\big\vert \tDefoPosi(\tz)-\tDefoPosi(\tz+\ty)\big\vert^2
- \big\vert y \big\vert^2
\Big) \nonumber\\
&\simeq& \frac{1}{2} \int d\tz
\big( \partial _i \tDefoPosi_l(\tz)\partial_j \tDefoPosi_l(\tz) -\delta_{ij}
\big) \int d\ty \,\, \ty_{i}\,\ty_{j} \, \tNLM(\tz,\ty),\end{aligned}$$ where the factor $\Jaco(z_1)^{-1}\Jaco(z_2)^{-1}$ is ignored because its difference from unity is of $O(1/\BulkModuZero)$. Now it is straightforward to express the elastic energy $\FreeEnerPhen_{\textrm{nonlocal}}$ in the standard form of Lagrangian elasticity using the strain tensor $\tStraTens_{ij}(\tz)=\frac{1}{2} \big( \partial _i \tDefoPosi_l(\tz)\partial_j \tDefoPosi_l(\tz) -\delta_{ij} \big)$.
The complete expression of the local form of the elastic energy for deformations relative to the relaxed state, including the contribution from the bulk term, will be calculated in Appendix \[APP:Lagr\].
Expansion of the bulk term
--------------------------
The bulk term in the elastic free energy Eq. (\[EQ:REExpaEner\]) is given by $$\begin{aligned}
\FreeEnerPhen_{\textrm{bulk}} \equiv \frac{\BulkModuZero}{2} \int d\tz \Jaco(z)^{-1}
\Big\lbrace
\Jaco(z)\, \textrm{det} \Big( \frac{\partial \tDefoPosi_i(\tz)}{\partial \tz_j}
\Big)-1
\Big\rbrace^2 .\end{aligned}$$ The determinant in this equation can be expanded using the strain tensor $\tStraTensT$: $$\begin{aligned}
\label{EQ:EXPDet}
\textrm{det} \Big( \frac{\partial \tDefoPosi_i(\tz)}{\partial \tz_j}\Big)
&=& \textrm{det} \big(\tDefoGradT(\tz)\big)
= \big\lbrace \textrm{det} \big(\IdenT+2\tStraTensT(\tz)\big)\big\rbrace^{1/2}
= e^{\frac{1}{2}\textrm{Tr}\, \ln \big(\IdenT+2\tStraTensT(\tz)\big)} \nonumber\\
&=& 1+ \textrm{Tr} \tStraTensT(\tz) -\textrm{Tr} \tStraTensT(\tz)^2
+\frac{1}{2} \big(\textrm{Tr} \tStraTensT(\tz)\big)^2 + O(\tStraTensT(\tz)^3).\end{aligned}$$ Thus, we have $$\begin{aligned}
\FreeEnerPhen_{\textrm{bulk}}
=&\, \frac{\BulkModuZero}{2} \int d\tz \Jaco(z)^{-1}
\Big\lbrace
\Jaco(z) \textrm{det} \Big( \frac{\partial \tDefoPosi_i(\tz)}{\partial \tz_j}
\Big)-1
\Big\rbrace^2 \nonumber\\
\simeq&\, \frac{\BulkModuZero}{2} \int d\tz \Jaco(z)^{-1}
\Big\lbrace
(\Jaco(z)-1)^2 +2(\Jaco(z)-1)\Jaco(z) \textrm{Tr} \tStraTensT (\tz)
\nonumber\\
& \quad\quad\quad - 2(\Jaco(z)-1)\Jaco(z) \textrm{Tr} \tStraTensT (\tz)^2
+ (2\Jaco(z)-1)\Jaco(z) (\textrm{Tr} \tStraTensT (\tz))^2
\Big\rbrace .\end{aligned}$$ Inserting the solutions for $\Contraction$ and $\RelaRand$, given in Eqs. (\[EQ:SoluCont\], \[EQ:SoluV\]), into the Jacobian $\Jaco (z) \equiv \big\vert \frac{\partial \tz_i}{\partial z_j}\big\vert$, we arrive at $$\begin{aligned}
\FreeEnerPhen_{\textrm{bulk}}
= \int d\tz \big\lbrace
\textrm{Tr}(\StreBulkT(\tz)\cdot\tStraTensT(\tz))
+ \SheaModu(\tz) \textrm{Tr} \tStraTensT (\tz)^2
+ \frac{\BulkModu(\tz)}{2} (\textrm{Tr} \tStraTensT (\tz))^2
\big\rbrace ,\end{aligned}$$ with the elastic parameters (in momentum space) being given by
$$\begin{aligned}
\StreBulk_{ij,p}
&=&\delta_{ij}\Big\lbrace
\frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\tp^2}
-\MeanSheaModu\Volu_0\delta_{\tp}
\Big\rbrace , \\
\SheaModu_{\tp}
&=& \MeanSheaModu\Volu_0\delta_{\tp} -
\frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\tp^2} , \\
\BulkModu_{\tp}
&=& \BulkModuZero\Volu_0\delta_{\tp} + 2\Big\lbrace
\frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\tp^2}
-\MeanSheaModu\Volu_0\delta_{\tp}
\Big\rbrace .\end{aligned}$$
Local form of the elastic energy relative to the relaxed state {#APP:Lagr}
--------------------------------------------------------------
Summing up the contributions from the nonlocal term $\FreeEnerPhen_{\textrm{nonlocal}}$ and the bulk term $\FreeEnerPhen_{\textrm{bulk}}$ to the elastic free energy (\[EQ:REExpaEner\]), we arrive at the local form of the elastic energy for deformations relative to the relaxed state: $$\begin{aligned}
\FreeEnerPhen
= \int d\tz \big\lbrace
\textrm{Tr}(\StreT(\tz)\cdot\tStraTensT(\tz))
+ \SheaModu(\tz) \textrm{Tr} \tStraTensT (\tz)^2
+ \frac{\BulkModu(\tz)}{2} (\textrm{Tr} \tStraTensT (\tz))^2
\big\rbrace ,\end{aligned}$$ with the elastic parameters being given by
$$\begin{aligned}
\Stre_{ij,\tp}
=&\, -\frac{\partial^2}{\partial \tq_i \partial \tq_j}
\Big\vert_{q=0} \NonLocaKernOne_{\tp-\tq,\tp}
+ i \delta_{ij} \frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\vert \tp \vert ^2}
-\frac{\RandForc_{a,\tp}}{\vert \tp \vert ^2}\big(
\tp_{i} \PPerp_{ja,\tp} + \tp_{j} \PPerp_{ia,\tp}
\big) , \\
\SheaModu_{\tp}
=&\, \MeanSheaModu \Volu_0 \delta_{\tp} -
\frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\vert \tp \vert ^2} , \\
\BulkModu_{\tp}
=&\, \BulkModuZero \Volu_0 \delta_{\tp} + 2\Big\lbrace
\frac{i\tp\cdot\vec{\RandForc}_{\tp}}{\vert \tp \vert ^2}
-\MeanSheaModu \Volu_0 \delta_{\tp}
\Big\rbrace .\end{aligned}$$
Relaxation of the deformed state {#APP:DefoRela}
================================
In this section we solve the stationarity equation with a given macroscopic deformation $\DefoGrad$, as discussed in Section \[SEC:NAD\], and thus obtain information about nonaffine deformations. The stationarity condition is given by
$$\begin{aligned}
2(\Contraction \DefoGrad_{ai}z_i+(\RelaRandL)_{a}(z)) \int dz_2 \NonLocaKern (z,z_2)
-2 \int dz_2 \NonLocaKern (z,z_2) (\Contraction \DefoGrad_{ai}z_{2,i}+(\RelaRandL)_{a}(z_2))
- \BulkModuZeroP \DefoGrad^{-1}_{ia}\DefoGrad^{-1}_{jb}\partial_i \partial_j (\RelaRandL)_{b}(z) =0 .\end{aligned}$$
We take $\NonLocaKernZero$ to be of zeroth order, and $\NonLocaKernOne$ and $\RelaRandL(z)$ to be first-order quantities. Thus, the zeroth order equation reads $$\begin{aligned}
0=2\DefoGrad_{ai}z_i\int dz_2 \NonLocaKernZero(z-z_2)-2\DefoGrad_{ai}\int dz_2 \NonLocaKernZero(z-z_2)z_{2,i} ,\end{aligned}$$ which is already satisfied given $\NonLocaKernZero(z-z_2)$ is even in $(z-z_2)$.
The first-order equation reads $$\begin{aligned}
&& \Contraction \DefoGrad_{ai}z_i \int dz_2 \NonLocaKernOne (z,z_2)
+(\RelaRandL)_{a}(z) \int dz_2 \NonLocaKernZero (z,z_2)
- \int dz_2 \NonLocaKernOne (z,z_2) \Contraction \DefoGrad_{ai}z_{2,i}
- \int dz_2 \NonLocaKernZero (z,z_2) (\RelaRandL)_{a}(z_2) \nonumber\\
&& - \frac{\BulkModuZeroP}{2} \DefoGrad^{-1}_{ia}\DefoGrad^{-1}_{jb}\partial_i \partial_j (\RelaRandL)_{b}(z) =0 .\end{aligned}$$ We can solve this equation in momentum space, where it is expressed as $$\begin{aligned}
0 &=& i \Contraction \DefoGrad_{ai} \frac{\partial}{\partial p_{1,i}} \NonLocaKernOne_{p_1,0}
-i \Contraction \DefoGrad_{ai} \frac{\partial}{\partial p_{2,i}} \Big\vert_{p_2=0}
\NonLocaKernOne_{p_1,p_2}
+ (\NonLocaKernZero_{0}-\NonLocaKernZero_{p_1})(\RelaRandL)_{a,p_1}
+ \frac{\BulkModuZeroP}{2} \DefoGrad^{-1}_{ia}\DefoGrad^{-1}_{jb} p_{1,i} p_{1,j}\RelaRand_{b,p_1} .\end{aligned}$$ Writing this equation in tensorial form, we have $$\begin{aligned}
\label{EQ:DefoRelaOOne}
\Bigg\lbrace
2\Big( \frac{D_p}{\vert p \vert ^2}\Big) \IdenT
+\BulkModuZeroP \DefoGradT^{-T} \PLongT \DefoGradT^{-1}
\Bigg\rbrace \cdot \vert p \vert ^2 (\Vect{\RelaRandL})_p
= (\Vect{\RandForcL})_p ,\end{aligned}$$ where $\MetrTens=\DefoGrad^{T} \DefoGrad$ is the metric tensor, and
$$\begin{aligned}
D_p &\equiv& \NonLocaKernZero_{0}-\NonLocaKernZero_{p}, \\
(\RandForcL)_{a,p_1} &\equiv& -2\Contraction \Big(
i \DefoGrad_{ai} \frac{\partial}{\partial p_{1,i}} \NonLocaKernOne_{p_1,0}
-i \DefoGrad_{ai} \frac{\partial}{\partial p_{2,i}} \NonLocaKernOne_{p_1,p_2}
\Big) .\end{aligned}$$
To solve Eq. (\[EQ:DefoRelaOOne\]), it is useful to define the $\DefoGradT$-deformed versions of the projection operators, i.e.,
$$\begin{aligned}
\PLongL &\equiv& \frac{1}{\textrm{Tr}(\PLongT \MetrTens^{-1})}
\DefoGradT^{-\textrm{T}} \PLongT \DefoGradT^{-1} , \\
\PPerpL &\equiv& \IdenT - \PLongL .\end{aligned}$$
It is straightforward to verify that they obey $$\begin{aligned}
(\PLongL)^2 = \PLongL, \quad
(\PPerpL)^2 = \PPerpL, \quad
\PLongL \cdot \PPerpL =0 .\end{aligned}$$ By using these projection operators we can write Eq. (\[EQ:DefoRelaOOne\]) as $$\begin{aligned}
\Big\lbrace
\frac{2 D_p}{\vert p \vert ^2} \PPerpL
+\Big(\frac{2 D_p}{\vert p \vert ^2}+\BulkModuZeroP \trOne \Big) \PLongL
\Big\rbrace\! \cdot \vert p \vert ^2 (\Vect{\RelaRandL})_p
= (\Vect{\RandForcL})_p ,\end{aligned}$$ where we have defined $$\begin{aligned}
\trOne \equiv \textrm{Tr}(\PLongT\MetrTens^{-1}) .\end{aligned}$$ Thus, it is straightforward to arrive at the solution: $$\begin{aligned}
(\Vect{\RelaRandL})_p = \Bigg\lbrace \frac{\PPerpL}{2 D_p}
+ \frac{\PLongL}{\BulkModuZeroP \trOne \vert p \vert ^2+2 D_p}
\Bigg\rbrace \cdot (\Vect{\RandForcL})_p .\end{aligned}$$
Correlation functions of the elastic parameters in the equilibrium reference state
==================================================================================
Correlation function of the non-local kernel in the equilibrium reference state {#APP:MM}
-------------------------------------------------------------------------------
The nonlocal kernel in the equilibrium reference state $\tNonLocaKern$ is related to the nonlocal kernel in the state right after linking $\NonLocaKern$ via Eq. (\[EQ:tGtu\]); to leading-order in the small quantity $\RelaRand$ we have
$$\begin{aligned}
\tNonLocaKern_{\tp_1,\tp_2}
\simeq \NonLocaKernZero_{\tp_1,\tp_2}+\NonLocaKernOne_{\tp_1,\tp_2}
-i \big(
\tp_1\cdot\vec{\RelaRand}_{(\tp_1+\tp_2)} \NonLocaKernZero_{\tp_2}
+\tp_2\cdot\vec{\RelaRand}_{(\tp_1+\tp_2)} \NonLocaKernZero_{\tp_1}
\big) .\end{aligned}$$
By using this relation, we can derive the correlation function of $\tNonLocaKern$ from the correlation function of $\NonLocaKern$ \[which is given in Eq. (\[EQ:KTwo\])\], and thus we arrive at the correlation function $$\begin{aligned}
&& \lda \tNLM_{p_1,q_1}\tNLM_{p_2,q_2} \rda_c \nonumber\\
&=& \delta_{p_1+p_2} \Big\lbrace
-\frac{\PartNumb \LinkDens \LocaPart^2}{2\BoltCons T} \int_{\ISLL_1,\ISLL_2}
\Big(\frac{1}{\ISLL_1}+\frac{1}{\ISLL_2}+\LinkScal^2\Big)^{-2}
\Big(
e^{-\frac{1}{2}\big(\frac{1}{\ISLL_1}
+\frac{1}{\ISLL_2}+\LinkScal^2 \big)\vert q_1+q_2\vert^2}
+ e^{-\frac{1}{2}\big(\frac{1}{\ISLL_1}
+\frac{1}{\ISLL_2}+\LinkScal^2 \big)\vert p_1-q_1+q_2\vert^2}
\Big)/2 \nonumber\\
&& \quad + \frac{\PartNumb}{(\BoltCons T)^2}
e^{-\LinkDens \LocaPart} \sum_{m=2}^{\infty} \frac{(\LinkDens \LocaPart)^m}{m!}
\frac{m(m-1)}{2}\int_{\ISLL_1,\ldots,\ISLL_m}\Big(
\frac{\tISLL_1 \tISLL_2}{\tISLL_1+\cdots+\tISLL_m}\Big)^2
\Big(
e^{-\frac{(\tISLL_1+\tISLL_2)\vert q_1+q_2\vert^2}{2\tISLL_1\tISLL_2}}
+ e^{-\frac{(\tISLL_1+\tISLL_2)\vert p_1-q_1+q_2\vert^2}{2\tISLL_1\tISLL_2}}
\Big)/2 \nonumber\\
&& \quad + \frac{\PartNumb}{(\BoltCons T)^2}
e^{-\LinkDens \LocaPart} \sum_{m=3}^{\infty} \frac{(\LinkDens \LocaPart)^m}{m!}
m(m-1)(m-2)\int_{\ISLL_1,\ldots,\ISLL_m}
\frac{\tISLL_1 \tISLL_2^2 \tISLL_3}{(\tISLL_1+\cdots+\tISLL_m)^2}
\Big(
e^{-\frac{1}{2}
\big(\frac{\vert p_1-q_1\vert^2}{\tISLL_1}
+\frac{\vert p_1-q_1+q_2\vert^2}{\tISLL_2}
+\frac{\vert q_2 \vert^2}{\tISLL_3}\big)} \nonumber\\
&& \quad\quad + e^{-\frac{1}{2}
\big(\frac{\vert q_1\vert^2}{\tISLL_1}
+\frac{\vert q_1+q_2\vert^2}{\tISLL_2}
+\frac{\vert q_2 \vert^2}{\tISLL_3}\big)}
+ e^{-\frac{1}{2}
\big(\frac{\vert p_1-q_1\vert^2}{\tISLL_1}
+\frac{\vert q_1+q_2\vert^2}{\tISLL_2}
+\frac{\vert p_1+q_2 \vert^2}{\tISLL_3}\big)}
+e^{-\frac{1}{2}
\big(\frac{\vert q_1\vert^2}{\tISLL_1}
+\frac{\vert p_1-q_1+q_2\vert^2}{\tISLL_2}
+\frac{\vert p_1+q_2 \vert^2}{\tISLL_3}\big)}
\Big)/4 \nonumber\\
&& \quad + \frac{\PartNumb}{(\BoltCons T)^2}
e^{-\LinkDens \LocaPart} \sum_{m=4}^{\infty} \frac{(\LinkDens \LocaPart)^m}{m!}
\frac{m(m-1)(m-2)(m-3)}{4}\nonumber\\
&&\quad\times \int_{\ISLL_1,\ldots,\ISLL_m}
\frac{\tISLL_1 \tISLL_2 \tISLL_3 \tISLL_4}{(\tISLL_1+\cdots+\tISLL_m)^2}
e^{-\frac{1}{2}
\big(\frac{\vert p_1-q_1\vert^2}{\tISLL_1}
+\frac{\vert q_1\vert^2}{\tISLL_2}
+\frac{\vert p_1+q_2 \vert^2}{\tISLL_3}
+\frac{\vert q_2 \vert^2}{\tISLL_4}\big)}\Big\rbrace \nonumber\\
&& +2i \delta_{p_1+p_2} q_1 \cdot \PPerpT_1 \cdot q_2 \Big\lbrace
-\frac{\PartNumb \LinkDens \LocaPart^2}{2\BoltCons T} \int_{\ISLL_1,\ISLL_2}
\Big(\frac{1}{\ISLL_1}+\frac{1}{\ISLL_2}+\LinkScal^2\Big)^{-1} \nonumber\\
&& \quad\quad \times
\Big\lbrack t_{p_1,q_1}
\big(
e^{-\frac{1}{2}\big(\frac{1}{\ISLL_1}
+\frac{1}{\ISLL_2}+\LinkScal^2 \big)\vert q_2\vert^2}
+ e^{-\frac{1}{2}\big(\frac{1}{\ISLL_1}
+\frac{1}{\ISLL_2}+\LinkScal^2 \big)\vert p_1+q_2\vert^2}
\big)/2 \nonumber\\
&&\quad -t_{-p_1,q_2}
\big(
e^{-\frac{1}{2}\big(\frac{1}{\ISLL_1}
+\frac{1}{\ISLL_2}+\LinkScal^2 \big)\vert q_1\vert^2}
+ e^{-\frac{1}{2}\big(\frac{1}{\ISLL_1}
+\frac{1}{\ISLL_2}+\LinkScal^2 \big)\vert -p_1+q_1\vert^2}
\big)/2
\Big\rbrack \nonumber\\
&& \quad + \frac{\PartNumb}{(\BoltCons T)^2}
e^{-\LinkDens \LocaPart} \sum_{m=2}^{\infty} \frac{(\LinkDens \LocaPart)^m}{m!}
\frac{m(m-1)}{2}\int_{\ISLL_1,\ldots,\ISLL_m}
\frac{\tISLL_1 \tISLL_2(\tISLL_1 + \tISLL_2)}{(\tISLL_1+\cdots+\tISLL_m)^2}
\nonumber\\
&& \quad\quad\times \Big\lbrack t_{p_1,q_1}
\big(e^{-\frac{(\tISLL_1+\tISLL_2)\vert q_2\vert^2}{2\tISLL_1\tISLL_2}}
+ e^{-\frac{(\tISLL_1+\tISLL_2)\vert p_1+q_2\vert^2}{2\tISLL_1\tISLL_2}}
\big)/2
- t_{-p_1,q_2}
\big(e^{-\frac{(\tISLL_1+\tISLL_2)\vert q_1\vert^2}{2\tISLL_1\tISLL_2}}
+ e^{-\frac{(\tISLL_1+\tISLL_2)\vert -p_1+q_1\vert^2}{2\tISLL_1\tISLL_2}}
\big)/2
\Big\rbrack \nonumber\\
&& \quad + \frac{\PartNumb}{(\BoltCons T)^2}
e^{-\LinkDens \LocaPart} \sum_{m=3}^{\infty} \frac{(\LinkDens \LocaPart)^m}{m!}
m(m-1)(m-2)\int_{\ISLL_1,\ldots,\ISLL_m}
\frac{\tISLL_1 \tISLL_2^2 \tISLL_3}{(\tISLL_1+\cdots+\tISLL_m)^2} \nonumber\\
&& \quad\quad\times \Big\lbrack t_{p_1,q_1} \big(
- e^{-\frac{1}{2} \big(\frac{1}{\tISLL_2}
+\frac{1}{\tISLL_3}\vert q_2\vert^2 \big)}
+ e^{-\frac{1}{2} \big(\frac{1}{\tISLL_2}
+\frac{1}{\tISLL_3}\vert p_1+q_2\vert^2 \big)}\big)/4 \nonumber\\
&&\quad - t_{-p_1,q_1} \big(
- e^{-\frac{1}{2} \big(\frac{1}{\tISLL_2}
+\frac{1}{\tISLL_3}\vert q_2\vert^2 \big)}
+ e^{-\frac{1}{2} \big(\frac{1}{\tISLL_2}
+\frac{1}{\tISLL_3}\vert - p_1+q_2\vert^2 \big)}\big)/4 \Big\rbrack
\Big\rbrace ,\end{aligned}$$
where we have used the notation $\NLM(z,y)\equiv\tNonLocaKern(z_1,z_2)$ defined in Appendix \[APP:REHT\].
Disorder correlators of the elastic parameters in the local form {#APP:CFLL}
----------------------------------------------------------------
In this appendix we calculate the disorder correlators of the quenched random elastic parameters in the local form of the elastic energy for deformations relative to the relaxed state.
First, we calculate the disorder correlator of the residual stress $\lda \Stre\Stre \rda_c$. The residual stress $\StreT$ in the relaxed state is related to the nonlocal kernel $\NonLocaKern$ via Eq. (\[EQ:Stress\]). Thus, the correlator of the residual stress can be expressed as
$$\begin{aligned}
\label{EQ:SSNN}
&& \lda \Stre_{ij,p_1}\Stre_{kl,p_2} \rda_c \nonumber\\
\!\! &=& \!\!\frac{\partial}{\partial q_{1,i}}\Big\vert_{q_1=0}
\frac{\partial}{\partial q_{2,j}}\Big\vert_{q_2=0} \lda N_{j,p_1,q_1} N_{l,p_2,q_2} \rda_c \nonumber\\
&&\!\!
- \frac{2}{\vert p_1 \vert^2} \big(p_{1,k}\PPerpT_{bl}(p_1)+p_{1,l}\PPerpT_{bk}(p_1)+p_{1,b}\PPerpT_{kl}(p_1)\big)
\frac{\partial}{\partial q_{1,i}}\Big\vert_{q_1=0} \lda N_{j,p_1,q_1} N_{l,p_2,0} \rda_c \nonumber\\
&&\!\!
+ \frac{2}{\vert p_1 \vert^2} \big(p_{1,i}\PPerpT_{aj}(p_1)+p_{1,j}\PPerpT_{ai}(p_1)+p_{1,a}\PPerpT_{ij}(p_1)\big)
\frac{\partial}{\partial q_{2,k}}\Big\vert_{q_2=0} \lda N_{j,p_1,0} N_{l,p_2,q_2} \rda_c \nonumber\\
&&\!\! - \frac{2}{(\vert p_1 \vert^2)^2}
\big(p_{1,i}\PPerpT_{aj}(p_1)\!+p_{1,j}\PPerpT_{ai}(p_1)\!+p_{1,a}\PPerpT_{ij}(p_1)\big) \!
\big(p_{1,k}\PPerpT_{bl}(p_1)\!+p_{1,l}\PPerpT_{bk}(p_1)\!+p_{1,b}\PPerpT_{kl}(p_1)\big)
\lda N_{j,p_1,0} N_{l,p_2,0} \rda_c , \,\end{aligned}$$
where we have defined $N_{j,p,q} \equiv {\partial \NLM_{p,q}}/{\partial q_j}$, and the notation $\NLM(z,y)\equiv\tNonLocaKern(z_1,z_2)$ is defined in Appendix \[APP:REHT\].
We then insert in the disorder correlator $\lda \NLM_{p_1,q_1}\NLM_{p_2,q_2} \rda _c$, given in Eq. (\[EQ:KTwo\]) (in the form of $\lda\NonLocaKern\NonLocaKern\rda$) into Eq. (\[EQ:SSNN\]). After a lengthy calculation, and making use of the identity $$\begin{aligned}
&& m\int_{\ISLL_1,\ldots,\tISLL_m}\frac{\tISLL_1^2}{(\tISLL_1+\cdots+\tISLL_m)^2} \nonumber\\
&& + m(m-1)\int_{\ISLL_1,\ldots,\tISLL_m}
\frac{\tISLL_1\,\tISLL_2}{(\tISLL_1+\cdots+\tISLL_m)^2}=1,\end{aligned}$$ we arrive at the correlator $$\begin{aligned}
&& \lda \Stre_{ij,p_1}\Stre_{kl,p_2}\rda_c
\nonumber\\
&=& \delta_{p_1+p_2} \frac{\PartNumb \UnivPara}{(\BoltCons T)^2}
(2\PPerpT_{ij}\PPerpT_{kl}+\PPerpT_{il}\PPerpT_{jk}+\PPerpT_{ik}\PPerpT_{jl}),\end{aligned}$$ where $\UnivPara\equiv -\frac{1}{2}{\LinkDens \LocaPart^2}+\LinkDens \LocaPart
+e^{-\LinkDens \LocaPart}-1$ is given in Eq. (\[EQ:UnivParaDef\]).
By following a similar scheme, we have also calculated the other disorder correlators and cross-correlators of the quenched random elastic parameters in the local form of elasticity of the relaxed state. Hence, we arrive at the correlators of the shear modulus and bulk modulus, which read
$$\begin{aligned}
\lda \SheaModu_{p_1}\SheaModu_{p_2} \rda_c =& \Corr \,\delta_{p_1+p_2} \PartNumb(\BoltCons T)^2 , \\
\lda \BulkModu_{p_1}\BulkModu_{p_2} \rda_c =& 4 \Corr\, \delta_{p_1+p_2} \PartNumb(\BoltCons T)^2 ,\end{aligned}$$
in which the dimensionless scale factor $\Corr$ is given by $$\begin{aligned}
\label{EQ:AppDefiCorr}
\Corr \equiv -\frac{3}{2}\LinkDens \LocaPart^2 + e^{\LinkDens \LocaPart}-1+\LinkDens \LocaPart+(\LinkDens \LocaPart)^2 .\end{aligned}$$ We also arrive at the cross-correlators, which are given by
$$\begin{aligned}
\lda \Stre_{ij,p_1} \SheaModu_{p_2}\rda_c &=&
-2 \PartNumb (\BoltCons T)^2 \UnivPara\, \delta_{p_1+p_2}\PPerpT_{ij}(p_1), \\
\lda \Stre_{ij,p_1} \BulkModu_{p_2} \rda_c &=&
4 \PartNumb (\BoltCons T)^2 \UnivPara\, \delta_{p_1+p_2} \PPerpT_{ij}(p_1), \\
\lda \SheaModu_{p_1} \BulkModu_{p_2} \rda_c &=&
-2 \PartNumb(\BoltCons T)^2\Corr\,\delta_{p_1+p_2};\end{aligned}$$
the scale factor $\UnivPara$ is defined in Eq. (\[EQ:UnivParaDef\]).
[31]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ed., ** (, ).
, ** (, ).
, , , ****, ().
, ** (, ).
, ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , ** (, ).
, ****, ().
, ** (, ).
, ****, ().
, ** (, ).
, , , ****, ().
, , , , ****, ().
, ** (, ).
, , , , ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, eds., ** (, ).
, , , ****, ().
, , , ****, ().
P. Erd[ő]{}s, A. R[é]{}nyi, Magyar Tud. Akad. Mat. Kut. Int. K[ö]{}zl. [**5**]{}, 17 (1960).
, , , , ****, ().
[^1]: This realization of rigidity should not be confused with *rigidity percolation*, which captures the rigidity of athermal (i.e., mechanical, rather than thermodynamic) networks [@Thorpe1999]. In the case of soft random solids, shear rigidity results from the entropy of thermal fluctuations of the positions of the constituent particles, and is proportional to the temperature
[^2]: [[This handshaking is the analog of that between the Born-Huang expansion for crystals and the continuum theory of elasticity, or that between the Newtonian equations of motion for particles and the Navier-Stokes equations of hydrodynamics. ]{}]{}
[^3]: [[In this preparation state, the temperature and the strength of the excluded-volume interaction can differ from those characterizing the measurement ensemble, and this has been discussed in Ref. [@Xing2004]]{}]{}
[^4]: For (unswelled) rubbery materials this is an appropriate assumption.
[^5]: Strictly speaking, the localized infinite cluster can undergo global translations or rotations from replica to replica, and this corresponds to a distinct equilibrium states of the field theory, which are connected by relative translations and/or rotations of the replicas.
[^6]: This relation is correct to linear order in $\GSDF$ or linear order in $n$.
[^7]: In earlier work (see Refs. [@Goldbart2004; @Mukhopadhyay2004]), the form of the Goldstone fluctuation was taken to be $ \VOPGS(\hat{x}) = \LocaPart \int d\ISLL \DistISLL(\ISLL)
\big(\ISLL/2\pi\big)^{(1+n)\Dime/2}
e^{-\frac{\ISLL}{2} \vert \hat{x}_{\tau} - \hat{\GSDF}_{\tau}(x_{\lambda})\vert^2}
- \LocaPart/\big(\Volu_0\Volu^n\big)$. This form (which we shall call the old Goldstone deformation) differs from the Goldstone deformation that we are currently using (the new Goldstone deformation) in two ways: (i) In the old Goldstone deformation, the deformation field was taken to be a function of $x_{\lambda}$ \[and, as a result, $z$ can be integrated out, as in the stationary-point form (\[EQ:OPInv\])\]. However, in the new Goldstone deformation, the deformation is instead taken to be function of $z$. The new Goldstone deformation is more physical, in the sense that the deformation field should be defined in terms of the the *mean* positions $z$ during thermal fluctuations, not the *instantaneous* positions of the particles $\hat{x}$. In the new Goldstone deformation, it is clear that it is the mean positions that are deformed, $\hat{z} \to \hat{\DefoPosi}(z)$, but the shape of the thermal cloud, which corresponds to a massive mode, is left untouched. (The point was already made in Ref. [@Ulrich2006], but we make it again, here, for the sake of completeness.)(ii) The deformation field in the old Goldstone deformation lies in the $\hat{x}_{\tau}$ direction, whereas the new Goldstone deformation has a deformation field in replicas $1$ through $n$. This issue has already been discussed in Sec. \[SEC:GSFT\], as has been the point that these two representations are related by a change of variables; see Eq. (\[EQ:VOPGSRTWO\]). The new Goldstone structure is more physical, in the sense that the preparation state (replica $0$) cannot be deformed once the sample has been made.
[^8]: To be consistent with the RLPM, we have used finite-volume versions of the Fourier transform and Kronecker delta function in momentum space; we shall take the continuum limit later on, in the final results in Section \[SEC:Local\]. Strictly speaking, the differentiations in Eq. (\[EQ:Stress\]) should be understood as the corresponding difference quotients.
[^9]: The dependence on the large-momentum cutoff is a result of keeping only terms of leading order at small momentum $p$ in the calculation of the disorder correlators given in Table \[TABLE:CorrTable\]. This enables us to extract the small-momentum behavior in momentum-space, which corresponds to the large-distance behavior in real-space.
[^10]: A similar expansion, but to higher order in $\RelaRand$, is performed in Eq. (\[EQ:EXPDet\]), in terms of the strain tensor $\StraTensT$.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The level of control that one has over neutral molecules in beams dictates their possible applications. Here we experimentally demonstrate that state-selected, neutral molecules can be kept together in a few mm long packet for a distance of over one mile. This is accomplished in a circular arrangement of 40 straight electrostatic hexapoles through which the molecules propagate over 1000 times. Up to 19 packets of molecules have simultaneously been stored in this ring structure. This brings the realization of a molecular low-energy collider within reach.'
author:
- 'Peter C. Zieger'
- 'Sebastiaan Y.T. van de Meerakker'
- 'Cynthia E. Heiner'
- 'Hendrick L. Bethlem'
- 'Andr[é]{} J.A. van Roij'
- Gerard Meijer
title: Multiple packets of neutral molecules revolving for over a mile
---
The important role that molecular beams have played ever since the early days of quantum mechanics stems from the exquisite control that one can exert over the internal and external degrees of freedom of neutral molecules in these beams. Advances in the control over molecular beams have gone hand in hand with new applications. In the 1930s, Rabi invented a method to unravel the quantum structure of molecules, based on the controlled manipulation of the trajectories of molecules on their way from an effusive source to the detector [@Rabi1939]. When the deflection fields from these original experiments were replaced by focusing elements, better control and higher densities were obtained in the molecular beam experiments, leading to the invention of the maser [@Gordon1955]. Improved control has also been obtained by upgrading the sources. Molecular beams with kinetic energies above one eV – sufficient to study and overcome reaction barriers in selected systems – were obtained by gas dynamic acceleration, i.e., by seeding small amounts of heavier molecules in light carrier gases [@Abuaf1967]. The cooling of the translational, rotational and vibrational degrees of freedom of molecules in these seeded mixed-gas expansions, has greatly reduced the complexity in molecular spectra, and has thereby revolutionized molecular spectroscopy [@Levy1981].
During the last decade, various methods have been demonstrated to control the forward velocity of molecules in beams. Decelerated beams of neutral molecules have been used, for instance, to reduce the transit-time broadening in high-resolution spectroscopy, to tune the collision energy in scattering studies, and to enable accurate measurements of lifetimes of long-lived states [@Meerakker2008]. Here, we report that the distance that these decelerated molecular beams travel can be extended to over one mile. This has been achieved by arranging a large number of straight electrostatic hexapoles in a circle. The motion of a compact packet of state-selected molecules through this circular arrangement is fully under control; the long distance travelled testifies to the intrinsic stability of the molecular trajectories.
The motivation for constructing the circular arrangement of straight hexapoles is its anticipated use as a low-energy collider; counter-propagating molecules meet many times when they revolve around the ring. This makes it possible to study low-energy (below $\approx$20 K) molecule-molecule collisions, probing the inter-molecular interaction potential in great detail [@RefColdMol]. Molecular storage rings of various designs have been theoretically analyzed [@Katz1997; @Nishimura2003], and a compact electrostatic storage ring, obtained by bending a linear hexapole focuser into the shape of a torus, has been experimentally demonstrated [@Crompvoets2001]. In this prototype storage ring, the molecules were transversely confined but longitudinally, i.e., along the circle, they were essentially free. Therefore, an injected packet of molecules spread out along the flight direction and eventually filled the entire ring. Collision studies can be most sensitively performed when the neutral molecules revolve in bunches, i.e., when they are both transversely and longitudinally confined. It has been recently demonstrated that the latter can be achieved in a storage ring composed of two half rings, separated by small gaps; packets of ammonia molecules were kept together by changing the electric fields synchronous with the molecules’ passage through the gaps [@Heiner2007].
In this prototype molecular synchrotron, the molecules could ultimately be kept together for about hundred round trips, implying that the molecules passed about 200 times through a gap, traveling a distance of about 80 meters during a total storage time of about one second [@Heiner2009a]. At this time the density of the stored packet was about two orders of magnitude below the value at the time of injection. In collision studies involving counter-propagating packets of molecules, losses in the number of stored molecules due to (in)elastic collisions accumulate over the interaction events. Let us consider $n$ packets of molecules revolving clockwise and $n$ packets revolving anti-clockwise, all with the same speed; this would require a structure consisting out of $n$ segments. After $m$ round trips, each packet of molecules would have interacted $2\cdot n\cdot m$ times with counter-propagating packets. As the summed intensity of the $n$ packets (revolving in either direction) can be recorded, an additional factor $\sqrt{n}$ will be gained in sensitivity. By tuning the speed with which the packets revolve, the total collision cross-section can be measured as a function of collision energy. Therefore, to operate it as a low-energy collider, the number of simultaneously stored packets – and thus the number of ring segments – as well as the number of completed round trips need to be made as large as possible.
Attempts to scale up the prototype molecular synchrotron to one consisting of many segments met with a number of serious problems: (i) Curved electrodes (as used in [@Heiner2007]) are difficult to machine to a high precision. This causes the alignment of the ring to be rather poor, and thereby limits the lifetime of the stored molecules. (ii) The original scheme to bunch the molecules involved switching between four different voltage configurations every time a packet passed through a gap [@Heiner2007]. For a ring consisting of many short segments the load on the high voltage switches would be unacceptable. Moreover, (iii) we have come to realize that this bunching scheme led to additional losses from the ring, either due to instabilities of the molecular trajectories or due to non-adiabatic transitions to non-trapped states. The molecular synchrotron presented here overcomes all these limitations.
A hexapole torus can be approximated by many short straight hexapole segments. From numerical simulations, we expect that the number of straight hexapoles can be reduced to below 100 without suffering a significant reduction in the transverse acceptance of the ring. As a compromise between construction demands and the expected performance, we have built a ring out of 40 straight hexapoles. A photograph of this ring is shown in the upper panel of Figure \[fig:BunchingScheme\], along with a zoom-in of three of the hexapoles. The diameter of the ring is 500 mm. Individual hexapoles are made by placing (nominally) 37.4 mm long electrodes, rounded off at each end, with a diameter of 4 mm on the outside of a circle with a radius of 3.54 mm [@Anderson1997]. Each hexapole spans an angle of 9 degrees of the ring. A difference of 1.5 mm between the lengths of the inner and outer electrodes is implemented to maintain a constant gap of 2 mm between adjacent hexapoles.
![(Color online) Upper panel: Photograph of the 0.5 m diameter molecular synchrotron with a zoom-in of three hexapoles. Middle panel, left: Voltage configuration used to transversely confine neutral molecules, overlayed with a contour plot of the resulting electric field strength. Right: Electric field strength (both for confinement and bunching) as a function of position along the equilibrium orbit. Lower panel: Scheme of the time-dependence of the voltage difference between adjacent hexapole electrodes during operation of the synchrotron. By temporarily switching to a higher voltage-difference whenever the packet passes through a gap, ammonia molecules with a velocity spread of one m/s are kept in a tight bunch while revolving. The green dashed curve indicates the equilibrium orbit of ND$_3$ molecules with a forward velocity of 125 m/s. \[fig:BunchingScheme\]](BunchingSchemeAndPicture.pdf){width="50.00000%"}
To confine the molecules transversely, a voltage difference of 6 kV is applied between adjacent hexapole rods, as shown in the middle panel of Figure \[fig:BunchingScheme\]. In this situation, the resulting electric field is cylindrically symmetric, increasing quadratically with distance from the center. Molecules in states that have a positive energy shift in the applied electric field – so-called low-field seekers – will experience a force towards the center of the hexapole; this force increases linearly with the radial distance when the Stark shift is linear with the electric field. In the ring, the molecules also experience a centrifugal force that depends on their forward velocity, and these forces will cancel at a certain position. In a perfect hexapole torus, a deuterated ammonia (ND$_3$) molecule in the low-field-seeking component of the $\mid$J,K$\rangle$=$\mid$1,1$\rangle$ state with a forward velocity of 125 m/s would be displaced radially outwards by 2.4 mm from the geometric center of the hexapole. In the present segmented ring, it is still a sufficiently good approximation that the equilibrium orbit for this molecule is a circle with a radius of 252.4 mm. Molecules flying with the same forward velocity but with a different radial position or with a nonzero radial velocity will oscillate around this equilibrium orbit with a transverse (betatron) frequency of around 1 kHz, i.e., these molecules will make 11.6 betatron oscillations per round trip.
Let us next look at the longitudinal motion of molecules on the equilibrium orbit. As is seen in the middle panel of Figure \[fig:BunchingScheme\], the electric field strength varies strongly along this orbit. This arises from the change in the distance of the equilibrium orbit from the centerline of the straight hexapole segment and from the presence of the gaps. The electric field in the middle of a gap drops to about half of the value near the end of a hexapole segment. Therefore, molecules in a low-field-seeking state will be accelerated in the fringe fields at the end of a hexapole segment, and will be decelerated in the fringe fields at the next hexapole segment. If the voltages are kept at a constant value, the deceleration will exactly cancel the acceleration and the molecules will keep their original velocity. In order to confine molecules in the longitudinal direction, we temporarily increase the positive voltages from +2 kV to +4 kV every time the molecules pass through a gap. As schematically indicated in Figure \[fig:BunchingScheme\], we increase the fields when a so-called ’synchronous’ molecule is a distance $s$ before the middle of the gap and decrease the fields when this molecule is the same distance $s$ after the middle of the gap. As before, the accelerating and decelerating forces on the synchronous molecule will be equal, and the velocity of the synchronous molecule will be the same after passing the gap. By contrast, a molecule that is slightly behind the synchronous molecule will enter the gap when the voltages are high, and it will leave the gap when the voltages are low. As a consequence, this molecule will experience a net acceleration. Vice versa, a molecule that is in front of the synchronous molecule will enter the gap when the voltages are low, and it will leave the gap when the voltages are high. This molecule will experience a net deceleration. This argument shows that molecules within a small position interval will experience a force towards the synchronous molecule and will oscillate around it; the molecules are hence trapped in a travelling potential well that revolves around the ring. From measurements and simulations of this bunching scheme, we have determined the longitudinal trap depth to be about 1 mK, and the trap (synchrotron) frequency to be on the order of 35 Hz, i.e., the molecules make 0.4 synchrotron oscillation per round trip. Note that in this bunching scheme the molecules experience a transverse confinement force at all times.
In our experiments, packets of deuterated ammonia ($^{14}$ND$_3$) molecules with a forward velocity of around 125 m/s are tangentially injected into the 40-segment molecular synchrotron. These packets are produced by Stark deceleration, followed by transverse focusing and longitudinal cooling, of a pulsed beam of ammonia, as described in detail elsewhere [@Heiner2006]. At the entrance of the synchrotron, each packet is several mm long and contains about a million molecules. All of these molecules are in the upper inversion doublet component of a single rotational level, $\mid$J,K$\rangle$=$\mid$1,1$\rangle$, in the vibrational and electronic ground state. At the injection point, four hexapoles are connected (in two sets of two hexapoles) to two separate high-voltage switches to allow for the injection of new packets of molecules without interfering with the ones already stored. These segments are switched off during injection of a new packet, and are switched to a small bias voltage during detection of the molecules. The remaining 36 hexapoles are connected to three other high voltage switches.
![Measurements of the density of ND$_3$ molecules as a function of time (in seconds) for a selected number of round trips. The observed temporal width of 21 $\mu$s after 1025 round trips corresponds to a packet length of 2.6 mm after a flight distance of more than a mile. The inset shows the exponential decay of the signal with time. \[fig:1milemeasurement\]](1mile-measurement-ohneVergleich.pdf){width="50.00000%"}
Figure \[fig:1milemeasurement\] shows the density of ammonia molecules as a function of time after injection into the synchrotron. Thirteen packets are injected at a rate of about 10 Hz, after which the loading is stopped. These packets then trail each other by a distance of three hexapoles; the first and the last injected packet are four hexapoles apart. The summed signal of the packets is recorded after a selected number of round trips, in time steps of 2.6 $\mu$s, using ionization with a focused laser beam at the injection point. The time-of-flight profiles shown are averaged over 16 of such measurement cycles. Even after 1025 round trips, i.e., after the molecules have travelled further than a mile and have passed through a gap 41,000 times, their signal can be clearly recognized; the temporal width of 21 $\mu$s corresponds to a packet length of 2.6 mm.
The density of ammonia molecules is seen to exponentially decay with time at a rate of 0.31 $s^{-1}$. This is the lowest decay rate that has been observed for neutral ammonia molecules in any trap to date. In all the early electrostatic trapping experiments, $1/e$-lifetimes of only a small fraction of a second have been observed. These were probably limited by non-adiabatic transitions to non-trappable states near the trap center, as was only realized when substantially longer lifetimes of up to 1.9 seconds were measured in an electrostatic trap with a non-zero electric field at the center [@Kirste2009]. With the present confinement and bunching scheme, the molecules are never close to a zero electric field in the ring either. In addition, we never change the direction – but only the magnitude – of the electric field in the ring, effectively preventing the occurrence of non-adiabatic transitions [@Wall2010]. A major contribution to the observed loss-rate is optical pumping of the ammonia molecules out of the $\mid$J,K$\rangle$=$\mid$1,1$\rangle$ level by blackbody radiation, calculated to occur at a rate of 0.14 $s^{-1}$ in the room-temperature chamber [@Hoekstra2007]. The remaining loss-rate of 0.17 $s^{-1}$ is well explained by collisions with background gas at the approximately 5$\cdot$10$^{-9}$ mbar pressure in the vacuum chamber.
![Time-of-flight measurement showing 19 packets of ammonia molecules revolving simultaneously in the molecular synchrotron. The horizontal axis shows the time relative to the time at which the first injected packet has completed 150 round trips. Each packet is labelled with two numbers: the main number labels the order of injection whereas the subscript labels the number of completed round trips. \[fig:19packets\]](19packets.pdf){width="50.00000%"}
Figure \[fig:19packets\] shows a measurement of nineteen packets of ammonia molecules with a forward velocity of 120 m/s being injected at about 10 Hz into the ring. The horizontal axis shows the time relative to the time at which the first injected packet has completed 150 round trips. The last injected packet has been in the ring for 320 ms and has made 24 round trips while the first injected packet has by then already been stored for more than 2 seconds. The loading scheme is set up such that each packet completes 7 round trips plus the length of 2 hexapoles before the next packet is injected. In the measurement the packets are seen to trail each other by 666 $\mu$s – precisely the time it takes them to fly through two hexapoles. In principle, it would be straightforward to store 40 packets simultaneously in the present circular arrangement, but in our current setup, we need to turn off at least two hexapoles to allow for new packets to enter the ring.
In conclusion, a synchrotron for neutral molecules consisting of 40 straight hexapoles together with a simple bunching scheme has been presented. Up to 19 packets of ammonia molecules have simultaneously been stored in this structure. The stored molecules can still be readily detected after they have completed a distance of over one mile, corresponding to a storage time of over 13 seconds. They have then passed – confined in a compact packet – 41,000 times through a gap, which explicitly demonstrates the stability of their trajectories. These measurements epitomize the level of control that can now be achieved over molecular beams and set the stage for novel experiments to come. For collision studies, the sensitivity of this ring is increased by more than three orders of magnitude over that of the prototype molecular synchrotron [@Heiner2007]. We are currently setting up a second beamline to inject counter-propagating packets into the ring.
We acknowledge the design of the electronics by G. Heyne, V. Platschkowski, and T. Vetter and the software support by U. Hoppe. H.L.B. acknowledges financial support from NWO via a VIDI-grant, and from the ERC via a Starting Grant.
I.I. Rabi, S. Millman, P. Kusch, and J.R. Zacharias, Phys. Rev., [**55**]{}, 526 (1939). J.P. Gordon, H.J. Zeiger, and C.H. Townes, Phys. Rev., [**99**]{}, 1264 (1955). N. Abuaf, J.B. Anderson, R.P. Andres, J.B. Fenn, and D.G.H. Marsden, Science, [**155**]{}, 997 (1967). D.H. Levy, Science, [**214**]{}, 263 (1981). S.Y.T. van de Meerakker, H.L. Bethlem, and G. Meijer, Nature Physics, [**4**]{}, 595 (2008), and references therein. R.V. Krems, W.C. Stwalley, and B. Friedrich, eds., [*Cold Molecules: Theory, Experiment, Applications*]{} (Taylor & Francis Group, Boca Raton, FL, 2009). D.P. Katz, J. Chem. Phys., [**107**]{}, 8491 (1997). H. Nishimura, G. Lambertson, J.G. Kalnins, and H. Gould, Rev. Sci. Instrum., [**74**]{}, 3271 (2003). F.M.H. Crompvoets, H.L. Bethlem, R.T. Jongma, and G. Meijer, Nature, [**411**]{}, 174 (2001). C.E. Heiner, D. Carty, G. Meijer, and H.L. Bethlem, Nature Physics, [**3**]{}, 115 (2007). C.E. Heiner, H.L. Bethlem, and G. Meijer, Chem. Phys. Lett., [**473**]{}, 1 (2009). R.W. Anderson, J. Phys. Chem. A, [**101**]{}, 7664 (1997). C.E. Heiner, H.L. Bethlem, and G. Meijer, Phys. Chem. Chem. Phys., [**8**]{}, 2666 (2006). M. Kirste, B.G. Sartakov, M. Schnell, and G. Meijer, Phys. Rev. A, [**79**]{}, 051401(R) (2009). T.E. Wall, S.K. Tokunaga, E.A. Hinds, and M.R. Tarbutt, Phys. Rev. A, [**81**]{}, 033414 (2010). S. Hoekstra, J.J. Gilijamse, B. Sartakov, N. Vanhaecke, L. Scharfenberg, S.Y.T. van de Meerakker, and G. Meijer, Phys. Rev. Lett., [**98**]{}, 133001 (2007).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Automatic Service Composition is a research direction aimed at facilitating the usage of atomic web services. Particularly, the goal is to build workflows of services that solve specific queries, which cannot be resolved by any single service from a known repository. Each of these services is described independently by their providers that can have no interaction with each other, therefore some common standards have been developed, such as WSDL, BPEL, OWL-S. Our proposal is to use such standards together with JSON-LD to model a next level of semantics, mainly based on binary relations between parameters of services. Services relate to a public ontology to describe their functionality. Binary relations can be specified between input and/or output parameters in service definition. The ontology includes some relations and inference rules that help to deduce new relations between parameters of services. To our knowledge, it is for the first time that parameters are matched not only based on their type, but on a more meaningful semantic context considering such type of relations. This enables the automation of a large part of the reasoning that a human person would do when manually building a composition. Moreover, the proposed model and the composition algorithm can work with multiple objects of the same type, a fundamental feature that was not possible before. We believe that the poor model expressiveness is what is keeping service composition from reaching large-scale application in practice.'
address: |
Alexandru Ioan Cuza Universy of Iaşi\
Faculty of Computer Science, Iaşi, Romania
author:
- Paul Diac
- Liana Ţucăr
- Andrei Netedu
bibliography:
- 'kes\_relational.bib'
title: |
Relational Model for Parameter Description in\
Automatic Semantic Web Service Composition
---
web service composition; semantic web; relational concepts; automatic composition algorithm; knowledge-based systems
Introduction {#sec:intro}
============
Semantic models used for Automatic Service Composition were developed some years ago, but their popularity slightly decreased more recently. We believe this is motivated by the lack of a model that is both expressive enough to allow the reasoning involved in manual composition, and simple enough to be easily adopted by service developers. One of the main shortcomings of previous models is that they do not define relationships over service parameters, which are intuitive for users creating manual compositions. Our contribution intends to take this next step.
The proposed model is designed in the context of stateless, information-providing services; and in applications where services parameter information is well structured. Services are known ahead of the composition, so *discovery* is not needed and providers adhere to the same ontological model. Particularly, service parameters are defined over a known ontology’s set of concepts enhanced with a set of *relations* specified in service definitions. Possible parameter *relations* are declared in the common reference ontology. Relations between inputs are restrictions or preconditions, and relations between output parameters are generated after calling the service; similar to postconditions. The ontology also includes a set of *inference rules*, conceptually similar with rules described in fundamental works like [@berners2001semantic] or [@sintek2002triple]. Rules are defined at ontology level only, and services cannot declare new rules. Rules are generally available and applicable anywhere as long as their *premise*s hold. These rules also help to achieve the desired expressiveness and model non-trivial examples.
Our main contribution is adding relations to the model, with properties and inference rules. Therefore we extended the semantics in a natural manner increasing the expressiveness of classic models first described in [@mcilraith2001semantic] and used in many other works, with variations such as [@bansal2008wsc], [@weise2008different] or [@bansal2016generalized]. Addressing the relational aspect of parameters was not tackled before in any work we know of, for example it is not mentioned in survey [@klusch2016semantic] on semantic services. There are a few works like [@lee2011scalable] that use relational databases to resolve the composition queries, however, the “relations” are only at services level. A more recent formalism presented in [@viriyasitavat2019extension] extends the semantic model with elaborate constructs, that seem hard to be implemented in practice, and yet does not introduce our type of parameter relations. Our paper also describes the implementation and synthetic evaluation of a proposed algorithm that searches a valid composition on this model. We motivate our approach with examples where classical composition fails to model human reasoning.
Section \[sec:example\] presents such a motivating example. The problem definition constitutes the following Section, \[sec:problem\], that first describes our model textually and then formally states the problem requirements. In Section \[sec:standards\] we describe the standards we use to serialize services. The algorithm that computes the composition on the defined model is presented in Section \[sec:algorithm\] and performance is evaluated in Section \[sec:evaluation\]. The last Section \[sec:conclusion\] concludes the paper and suggests some ideas for continuation.
Motivating Example {#sec:example}
==================
We analyze the following use case where **relational semantic** composition can model more complex interactions than simple or syntactic models based for example just on a hierarchy of concepts. More precisely, our addition of *relationships* and *inference rules* has the most important role, enabling the required expressiveness.
Assume a researcher is trying to schedule a meeting at some collaborating university. We are provided with the following information: the person’s name, the name of the university where the researcher works, and the name of the university to visit. We consider the person and the two universities as instances of two concepts: **Person** and **University**. As we can already see, there is a need to distinguish between the two different instances of **University** and we are going to model that by the use of two different types of relations between our **Person** and each : **isEmployeeOf** and **hasDestination**. Finally, we use a third relation: **isLocatedIn** and two inference rules that can help to expand relations. For example, if **X** is an employee of **Y** and **Y** is located at **Z** then **X** is located in **Z** (this may not be true in any situation, but it is a reasonable assumption that works for our example). In the composition model and in the further presented algorithm we can handle inference rules in the same manner as web services that “return” only new *relations* defined on already known “parameters”, for simplicity. If the services cost, execution time, throughput or other Quality of Service is considered, rules are services with zero cost or instant runtime, unlimited throughput.
More precisely, we have the following web services and inference rules (with names ending in $Rule$). We specify in a service definition the required relations between input parameters and generated relations between output and possibly input parameters:
$\\ getUniversityLocation_{\textstyle\hspace{-1cm} output=\Big\{\begin{tabular}{c} $city : City,$ \\ $isLocatedIn(univ, city)$ \end{tabular}\Big\}}^{\textstyle \hspace{5pt} input=\{univ : University\}} \hspace{1cm} getAirplaneTicket_{\textstyle output=\{airplaneTicket : Ticket\}}^{\textstyle\hspace{-1.3cm}input=\Bigg\{\begin{tabular}{c}$pers : Person, souce, dest : City,$ \\ $isLocatedIn(pers, source),$ \\ $hasDestination(person, dest)$\end{tabular}\Bigg\}} $
$\\ locatedAtWorkRule_{\textstyle \hspace{5pt} output=\{isLocatedIn(X, Z)\}}^{\textstyle input=\Bigg\{\begin{tabular}{c} $X, Y, Z,$ \\ $isEmployeeOf(X, Y),$ \\ $isLocatedIn(Y, Z)$ \end{tabular}\Bigg\}} \hspace{1cm} destinationGenRule_{\textstyle output=\{hasDestination(X, Z)\}}^{\textstyle \hspace{-0.3cm} input=\Bigg\{\begin{tabular}{c} $X, Y, Z,$ \\ $hasDestination(X, Y),$ \\ $isLocatedIn(Y, Z)$ \end{tabular}\Bigg\}} $
The solution for a composition is a list of services that can be called in the order from the list, and after all calls, the information required by the user is known. We also specify how to pass the output of a service to the input of another. This was trivial on previous models, but now as multiple instances of the same concept can be known, we need to distinguish between them, based on their *relations* with other objects.
In Figure \[fig:largeExample\], solid boxes represent services and the user request (one for input - the initially known concepts and their relations, and one for the required output). Dashed boxes represent inference rules. Edges show information flow: solid edges - parameters and the dashed edges - relationships among them. Not all possible relations are used. Parameters are matched to rule variables (rule “parameters”), or other service parameters, based on the specification in gray in parenthesis. Multiple calls to the same service can be handled and are shown with double edges.
One composition representing a valid solution for the above instance would be the following list of service invocations, in order:
**getInput**($\emptyset$) $\Longrightarrow$ pers, homeUniv, foreignUniv, isEmployeeOf(pers, homeUniv), hasDestination(pers, foreignUniv);
**getUniversityLocation**(homeUniv) $\Longrightarrow$ homeCity, isLocatedIn(homeUniv, homeCity);
**getUniversityLocation**(foreignUniv) $\Longrightarrow$ foreignCity, isLocatedIn(foreignUniv, foreignCity);
The two cities: *homeCity* and *foreignCity* are differentiated based on their relations, the names are not relevant for the composition (there is no restriction on what names they get if they would be automatically created; i.e. any distinct strings would work).
**locatedAtWorkRule**(pers, homeUniv, homeCity, isEmployeeOf(pers, homeUniv), isLocatedIn(homeUniv, homeCity)) $\Longrightarrow$ isLocatedIn(pers, homeCity);
**destinationGenRule**(pers, foreignUniv, foreignCity, hasDestination(pers, foreignUniv), isLocatedIn( foreignUniv, foreignCity)) $\Longrightarrow$ hasDestination(pers, foreignCity);
**getAirplaneTicket**(pers, homeCity, foreignCity, isLocatedIn(pers, homeCity), hasDestination(pers, foreignCity) $\Longrightarrow$ airplaneTicket;
We can immediately notice the usefulness of semantic relations as the pairs *homeUniv* and *foreignUniv*, as well as *homeCity* and *foreignCity* are essentially indistinguishable between themselves otherwise: if knowledge would consist only of known types as before.
If we did not use semantic relations, to get the same desired functionality we would copy same services separating them for each possible instance of input parameters, for example **getUniversityLocation** could be split into two different services: **getForeignUniversityLocation** and **getHomeUniversityLocation**. We can imagine cases where this workaround raises the number of services provided at input by a high amount for each group of different input parameters.
In this example we did not use hierarchical concepts. However, *subsumption* can be used together with presented additions to the model of composition, as defined in the next Section \[sec:problem\]. Overall, the simple hierarchy model proves to be easier to design and solve than our proposal.
Problem Model Formulation {#sec:problem}
=========================
Informal problem description
----------------------------
The first method of adding semantics to Web Service Composition was done by mapping parameters of services to a common *ontology* of concepts. More precisely, this was a simple *taxonomy* or tree of concepts, like in the first Semantic Composition Challenge [@bansal2008wsc]. Parameter matching, in this case, is based not just on name - string equality, as in the previous model; but on the *subsumes* relation. Intuitively, this means that one output of a service can be used as input for another service if the latter is a superconcept of the former. A subconcept can then replace a more generic concept, e.g., a species of a tree like *oak* can be used as an input of a service that requires any type of *plant*.
The proposed approach is increasing the expressiveness of the classic model by introducing *binary relations*, defined over instances of concepts. *Relations* are identified by a unique name that is a string. Within the ontology, there are also a set of relation *properties* and *inference rules*. The properties and rules are important because they enable the inference of new relations over parameter instances that are not from within a single service definition.
From one service definition, some relations between parameters are easily deduced by a human person because they are semantically implied by the name and description of the service and the name of parameters. A person would then naturally reason with the supposed relations when composing services. This must be considered by automatic composition as well.
Formal problem definition
-------------------------
**Concept**. A concept $c$ identified by a *conceptName* is an element of the given set of concepts $C$. Concepts are arranged in a hierarchy, by the use of inheritance. Specialization concepts (or sub-types) can replace their more generic version (or super-types).
**Object**. An object is defined by a name *objectName* and a type *objectType* $ \in C$. Intuitively, objects exist in the dynamic context of a structure of services that have been called. It is similar to an instance of a *Concept*, for which we know how it was produced and used in a workflow, or a series of service calls with their inputs matched by objects. Let $O$ be the set of all objects known in the current composition.
**Relation**. A binary relation over the objects is identified by a unique *relationName* and is a subset of $O \times O$. Relations are not restricted on types.
**Relation properties**. A relation can have none, one or both of the properties: *symmetry* and *transitivity*. A symmetric relation *rel* it is not oriented from the first object to the second, i.e., if *rel(o1, o2)* holds then *rel(o2, o1)* holds as well. Transitivity implies that if *rel(o1, o2)* and *rel(o2, o3)* hold then *rel(o1, o3)* holds.
**Inference rules**. An inference rule is composed of a *premise* and a *conclusion*. Both are a set of conjunction connected relations over variables of the set of objects $O$. Variables use a local name that has relevance only within the definition of this rule.
Inference rules can specify properties of relations, but we highlighted the frequently used *symmetry* and *transitivity* as properties, for simplicity. This also helps to optimize the composition search algorithm.
**Ontology**. An ontology is constituted by concepts $C$, a set of $relations$ names with $properties$ and $rules$. Only names and properties of *relations* are part of the ontology; the *objects* that are in a given *relation* depend on the context of the further defined composition.
**Web Service**. A Web Service is a triple $(input$, $output$, $relations$). The $input$ and $output$ are **disjoint** sets of **parameters**. One parameter is identified by a *parameterName* and has a type $\in C$. $relations$ are restricted to local service parameters: $\subseteq$ $(input $ $\cup$ $ output) \times (input $ $ \cup $ $ output)$, i.e. a service defines relations only for its set of parameters. Similar to *rule* variable names, *service* parameter names have relevance only within the current service definition. The relations between input parameters of a service have to hold before the service is called, or, more precisely added to the composition and can be understood as *preconditions*. All the rest of the relations within a service definition hold after the service is called, they can be between objects created by this service call and are similar to *postconditions*.
The **user request** structure is similar to a web service. The *input* specifies what the user initially knows, that are objects with types and possible relations between them, and the *output* is the user’s desired objects with required relations. Like in a service, there can be relations between inputs and outputs, specifying restrictions on what the user required outputs relative to his initially known object types.
**Relation based Web Service Composition Problem**. Given an ontology defining *concepts*, *relations* with *properties* and *inference rules*, a user request and repository of services, defined on the ontology; find an ordered list of services that are callable in that order, that solves the user request, starting with the initial information. For each service in the composition, the source of its input parameter must be clearly specified: resolved of bound to an output of a specified previously called service or user query. **Further clarifications** on the proposed model:
### Services and Rules
Services and inference rules are structurally similar, with the distinction that: services must output at least some new parameter and cannot generate only new relations. Services do not “generate” relations between their input ($input \times input$) parameters. If their definition specifies relations between input parameters, they represent restrictions, i.e., conditioning the calling of the service. Rules however never “generate” any parameters but only new relations based on outputs of previously called services. Rules variables are not restricted by type i.e., they do not have types.
### Parameters and Types
Each parameter of a service has a type, but unlike previous models, objects of the same type can be differentiated by their relations. This is fundamental in manual compositions, or other types of workflows, and we believe the impossibility to express this was the main flaw of previous models. This greatly increases the problem computational complexity, but as we will see in Section \[sec:algorithm\], our proposed algorithm is still able to find compositions on non-trivial instances of significant size, that a human user could not work with.
### Objects
Objects are defined by their types, and all relations they are in. An object is similar to an instance of that type but is not an actual value or an instantiation. Objects are at the abstract level, they are dynamic elements that pass through the composition stages, and we keep information about their semantic context, through relations. On the current model, it is useless to have multiple objects of the same type with the same set of relations, so we keep only one for each possible context. Even with this reduction, the total possible number of objects is exponential compared to the number of relations (any subset of relations can define a different object). Moreover, when considering a relation of the current object, the type of the pair object on which the relation is defined, is also relevant to the current object state. This is motivated by parameter matching mechanism, without considering this the model (or algorithm) might fail to find valid compositions even if they would exist.
Standards {#sec:standards}
=========
The information required to work with data on this model is serialized by a set of WSDL and JSON-LD files and an XML extension file. In order to add rules over parameters, the WSDL is extended with a type of field that defines relations. JSON-LD is used to describe types, and for inference rules an additional XML file is added. In short there are only three files: the semantic description of the concepts (types and relations), the inferences and the repository. The user query is serialized exactly as a service.
**Ontology**. Below we show the serialization of *Person*, as subtype of *Thing*. The JSON-LD file describes both types and relations. A relation has two fields that describe if it has the properties *transitivity* or *symmetry*.
**Inference Rules** are similar to a service since they have input and output variables, equivalent of parameters. Since it makes no sense to use WSDL, the input and output XML nodes are not necessary nodes.
**Repository**: below is the *getAirplaneTicketService*. The WSDL description is extended in order to incorporate rules in the definition. The first message to appear in the service contains the input parameters for the service, while the next contains the output parameters. Lastly, a relation is described using a source, a target and name.
[ c | c | c ]{} **Ontology Format** & **Inference Rules Format** & **Repository Format**\
``` {language="json"}
{ "@graph": [
{ "@id": "Person",
"@type": "rdfs:Class"
"rdfs:subClassOf": {
"@id": "Thing"
}
},
{ "@id": "IsLocatedIn",
"@type": "rdfs:Class"
"rdfs:subClassOf": {
"@id": "Relation"
},
"isTransitive": true,
"isSymetric": false
}, ...
] }
```
&
``` {.xml language="XML"}
<inference
name = "locatedAtWorkRule">
<input>
<part name = "X"/>
<part name = "Y"/>
<part name = "Z"/>
<relation name = "IsEmployeeOf"
source = "X"
target = "Z">
<relation name = "IsLocatedIn"
source = "Y"
target = "Z">
</input>
<output>
<relation name = "IsLocatedIn"
source = "X"
target = "Z">
</output>
</inference>
```
&
``` {.xml language="XML"}
<service name = "getAirplaneTicket">
<message
name = "getAirplaneTicketInput">
<part name = "pers" type = "Person"/>
<part name = "source" type = "City"/>
<part name = "dest" type = "City"/>
</message>
<message
name = "getAirplaneTicketOutput">
<part name = "airplaneTicket" =
"xsd:Reservation"/>
</message>
<relation source = "pers"
target = "source"
name = "IsLocatedIn"/>
<relation source = "pers"
target = "dest"
name = "HasDestination"/>
</service>
```
Composition algorithm {#sec:algorithm}
=====================
The proposed algorithm is inspired from our previous work [@ctucuar2018semantic]. Previously, the focus was on the computational efficiency. The classic model from [@bansal2008wsc] was solved, with a simple hierarchy of types. To solve the composition on the new model, we prioritized less the efficiency, and more the inclusion of the extra functionality of the much more complicated semantic inference stage and parameter matching. Another choice was to avoid reducing the instance to any other known problem like *planning* or other solvers with reasoning but directly solve the instance by building the composition step by step and expanding information.
(60.43,34.97) .. controls (60.43,30.02) and (64.44,26) .. (69.4,26) – (136.39,26) .. controls (141.35,26) and (145.36,30.02) .. (145.36,34.97) – (145.36,99.03) .. controls (145.36,103.98) and (141.35,108) .. (136.39,108) – (69.4,108) .. controls (64.44,108) and (60.43,103.98) .. (60.43,99.03) – cycle ; (59.5,54) – (146.5,53.67) ;
(252,17.3) .. controls (252,12.16) and (256.16,8) .. (261.3,8) – (327.73,8) .. controls (332.87,8) and (337.03,12.16) .. (337.03,17.3) – (337.03,117.37) .. controls (337.03,122.5) and (332.87,126.67) .. (327.73,126.67) – (261.3,126.67) .. controls (256.16,126.67) and (252,122.5) .. (252,117.37) – cycle ; (252,34.58) – (337.5,34.58) ;
(269.85,61.16) .. controls (269.85,54.35) and (282.58,48.82) .. (298.27,48.82) .. controls (313.97,48.82) and (326.7,54.35) .. (326.7,61.16) .. controls (326.7,67.98) and (313.97,73.5) .. (298.27,73.5) .. controls (282.58,73.5) and (269.85,67.98) .. (269.85,61.16) – cycle ; (269.85,101.98) .. controls (269.85,95.17) and (282.58,89.64) .. (298.27,89.64) .. controls (313.97,89.64) and (326.7,95.17) .. (326.7,101.98) .. controls (326.7,108.8) and (313.97,114.33) .. (298.27,114.33) .. controls (282.58,114.33) and (269.85,108.8) .. (269.85,101.98) – cycle ; (430.89,22.13) .. controls (430.89,16.53) and (435.43,12) .. (441.02,12) – (513.37,12) .. controls (518.97,12) and (523.5,16.53) .. (523.5,22.13) – (523.5,114.87) .. controls (523.5,120.47) and (518.97,125) .. (513.37,125) – (441.02,125) .. controls (435.43,125) and (430.89,120.47) .. (430.89,114.87) – cycle ; (430,39) – (523.5,39) ;
(450.79,65) .. controls (450.79,57.82) and (462.91,52) .. (477.86,52) .. controls (492.8,52) and (504.92,57.82) .. (504.92,65) .. controls (504.92,72.18) and (492.8,78) .. (477.86,78) .. controls (462.91,78) and (450.79,72.18) .. (450.79,65) – cycle ; (450.79,99) .. controls (450.79,91.82) and (462.91,86) .. (477.86,86) .. controls (492.8,86) and (504.92,91.82) .. (504.92,99) .. controls (504.92,106.18) and (492.8,112) .. (477.86,112) .. controls (462.91,112) and (450.79,106.18) .. (450.79,99) – cycle ; (326.7,61.16) .. controls (369.48,75.76) and (365.54,52.9) .. (449.52,64.82) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(326.7,98.98) .. controls (379.24,114.59) and (365.64,92.89) .. (449.52,98.91) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(129.58,82) .. controls (171.29,75.7) and (154.67,56.86) .. (268.13,61.1) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(129.58,82) .. controls (187.21,119.48) and (181.56,86) .. (268.53,101.74) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(73.42,82) .. controls (73.42,74.82) and (85.99,69) .. (101.5,69) .. controls (117.01,69) and (129.58,74.82) .. (129.58,82) .. controls (129.58,89.18) and (117.01,95) .. (101.5,95) .. controls (85.99,95) and (73.42,89.18) .. (73.42,82) – cycle ;
(106.84,41) node \[align=left\] [**Person**]{}; (101.5,82) node \[align=left\] [pers1]{}; (296.16,22.24) node \[align=left\] [**University**]{}; (298.27,61.16) node \[align=left\] [univ1]{}; (298.27,101.98) node \[align=left\] [univ2]{}; (476.53,27) node \[align=left\] [**City**]{}; (477.86,65) node \[align=left\] [city1]{}; (477.86,99) node \[align=left\] [city2]{}; (383.5,50.32) node \[scale=0.7,color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-354.18\] \[align=left\] [isLocatedIn\
(univ1, city1)]{}; (386,88) node \[scale=0.7,color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-352.81\] \[align=left\] [isLocatedIn\
(univ2, city2)]{}; (199,48) node \[scale=0.7,color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-355.2\] \[align=left\] [isEmployeeOf\
(pers1,univ1)]{}; (202,86) node \[scale=0.7,color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-357.81\] \[align=left\] [hasDestination\
(pers1, univ2)]{};
During the construction of the composition, a **knowledge base** is kept in memory, that is constituted of **objects**. An object has a known type, that is a concept from the ontology, and also a set of *relations* defined on the object. For example objects in Figure \[fig:knowledgeFig\] represent a knowledge base. For each relation, the relation type and the pair object with which the relation is defined are kept in memory. Based on this information, the next service is chosen together with a matching between known objects and service inputs. The matching must satisfy the defined relations between service input parameters. After this fictive “call” (conceptually, the call is done by adding the service to the composition), the knowledge base is updated with the output of the service. New objects can be created after a call, and for known objects, new relations can be added. After calling services, all inference rules are processed to see if they can be applied based on new objects or relations.
To check if in a given **knowledge** state some service can be validly called, a backtracking procedure is implemented. For each service input parameter in order, all *objects* of type or sub-types of the parameter type are iterated. The algorithm checks all the relations with objects that have already been matched with service parameters like shown from left to right in Figure \[fig:inputmatching\]. Branches that do not satisfy any relation between completed levels are dropped. If finally, the last parameter could be matched with some object, then the match is complete, and the service can be called. Also, to avoid repeated loops, a history of calls is kept for every service, that includes matched objects (by using a hash value computed over the objects). Using the history of calls, calling services with parameters that have been used before is avoided.
(307.79,39.71) .. controls (309.46,41.38) and (309.46,43.04) .. (307.79,44.71) .. controls (306.12,46.38) and (306.12,48.04) .. (307.79,49.71) .. controls (309.46,51.38) and (309.46,53.04) .. (307.79,54.71) .. controls (306.12,56.38) and (306.12,58.04) .. (307.79,59.71) .. controls (309.46,61.38) and (309.46,63.04) .. (307.79,64.71) .. controls (306.12,66.38) and (306.12,68.04) .. (307.79,69.71) .. controls (309.46,71.38) and (309.46,73.04) .. (307.79,74.71) .. controls (306.12,76.38) and (306.12,78.04) .. (307.79,79.71) .. controls (309.46,81.38) and (309.46,83.04) .. (307.79,84.71) .. controls (306.12,86.38) and (306.12,88.04) .. (307.79,89.71) .. controls (309.46,91.38) and (309.46,93.04) .. (307.79,94.71) .. controls (306.12,96.38) and (306.12,98.04) .. (307.79,99.71) .. controls (309.46,101.38) and (309.46,103.04) .. (307.79,104.71) .. controls (306.12,106.38) and (306.12,108.04) .. (307.79,109.71) .. controls (309.46,111.38) and (309.46,113.04) .. (307.79,114.71) .. controls (306.12,116.38) and (306.12,118.04) .. (307.79,119.71) .. controls (309.46,121.38) and (309.46,123.04) .. (307.79,124.71) .. controls (306.12,126.38) and (306.12,128.04) .. (307.79,129.71) .. controls (309.46,131.38) and (309.46,133.04) .. (307.79,134.71) .. controls (306.12,136.38) and (306.12,138.04) .. (307.79,139.71) .. controls (309.46,141.38) and (309.46,143.04) .. (307.79,144.71) .. controls (306.12,146.38) and (306.12,148.04) .. (307.79,149.71) .. controls (309.46,151.38) and (309.46,153.04) .. (307.79,154.71) .. controls (306.12,156.38) and (306.12,158.04) .. (307.79,159.71) .. controls (309.46,161.38) and (309.46,163.04) .. (307.79,164.71) .. controls (306.12,166.38) and (306.12,168.04) .. (307.79,169.71) .. controls (309.46,171.38) and (309.46,173.04) .. (307.79,174.71) .. controls (306.12,176.38) and (306.12,178.04) .. (307.79,179.71) .. controls (309.46,181.38) and (309.46,183.04) .. (307.79,184.71) .. controls (306.12,186.38) and (306.12,188.04) .. (307.79,189.71) .. controls (309.46,191.38) and (309.46,193.04) .. (307.79,194.71) .. controls (306.12,196.38) and (306.12,198.04) .. (307.79,199.71) .. controls (309.46,201.38) and (309.46,203.04) .. (307.79,204.71) .. controls (306.12,206.38) and (306.12,208.04) .. (307.79,209.71) .. controls (309.46,211.38) and (309.46,213.04) .. (307.79,214.71) .. controls (306.12,216.38) and (306.12,218.04) .. (307.79,219.71) .. controls (309.46,221.38) and (309.46,223.04) .. (307.79,224.71) .. controls (306.12,226.38) and (306.12,228.04) .. (307.79,229.71) .. controls (309.46,231.38) and (309.46,233.04) .. (307.79,234.71) .. controls (306.12,236.38) and (306.12,238.04) .. (307.79,239.71) .. controls (309.46,241.38) and (309.46,243.04) .. (307.79,244.71) .. controls (306.12,246.38) and (306.12,248.04) .. (307.79,249.71) .. controls (309.46,251.38) and (309.46,253.04) .. (307.79,254.71) – (307.79,255.71) – (307.79,255.71) ;
(203,213) .. controls (227.01,243.87) and (261.58,235.56) .. (278.96,214.5) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(298,213.2) .. controls (305.43,257.92) and (452.02,258.98) .. (481.15,213.59) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(373,212.2) .. controls (356.34,231.93) and (332,233.16) .. (313.15,214.37) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(45,73) – (57.73,57.54) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(85,73) – (68.45,57.37) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(45,134) – (45,118.67) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(239,151.2) .. controls (257.04,159.49) and (272.7,158.79) .. (288.76,142.53) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(239,91.2) – (260.14,99.66) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(395,73) .. controls (348.21,60.39) and (328.59,67.29) .. (301.25,87.56) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(306,141.25) .. controls (351.54,182.83) and (462.75,174.38) .. (487.28,126.66) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(285,89) .. controls (314.85,40.91) and (430.83,28) .. (487.16,101.09) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(276.5,33) – (149,33) ; (14.21,-6.37) .. controls (9.04,-2.99) and (4.3,-0.87) .. (0,0) .. controls (4.3,0.87) and (9.04,2.99) .. (14.21,6.37) ;
(339.5,33) – (440.5,33) ; (14.21,-6.37) .. controls (9.04,-2.99) and (4.3,-0.87) .. (0,0) .. controls (4.3,0.87) and (9.04,2.99) .. (14.21,6.37) ;
(49.5,186) .. controls (49.5,184.9) and (50.4,184) .. (51.5,184) – (148.5,184) .. controls (149.6,184) and (150.5,184.9) .. (150.5,186) – (150.5,222) .. controls (150.5,223.1) and (149.6,224) .. (148.5,224) – (51.5,224) .. controls (50.4,224) and (49.5,223.1) .. (49.5,222) – cycle ; (100,204) node \[align=left\] [ WebService1\
**Input** Params]{}; (165,193) .. controls (165,191.9) and (165.9,191) .. (167,191) – (237,191) .. controls (238.1,191) and (239,191.9) .. (239,193) – (239,211) .. controls (239,212.1) and (238.1,213) .. (237,213) – (167,213) .. controls (165.9,213) and (165,212.1) .. (165,211) – cycle ; (202,202) node \[align=left\] [p1 : type1]{}; (157,202) node \[align=left\] [[:]{}]{}; (263,193) .. controls (263,191.9) and (263.9,191) .. (265,191) – (335,191) .. controls (336.1,191) and (337,191.9) .. (337,193) – (337,211) .. controls (337,212.1) and (336.1,213) .. (335,213) – (265,213) .. controls (263.9,213) and (263,212.1) .. (263,211) – cycle ; (300,202) node \[align=left\] [p2 : type2]{}; (355,192) .. controls (355,190.9) and (355.9,190) .. (357,190) – (427,190) .. controls (428.1,190) and (429,190.9) .. (429,192) – (429,210) .. controls (429,211.1) and (428.1,212) .. (427,212) – (357,212) .. controls (355.9,212) and (355,211.1) .. (355,210) – cycle ; (392,201) node \[align=left\] [p3 : type3]{}; (448,192) .. controls (448,190.9) and (448.9,190) .. (450,190) – (520,190) .. controls (521.1,190) and (522,190.9) .. (522,192) – (522,210) .. controls (522,211.1) and (521.1,212) .. (520,212) – (450,212) .. controls (448.9,212) and (448,211.1) .. (448,210) – cycle ; (485,201) node \[align=left\] [p4 : type4]{}; (165,143) .. controls (165,141.9) and (165.9,141) .. (167,141) – (237,141) .. controls (238.1,141) and (239,141.9) .. (239,143) – (239,161) .. controls (239,162.1) and (238.1,163) .. (237,163) – (167,163) .. controls (165.9,163) and (165,162.1) .. (165,161) – cycle ; (202,152) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o1 : type1]{}; (165,112) .. controls (165,110.9) and (165.9,110) .. (167,110) – (237,110) .. controls (238.1,110) and (239,110.9) .. (239,112) – (239,130) .. controls (239,131.1) and (238.1,132) .. (237,132) – (167,132) .. controls (165.9,132) and (165,131.1) .. (165,130) – cycle ; (202,121) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o2 : type4]{}; (165,82) .. controls (165,80.9) and (165.9,80) .. (167,80) – (237,80) .. controls (238.1,80) and (239,80.9) .. (239,82) – (239,100) .. controls (239,101.1) and (238.1,102) .. (237,102) – (167,102) .. controls (165.9,102) and (165,101.1) .. (165,100) – cycle ; (202,91) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o3 : type5]{}; (165,51) .. controls (165,49.9) and (165.9,49) .. (167,49) – (237,49) .. controls (238.1,49) and (239,49.9) .. (239,51) – (239,69) .. controls (239,70.1) and (238.1,71) .. (237,71) – (167,71) .. controls (165.9,71) and (165,70.1) .. (165,69) – cycle ; (202,60) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o4 : type1]{}; (262,121) .. controls (262,119.9) and (262.9,119) .. (264,119) – (334,119) .. controls (335.1,119) and (336,119.9) .. (336,121) – (336,139) .. controls (336,140.1) and (335.1,141) .. (334,141) – (264,141) .. controls (262.9,141) and (262,140.1) .. (262,139) – cycle ; (299,130) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o5 : type2]{}; (262,91) .. controls (262,89.9) and (262.9,89) .. (264,89) – (334,89) .. controls (335.1,89) and (336,89.9) .. (336,91) – (336,109) .. controls (336,110.1) and (335.1,111) .. (334,111) – (264,111) .. controls (262.9,111) and (262,110.1) .. (262,109) – cycle ; (299,100) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o6 : type2]{}; (355,136) .. controls (355,134.9) and (355.9,134) .. (357,134) – (427,134) .. controls (428.1,134) and (429,134.9) .. (429,136) – (429,154) .. controls (429,155.1) and (428.1,156) .. (427,156) – (357,156) .. controls (355.9,156) and (355,155.1) .. (355,154) – cycle ; (392,145) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o7 : type3]{}; (355,105) .. controls (355,103.9) and (355.9,103) .. (357,103) – (427,103) .. controls (428.1,103) and (429,103.9) .. (429,105) – (429,123) .. controls (429,124.1) and (428.1,125) .. (427,125) – (357,125) .. controls (355.9,125) and (355,124.1) .. (355,123) – cycle ; (392,114) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o8 : type6]{}; (355,75) .. controls (355,73.9) and (355.9,73) .. (357,73) – (427,73) .. controls (428.1,73) and (429,73.9) .. (429,75) – (429,93) .. controls (429,94.1) and (428.1,95) .. (427,95) – (357,95) .. controls (355.9,95) and (355,94.1) .. (355,93) – cycle ; (392,84) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o9 : type3]{}; (448,105) .. controls (448,103.9) and (448.9,103) .. (450,103) – (520,103) .. controls (521.1,103) and (522,103.9) .. (522,105) – (522,123) .. controls (522,124.1) and (521.1,125) .. (520,125) – (450,125) .. controls (448.9,125) and (448,124.1) .. (448,123) – cycle ; (485,114) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [o2 : type4]{}; (235,240) node \[align=left\] [[rel1(p1,p2)]{}]{}; (388,255.79) node \[scale=0.7\] \[align=left\] [[rel2(p2,p4)]{}]{}; (364,230.79) node \[scale=0.7,rotate=-351.77\] \[align=left\] [[rel3(p3,p2)]{}]{}; (45.5,40) .. controls (45.5,38.9) and (46.4,38) .. (47.5,38) – (80.5,38) .. controls (81.6,38) and (82.5,38.9) .. (82.5,40) – (82.5,54) .. controls (82.5,55.1) and (81.6,56) .. (80.5,56) – (47.5,56) .. controls (46.4,56) and (45.5,55.1) .. (45.5,54) – cycle ; (64,47) node \[scale=0.8\] \[align=left\] [type1]{}; (73.5,110) .. controls (73.5,108.9) and (74.4,108) .. (75.5,108) – (108.5,108) .. controls (109.6,108) and (110.5,108.9) .. (110.5,110) – (110.5,124) .. controls (110.5,125.1) and (109.6,126) .. (108.5,126) – (75.5,126) .. controls (74.4,126) and (73.5,125.1) .. (73.5,124) – cycle ; (92,117) node \[scale=0.8\] \[align=left\] [type2]{}; (68.5,75) .. controls (68.5,73.9) and (69.4,73) .. (70.5,73) – (103.5,73) .. controls (104.6,73) and (105.5,73.9) .. (105.5,75) – (105.5,89) .. controls (105.5,90.1) and (104.6,91) .. (103.5,91) – (70.5,91) .. controls (69.4,91) and (68.5,90.1) .. (68.5,89) – cycle ; (87,82) node \[scale=0.8\] \[align=left\] [type5]{}; (24.5,75) .. controls (24.5,73.9) and (25.4,73) .. (26.5,73) – (59.5,73) .. controls (60.6,73) and (61.5,73.9) .. (61.5,75) – (61.5,89) .. controls (61.5,90.1) and (60.6,91) .. (59.5,91) – (26.5,91) .. controls (25.4,91) and (24.5,90.1) .. (24.5,89) – cycle ; (43,82) node \[scale=0.8\] \[align=left\] [type4]{}; (25.5,136) .. controls (25.5,134.9) and (26.4,134) .. (27.5,134) – (60.5,134) .. controls (61.6,134) and (62.5,134.9) .. (62.5,136) – (62.5,150) .. controls (62.5,151.1) and (61.6,152) .. (60.5,152) – (27.5,152) .. controls (26.4,152) and (25.5,151.1) .. (25.5,150) – cycle ; (44,143) node \[scale=0.8\] \[align=left\] [type6]{}; (65,15) node \[align=left\] [ types\
hierarchy]{}; (264,164) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-354.33\] \[align=left\] [[rel1(o1,o5)]{}]{}; (263,81) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-10.23\] \[align=left\] [[rel1(o3,o6)]{}]{}; (383,41) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-1.4\] \[align=left\] [[rel2(o6,o2)]{}]{}; (363,62) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-1.4\] \[align=left\] [[rel3(o9,o6)]{}]{}; (380,174) node \[color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 ,rotate=-1.4\] \[align=left\] [[rel2(o5,o2)]{}]{}; (26.5,101) .. controls (26.5,99.9) and (27.4,99) .. (28.5,99) – (61.5,99) .. controls (62.6,99) and (63.5,99.9) .. (63.5,101) – (63.5,115) .. controls (63.5,116.1) and (62.6,117) .. (61.5,117) – (28.5,117) .. controls (27.4,117) and (26.5,116.1) .. (26.5,115) – cycle ; (45,108) node \[scale=0.8\] \[align=left\] [type3]{}; (309,22) node \[scale=0.9,color=[rgb, 255:red, 0; green, 0; blue, 0 ]{} ,opacity=0.7 \] \[align=left\] [current\
level]{}; (221,20) node \[scale=0.9,color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [completed levels]{}; (390,20) node \[scale=0.9,color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [next levels]{};
For *inference rules*, the only distinction is that variables do not have types, so any level of backtracking can take any object. Rules with large premises and few relation restrictions would significantly slow down the search. This is another reason for disallowing provides to add rules.
The problem of matching input parameters to objects is **NP-Complete** on the defined model, both for services and for rules. This is because the problem is equivalent with *labeled subgraph isomorphism*, that is known to be **NP-Complete** as stated, for example in [@cordella2004sub]. The equivalence is obvious: graph nodes are objects and directed labeled edges are relations with their types as labels. There are four main structures used in the algorithm: **ontology**, **repository**, **knowledge**, and **query**. First of all, the ontology is loaded, then the repository and the query that are verified to use the correct terms of the ontology. With the user query, the knowledge is instantiated with initially known objects and relations, and inference rules are applied for the first time.
The high-level view of the main Algorithm \[algo:main\] is simple: iterate all services and verify for each if the service can be called with the current knowledge, and do so if possible. After looping all services, the inference rules are applied, similarly. If at any time the user required output is satisfied with all specified relations, the composition is complete, thus returned and the algorithm ends. If in some loop no new service or rule could be applied, then the algorithm is blocked in the current state; thus the instance is unsolvable. The use of $query.out$ and $query.in$ as input and, respectively, output parameters of services is an implementation shortcut that can be done based on the structural similarity: $(\emptyset , query.in)$ is a fictive service that has $\emptyset$ as input and $query.in$ as output. [[<span style="font-variant:small-caps;">callService</span>]{.nodecor}]{}($(\emptyset, query.in), \emptyset$) is just adding initially known objects to the knowledge, using the same method that adds service outputs later. The [[<span style="font-variant:small-caps;">applyInferenceRules()</span>]{.nodecor}]{} is very similar to the core part of the composition search, Algorithm \[algo:main\]. It iterates all rules and applies any applicable rule until no rule can be applied for objects it was not already applied. Also, at the end the algorithm checks if there are any services generating only unused parameters and relations and removes them.
; ; $newCall \gets true$; $newCall \gets false$; $matchObjects[$ $] \gets \Call{findMatch}{service}$; $newCall \gets true$; ; ; $ $ ; $ $ ;
[[<span style="font-variant:small-caps;">callService()</span>]{.nodecor}]{} - Algorithm \[algo:callservice\] takes as parameters the $service$ to call and the already found matching objects. Output objects are created with their name generated from the joined service name with the parameter name and also the number of times the service was called in total. This can help for debugging and getting provenance [@ding2010vipen] information about objects later (though this is not yet implemented). After objects are created, all relations between input and output, or between output parameters, defined in the service are added to the matched or new objects according to the order known by their names.
$hash \gets \Call{hashValue}{matchObjects}$; $service.history \gets service.history \cup \{hash\}$; $newObjects[$ $]\gets \emptyset$; $newObjects.add($new object of type $parameter.type)$; $\Call{addRelation}{newObjects[x],newObjects[y]}$; $knowledge \gets knowledge \cup newObjects$;
The most runtime consuming method is [[<span style="font-variant:small-caps;">findMatch()</span>]{.nodecor}]{} for services and its equivalent for rules that have exponential complexity. Searching for objects that match variables in rule premises is similar to services, and simpler to implement, as variables are not typed like parameters. The matching was defined on problem model, exemplified in Figure \[fig:inputmatching\] and is implemented in Algorithm \[algo:findmatch\]. For each level of the backtracking, all possible objects of the type are tried, and object relations are checked to match any parameter relation from the current level to any previous level. Both possible orientations of relations are checked. Reaching one level higher than the number of input parameters indicates a complete match, for which we also lookup in service history. If the match is also new, $matchFound$ is set to $true$, and the search stops for all branches. There are a few intuitive special structures and methods used to get information faster. For example, $ontology.subTypes(...)$ returns all subtypes for a type including itself and $knowledge.objectsOfTypes(...)$ returns all objects known of specified types.
The algorithm presented is more a proof of concept implementation for the composition model. There are many possible enhancements and extra features that could be useful, some of which will be discussed in Section \[sec:conclusion\].
$matchFound \gets false$; $matchObjects \gets new $ $Object[$ $]$; $\Call{bktParam}{0, service, matchObjects}$; $null$; $ $ $matchObjects$; $hash \gets \Call{hashValue}{matchObjects}$; $matchFound \gets (hash \notin service.history)$; $type \gets service.in[level].type$; $subTypes \gets ontology.subTypes(type)$; $candidates \gets knowledge.objectsOfTypes(subTypes)$; $matchObjects.add(candidate)$; $\Call{bktParam}{level+1, service, matchObjects}$;
Evaluation {#sec:evaluation}
==========
In order to evaluate the algorithm presented in Section \[sec:algorithm\], we implemented a test generator. The generator produces a problem instance: the repository, ontology and user query that has a high probability of being solvable. In the first phase, it generates a set of services from which a composition can be extracted. To generate services with this property, we use this observation: as services are added in the composition, the knowledge is enhanced: more objects are obtained with their corresponding relations. So, to generate the repository, we start by incrementally generating the knowledge i.e., objects and relations, and saving all intermediate stages of this process. To generate the services we consider that every intermediate stage in the knowledge is the information gained during composition, thus, we need to create services that can generate these stages. Between each two stages $K_i$, $K_{i+1}$ we generate a “layer” of services with each service having as input parameters objects and relations between them from $K_i$ and output parameters in a similar way from $K_{i+1}$, as shown in Figure \[fig:testgenerator\]. The user query is generated as following: the input parameters and relations are a subset of the first stage, while the output parameters and relations are a subset of the last stage of the knowledge. “Noise” is added in the repository and ontology, i.e., random concepts and services with the aim of hiding the composition.
(96.5,204.2) – (102.7,198) – (149.3,198) – (155.5,204.2) – (155.5,229) – (155.5,229) – (96.5,229) – (96.5,229) – cycle ; (186.5,204.2) – (192.7,198) – (239.3,198) – (245.5,204.2) – (245.5,229) – (245.5,229) – (186.5,229) – (186.5,229) – cycle ; (96.5,188) – (257.5,188) ;
(278.83,24.33) – (277.5,260) ;
(85.33,24.67) – (83.5,263) ;
(127.5,23) .. controls (127.5,21.9) and (128.4,21) .. (129.5,21) – (226.5,21) .. controls (227.6,21) and (228.5,21.9) .. (228.5,23) – (228.5,41) .. controls (228.5,42.1) and (227.6,43) .. (226.5,43) – (129.5,43) .. controls (128.4,43) and (127.5,42.1) .. (127.5,41) – cycle ; (178,32) node \[scale=1\] \[align=left\] [ WebService1]{}; (97.5,65) .. controls (97.5,63.9) and (98.4,63) .. (99.5,63) – (156.5,63) .. controls (157.6,63) and (158.5,63.9) .. (158.5,65) – (158.5,79) .. controls (158.5,80.1) and (157.6,81) .. (156.5,81) – (99.5,81) .. controls (98.4,81) and (97.5,80.1) .. (97.5,79) – cycle ; (128,72) node \[scale=0.8\] \[align=left\] [p1 : type1]{}; (97.5,115) .. controls (97.5,113.9) and (98.4,113) .. (99.5,113) – (156.5,113) .. controls (157.6,113) and (158.5,113.9) .. (158.5,115) – (158.5,129) .. controls (158.5,130.1) and (157.6,131) .. (156.5,131) – (99.5,131) .. controls (98.4,131) and (97.5,130.1) .. (97.5,129) – cycle ; (128,122) node \[scale=0.8\] \[align=left\] [p2 : type2]{}; (97.83,166.33) .. controls (97.83,165.23) and (98.73,164.33) .. (99.83,164.33) – (156.83,164.33) .. controls (157.94,164.33) and (158.83,165.23) .. (158.83,166.33) – (158.83,180.33) .. controls (158.83,181.44) and (157.94,182.33) .. (156.83,182.33) – (99.83,182.33) .. controls (98.73,182.33) and (97.83,181.44) .. (97.83,180.33) – cycle ; (128.33,173.33) node \[scale=0.8\] \[align=left\] [p4 : type4]{}; (7.67,59.67) .. controls (7.67,58.56) and (8.56,57.67) .. (9.67,57.67) – (60.67,57.67) .. controls (61.77,57.67) and (62.67,58.56) .. (62.67,59.67) – (62.67,73.67) .. controls (62.67,74.77) and (61.77,75.67) .. (60.67,75.67) – (9.67,75.67) .. controls (8.56,75.67) and (7.67,74.77) .. (7.67,73.67) – cycle ; (35.17,66.67) node \[scale=0.8,color=[rgb, 255:red, 2; green, 2; blue, 2 ]{} ,opacity=1 \] \[align=left\] [[o3 : type1]{}]{}; (7.17,110) .. controls (7.17,108.9) and (8.06,108) .. (9.17,108) – (60.17,108) .. controls (61.27,108) and (62.17,108.9) .. (62.17,110) – (62.17,124) .. controls (62.17,125.1) and (61.27,126) .. (60.17,126) – (9.17,126) .. controls (8.06,126) and (7.17,125.1) .. (7.17,124) – cycle ; (34.67,117) node \[scale=0.8,color=[rgb, 255:red, 5; green, 5; blue, 5 ]{} ,opacity=1 \] \[align=left\] [[o6 : type2]{}]{}; (8,160.67) .. controls (8,159.56) and (8.9,158.67) .. (10,158.67) – (61,158.67) .. controls (62.1,158.67) and (63,159.56) .. (63,160.67) – (63,174.67) .. controls (63,175.77) and (62.1,176.67) .. (61,176.67) – (10,176.67) .. controls (8.9,176.67) and (8,175.77) .. (8,174.67) – cycle ; (35.5,167.67) node \[scale=0.8,color=[rgb, 255:red, 2; green, 2; blue, 2 ]{} ,opacity=1 \] \[align=left\] [[o2 : type4]{}]{}; (62.33,86.17) node \[scale=0.9\] \[align=left\] [[rel1(o3,06)]{}]{}; (60.83,136.17) node \[scale=0.9\] \[align=left\] [[rel2(o6,02)]{}]{}; (155.33,141.17) node \[scale=0.9\] \[align=left\] [[rel2(p2,p4)]{}]{}; (154.33,91.5) node \[scale=0.9\] \[align=left\] [[rel1(p1,p2)]{}]{}; (281.67,62) .. controls (281.67,60.9) and (282.56,60) .. (283.67,60) – (334.67,60) .. controls (335.77,60) and (336.67,60.9) .. (336.67,62) – (336.67,76) .. controls (336.67,77.1) and (335.77,78) .. (334.67,78) – (283.67,78) .. controls (282.56,78) and (281.67,77.1) .. (281.67,76) – cycle ; (309.17,69) node \[scale=0.8,color=[rgb, 255:red, 0; green, 0; blue, 0 ]{} ,opacity=1 \] \[align=left\] [[o7 : type8]{}]{}; (281.17,112.33) .. controls (281.17,111.23) and (282.06,110.33) .. (283.17,110.33) – (334.17,110.33) .. controls (335.27,110.33) and (336.17,111.23) .. (336.17,112.33) – (336.17,126.33) .. controls (336.17,127.44) and (335.27,128.33) .. (334.17,128.33) – (283.17,128.33) .. controls (282.06,128.33) and (281.17,127.44) .. (281.17,126.33) – cycle ; (308.67,119.33) node \[scale=0.8,color=[rgb, 255:red, 0; green, 0; blue, 0 ]{} ,opacity=1 \] \[align=left\] [[o9 : type2]{}]{}; (330.33,87.5) node \[scale=0.9\] \[align=left\] [[rel3(o7,09)]{}]{}; (186.5,66) .. controls (186.5,64.9) and (187.4,64) .. (188.5,64) – (245.5,64) .. controls (246.6,64) and (247.5,64.9) .. (247.5,66) – (247.5,80) .. controls (247.5,81.1) and (246.6,82) .. (245.5,82) – (188.5,82) .. controls (187.4,82) and (186.5,81.1) .. (186.5,80) – cycle ; (217,73) node \[scale=0.8\] \[align=left\] [p5 : type8]{}; (186.5,116) .. controls (186.5,114.9) and (187.4,114) .. (188.5,114) – (245.5,114) .. controls (246.6,114) and (247.5,114.9) .. (247.5,116) – (247.5,130) .. controls (247.5,131.1) and (246.6,132) .. (245.5,132) – (188.5,132) .. controls (187.4,132) and (186.5,131.1) .. (186.5,130) – cycle ; (217,123) node \[scale=0.8\] \[align=left\] [p9 : type2]{}; (243.33,92.5) node \[scale=0.9\] \[align=left\] [[rel3(p5,p9)]{}]{}; (132,53) node \[scale=0.9\] \[align=left\] [Input params:]{}; (229,53) node \[scale=0.9\] \[align=left\] [Output params:]{}; (281,159.33) .. controls (281,158.23) and (281.9,157.33) .. (283,157.33) – (334,157.33) .. controls (335.1,157.33) and (336,158.23) .. (336,159.33) – (336,173.33) .. controls (336,174.44) and (335.1,175.33) .. (334,175.33) – (283,175.33) .. controls (281.9,175.33) and (281,174.44) .. (281,173.33) – cycle ; (308.5,166.33) node \[scale=0.8,color=[rgb, 255:red, 155; green, 155; blue, 155 ]{} ,opacity=1 \] \[align=left\] [[o2 : type4]{}]{}; (126,213.5) node \[scale=0.7\] \[align=left\] [WebService]{}; (216,213.5) node \[scale=0.7\] \[align=left\] [WebService]{}; (36,182.67) node \[align=left\] [...]{}; (171,199) node \[align=left\] [...]{}; (307,145) node \[align=left\] [...]{}; (39,238.33) node \[scale=0.8\] \[align=left\] [knowledge\
stage 1]{}; (314,240.67) node \[scale=0.8\] \[align=left\] [knowledge\
stage 2]{}; (172,248) node \[scale=0.8\] \[align=left\] [generated services\
stage 1]{}; (9,201.67) .. controls (9,200.56) and (9.9,199.67) .. (11,199.67) – (62,199.67) .. controls (63.1,199.67) and (64,200.56) .. (64,201.67) – (64,215.67) .. controls (64,216.77) and (63.1,217.67) .. (62,217.67) – (11,217.67) .. controls (9.9,217.67) and (9,216.77) .. (9,215.67) – cycle ; (36.5,208.67) node \[scale=0.8,color=[rgb, 255:red, 97; green, 97; blue, 97 ]{} ,opacity=1 \] \[align=left\] [[o4 : type4]{}]{}; (35.08,75.67) – (34.78,106) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(34.81,126) – (35.32,156.67) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(128,81) – (128,111) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(128.06,131) – (128.26,162.33) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(309.08,78) – (308.78,108.33) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
(217,82) – (217,112) ; (8.93,-4.29) – (0,0) – (8.93,4.29) – cycle ;
\[table:eval\_generator\]
\[table:eval\_2008\]
The algorithm provided the expected composition on the example in Section \[sec:example\]. We evaluated its efficiency on tests generated as above, with and without inference rules (and also, ignoring them). Results in Table \[table:eval\_generator\] show that the use of inference rules improves the composition, that becomes shorter on two tests; and that the algorithm is efficient. The slowest run is observed on a testcase where more rules are applied, which is expected. To compare our solution with others, we also tested on the converted benchmark from the composition challenge [@bansal2008wsc]. These tests are obviously without any relations or rules, and these results are shown in Table \[table:eval\_2008\]. Even if our algorithm is designed for the extended model, it finds a composition of size comparable with the challenge winners (relative to the size of the repository) and in shorter amount of time. This also shows that the algorithm is compatible with the previous models.
Conclusion and Future Work {#sec:conclusion}
==========================
The essence of our contribution is the definition of the new composition model. Without the contextual-aware representation of elements involved in a composition at the parameter level, the composition is limited and cannot allow automation of the reasoning involved in manual composition. A simple hierarchy of types is insufficient. The implemented algorithm proves the problem is still solvable even on larger cases with many relations and rules.
There are many paths of continuation. The obvious is to mature the algorithm, for example by improving the strategy to reduce composition length or detection of useless services, improve service selection, prune useless paths of parameter matching or composition, by adding more conditions; or to consider more elaborate Quality of Service metrics. Also, it would help to develop ideas for aiding service providers to adhere to the model, including to convert definitions of existing services to the relational model. This process should be automated as well at least partially.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
Marcus, Spielman, and Srivastava in their seminal work [@MSS13] resolved the Kadison-Singer conjecture by proving that for any set of finitely supported independently distributed random vectors $v_1,\dots, v_n$ which have “small” expected squared norm and are in isotropic position (in expectation), there is a positive probability that the sum $\sum v_i v_i^\intercal$ has small spectral norm. Their proof crucially employs real stability of polynomials which is the natural generalization of real-rootedness to multivariate polynomials.
Strongly Rayleigh distributions are families of probability distributions whose generating polynomials are real stable [@BBL09]. As independent distributions are just special cases of strongly Rayleigh measures, it is a natural question to see if the main theorem of [@MSS13] can be extended to families of vectors assigned to the elements of a strongly Rayleigh distribution.
In this paper we answer this question affirmatively; we show that for any homogeneous strongly Rayleigh distribution where the marginal probabilities are upper bounded by $\eps_1$ and any isotropic set of vectors assigned to the underlying elements whose norms are at most $\sqrt{\eps_2}$, there is a set in the support of the distribution such that the spectral norm of the sum of the natural quadratic forms of the vectors assigned to the elements of the set is at most $O(\eps_1+\eps_2)$. We employ our theorem to provide a sufficient condition for the existence of spectrally thin trees. This, together with a recent work of the authors [@AO14], provides an improved upper bound on the integrality gap of the natural LP relaxation of the Asymmetric Traveling Salesman Problem.
author:
- 'Nima Anari [^1]'
- 'Shayan Oveis Gharan [^2]'
bibliography:
- 'ref2.bib'
title: |
The Kadison-Singer Problem for Strongly Rayleigh Measures\
and Applications to Asymmetric TSP
---
Introduction
============
Marcus, Spielman and Srivastava [@MSS13] in a breakthrough work proved the Kadison-Singer conjecture [@KS59] by proving Weaver’s [@Wea04] conjecture $\text{KS}_2$ and the Akemann and Anderson’s Paving conjecture [@AA91]. The following is their main technical contribution.
\[thm:mss\] If $\eps>0$ and $v_1,\ldots,v_m$ are independent random vectors in $\R^d$ with finite support where, $$\sum_{i=1}^m {{\mathbb{E}}}v_iv_i^\intercal = I,$$ such that for all $i$, $${{\mathbb{E}}}{\left\lVertv_i\right\rVert}^2 \leq \eps,$$ then $${{\mathbb{P}}\left[{\left\lVert\sum_{i=1}^m v_iv_i^\intercal\right\rVert} \leq (1+\sqrt{\eps})^2\right]} >0.$$
In this paper, we prove an extension of the above theorem to families of vectors assigned to elements of a not necessarily independent distribution.
Let $\mu:2^{[m]}\to R_+$ be a probability distribution on the subsets of the set $[m]=\{1,2,\ldots,m\}$. In particular, we assume that $\mu(.)$ is nonnegative and, $$\sum_{S\subseteq [m]} \mu(S) = 1.$$ We assign a multi-affine polynomial with variables $z_1,\ldots,z_m$ to $\mu$, $$g_\mu(z) = \sum_{S\subseteq [m]} \mu(S)\cdot z^S,$$ where for a set $S\subseteq[m]$, $z^S=\prod_{i\in S} z_i$. The polynomial $g_\mu$ is also known as the [*generating polynomial*]{} of $\mu$. We say $\mu$ is a [*homogeneous*]{} probability distribution if $g_\mu$ is a homogeneous polynomial.
We say that $\mu$ is a [*strongly Rayleigh*]{} distribution if $g_\mu$ is a real stable polynomial. See for the definition of real stability. Strongly Rayleigh measures are introduced and deeply studied in the seminal work of Borcea, Brändén and Liggett [@BBL09]. They are natural generalizations of product distributions and cover several interesting families of probability distributions including determinantal measures and random spanning tree distributions. We refer interested readers to [@OSS11; @PP14] for applications of these probability measures. Our main theorem extends to families of vectors assigned to the elements of a strongly Rayleigh distribution. This can be seen as a generalization because independent distributions are special classes of strongly Rayleigh measures. To state the main theorem we need another definition. The [*marginal*]{} probability of an element $i$ with respect to a probability distribution, $\mu$, is the probability that $i$ is in a sample of $\mu$, $$\label{eq:marginaldef}
{{\mathbb{P}}_{S\sim\mu}\left[i\in S\right]} = \partial_{z_i} g_\mu(z) \big|_{z_1=\ldots=z_m=1}.$$
\[thm:main\] Let $\mu$ be a homogeneous strongly Rayleigh probability distributions on $[m]$ such that the marginal probability of each element is at most $\eps_1$, and let $v_1,\ldots,v_m\in\R^d$ be vectors in an isotropic position, $$\sum_{i=1}^m v_iv_i^\intercal =I,$$ such that for all $i$, ${\left\lVertv_i\right\rVert}^2 \leq \eps_2$. Then, $${{\mathbb{P}}_{S\sim \mu}\left[{\left\lVert\sum_{i\in S} v_iv_i^\intercal\right\rVert} \leq 4(\eps_1+\eps_2)+2(\eps_1+\eps_2)^2 \right]} >0.$$
The above theorem does not directly generalize , but it can be seen as a variant of to the case where the vectors $v_1,\dots,v_m$ are negatively dependent. We expect to see several applications of our main theorem that are not realizable by the original proof of [@MSS13]. In the following subsections we describe our main motivation for studying the above statement, which is to design approximation algorithms for the Asymmetric Traveling Salesman Problem (ATSP).
Let us conclude this part by proving a simple application of the above theorem to prove $\text{KS}_r$ for $r\geq 5$.
\[cor:KSr\] Given a set vectors $v_1,\dots,v_m\in\R^d$ in isotropic position, $$\sum_{i=1}^m v_iv_i^\intercal = I,$$ if for all $i$, ${\left\lVertv_i\right\rVert}^2\leq \eps$, then for any $r$, there is an $r$ partitioning of $[m]$ into $S_1,\dots,S_r$ such that for any $j\leq r$, $${\left\lVert\sum_{i\in S_j} v_iv_i^\intercal\right\rVert} \leq 4(1/r+\eps)+2(1/4+\eps)^2.$$
The proof is inspired by the lifting idea in [@MSS13]. For $i\in [m]$ and $j\in [r]$ let $w_{i,j}\in\R^{d\cdot r}$ be the directed sum of $r$ vectors all of which are $0^d$ except the $j$-th one which is $v_i$, i.e., $$w_{i,1}=\begin{pmatrix}v_i\\ 0^d \\ \vdots \\ 0^d\end{pmatrix}, w_{i,2} = \begin{pmatrix}0^d \\ v_i \\ \vdots \\ 0^d\end{pmatrix}, \text{ and so on.}$$ Let $E=\{(i,j): i\in [m],j\in[r]\}$ and let $\mu:2^E\to\R_+$ be a product distribution defined in a way that selects exactly one pair $(i,j)\in E$ for any $i\in [m]$ uniformly at random. Observe that there are $m^r$ sets in the support of $\mu$ each of size exactly $m$ and each has probability $1/r^m$. Therefore, $\mu$ is a homogeneous probability distribution and the marginal probability of each element of $E$ is exactly $1/r$. In addition, since product distributions are strongly Rayleigh, $\mu$ is strongly Rayleigh. Therefore, by , there is a set $S$ in the support of $\mu$ such that $${\left\lVert\sum_{(i,j)\in S} w_{i,j}w_{i,j}^\intercal\right\rVert} \leq \alpha,$$ for $\alpha=4(1/r+\eps) + 2(1/r+\eps)^2$. Now, let $S_j=\{i: (i,j)\in S\}$. It follows that for any $j\in [r]$, $${\left\lVert\sum_{i\in S_j} v_iv_i^\intercal\right\rVert} \leq \alpha,$$ as desired.
The Thin Basis Problem
----------------------
In this section we use the main theorem to prove the existence of a thin basis among a given set of isotropic vectors. In the next section, we will use this theorem to prove the existence of thin trees in graphs, i.e., trees which are “sparse” in all cuts of a given graph.
Given a set of vectors $v_1,\dots,v_m\in \R^d$ in the isotropic position, $$\sum_{i=1}^m v_iv_i^\intercal = I,$$ we want to find a sufficient condition for the existence of a [*thin basis*]{}. Recall that a set $T\subset [m]$ is a basis if $|T|=d$ and all vectors indexed by $T$ are linearly independent. We say $T$ is $\alpha$-thin if $${\left\lVert\sum_{i\in T} v_iv_i^\intercal\right\rVert} \leq \alpha.$$ An obvious necessary condition for the existence of an $\alpha$-thin basis is that the set $$V(\alpha):=\{v_i: {\left\lVertv_i\right\rVert}^2 \leq \alpha\},$$ contains a basis. We show that there exist universal constants $C_1,C_2>0$ such that the existence of $C_1/\alpha$ disjoint bases in $V(C_2\cdot \alpha)$ is a sufficient condition.
\[thm:thinbases\] Given a set of vectors $v_1,\dots,v_m\in\R^d$ in the sub-isotropic position $$\sum_{i=1}^m v_iv_i^\intercal \preceq I,$$ if for all $1\leq i\leq m$, ${\left\lVertv_i\right\rVert}^2\leq \eps$, and the set $\{v_1,\dots,v_m\}$ contains $k$ disjoint bases, then there exists an $O(\eps+1/k)$-thin basis $T\subseteq [m]$.
We will use to prove the above theorem. To use we need to define a strongly Rayleigh distribution on $[m]$ with small marginal probabilities. This is proved in the following proposition.
\[prop:determinantalconst\] Given a set of vectors $v_1,\dots,v_m\in\R^d$ that contains $k$ disjoint bases, there is a strongly Rayleigh probability distribution $\mu:2^{[m]}\to\R_+$ supported on the bases such that the marginal probability of each element is at most $O(1/k)$.
Now, follows simply from the above proposition. Letting $\mu$ be defined as above, we get $\eps_1=\eps$ and $\eps_2=O(1/k)$ in which implies the existence of a basis $T\subseteq [m]$ such that $${\left\lVert\sum_{i\in T} v_iv_i^\intercal\right\rVert} \leq O(\eps+1/k),$$ as desired.
In the rest of this section we prove the above proposition. In our proof $\mu$ will in fact be a homogeneous *determinantal* probability distribution. We say $\mu:2^{[m]}\to\R_+$ is a determinantal probability distribution if there is a PSD matrix $M\in\R^{m\times m}$ such that for any set $T\subseteq [m]$, $${{\mathbb{P}}_{S\sim\mu}\left[T\subseteq S\right]} = \det(M_{T,T}),$$ where $M_{T,T}$ is the principal submatrix of $M$ whose rows and columns are indexed by $T$. It is proved in [@BBL09] that any determinantal probability distribution is a strongly Rayleigh measure, so this is sufficient for our purpose. In fact, we will find nonnegative weights $\lambda:[m]\to\R_+$ and for any basis $T$ we will let $$\mu_\lambda(T) \propto \det\left(\sum_{i\in T} \lambda_i v_iv_i^\intercal\right).\label{eq:mulambda}$$ It follows by the Cauchy-Binet identity that for any $\lambda$, such a distribution is determinantal with respect to the gram matrix $$M(i,j)=\sqrt{\lambda_i\lambda_j}\left\langle B^{-1/2} v_i,B^{-1/2} v_j\right\rangle$$ where $B=\sum_{i=1}^m \lambda_i v_iv_i^\intercal$. So, all we need to do is find $\{\lambda_i\}_{1\leq i\leq m}$ such that the marginal probability of each element in $\mu_\lambda$ is $O(1/k)$.
For any basis $T\subset [m]$ let $\bone_T\in \R^m$ be the indicator vector of the set $T$. Let $P$ be the convex hull of bases’ indicator vectors, $$P:=\operatorname{conv}\{\bone_T: T\text{ is a basis}\}.$$ Recall that a point $x$ is in the *relative interior* of $P$, $x\in \operatorname{relint}P$, if and only if $x$ can be written as a convex combination of all of the vertices of $P$ with strictly positive coefficients.
We find the weights in two steps. First, we show that for any point $x\in \operatorname{relint}P$, there exist weights $\lambda:[m]\to\R$ such that for any $i$, $${{\mathbb{P}}_{S\sim\mu_\lambda}\left[i\in S\right]} = x(i),$$ where $x(i)$ is the $i$-th coordinate of $x$ and $\mu_\lambda$ is defined as in . Then, we show that there exists a point $x\in \operatorname{relint}P$ such that for all $i$, $x(i) \leq O(1/k)$.
For any $x\in\operatorname{relint}P$ there exist $\lambda:[m]\to\R_+$ such that the marginal probability of each element $i$ in $\mu_\lambda$ is $x(i)$.
Let $\mu^*:=\mu_\bone$ be the (determinantal) distribution where $\lambda_i=1$ for all $i$. The idea is to find a distribution $p(.)$ maximizing the relative entropy with respect to $\mu^*$ and preserves $x$ as the marginal probabilities. This is analogous to the recent applications of maximum entropy distributions in approximation algorithms [@AGMOS10; @SV14].
Consider the following entropy maximization convex program. $$\begin{aligned}
\min \hspace{5ex} & \sum_{T} p(T)\cdot \log \frac{p(T)}{\mu^*(T)} &\\
{\mbox{\rm s.t. }}\hspace{2ex} & \sum_{T: i\in T} p(T) = x(i) & \forall i,\\
& p(T) \geq 0.
\end{aligned}
\label{cp:maxentropy}$$ Note that any feasible solution satisfies $\sum_T p(T)=1$ so we do not need to add this as a constraint. First of all, since $x\in\operatorname{relint}P$, there exists a distribution $p(.)$ such that for all bases $T$, $p(T)>0$. So, the Slater condition holds and the duality gap of the above program is zero.
Secondly, we use the Lagrange duality to characterize the optimum solution of the above convex program. For any element $i$ let $\gamma_i$ be the Lagrange dual variable of the first constraint. The Lagrangian $L(p,\gamma)$ is defined as follows: $$L(p,\gamma) = \inf_{p\geq 0} \sum_T p(T)\cdot \log\frac{p(T)}{\mu^*(T)} - \sum_i \gamma_i \sum_{T:e\in T} (p(T)-x(i))$$ Let $p^*$ be the optimum $p$, letting the gradient of the RHS equal to zero we obtain, for any bases $T$, $$\log\frac{p^*(T)}{\mu^*(T)} + 1 = \sum_{i\in T} \gamma_i.$$ For all $i$, let $\lambda_i = \exp(\gamma_i - 1/d)$, where $d$ is the dimension of the $v_i$’s. Then, we get $$\begin{aligned}
p^*(T) &=& \prod_{i\in T} \lambda_i \cdot \mu^*(T) \\
&=& \prod_{i\in T} \lambda_i \cdot \det\left(\sum_{i \in T} v_iv_i^\intercal\right) \\
&=& \det\left(\sum_{i\in T} \lambda_i v_iv_i^\intercal\right).\end{aligned}$$ Therefore $p^*\equiv\mu_\lambda$. Since the duality gap is zero, the above $p^*$ is indeed an optimal solution of the convex program . Therefore, the marginal probability of every element $i$ with respect to $p^*$ ($\mu_\lambda$) is equal to $x(i)$.
If $\{v_1,\dots, v_m\}$ contains $k$ disjoint bases, then there exists a point $x\in\operatorname{relint}P$, such that $x(i)=O(1/k)$ for all $i$.
Let $T_1,\dots, T_k$ be the promised disjoint bases. Let $$x_0=\frac{\bone_{T_1}+\dots+\bone_{T_k}}{k}.$$ The above is a convex combination of the vertices of $P$; so $x_0\in P$. We now perturb $x_0$ by a small amount to find a point in $\operatorname{relint}P$. Let $x_1$ be an arbitrary point in $\operatorname{relint}P$ (such as the average of all vertices). For any $\epsilon>0$, the point $x=(1-\epsilon)x_0+\epsilon x_1\in \operatorname{relint}P$. If $\epsilon$ is small enough, we get $x(i)=O(1/k)$ which proves the claim.
This completes the proof of .
Spectrally Thin Trees
---------------------
For a graph $G=(V,E)$, the Laplacian of $G$, $L_G$, is defined as follows: For a vertex $i\in V$ let $\bone_i\in\R^V$ be the vector that is one at $i$ and zero everywhere else. Fix an arbitrary orientation on the edges of $E$ and let $b_e=\bone_i-\bone_j$ for an edge $e$ oriented from $i$ to $j$. Then, $$L_G=\sum_{e\in E} b_eb_e^\intercal.$$ We use $L_G^{\dagger}$ to denote the pseudo-inverse of $L_G$. Also, for a set $T\subseteq E$, we write $$L_T=\sum_{e\in T} b_eb_e^\intercal.$$ We say a spanning tree $T\subseteq E$ is $\alpha$-thin with respect to $G$ if for any set $S\subset V$, $$|T(S,\overline{S})|\leq \alpha \cdot |E(S,\overline{S})|,$$ where $T(S,\overline{S}),E(S,\overline{S})$ are the set of edges cross the cut $(S,\overline{S})$ in $T,G$ respectively. We say a spanning tree $T$ is $\alpha$-spectrally thin with respect to $G$ if $$L_T \preceq \alpha\cdot L_G.$$ It is easy to see that spectral thinness is a generalization of the combinatorial thinness, i.e., if $T$ is $\alpha$-spectrally thin it is also $\alpha$-thin.
We say a graph $G$ is $k$-edge connected if it has at least $k$ edges in any cut. In recent works on Asymmetric TSP [@AGMOS10; @OS11] it was shown that the existence of (combinatorially) thin trees in $k$-edge connected graphs plays an important role in bounding the integrality gap of the natural linear programming relaxation of the Asymmetric TSP [@AO14].
It turns out that the existence of spectrally thin trees is significantly easier to prove than combinatorially thin trees thanks to of [@MSS13]. Given a graph $G=(V,E)$, Harvey and Olver [@HO14] employ a recursive application of [@MSS13] and show that if for all edges $e\in E$, $b_e^\intercal L_G^{\dagger}b_e \leq \alpha$, then $G$ has an $O(\alpha)$-spectrally thin tree. The quantity $b_e L_G^{\dagger}b_e$ is the effective resistance between the endpoints of $e$ when we replace every edge of $G$ with a resistor of resistance 1 [@LP13 Ch. 2]. Unfortunately, $k$-edge connectivity is a significantly weaker property than $\max_e b_e L_G^\dagger b_e\leq \alpha$ [@AO14]. So, this does not resolve the thin tree problem.
The main idea of [@AO14] is to slightly change the graph $G$ in order to decrease the effective resistance of edges while maintaining the size of the cuts intact. More specifically, to add a “few” edges $E'$ to $G$ such that in the new graph $G'=(V,E\cup E')$, the effective resistance of every edge of $E$ is small and the size of every cut of $G'$ is at most twice of that cut in $G$. If we can prove that $G'$ has a spectrally thin tree $T\subseteq E$ such a tree is combinatorially thin with respect to $G$ because $G,G'$ have the same cut structure. To show that $G'$ has a spectrally thin tree we need to answer the following question.
\[pr:Feffres\] Given a graph $G=(V,E)$, suppose there is a set $F\subseteq E$ such that $(V,F)$ is $k$-edge connected, and that for all $e\in F$, $b_e^\intercal L_G^\dagger b_e \leq \alpha$. Can we say that $G$ has a $C\cdot \max\{\alpha,1/k\}$-spectrally thin tree for a universal constant $C$?
We use to answer the above question affirmatively. Note that the above question cannot be answered by . One can use to show that the set $F$ can be partitioned into two sets $F_1,F_2$ such that each $F_i$ is $1/2+O(\alpha)$-spectrally thin, but gives no guarantee on the connectivity of $F_i$’s. On the other hand, once we apply our main theorem to a strongly Rayleigh distribution supported on connected subgraphs of $G$, e.g. the spanning trees of $G$, we get connectivity for free.
Given a graph $G=(V,E)$ and a set $F\subseteq E$ such that $(V,F)$ is $k$-edge connected, if for $\eps>0$ and any edge $e\in F$, $b_e^\intercal L_G^{\dagger} b_e \leq \eps$, then $G$ has an $O(1/k+\eps)$ spectrally thin tree.
Let $L_G^{\dagger/2}$ be the square root of $L_G^{\dagger}$. Note that since $L_G^\dagger \succeq 0$, its square root is well defined. For all $e\in F$, let $$v_e = L_G^{\dagger/2} b_e.$$ Then, by the corollary’s assumption, for each $e\in F$, $${\left\lVertv_e\right\rVert}^2 = b_e L_G^\dagger b_e \leq \eps,$$ and the vectors $\{v_e\}_{e\in F}$ are in sub-isotropic position, $$\begin{aligned}
\sum_{e\in F} v_e v_e^\intercal &=& L_G^{\dagger/2} \left(\sum_{e\in F} v_e v_e^\intercal\right) L_G^{\dagger/2}\\
&=& L_G^{\dagger/2} L_F L_G^{\dagger/2} \preceq I.
\end{aligned}$$ In addition, we can show that $\{v_e\}_{e\in F}$ contains $k/2$ disjoint bases. First of all, note that each basis of the vectors $\{v_e\}_{e\in F}$ corresponds to a spanning tree of the graph $(V,F)$. Nash-Williams [@NW61] proved that any $k$-edge connected graph has $k/2$ edge-disjoint spanning trees. Since $(V,F)$ is $k$-edge connected, it has $k/2$ edge-disjoint spanning trees, and equivalently, $\{v_e\}_{e\in F}$ contains $k/2$ disjoint bases.
Therefore, by , there exists a basis (i.e., a spanning tree) $T\subseteq F$ such that $${\left\lVert\sum_{e\in T} v_ev_e^\intercal \right\rVert} \leq \alpha,
\label{eq:LTnorm}$$ for $\alpha=O(\eps+1/k)$. Fix an arbitrary vector $y\in\R^V$. We show that $$y^\intercal L_T y \leq \alpha\cdot y^\intercal L_G y,
\label{eq:LTnormy}$$ and this completes the proof. By for any $x\in\R^V$, $$x^\intercal \left(\sum_{e\in T} v_ev_e^\intercal\right) x \leq \alpha\cdot {\left\lVertx\right\rVert}^2.$$ Let $x=L_G^{1/2}y$, we get $$y^\intercal L_G^{1/2} \left(L_G^{\dagger/2}\sum_{e\in T} b_eb_e^\intercal L_G^{\dagger/2}\right) L_G^{1/2} y \leq \alpha\cdot y^\intercal L_G y.$$ The above is the same as and we are done.
The above corollary completely answers but it is not enough for our purpose in [@AO14]; we need a slightly stronger statement. For a matrix $D\in\R^{V\times V}$ we say $D\preceq_\square L_G,$ if for any set $S\subset V$, $$\bone_S^\intercal D \bone_S \leq \bone_S^\intercal L_G \bone_S,$$ where as usual $\bone_S\in\R^V$ is the indicator vector of the set $S$. In the main theorem of [@AO14] we show that for any $k$-edge-connected graph $G$ (for $k=7\log n$) there is a positive definite (PD) matrix $D\preceq_\square L_G$ and a set $F\subseteq E$ such that $(V,F)$ is $\Omega(k)$-edge-connected and $$\max_{e\in F} b_e^\intercal D^{-1} b_e \leq \frac{\operatorname{polylog}(k)}{k}.$$ To show that $G$ has a combinatorially thin tree it is enough to show that there is a tree $T\subseteq E$ that is $\alpha$-spectrally thin w.r.t. $L_G+D$ for $\alpha=\operatorname{polylog}(k)/k$, i.e., $$L_T \preceq \frac{\operatorname{polylog}(k)}{k} (L_G+D).$$ Such a tree is $2\alpha$-combinatorially thin w.r.t. $G$ because $D\preceq_\square L_G$. Note that the above corollary does not imply $L_G+D$ has a spectrally thin tree because $D$ is not necessarily a Laplacian matrix. Nonetheless, we can prove the existence of a spectrally thin tree with another application of .
\[thm:spectralthintree\] Given a graph $G=(V,E)$, a PD matrix $D$, and $F\subseteq E$ such that $(V,F)$ is $k$-edge connected, if for any edge $e \in F$, $$b_e^\intercal D^{-1} b_e \leq \eps,$$ then $G$ has a spanning tree $T\subseteq F$ such that $$L_T \preceq O(\eps+1/k)\cdot (L_G + D).$$
The proof is very similar to . For any edge $e\in F$, let $v_e = (D+L_G)^{-1/2} b_e$. Note that since $D$ is PD, $D+L_G$ is PD and $(D+L_G)^{-1/2}$ is well defined. By the assumption, $${\left\lVertv_e\right\rVert}^2 = b_e^\intercal (D+L_G)^{-1} b_e \leq b_e^\intercal D^{-1} b_e = \eps,$$ where the inequality uses . In addition, the vectors are in sub-isotropic position, $$\sum_{e\in F} v_e v_e^\intercal = (D+L_G)^{\dagger/2} L_F (D+L_G)^{\dagger/2} \preceq I.$$ The matrix PSD inequality uses that $L_F \preceq L_G \preceq D+L_G$. Furthermore, every basis of $\{v_e\}_{e\in E}$ is a spanning tree of $G$ and by $\Omega(k)$-connectivity of $F$, there are $\Omega(k)$-edge disjoint bases. Therefore, by , there is a tree $T\subseteq F$ such that $${\left\lVert\sum_{e\in T} v_ev_e^\intercal\right\rVert} \leq \alpha,$$ for $\alpha=O(\eps+1/k)$. Similar to this tree satisfies $$L_T \preceq \alpha \cdot (L_G+D),$$ and this completes the proof.
To apply our main theorem first we need to construct a strongly Rayleigh distribution of spanning trees of $G$ such that the marginal probability of each edge is “small”.
For $\gamma:E\to\R$, a $\gamma$-uniform random spanning tree distribution is a distribution on spanning trees of $G$ such that for each tree $T$, $${{\mathbb{P}}_{\mu}\left[T\right]}\propto \prod_{e\in T} \exp(\gamma_e).$$ Borcea, Brändén, and Liggett showed that any $\gamma$-uniform distribution of spanning trees is strongly Rayleigh [@BBL09]. So, it is enough to find a $\gamma$-uniform random spanning tree distribution such that the marginal probability of each edge is small. Asadpour, Goemans, Madry, the second author, and Saberi [@AGMOS10] proved a necessary and sufficient condition for the existence of such a distribution. Recall that the spanning tree polytope of $G$ is the convex hull of the characteristic vectors of spanning trees of $G$.
For any graph $G=(V,E)$ and any point $x$ in the interior of the spanning tree polytope of $G$, there is a $\gamma$-uniform random spanning tree distribution $\mu$, such that for any edge $e\in E$, ${{\mathbb{P}}_{T\sim\mu}\left[e\in T\right]}=z_e$.
Nash-Williams [@NW61] proved that any $k$-edge connected graph has $k/2$ edge-disjoint spanning trees. Let $T_1,\ldots,T_{k/2}$ be $k/2$ edge-disjoint spanning trees of the $k$-edge connected graph, $G'=(V,F)$. Define $x:F\to\R_+$ as follows: $x_e=2/k$ for $e\in \cup_{i=1}^{k/2} T_i$ and $x_e=0$ otherwise. By definition, $x$ is in the spanning tree polytope of $G'$, so a slight perturbation of $x$ is in the [*interior*]{} of the spanning tree polytope. By the above theorem there is a $\gamma$-uniform distribution $\mu$ such that for each edge $e\in F$, $${{\mathbb{P}}_{\mu}\left[e\right]} \leq \frac{2}{k-1}.$$
To apply we just need to add a set of vectors $\{v_e\}_{e\in F'}$ of squared norm at most $\eps$ such that $$\sum_{e\in F'} v_ev_e^\intercal = I - \sum_{e\in E} v_ev_e^\intercal.$$ This is always possible because the RHS of the above identity is positive semidefinite. Note that any $e\in F'$ has zero marginal probability in $\mu$. Applying to $\mu$ and vectors $\{v_e\}_{e\in F\cup F'}$, there is a spanning tree $T$ such that $$\sum_{e\in T} v_ev_e^\intercal \preceq 4\alpha+2\alpha^2,$$ for $\alpha=\eps+2/(k-1)$ as desired.
Proof Overview
--------------
We build on the method of interlacing polynomials of [@MSS12; @MSS13]. Recall that an interlacing family of polynomials has the property that there is always a polynomial whose largest root is at most the largest root of the sum of the polynomials in the family. First, we show that for any set of vectors assigned to the elements of a homogeneous strongly Rayleigh measure, the characteristic polynomials of natural quadratic forms associated with the samples of the distribution form an interlacing family. This implies that there is a sample of the distribution such that the largest root of its characteristic polynomial is at most the largest root of the average of the characteristic polynomials of all samples of $\mu$. Then, we use the multivariate barrier argument of [@MSS13] to upper-bound the largest root of our expected characteristic polynomial.
Our proof has two main ingredients. The first one is the construction of a new class of expected characteristic polynomials which are the weighted average of the characteristic polynomials of the natural quadratic forms associated to the samples of the strongly Rayleigh distribution, where the weight of each polynomial is proportional to the probability of the corresponding sample set in the distribution. To show that the expected characteristic polynomial is real rooted we appeal to the theory of real stability. We show that our expected characteristic polynomial can be realized by applying $\prod_{i=1}^m(1-\partial/\partial^2_{z_i})$ operator to the real stable polynomial $g_\mu(z) \cdot \det(\sum_{i=1}^m z_i v_iv_i^\intercal)$, and then projecting all variables onto $x$. Our second ingredient is the extension of the multivariate barrier argument. Unlike [@MSS13], here we need to prove an upper bound on the largest root of the mixed characteristic polynomial which is very close to zero. It turns out that the original idea of [@BSS14] that studies the behavior of the roots of a (univariate) polynomial $p(x)$ under the operator $1-\partial/\partial_x$ cannot establish upper bounds that are less than one. Fortunately, here we need to study the behavior of the roots of a (multivariate) polynomial $p(z)$ under the operators $1-\partial/\partial^2_{z_i}$. The $1-\partial/\partial^2_{z_i}$ operators allow us to impose very small shifts on the multivariate upper barrier assuming the barrier functions are sufficiently small. The intuition is that, since $$1-\frac{\partial}{\partial^2_{z_i}} = \left(1-\frac{\partial}{\partial_{z_i}}\right)\cdot \left(1+\frac{\partial}{\partial_{z_i}}\right),$$ we expect $(1-\partial/\partial_{z_i})$ to shift the upper barrier by $1+\Theta(\delta)$ (for some $\delta$ depending on the value of the $i$-th barrier function) as proved in [@MSS13] and $(1+\partial/\partial_{z_i})$ to shift the upper barrier by $1-\Theta(\delta)$. Therefore, applying both operators the upper barrier must be moved by no more than $\Theta(\delta)$. We build on the method of interlacing polynomials of [@MSS12; @MSS13]. Recall that an interlacing family of polynomials has the property that there is always a polynomial whose largest root is at most the largest root of the sum of the polynomials in the family. In we show that for any set of vectors assigned to the elements of a homogeneous strongly Rayleigh measure, the characteristic polynomials of natural quadratic forms associated with the samples of the distribution form an interlacing family.
Analogously to [@MSS13], we define a [*mixed characteristic polynomial*]{} as the weighted average of the characteristic polynomials of the natural quadratic forms associated to the samples of the strongly Rayleigh distribution, where the weight of each polynomial is proportional to the probability of the corresponding sample set in the distribution. We show that there is a sample of the distribution such that the largest root of its characteristic polynomial is at most the largest root of the mixed characteristic polynomial.
In we extend the multivariate barrier argument of [@MSS13] to upper-bound the largest root of our mixed characteristic polynomial. Unlike [@MSS13], here we need to prove an upper bound on the largest root of the mixed characteristic polynomial which is very close to zero. It turns out that the original idea of [@BSS14] that studies the behavior of the roots of a (univariate) polynomial $p(x)$ under the operator $1-\partial/\partial_x$ cannot establish upper bounds that are less than one. Therefore, we study the behavior of the roots of a (multivariate) polynomial $p(z)$ under the operators $1-\partial/\partial^2_{z_i}$. Similar to [@BSS14; @MSS13] we keep track of an upper bound on the roots of a polynomial, along with a measure of how far above the roots this upper bound is. The $1-\partial/\partial^2_{z_i}$ operators allow us to impose very small shifts on this upper bound assuming it is sufficiently far from the roots. This, together with the observation that at the start of the argument the upper bound is close to zero, will complete our proof.
Preliminaries
=============
We adopt a notation similar to [@MSS13]. We write ${[m]\choose k}$ to denote the collection of subsets of $[m]$ with exactly $k$ elements. We write $2^{[m]}$ to denote the family of all subsets of the set $[m]$. We write $\partial_{z_i}$ to denote the operator that performs partial differentiation with respect to $z_i$. We use ${\left\lVertv\right\rVert}$ to denote the Euclidean $2$-norm of a vector $x$. For a matrix $M$, we write ${\left\lVertM\right\rVert}=\max_{{\left\lVertx\right\rVert}=1}{\left\lVertMx\right\rVert}$ to denote the operator norm of $M$. We use $\bone$ to denote the all $1$ vector.
Interlacing Families
--------------------
We recall the definition of interlacing families of polynomials from [@MSS12], and its main consequence.
We say that a real rooted polynomial $g(x) = \alpha_0 \prod_{i=1}^{m-1} (x-\alpha_i)$ interlaces a real rooted polynomial $f(x) = \beta_0 \prod_{i=1}^m (x-\beta_i)$ if $$\beta_1 \leq \alpha_1 \leq \beta_2 \leq \alpha_2 \leq \ldots \leq \alpha_{m-1} \leq \beta_m.$$
We say that polynomials $f_1,\ldots,f_k$ have a common interlacing if there is a polynomial $g$ such that $g$ interlaces all $f_i$. The following lemma is proved in [@MSS12].
\[lem:interlacringroot\] Let $f_1,\ldots,f_k$ be polynomials of the same degree that are real rooted and have positive leading coefficients. Define $$f_{\emptyset} = \sum_{i=1}^k f_i.$$ If $f_1,\ldots,f_k$ have a common interlacing, then there is an $i$ such that the largest root of $f_i$ is at most the largest root of $f_\emptyset$.
\[def:interlacingfamily\] Let $\cF\subseteq 2^{[m]}$ be nonempty. For any $S\in\cF$, let $f_S(x)$ be a real rooted polynomial of degree $d$ with a positive leading coefficient. For $s_1,\ldots,s_k\in\{0,1\}$ with $k<m$, let $$\cF_{s_1,\ldots,s_k}:=\{S\in\cF: i\in S \Leftrightarrow s_i=1\}.$$ Note that $\cF=\cF_{\emptyset}$. Define $$f_{s_1,\ldots,s_k} = \sum_{S\in\cF_{s_1,\ldots,s_k}} f_S,$$ and $$f_{\emptyset} = \sum_{S\in\cF} f_S.$$ We say polynomials $\{f_S\}_{S\in\cF}$ form an [*interlacing family*]{} if for all $0\leq k<m$ and all $s_1,\ldots,s_k\in\{0,1\}$ the following holds: If both of $\cF_{s_1,\ldots,s_k,0}$ and $\cF_{s_1,\ldots,s_k,1}$ are nonempty, $f_{s_1,\ldots,s_k,0}$ and $f_{s_1,\ldots,s_k,1}$ have a common interlacing.
The following is analogous to [@MSS13 Thm 3.4].
\[thm:interlacingfamily\] Let $\cF\subseteq 2^{[m]}$ and let $\{f_S\}_{S\in\cF}$ be an interlacing family of polynomials. Then, there exists $S\in\cF$ such that the largest root of $f(S)$ is at most the largest root of $f_\emptyset$.
We prove by induction. Assume that for some choice of $s_1,\ldots,s_k\in\{0,1\}$ (possibly with $k=0$), $\cF_{s_1,\ldots,s_k}$ is nonempty and the largest root of $f_{s_1,\ldots,s_k}$ is at most the largest root of $f_\emptyset$. If $\cF_{s_1,\ldots,s_k,0}=\emptyset$, then $f_{s_1,\ldots,s_k}=f_{s_1,\ldots,s_k,1}$, so we let $s_{k+1}=1$ and we are done. Similarly, if $\cF_{s_1,\ldots,s_k,1}=\emptyset$, then we let $s_{k+1}=0$ and we are done with the induction. If both of these sets are nonempty, then $f_{s_1,\ldots,s_k,0}$ and $f_{s_1,\ldots,s_k,1}$ have a common interlacing. So, by , for some choice of $s_{k+1}\in\{0,1\}$, the largest root of $f_{s_1,\ldots,s_{k+1}}$ is at most the largest root of $f_\emptyset$.
We use the following lemma which appeared as Theorem 2.1 of [@Ded92] to prove that a certain family of polynomials that we construct in form an interlacing family.
\[lem:realrootedimplyinterlacing\] Let $f_1,\ldots,f_k$ be univariate polynomials of the same degree with positive leading coefficients. Then, $f_1,\ldots,f_k$ have a common interlacing if and only if $\sum_{i=1}^k \lambda_i f_i$ is real rooted for all convex combinations $\lambda_i\geq 0$, $\sum_{i=1}^k \lambda_i=1$.
Stable Polynomials {#sec:realstability}
------------------
Stable polynomials are natural multivariate generalizations of real-rooted univariate polynomials. For a complex number $z$, let $\operatorname{Im}(z)$ denote the imaginary part of $z$. We say a polynomial $p(z_1,\ldots,z_m)\in\C[z_1,\ldots,z_m]$ is [*stable*]{} if whenever $\operatorname{Im}(z_i)>0$ for all $1\leq i\leq m$, $p(z_1,\ldots,z_m)\neq 0$. We say $p(.)$ is real stable, if it is stable and all of its coefficients are real. It is easy to see that any univariate polynomial is real stable if and only if it is real rooted.
One of the most interesting classes of real stable polynomials is the class of determinant polynomials as observed by Borcea and Brändén [@BB08].
\[thm:motherrealstability\] For any set of positive semidefinite matrices $A_1,\ldots,A_m$, the following polynomial is real stable: $$\det\Big(\sum_{i=1}^m z_i A_i\Big).$$
Perhaps the most important property of stable polynomials is that they are closed under several elementary operations like multiplication, differentiation, and substitution. We will use these operations to generate new stable polynomials from the determinant polynomial. The following is proved in [@MSS13].
\[lem:partialrealstability\] If $p\in\R[z_1,\ldots,z_m]$ is real stable, then so are the polynomials $(1-\partial_{z_1})p(z_1,\ldots,z_m)$ and $(1+\partial_{z_1})p(z_1,\ldots,z_m)$.
The following corollary simply follows from the above lemma.
\[cor:partial2realstability\] If $p\in\R[z_1,\ldots,z_m]$ is real stable, then so is $$(1-\partial^2_{z_1})p(z_1,\ldots,z_m).$$
First, observe that $$(1-\partial^2_{z_1})p(z_1,\ldots,z_m) = (1-\partial_{z_1})(1+\partial_{z_1})p(z_1,\ldots,z_m).$$ So, the conclusion follows from two applications of .
The following closure properties are elementary.
\[lem:stabilitymultiplication\] If $p\in\R[z_1,\ldots,z_m]$ is real stable, then so is $p(\lambda\cdot z_1,\ldots,\lambda_m\cdot z_m)$ for real-valued $\lambda_1,\ldots,\lambda_m > 0$.
Say $(z_1,\ldots,z_m)\in\C^m$ is a root of $p(\lambda\cdot z_1,\ldots,\lambda_m\cdot z_m)$. Then $(\lambda_1\cdot z_1,\ldots,\lambda_m\cdot z_m)$ is a root of $p(z_1,\ldots,z_m)$. Since $p$ is real stable, there is an $i$ such that $\operatorname{Im}(\lambda_i\cdot z_i)\leq 0$. But, since $\lambda_i>0$, we get $\operatorname{Im}(z_i)\leq 0$, as desired.
\[lem:stabilityaddition\] If $p\in\R[z_1,\ldots,z_m]$ is real stable, then so is $p(z_1+x,\ldots,z_m+x)$ for a new variable $x$.
Say $(z_1,\ldots,z_m,x)\in\C^m$ is a root of $p(z_1+x,\ldots,z_m+x)$. Then $(z_1+x,\ldots,z_m+x)$ is a root of $p(z_1,\ldots,z_m)$. Since $p$ is real stable, there is an $i$ such that $\operatorname{Im}(z_i+x)\leq 0$. But, then either $\operatorname{Im}(x)\leq 0$ or $\operatorname{Im}(z_i)\leq 0$, as desired.
Facts from Linear Algebra
-------------------------
For a Hermitian matrix $M\in\C^{d\times d}$, we write the characteristic polynomial of $M$ in terms of a variable $x$ as $$\X[M](x)=\det(xI-M).$$ We also write the characteristic polynomial in terms of the square of $x$ as $$\X[M](x^2)=\det(x^2I-M).$$
For $1\leq k\leq n$, we write $\sigma_k(M)$ to denote the sum of all principal $k\times k$ minors of $M$, in particular, $$\X[M](x) = \sum_{k=0}^d x^{d-k} (-1)^k\sigma_k(M).$$
The following lemma follows from the Cauchy-Binet identity. See [@MSS13] for the proof.
\[lem:cauchbinet\] For vectors $v_1,\ldots,v_m\in\R^d$ and scalars $z_1,\ldots,z_m$, $$\det\left(xI + \sum_{i=1}^m z_iv_iv_i^\intercal\right) = \sum_{k=0}^d x^{d-k} \sum_{S\subseteq {[m]\choose k}} z^S \sigma_k\big(\sum_{i\in S} v_iv_i^\intercal\big).$$ In particular, for $z_1=\ldots=z_m=-1$, $$\det\left(xI - \sum_{i=1}^m v_iv_i^\intercal\right) = \sum_{k=0}^d x^{d-k} (-1)^k \sum_{S\subseteq {[m]\choose k}} \sigma_k\big(\sum_{i\in S} v_iv_i^\intercal\big).$$
The following is Jacboi’s formula for the derivative of the determinant of a matrix.
\[thm:jacobi\] For an invertible matrix $A$ which is a differentiable function of $t$, $$\partial_t \det(A) = \det(A)\cdot \operatorname{Tr}(A^{-1} \partial_t A).$$
\[lem:inverse\] For an invertible matrix $A$ which is a differentiable function of $t$, $$\frac{\partial A^{-1}}{\partial t} = -A^{-1} (\partial_t A) A^{-1}.$$
Differentiating both sides of the identity $A^{-1} A = I$ with respect to $t$, we get $$A^{-1} \frac{\partial A}{\partial t} + \frac{\partial A^{-1}}{\partial t} A = 0.$$ Rearranging the terms and multiplying with $A^{-1}$ gives the lemma’s conclusion.
The following two standard facts about trace will be used throughout the paper. First, for $A\in\R^{k\times n}$ and $B\in\R^{n\times k}$, $$\operatorname{Tr}(AB)=\operatorname{Tr}(BA).$$ Secondly, for positive semidefinite matrices $A,B$ of the same dimension, $$\operatorname{Tr}(AB)\geq 0.$$ Also, we use the fact that for any positive semidefinite matrix $A$ and any Hermitian matrix $B$, $BAB$ is positive semidefinite.
\[lem:matrixinvineq\] If $A,B\in\R^{n\times n}$ are PD matrices and $A\preceq B$, then $B^{-1} \preceq A^{-1}.$
Since $A\preceq B$, $$B^{-1/2} A B^{-1/2} \preceq B^{-1/2} B B^{-1/2} = I.$$ So, $$B^{1/2} A^{-1} B^{1/2} = (B^{-1/2} A B^{-1/2})^{-1} \succeq I$$ Multiplying both sides of the above by $B^{-1/2}$, we get $$A^{-1} = B^{-1/2} B^{1/2} A^{-1} B^{1/2} B^{-1/2} \succeq B^{-1/2} I B^{-1/2} = B^{-1}.$$
The Mixed Characteristic Polynomial {#sec:mixedpoly}
===================================
For a probability distribution $\mu$, let $d_\mu$ be the degree of the polynomial $g_\mu$.
For $v_1,\ldots,v_m\in\R^d$ and a homogeneous probability distribution $\mu:[m]\to\R_+$, $$\label{eq:mixedcharpoly}
x^{d_\mu-d}{\underset{S\sim \mu}{{{\mathbb{E}}}}}\X\left[\sum_{i\in S} 2v_iv_i^\intercal\right] (x^2) = \prod_{i=1}^m \left(1-\partial^2_{z_i}\right) \left(g_{\mu}(x\bone +z) \cdot \det\left(xI+\sum_{i=1}^m z_i v_iv_i^\intercal \right)\right) \Bigg|_{z_1=\ldots=z_m=0}.$$
We call the polynomial ${{\mathbb E}_{S\sim\mu}}\X[\sum_{i\in S} 2v_iv_i^\intercal](x^2)$ the [*mixed characteristic polynomial*]{} and we denote it by $\mu[v_1,\ldots,v_m](x)$.
For $S\subseteq [m]$, let $z^{2S}=\prod_{i\in S} z_i^2.$ By , the coefficient of $z^{2S}$ in $$g_\mu(x\bone +z)\cdot \det(xI+\sum_{i=1}^m z_iv_iv_i^\intercal)$$ is equal to $$\left(\prod_{i\in S} \partial^2_{z_i}\right) \left(g_{\mu}(x\bone +z)\cdot \det\left(xI+\sum_{i=1}^m z_i v_iv_i^\intercal\right)\right)\Bigg|_{z_1=\ldots=z_m=0}.
$$ Each of the two polynomials $g_\mu(x\bone +z)$ and $\det(xI+\sum_{i=1}^m z_i v_iv_i^\intercal)$ is multi-linear in $z_1,\ldots,z_m$. Therefore, for $k=|S|$, the above is equal to $$\label{eq:firstderivmixedpoly}
2^k\cdot \left(\prod_{i\in S} \partial_{z_i}\right) g_\mu(x\bone +z) \Bigg|_{z_1=\ldots=z_m=0} \cdot \left(\prod_{i\in S} \partial_{z_i}\right) \det\left(xI+\sum_{i=1}^m z_i v_iv_i^\intercal\right)\Bigg|_{z_1=\ldots=z_m=0}.$$ Since $g_\mu$ is a homogeneous polynomial of degree $d_\mu$, the first term in the above is equal to $$x^{d_\mu-k} {{\mathbb{P}}_{T\sim\mu}\left[S\subseteq T\right]}.$$ And, by , the second term of is equal to $$x^{d-k} \sigma_k\left(\sum_{i\in S} z_iv_iv_i^\intercal\right).$$ Applying the above identities for all $S\subseteq [m]$, $$\begin{aligned}
\prod_{i=1}^m \left(1-\partial^2_{z_i}\right) \Bigg(g_{\mu}(x\bone +z) &\cdot \det\left(xI+\sum_{i=1}^m z_i v_iv_i^\intercal \right) \Bigg)\Bigg|_{z_1=\ldots=z_m=0}\\
&=\sum_{k=0}^m (-1)^k \sum_{S\subseteq {[m]\choose k}} \left(\prod_{i\in S} \partial^2_{z_i}\right) \left(g_\mu(x\bone +z)\cdot\det\left(xI+\sum_{i=1}^m z_i v_iv_i^\intercal\right)\right)\Bigg|_{z_1=\ldots=z_m=0}\\
&=\sum_{k=0}^d (-1)^k 2^k x^{d_\mu+d-2k}\sum_{S\in {[m]\choose k}} {{\mathbb{P}}_{T\sim\mu}\left[S\subseteq T\right]}\cdot\sigma_k\left(\sum_{i\in S} v_iv_i^\intercal\right)\\
&=x^{d_\mu-d} {\underset{S\sim\mu}{{{\mathbb{E}}}}}\X\left[\sum_{i\in S}2v_iv_i^\intercal\right](x^2).\end{aligned}$$ The last identity uses .
\[cor:mixedrealstability\] If $\mu$ is a strongly Rayleigh probability distribution, then the mixed characteristic polynomial is real-rooted.
First, by , $$\det\left(xI+\sum_{i=1}^m z_iv_iv_i^\intercal\right)$$ is real stable. Since $\mu$ is strongly Rayleigh, $g_\mu(z)$ is real stable. So, by , $g_\mu(x\bone+z)$ is real stable. The product of two real stable polynomials is also real stable, so $$g_\mu(x\bone +z)\cdot\det\left(xI+\sum_{i=1}^m z_iv_iv_i^\intercal\right)$$ is real stable. implies that $$\prod_{i=1}^m \left(1-\partial^2_{z_i}\right) \left(g_\mu(x\bone +z)\cdot\det\left(xI+\sum_{i=1}^mz_iv_iv_i^\intercal\right)\right)$$ is real stable as well. Wagner [@Wag11 Lemma 2.4(d)] tells us that real stability is preserved under setting variables to real numbers, so $$\prod_{i=1}^m \left(1-\partial^2_{z_i}\right) \left(g_\mu(x\bone +z)\cdot\det\left(xI+\sum_{i=1}^mz_iv_iv_i^\intercal\right)\right)\Bigg|_{z_1=\ldots=z_m=0}$$ is a univariate real-rooted polynomial. The mixed characteristic polynomial is equal to the above polynomial up to a term $x^{d_\mu-d}$. So, the mixed characteristic polynomial is also real rooted.
Now, we use the real-rootedness of the mixed characteristic polynomial to show that the characteristic polynomials of the set of vectors assigned to any set $S$ with nonzero probability in $\mu$ form an interlacing family. For a homogeneous strongly Rayleigh measure $\mu$, let $$\cF=\{S: \mu(S) >0\},$$ and for $s_1,\ldots,s_k\in\{0,1\}$ let $\cF_{s_1,\ldots,s_k}$ be as defined in . For any $S\in\cF$, let $$q_S(x) = \mu(S)\cdot \X\left[\sum_{i\in S} 2v_iv_i^\intercal \right](x^2).$$
\[thm:mixedinterlacing\] The polynomials $\{q_S\}_{S\in\cF}$ form an interlacing family.
For $1\leq k\leq m$ and $s_1,\ldots,s_k\in\{0,1\}$, let $\mu_{s_1,\ldots,s_k}$ be $\mu$ conditioned on the sets $S\in\cF_{s_1,\ldots,s_k}$, i.e., $\mu$ conditioned on $i\in S$ for all $i\leq k$ where $s_i=1$ and $i\notin S$ for all $i\leq k$ where $s_i=0$. We inductively write the generating polynomial of $\mu_{s_1,\ldots,s_k}$ in terms of $g_\mu$. Say we have written $g_{\mu_{s_1,\ldots,s_k}}$ in terms of $g_\mu$. Then, we can write, $$\begin{aligned}
\label{eq:musk1}
g_{\mu_{s_1,\ldots,s_k,1}}(z) &=&
\frac{z_{k+1}\cdot\partial_{z_{k+1}} g_{\mu_{s_1,\ldots,s_k}}(z)}{\partial_{z_{k+1}}
g_{\mu_{s_1,\ldots,s_k}}(z)\big|_{z_i=1}},\\
g_{\mu_{s_1,\ldots,s_k,0}}(z) &=& \frac{g_{\mu_{s_1,\ldots,s_k}}(z) \big|_{z_{k+1}=0}}{g_{\mu_{s_1,\ldots,s_k}}(z)\big|_{z_{k+1}=0, z_i=1 \text{ for } i\neq k+1}}.
\label{eq:musk0}\end{aligned}$$ Note that the denominators of both equations are just normalizing constants. The above polynomials are well defined if the normalizing constants are nonzero, i.e., if the set $\cF_{s_1,\ldots,s_k,s_{k+1}}$ is nonempty. Since the real stable polynomials are closed under differentiation and substitution, for any $1\leq k\leq m$, and $s_1,\ldots,s_k\in\{0,1\}$, if $g_{\mu_{s_1,\ldots,s_k}}$ is well defined, it is real stable, so $\mu_{s_1,\ldots,s_k}$ is a strongly Rayleigh distribution.
Now, for $s_1,\ldots,s_k\in\{0,1\}$, let $$q_{s_1,\ldots,s_k}(x) = \sum_{S\in\cF_{s_1,\ldots,s_k}} q_S(x).$$ Since $\mu_{s_1,\ldots,s_k}$ is strongly Rayleigh, by , $q_{s_1,\ldots,s_k}(x)$ is real rooted.
By , to prove the theorem it is enough to show that if $\cF_{s_1,\ldots,s_k,0}$ and $\cF_{s_1,\ldots,s_k,1}$ are nonempty, then for any $0 < \lambda < 1$, $$\lambda\cdot q_{s_1,\ldots,s_k,1}(x) + (1-\lambda) \cdot q_{s_1,\ldots,s_k,0}(x)$$ is real rooted. Equivalently, by , it is enough to show that for any $0<\lambda<1$, $$\lambda\cdot g_{\mu_{s_1,\ldots,s_k,1}}(z) + (1-\lambda) \cdot g_{\mu_{s_1,\ldots,s_k,0}}(z)
\label{eq:lambdagrealstable}$$ is real stable. Let us write, $$\begin{aligned}
g_{\mu_{s_1,\ldots,s_k}}(z) &=& z_{k+1}\cdot\partial_{z_{k+1}} g_{\mu_{s_1,\ldots,s_k}}(z) + g_{\mu_{s_1,\ldots,s_k}}(z)\big|_{z_{k+1}=0} \\
&=& \alpha\cdot g_{\mu_{s_1,\ldots,s_k,1}}(z) + \beta\cdot g_{\mu_{s_1,\ldots,s_k,0}}(z),\end{aligned}$$ for some $\alpha,\beta>0$. The second identity follows by and . Let $\lambda_{k+1} > 0$ such that $$\frac{\lambda_{k+1}\cdot \alpha}{\lambda} = \frac{\beta}{1-\lambda}.
\label{eq:lambdakp1def}$$ Since $g_{\mu_{s_1,\ldots,s_k}}$ is real stable, by $$g_{\mu_{s_1,\ldots,s_k}}(z_1,\ldots,z_k,\lambda_{k+1}\cdot z_{k+1},z_{k+2},\ldots,z_m)$$ is real stable. But, by the above polynomial is just a multiple of . So, is real stable.
An Extension of [@MSS13] Multivariate Barrier Argument {#sec:barrier}
======================================================
In this section we upper-bound the roots of the mixed characteristic polynomial in terms of the marginal probabilities of elements of $[m]$ in $\mu$ and the maximum of the squared norm of vectors $v_1,\ldots,v_m$.
\[thm:barriermaxroot\] Given vectors $v_1,\ldots,v_m\in\R^d$, and a homogeneous strongly Rayleigh probability distribution $\mu:[m]\to\R_+$, such that the marginal probability of each element $i\in [m]$ is at most $\eps_1$, $\sum_{i=1}^m v_iv_i^\intercal=I$ and ${\left\lVertv_i\right\rVert}^2 \leq \eps_2$, the largest root of $\mu[v_1,\ldots,v_m](x)$ is at most $4(2\eps+\eps^2)$, where $\eps=\eps_1+\eps_2$,
First, similar to [@MSS13] we derive a slightly different expression.
For any probability distribution $\mu$ and vectors $v_1,\ldots,v_m\in\R^d$ such that $\sum_{i=1}^m v_iv_i^\intercal=I$, $$x^{d_\mu-d}\mu[v_1,\ldots,v_m](x)=\prod_{i=1}^m \left(1-\partial^2_{y_i}\right)\left(g_\mu(y)\cdot\det\left(\sum_{i=1}^m y_iv_iv_i^\intercal\right)\right)\Bigg|_{y_1=\ldots=y_m=x}.$$
This is because for any differentiable function $f$, $\partial_{y_i} f(y_i)|_{y_i=z_i+x} = \partial_{z_i} f(z_i+x).$
Let $$Q(y_1,\ldots,y_m)=\prod_{i=1}^m \left(1-\partial^2_{y_i}\right)\left(g_\mu(y) \cdot\det\left(\sum_{i=1}^my_iv_iv_i^\intercal\right)\right).$$ Then, by the above lemma, the maximum root of $Q(x,\ldots,x)$ is the same as the maximum root of $\mu[v_1,\ldots,v_m](x)$. In the rest of this section we upper-bound the maximum root of $Q(x,\ldots,x)$.
It directly follows from the proof of Theorem 5.1 in [@MSS13] that the maximum root of $Q(x,\ldots,x)$ is at most $(1+\sqrt{\eps})^2$. But, in our setting, any upper-bound that is more than 1 obviously holds, as for any $S\subseteq[m]$, $${\left\lVert\sum_{i=1}^m v_iv_i^\intercal\right\rVert} \leq 1.$$ The main difficulty that we are facing is to prove an upper-bound of $O(\eps)$ on the maximum root of $Q(x,\ldots,x)$.
We use an extension of the multivariate barrier argument of [@MSS13] to upper-bound the maximum root of $Q$. We manage to prove a significantly smaller upper-bound because we apply $1-\partial^2_{y_i}$ operators as opposed to the $1-\partial_{y_i}$ operators used in [@MSS13]. This allows us to impose significantly smaller shifts on the barrier upper-bound in our inductive argument.
For a multivariate polynomial $p(z_1,\ldots,z_m)$, we say $z\in\R^m$ is above all roots of $p$ if for all $t\in\R^m_+$, $$p(z+t)> 0.$$ We use $\Ab_p$ to denote the set of points which are above all roots of $p$.
We use the same barrier function defined in [@MSS13].
For a real stable polynomial $p$, and $z\in\Ab_p$, the barrier function of $p$ in direction $i$ at $z$ is $$\Phi^i_p(z) := \frac{\partial_{z_i}p(z)}{p(z)}= \partial_{z_i}\log p(z).$$ To analyze the rate of change of the barrier function with respect to the $1-\partial^2_{z_i}$ operator, we need to work with the second derivative of $p$ as well. We define, $$\Psi^i_p(z) := \frac{\partial^2_{z_i}p(z)}{p(z)}.$$ Equivalently, for a univariate restriction $q_{z,i}(t) = p(z_1,\ldots,z_{i-1},t,z_{i+1},\ldots,z_m)$, with real roots $\lambda_1,\ldots,\lambda_r$ we can write, $$\begin{aligned}
\Phi^i_p(z) &= \frac{q_{z,i}'(z_i)}{q_{z,i}(z_i)}=
\sum_{j=1}^r \frac{1}{z_i-\lambda_j},\\
\Psi^i_p(z)&= \frac{q_{z,i}''(z_i)}{q_{z,i}(z_i)} = \sum_{1\leq j<k\leq r} \frac{2}{(z_i-\lambda_j)(z_i-\lambda_k)}. \end{aligned}$$
The following lemma is immediate from the above definition.
\[lem:psiphi\] If $p$ is real stable and $z\in\Ab_p$, then for all $i\leq m$, $$\Psi^i_p(z) \leq \Phi^i_p(z)^2.$$
Since $z\in\Ab_p$, $z_i > \lambda_j$ for all $1\leq j\leq r$, so, $$\Phi^i_p(z)^2 - \Psi^i_p(z) = \left(\sum_{j=1}^r \frac{1}{z_i-\lambda_j}\right)^2 - \sum_{1\leq j<k\leq r}\frac{2}{(z_i-\lambda_j)(z_i-\lambda_k)} = \sum_{j=1}^r \frac{1}{(z_i-\lambda_j)^2} >0.$$
The following monotonicity and convexity properties of the barrier functions are proved in [@MSS13].
\[lem:barrierprops\] Suppose $p(.)$ is a real stable polynomial and $z\in\Ab_p$. Then, for all $i,j\leq m$ and $\delta\geq 0$, $$\begin{aligned}
\label{eq:monotonicity}
&\Phi^i_p(z+\delta \bone_j) \leq \Phi^i_p(z) \text{ and},~~~~~~~~~~~~~~~~~~~~~\text{(monotonicity)}\\
&\Phi^i_p(z+\delta\bone_j) \leq \Phi^i_p(z) + \delta\cdot\partial_{z_j}\Phi^i_p(z+\delta\bone_j)~~~~~~~\text{(convexity)}.
\label{eq:convexity}\end{aligned}$$
Recall that the purpose of the barrier functions $\Phi^i_p$ is to allow us to reason about the relationship between $\Ab_p$ and $\Ab_{p-\partial^2_{z_i}p}$; the monotonicity property and imply the following lemma.
\[lem:rootsandbarrier\] If $p$ is real stable and $z\in\Ab_p$ is such that $\Phi^i_p(z)<1$, then $z\in\Ab_{p-\partial^2_{z_i} p}$.
Fix a nonnegative vector $t$. Since $\Phi$ is nonincreasing in each coordinate, $$\Phi^i_p(z+t)\leq \Phi^i_p(z) <1.$$ Since $z+t\in\Ab_p$, by , $$\Psi^i_p(z+t)\leq \Phi^i_p(z+t)^2 < 1.$$ Therefore, $$\partial^2_{z_i} p(z+t) < p(z+t) \Rightarrow (1-\partial^2_{z_i})p(z+t) >0,$$ as desired.
We use an inductive argument similar to [@MSS13]. We argue that when we apply each operator $(1-\partial^2_{z_j})$, the barrier functions, $\Phi^i_p(z)$, do not increase by shifting the upper bound along the direction $\bone_j$. As we would like to prove a significantly smaller upper bound on the maximum root of the mixed characteristic polynomial, we may only shift along direction $\bone_j$ by a small amount. In the following lemma we show that when we apply the $(1-\partial^2_{z_j})$ operator we only need to shift the upper bound proportional to $\Phi^j_p(z)$ along the direction $\bone_j$.
\[lem:barriershift\] Suppose that $p(z_1,\ldots,z_m)$ is real stable and $z\in\Ab_p$. If for $\delta>0$, $$\frac2{\delta}\Phi^j_p(z)+\Phi^j_p(z)^2 \leq 1,$$ then, for all $i$, $$\Phi^i_{p-\partial^2_{z_j}p}(z+\delta\cdot\bone_j) \leq \Phi^i_p(z).$$
To prove the above lemma we first need to prove a technical lemma to upper-bound $\frac{\partial_{z_i}\Psi^j_p(z)}{\partial_{z_i}\Phi^j_p(z)}$. We use the following characterization of the bivariate real stable polynomials proved by Lewis, Parrilo, and Ramana [@LPR05]. The following form is stated in [@BB10 Cor 6.7].
\[lem:lax\] If $p(z_1,z_2)$ is a bivariate real stable polynomial of degree $d$, then there exist $d\times d$ positive semidefinite matrices $A,B$ and a Hermitian matrix $C$ such that $$p(z_1,z_2) = \pm \det(z_1A+z_2B+C).$$
\[lem:barrierhigherderivative\] Suppose that $p$ is real stable and $z\in\Ab_p$, then for all $i,j\leq m$, $$\frac{\partial_{z_i}\Psi^j_p(z)}{\partial_{z_i}\Phi^j_p(z)} \leq 2\Phi^j_p(z).$$
If $i=j$, then we consider the univariate restriction $q_{z,i}(z_i) = \prod_{k=1}^r (z_i-\lambda_k)$. Then, $$\begin{aligned}
\frac{\partial_{z_i}\sum_{1\leq k<\l\leq r} \frac{2}{(z_i-\lambda_k)(z_i-\lambda_\l)}}{\partial_{z_i}\sum_{k=1}^r \frac{1}{(z_i-\lambda_k)}} = \frac{\sum_{k\neq \l} \frac{-2}{(z_i-\lambda_k)^2(z_i-\lambda_\l)}}{\sum_{k=1}^r \frac{-1}{(z_i-\lambda_k)^2}} \leq \sum_{\l=1}^r \frac{2}{(z_i-\lambda_\l)} = 2\Phi^j_p(z). \end{aligned}$$ The inequality uses the assumption that $z\in\Ab_p$.
If $i\neq j$, we fix all variables other than $z_i,z_j$ and we consider the bivariate restriction $$q_{z,ij}(z_i,z_j)=p(z_1,\ldots,z_m).$$ By , there are Hermitian positive semidefinite matrices $B_i,B_j$, and a Hermitian matrix $C$ such that $$q_{z,ij}(z_i,z_j) = \pm \det(z_iB_i+z_jB_j+C).$$ Let $M=z_iB_i +z_jB_j+C$. Marcus, Spielman, and Srivastava [@MSS13 Lem 5.7] observed that the sign is always positive, that $B_i+B_j$ is positive definite. In addition, $M$ is positive definite since $B_i+B_j$ is positive definite and $z\in\Ab_p$.
By , the barrier function in direction $j$ can be expressed as $$\label{eq:Philax}
\Phi^j_p(z) = \frac{\partial_{z_j}\det(M)}{\det(M)} = \frac{\det(M)\operatorname{Tr}(M^{-1}B_j)}{\det(M)}=\operatorname{Tr}(M^{-1}B_j).$$ By another application of , $$\begin{aligned}
\Psi^j_p(z)=\frac{\partial^2_{z_j}\det(M)}{\det(M)} &=& \frac{\partial_{z_j}(\det(M)\operatorname{Tr}(M^{-1}B_j))}{\det(M)} \\
&=& \frac{\det(M)\operatorname{Tr}(M^{-1}B_j)^2}{\det(M)} + \frac{\det(M)\operatorname{Tr}((\partial_{z_j}M^{-1})B_j)}{\det(M)}\\
&=& \operatorname{Tr}(M^{-1}B_j)^2+\operatorname{Tr}(-M^{-1}B_jM^{-1}B_j)\\
&=&\operatorname{Tr}(M^{-1}B_j)^2 -\operatorname{Tr}((M^{-1}B_j)^2).\end{aligned}$$ The second to last identity uses . Next, we calculate $\partial_{z_i}\Phi^j_p$ and $\partial_{z_i}\Psi^j_p$. First, by another application of , $$\partial_{z_i} M^{-1}B_j = -M^{-1} B_i M^{-1} B_j=:L.$$ Therefore, $$\partial_{z_i} \Phi^j_p(z) = \partial_{z_i}\operatorname{Tr}(M^{-1}B_j) = \operatorname{Tr}(L),$$ and $$\begin{aligned}
\partial_{z_i} \Psi^j_p(z) &=& \partial_{z_i}\operatorname{Tr}(M^{-1}B_j)^2 - \partial_{z_i}\operatorname{Tr}((M^{-1}B_j)^2) \\
&=& 2\operatorname{Tr}(M^{-1}B_j)\operatorname{Tr}(L) - \operatorname{Tr}\left(L(M^{-1}B_j) + (M^{-1}B_j)L \right)\\
&=& 2\operatorname{Tr}(M^{-1}B_j)\operatorname{Tr}(L) - 2\operatorname{Tr}(LM^{-1}B_j).\end{aligned}$$ Putting above equations together we get $$\begin{aligned}
\frac{\partial_{z_i} \Psi^j_p(z)}{\partial_{z_i} \Phi^j_p(z)} &=& 2\frac{\operatorname{Tr}(M^{-1}B_j)\operatorname{Tr}(L) - \operatorname{Tr}(LM^{-1}B_j)}{\operatorname{Tr}(L)} \\
&=& 2\operatorname{Tr}(M^{-1}B_j) - 2\frac{\operatorname{Tr}(LM^{-1}B_j)}{\operatorname{Tr}(L)}\\
&=& 2\Phi^j_p(z) - 2\frac{\operatorname{Tr}(LM^{-1}B_j)}{\operatorname{Tr}(L)} $$ where we used .
To prove the lemma it is enough to show that $\frac{\operatorname{Tr}(LM^{-1}B_j)}{\operatorname{Tr}(L)} \geq 0$. We show that both the numerator and the denominator are nonpositive. First, $$\operatorname{Tr}(L)=-\operatorname{Tr}(M^{-1}B_iM^{-1}B_j)\leq 0$$ where we used that $M^{-1}B_iM^{-1}$ and $B_j$ are positive semidefinite and the fact that the trace of the product of positive semidefinite matrices is nonnegative. Secondly, $$\operatorname{Tr}(LM^{-1}B_j) = \operatorname{Tr}(-M^{-1}B_iM^{-1}B_jM^{-1}B_j) = -\operatorname{Tr}(B_iM^{-1}B_jM^{-1}B_jM^{-1}) \leq 0,$$ where we again used that $M^{-1}B_jM^{-1}B_jM^{-1}$ and $B_i$ are positive semidefinite and the trace of the product of two positive semidefinite matrices is nonnegative.
We write $\partial_i$ instead of $\partial_{z_i}$ for the ease of notation. First, we write $\Phi^i_{p-\partial^2_j p}$ in terms of $\Phi^i_p$ and $\Psi^j_p$ and $\partial_i\Psi^j_p$. $$\begin{aligned}
\Phi^i_{p-\partial^2_j p} &=& \frac{\partial_i(p-\partial^2_j p)}{p-\partial^2_j p}\\
&=& \frac{\partial_i((1-\Psi^j_p)p)}{(1-\Psi^j_p)p}\\
&=& \frac{(1-\Psi^j_p)(\partial_i p)}{(1-\Psi^j_p)p} + \frac{(\partial_i(1-\Psi^j_p))p}{(1-\Psi^j_p)p}\\
&=& \Phi^i_p - \frac{\partial_i\Psi^j_p}{1-\Psi^j_p}.\end{aligned}$$ We would like to show that $\Phi^i_{p-\partial^2_j p}(z+\delta\bone_j)\leq \Phi^i_p(z)$. Equivalently, it is enough to show that $$-\frac{\partial_i \Psi^j_p(z+\delta\bone_j)}{1-\Psi^j_p(z+\delta\bone_j)} \leq \Phi^i_p(z) - \Phi^i_p(z+\delta\bone_j).$$ By of , it is enough to show that $$-\frac{\partial_i\Psi^j_p(z+\delta\bone_j)}{1-\Psi^j_p(z+\delta\bone_j)} \leq \delta\cdot(-\partial_j\Phi^i_p(z+\delta\bone_j)).$$ By of , $\delta\cdot(-\partial_j\Phi^i_p(z+\delta\bone_j)) >0$ so we may divide both sides of the above inequality by this term and obtain $$\frac{-\partial_i\Psi^j_p(z+\delta\bone_j)}{-\delta\cdot\partial_i\Phi^j_p(z+\delta\bone_j)} \cdot \frac{1}{1-\Psi^j_p(z+\delta\bone_j)} \leq 1,$$ where we also used $\partial_j\Phi^i_p=\partial_i\Phi^j_p$. By , $ \frac{\partial_i \Psi^j_p}{\partial_i \Phi^j_p} \leq 2\Phi^j_p$. So, we can write, $$\frac2{\delta} \Phi^j_p(z+\delta\bone_j)\cdot \frac{1}{1-\Psi^j_p(z+\delta\bone_j)}\leq 1.$$ By and of , $$\begin{aligned}
\Phi^j_p(z+\delta\bone_j) &\leq \Phi^j_p(z),\\
\Psi^j_p(z+\delta\bone_j)&\leq \Phi^j_p(z+\delta\bone_j)^2\leq \Phi^j_p(z)^2.\end{aligned}$$ So, it is enough to show that $$\frac2{\delta}\Phi^j_p(z)\cdot\frac{1}{1-\Phi^j_p(z)^2} \leq 1$$ Using $\Phi^j_p(z) <1$ we may multiply both sides with $1-\Phi^j_p(z)$ and we obtain, $$\frac2{\delta}\Phi^j_p(z) + \Phi^j_p(z)^2 \leq 1,$$ as desired.
Now, we are read to prove .
Let $$p(y_1,\ldots,y_m)=g_\mu(y) \cdot \det\left(\sum_{i=1}^m y_iv_iv_i^\intercal\right).$$ Set $\eps=\eps_1+\eps_2$ and $$\delta=t=\sqrt{2\eps+\eps^2}.$$
For any $z\in \R^m$ with positive coordinates, $g_\mu(z)>0$, and additionally
$$\det\left(\sum_{i=1}^m z_iv_iv_i^\intercal\right)>0.$$ Therefore, for every $t>0$, $t\bone\in\Ab_{p}$.
Now, by , $$\begin{aligned}
\Phi^i_p(y) &=& \frac{(\partial_i g_{\mu}(y))\cdot\det(\sum_{i=1}^m y_iv_iv_i^\intercal)}{g_\mu(y)\cdot\det(\sum_{i=1}^m y_i v_iv_i^\intercal)}
+ \frac{g_\mu(y)\cdot(\partial_i\det(\sum_{i=1}^m y_iv_iv_i^\intercal))}
{g_\mu(y)\cdot\det(\sum_{i=1}^m y_iv_iv_i^\intercal)}\\
&=& \frac{\partial_i g_\mu(y)}{g_\mu(y)} + \operatorname{Tr}\left(\left(\sum_{i=1}^m y_iv_iv_i^\intercal\right)^{-1} y_iy_i^\intercal\right)
$$ Therefore, since $g_\mu$ is homogeneous, $$\begin{aligned}
\Phi^i_p(t\bone) &=& \frac{1}{t} \cdot \frac{\partial_i g_\mu(\bone)}{g_\mu(\bone)} +
\frac{{\left\lVertv_i\right\rVert}^2}{t} \\
&=& \frac{{{\mathbb{P}}_{S\sim\mu}\left[i\in S\right]}}{t} + \frac{{\left\lVertv_i\right\rVert}^2}{t} \leq \frac{\eps_1}{t} + \frac{\eps_2}{t} = \frac{\eps}{t}.\end{aligned}$$ The second identity uses . Let $\phi=\eps/t$. Using $t=\delta$, it follows that $$\frac2{\delta}\phi + \phi^2 = \frac{2\eps}{t^2} + \frac{\eps^2}{t^2} = 1.$$ For $k\in[m]$ define $$p_k(y_1,\ldots,y_m) = \prod_{i=1}^k \left(1-\partial^2_{y_i}\right) \left(g_\mu(y)\cdot\det\left(\sum_{i=1}^m y_iv_iv_i^\intercal\right)\right),$$ and note that $p_m=Q$. Let $x^0$ be the all-$t$ vector and $x^k$ be the vector that is $t+\delta$ in the first $k$ coordinates and $t$ in the rest. By inductively applying and for any $k\in[m]$, $x^k$ is above all roots of $p_k$ and for all $i$, $$\Phi^i_{p_k}(x_k) \leq \phi \Rightarrow \frac2{\delta}\Phi^i_{p_k}(x_i) + \Phi^i_{p_k}(x_i)^2 \leq 1.$$ Therefore, the largest root of $\mu[v_1,\ldots,v_m](x)$ is at most $$t+\delta = 2\sqrt{2\eps+\eps^2}.$$
Let $\eps=\eps_1+\eps_2$ as always. implies that the largest root of the mixed characteristic polynomial, $\mu[v_1,\ldots,v_m](x)$, is at most $2\sqrt{2\eps+\eps^2}.$ tells us that the polynomials $\{q_{S}\}_{S:\mu(S)>0}$ form an interlacing family. So, by there is a set $S\subseteq [m]$ with $\mu(S)>0$ such that the largest root of $$\det\left(x^2I-\sum_{i\in S} 2v_iv_i^\intercal\right)$$ is at most $2\sqrt{2\eps+\eps^2}$. This implies that the largest root of $$\det\left(xI-\sum_{i\in S} 2v_iv_i^\intercal\right)$$ is at most $(2\sqrt{2\eps+\eps^2})^2$. Therefore, $${\left\lVert\sum_{i\in S} v_iv_i^\intercal\right\rVert} =\frac12{\left\lVert\sum_{i\in S} 2v_iv_i^\intercal\right\rVert} \leq \frac12 (2\sqrt{2\eps+\eps^2})^2=4\eps+2\eps^2.$$
Discussion
==========
Similar to [@MSS13] our main theorem is not algorithmic, i.e., we are not aware of any polynomial time algorithm that for a given homogeneous strongly Rayleigh distribution with small marginal probabilities and for a set of vectors assigned to the underlying elements with small norm finds a sample of the distribution with spectral norm bounded away from 1. Such an algorithm can lead to improved approximation algorithms for the Asymmetric Traveling Salesman Problem.
Although our main theorem can be seen as a generalization of [@MSS13] the bound that we prove on the maximum root of the mixed characteristic polynomial is incomparable to the bound of . In we used our main theorem to prove Weaver’s $\text{KS}_r$ conjecture [@Wea04] for $r>4$. It is an interesting question to see if the dependency on $\eps$ in our multivariate barrier can be improved, and if one can reprove $\text{KS}_2$ using our machinery.
### Acknowledgement {#acknowledgement .unnumbered}
We would like to thank Adam Marcus, Dan Spielman, and Nikhil Srivastava for stimulating discussions regarding the main obstacles in generalizing their proof of the Kadison-Singer problem.
[^1]: Computer Science Division, UC Berkeley. Email: [[email protected]]([email protected]).
[^2]: Department of Computer Science and Engineering, University of Washington. This work was partly done while the author was a postdoctoral Miller fellow at UC Berkeley. Email: [[email protected]]([email protected]).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The (778) Theobalda asteroid family attracted little attention so far, but our study shows that it is important in several aspects. In this paper we investigate the origin and evolution of Theobalda family. Firstly, we identify the family as a statistically relevant group in the space of synthetic proper elements. Using the hierarchical clustering method and adopted cut-off velocity of $d_{cutoff}=85$ ms$^{-1}$ we found that Theobalda family currently consists of 128 members. This family is located in the outer belt, near proper semi-major axis $a_{p}$ $\approx$ 3.175 au. This region is crossed by several three-body mean motion resonances which give rise to significant chaotic zones. Consequently, the majority of family members reside on chaotic orbits. Using two independent methods, chaotic chronology and backward integration, we found Theobalda family to be only 6.9 $\pm$ 2.3 Myr old. We have also estimated, that the family was likely produced by the cratering impact on a parent body of diameter $D_{PB}$ $\approx78\pm9$ km.'
author:
- |
Bojan Novaković [^1]\
Department of Astronomy, Faculty of Mathematics, University of Belgrade, Studentski trg 16, 11000 Belgrade, Serbia
title: Portrait of Theobalda as a Young Asteroid Family
---
celestial mechanics, minor planets, asteroids, methods: numerical
Introduction {#s:intro}
============
Asteroid families are believed to originate by a catastrophic disruption of large asteroids. To identify an asteroid family, one looks for clusters of asteroids in the space of proper orbital elements [@MilKne90; @MilKne94]: the proper semi-major axis ($a_{p}$), proper eccentricity ($e_{p}$) and proper inclination ($I_{p}$). The orbital elements describe the size, shape and tilt of orbits. Proper orbital elements, being more constant over time than instantaneous ones, provide a dynamical criterion of whether or not a group of bodies has a common ancestor. Up to now, ejecta from a few tens of major collisions have been discovered in the main belt [e.g. @zappala94; @Diniz05].
The size and velocity distributions of the family members provide important constraints for testing our understanding of the break-up process, but erosion and dynamical evolution of the orbits over time can alter the original signature of the collision. It is nowadays well known that the kinematical structures of the asteroid families evolved over the time, with respect to the original post-impact situations, due to chaotic diffusion, gravitational and non-gravitational perturbations [@MilFar94; @bottke2001; @carruba03; @delloro04]. These mechanisms changed the original shapes of the families produced in collisions, and consequently complicated physical studies of high-velocity collisions.
Unfortunately, most of the observed asteroid families are old enough (older than 100 Myr [@nes2006]) to be substantially eroded and dispersed. On the other hand, young asteroid families (younger than 10 Myr) such as Karin, Veritas and Iannini [@nes2002; @nes2003] or even very young families (younger than 1 Myr) such as Datura, Emilkowalski, 1992YC2, and Lucascavin [@nesvok2006], suffer little erosion during the period of time after a breakup event. Thus, they provide a unique opportunity to study a collisional outcome almost unaffected by orbit evolution.
In this paper we study Theobalda asteroid family. We present its basic properties including the identification of its membership, and the study of cumulative absolute magnitude distribution of the family members. Moreover, the diameter of the parent body has been estimated. We studied in detail the dynamical characteristics in the region occupied by the Theobalda asteroid family and analyzed the role of the dynamics in shaping the family .
As was noted by @nov09 this family is a very good candidate to estimate its age by the method of chaotic chronology (MCC). In order to apply MCC the family has to be located in the region of the main asteroid belt where diffusion takes place. Also, it is necessary that diffusion is fast enough to cause measurable effects, but slow enough so that most of the family members are still forming a robust family structure. As we show, it turns out that Theobalda family is an excellent case in this respect. Given that, as our main result we have estimated the age of the family. Using two different methods, MCC [@menios07; @nov09] and backward integrations [@nes2002], we estimate the age of the family to be 6.9 $\pm$ 2.3 Myr. Thus, we establish it as another young asteroid family.
The paper is organized as follows: In Section \[s:basic\] we present the basic properties of the Theobalda family. We use the hierarchical clustering method (HCM), proposed by @zappala90, to identify family members. Next, the cumulative absolute magnitude distribution of identified family members is discussed, and the size of parent body is estimated. The dynamical characteristics of the region occupied by Theobalda family are presented and discussed, and main mean motion and secular resonances, in that region, identified. The dynamical stability of the family members is analyzed, in particular the stability of the largest member of the family (778) Theobalda. In Section \[s:age\] we estimate the age of the family. This is performed firstly by using the backward integration method, and then by using the method of chaotic chronology. The good agreement between these two results indicates a reliable age determination. Finally, in Section \[s:conclusions\] we summarize our results, discus some possibly interesting relations to other works, and draw our conclusions.
Theobalda family: the basic facts {#s:basic}
=================================
This asteroid family has attracted little attention so far, mainly because the number of asteroids associated with it was relatively small. However, the situation is different at present, and, as we will show later, Theobalda family now has over 100 known members. This number is large enough that the family characteristics can be reliably determined.
Identification of the family members {#ss:hcm}
------------------------------------
The identification of family members is the first step in our study of the Theobalda family. This is done by applying the HCM to the catalog of synthetic proper elements of numbered asteroids [@KneMil00; @synthpro2] from AstDys[^2] (database as of October 2009). The HCM requires that distances among the family members, in the proper elements space, are less than the so called *cut-off* distance ($d_{cutoff}$), which has dimension of velocity. As the cut-off distance is a free parameter of HCM, we tested different values ranging from $20$ to $135$$~ms^{-1}$. Also, we apply HCM using two different *central objects*: (i) (778) $Theobalda$ which has a chaotic orbit, and (ii) (84892) $2003QD_{79}$ which is on the relatively stable orbit. The results are shown in the top panel of Fig. \[f:cutoff\]. The HCM identified the family around (778) Theobalda for $d_{cutoff} \geq 60$$ms^{-1}$, while around (84892) $2003QD_{79}$ family exists even for lowest tested value of $d_{cutoff} = 20$$~ms^{-1}$. For $d_{cutoff} \geq 60$$~ms^{-1}$ resulting family is the same. This suggests that (778) Theobalda probably has been displaced from its original position due to the chaotic diffusion.[^3] In the bottom panel of Fig. \[f:cutoff\] the best-fit power-law index $\gamma$ of the form N($<$H) $\varpropto$ $10^{\gamma H}$ of the cumulative absolute magnitude (H) distribution in the range H$\in$\[13-15\], as a function of cut-off distance ($d_{cutoff}$), is shown[^4]. For $d_{cutoff}$ $\in$\[75,115\] ms$^{-1}$ the number of asteroids as well as index $\gamma$ are nearly constant, and probably each value from this interval can be safely used to identified family members by HCM.[^5] We adopted value of $d_{cutoff}=85$ ms$^{-1}$ to identify *nominal* family. For this value of $d_{cutoff}$, HCM linked 128 asteroids to Theobalda family. There are two main reasons for our choice. The first one is that this value of velocity cutoff corresponds to the center of the *plateau* which can be seen in Fig. \[f:cutoff\]. The second reason is very good agreement between the ages of family estimated applying MCC to two different groups of family members. We will explain this in more detail in Section \[s:age\].
Note that the values of $\gamma$ for family members are always larger than the value of the same index calculated for background asteroids. This is the first indication that the family is relatively young. On the other hand, @parker08 estimated $\gamma$=0.44 for H$\in$\[13.0-15.5\]. This value is much lower than ours, as we found $\gamma$=0.60$\pm$0.02 for the nominal family, and very close to the value that we found for background population in the region of Theobalda family. Probably, @parker08 underestimated this value due to the observational incompleteness, as they worked in the range H$\in$\[13-15.5\] and used smaller dataset for which Sloan Digital Sky Survey (SDSS) colors were available. Although they linked 100 asteroids with the family, a significant number of these asteroids are probably interlopers.
Size of the parent body {#ss:pb}
-----------------------
To estimate the diameter of the parent body ($D_{PB}$) of the Theobalda family, it is necessary to account for small and still undiscovered family members. In general, data set on asteroids below $H$=15 mag is basically complete [@gladman09]. However, as we are dealing with the family at the edge of outer belt, the family members are the C-type asteroids which are several times darker than the S-type asteroids, and, since we are using catalog of synthetic proper elements which does not include all known asteroids, the completeness limit for our sample has to be analyzed. An indication about the completeness limit can be obtained by simply looking at Fig. \[f:sfd\_bw\] for the value of $H$ where slopes of two distribution curves change. This is approximately at about $H$=14.5 mag. Somewhat better estimation of completeness limit can be inferred using the catalog of asteroids, which are not included in the catalog of synthetic proper elements we deal with, i.e. the catalog of multi-opposition objects maintained at AstDys web site. As about 99 per cent of multi-opposition objects, with osculating semi-major axes in the range \[3.15,3.20\] au, have $H\geq$14.2 mag (see Fig. \[f:MB\_aH\]), we assume that the catalog of synthetic proper elements of asteroids in this region is complete up to $H$=14.2 mag.
In order to overcome the problem of observational incompleteness, using the size-frequency distribution (SFD) of the main belt asteroids estimated by @gladman09[^6] and the fact that SFD of asteroid families is considered to be somewhat shallower than that of the background [@morby2003] we added some fictitious bodies with $H\in$\[14.2,17.0\] to the family. This was performed in such a way to make SFD of ”extended” family (real asteroids + fictitious objects), to be somewhat shallower than obtained SFD for background asteroids. More precisely, in order to be able to estimate the uncertainty of our approach, we generated 100 different sets of fictitious objects and then estimated the size of the parent body from each of the 100 sets.
$D_{PB}$ corresponds to a spherical body with volume equal to the estimated total volume of all the family members, including these added members with $H\in$\[14.2,17.0\]. Next, we assume all family members have the same geometric albedo ($p_{v}$) as (778) Theobalda, that is $p_{v}$=0.0589 according to @ted02. Having the values of $H$ and $p_{v}$, the radius $R$ of a body can be estimated, using the relation [e.g. @bowell1989] $$R~{\rm (km)} = 1329~ \frac{10^{\frac{-H}{5}}}{2\, \sqrt{p_{v}}}$$ This allows us to infer that the diameter[^7] of the parent body was $D_{PB}$ $\approx78\pm9$ km. The estimated uncertainty accounts for uncertainties in albedos, absolute magnitudes and SFD. Also it takes into account the dependence on the HCM cut-off. However, the real uncertainty is somewhat larger, e.g. because of the possible interlopers. According to our estimate, the largest remnant (778) Theobalda contains 87 per cent of the mass of the parent body. Although this should be considered as an upper bound, because it does not account for small family members with $H>$ 17, our result suggest that Theobalda family was produced by a cratering impact. The typical density ($\rho$) for C-type asteroids is $\rho=1500~$kg m$^{-3}$ [e.g. @broz05]. Given that, the escape velocity[^8] from Theobalda family parent body was $V_{esc}$$\approx$ 32 ms$^{-1}$.
Dynamical characteristics {#ss:dynamics}
-------------------------
The dynamics in the region of the phase space occupied by Theobalda family members is much like in the case of the Veritas family because two families stretch over the similar range of the proper semi-major axes. The dynamics in the region of Veritas family is very well studied [see e.g. @MilFar94; @KnePav02; @nes2003; @menios07]. Therefore, here we will focus only on some differences between dynamics in the regions occupied by two families. The differences arise from the fact that proper eccentricities of Theobalda family members ($e_{p}$$\approx$0.25) are significantly higher than those of Veritas family ($e_{p}$$\approx$0.06). Also, the proper inclinations are by about $5^{o}$ higher. These make Theobalda family members even more strongly chaotic than Veritas family members.
In Fig. \[f:LCE\] the Lyapounov Characteristic Exponents (LCEs), as a function of proper semi-major axis and eccentricity, in the region occupied by Theobalda family members, as well as in the surrounding area are shown. Most of the orbits in that region are unstable even for comparatively low values of proper eccentricity, while for eccentricity above 0.3 almost all chaotic zones are connected forming a wide *chaotic sea* with a fast diffusion therein. The chaos is also dominating in the region where Theobalda family is located (inside or close to the equivelocity ellipse). This can be better appreciated from Fig. \[f:tlyap\] where LCEs of Theobalda family members are shown. The vertical strip of the largest at $a_{p}$$\approx$3.174 au values of LCEs is associated with (5, -2, -2) three-body[^9] MMR, but it seems that this chaotic zone includes (3, 3, -2) and (7, -7, -2) three-body MMRs as well. Most of the bodies have LCE $\geq$1$\times$$10^{-4}$ $yr^{-1}$, which corresponds to the Lyapounov times $T_{lyap}$ $\leq$ 10,000 yr and these bodies are probably in the so-called *Chirikov regime* [@guzzo2002; @morby02].
Using the proper frequencies $g$ (average rate of the perihelion longitude ̟$\varpi$) and $s$ (average rate of the node longitude $\Omega$)[^10] we found that Theobalda family region is also crossed by two secular resonances $g+s-g_{5}-s_{6}$ and $g+s-g_{6}-s_{6}$ (Fig. \[f:secular\_resonances\]). Both are of the order 4, i.e. they arise from the perturbing terms of degree of at least 4 in eccentricity and inclination [@MilKne90; @kne1991]. However, we did not find evidence (see Section \[sss:diffusion\]) that these resonances increase diffusion speed. Probably this is because these resonances are effective only in narrow bands within the Theobalda family[^11]. Consequently, some of the family members, during their secular cycles, might be temporally trapped in one or both secular resonances, but most of the time these asteroids are outside the secular resonances. In Fig. \[f:secular\_resonances\] the time evolution of the critical angles $\sigma_{1}$=$\varpi$+$\Omega$-$\varpi_{5}$-$\Omega_{6}$, and $\sigma_{2}$=$\varpi$+$\Omega$-$\varpi_{6}$-$\Omega_{6}$, for asteroid (778) Theobalda are shown. This asteroid might be temporally trapped in both secular resonances. Although the short episodes of ”libration” are visible, these events may be related to resonance crossing rather then to the resonance trapping. Most of the time both critical arguments circulate.
In Fig. \[f:dis\_aei\] distributions of family members, as identified by HCM, are shown along with the positions of main mean motion and secular resonances. Obviously, the structure of the family is a result of dynamical mechanisms at work, which are mainly controlled by MMRs. The largest spread of family members, in both ($a_{p}$,$e_{p}$) and ($a_{p}$,$I_{p}$) planes, is associated to (5, -2, -2) resonance. Somewhat smaller spread is observable in the (3, 3, -2) resonance, while (7, -7, -2) resonance caused only small diffusion of asteroids. This agree very well with obtained values of LCEs (see Fig. \[f:tlyap\]).
It is interesting to note that there are gaps (without family members) between the (3, 3, -2) and (5, -2, -2) resonances, as well as between the (5, -2, -2) and (7, -7, -2) resonances. We suggest that this is another confirmation that all these three resonances are connected and make one wide chaotic zone. Because of this, all asteroids from $a_{p}$$\approx$3.167 au to $a_{p}$$\approx$3.181 au reside in one of these three resonances. The asteroids can switch from one resonances to another (see Fig. \[f:778\]), but on the time scale of a few Myr this is a rare event, so that each asteroid spends most of the time in one of the resonances. As a result, due to the some uncertainty in the procedure of computation of synthetic proper elements for resonant asteroids, i.e. averaging does not work well, all bodies appear to be located in (or close to) the center of one of the resonances. This can be verified by using the analytical proper elements[^12] of @MilKne90. These elements are calculated by means of analytical theory based on the series development of the perturbing Hamiltonian, and which does not include averaging. The distribution of Theobalda family members in the space of analytical proper elements (in the $a_{p},e_{p}$ plane) is shown in Fig. \[f:analytical\]. The shown distribution is roughly random and without gaps in terms of proper semi-major axis, what confirms our claim that the gaps appear due to the averaging procedure. Moreover, it means that switching from one resonance to another must be a rare event, but, the fact that not all of the asteroids are located in the center of one of the resonances, is another evidence that resonance switching is possible, i.e. these three resonances are connected.
The position of the largest remnant, asteroid (778) Theobalda, is not close to the center of the family. This is evident also in ($a_{p}$,$I_{p}$) plane, but it is more obvious in ($a_{p}$,$e_{p}$) plane. As we already mentioned above, this asteroid has probably been displaced from its original position due to the some dynamical mechanisms. It is located close to or inside the (7, -7, -2) three-body MMR, which might be responsible for its relatively high proper eccentricity. However, its proper semi-major axis is also larger than that of the center of family, and this could not be explained by (7, -7, -2) resonance. Because of that, we investigate dynamics of this asteroid in more detail. The orbit of (778) Theobalda is propagated for 100 Myr back in time[^13]. As we will show later, the family is about 6-7 Myr old. Why than it is meaningful to integrate 100 Myr? The answer is hidden in the chaotic motion of this asteroid. As we know, chaos is not predictable on the time scales of several times the inverse of LCE, which is in the case of (778) Theobalda $\approx$8,000 yr. All that we can achieve is to show what kind of behavior (i.e. motion) is possible. In this respect, our 100 Myr long integrations are equivalent to many shorter integrations with slightly different initial conditions. Similar technique was used by @laskar94 to study stability of the Solar system.
Fig. \[f:778\] shows 100 Myr of the back-in-time evolution of the mean semi-major axis of asteroid (778) Theobalda. Initially, its semi-major axis oscillates around 3.181 au (close to (7, -7, -2) resonance), but after about 15 Myr (in the past) the value of the semi-major axis drops to $\approx$3.174 au. Around this time it actually switches from (7, -7, -2) to (5, -2, -2) resonance. There is another switch of the resonance about 42 Myr, when the asteroid is temporally trapped in (3, 3, -2) resonance at $\approx$3.168 au. Finally, at about 95 Myr in the past, it went back from (5, -2, -2) to (7, -7, -2) resonance. This is the confirmation that these three resonances are connected at higher eccentricities ($e_{p}$ $\geq$ 0.25). But again, due to the chaoticity, the behavior of the mean semi-major axis of asteroid (778) Theobalda does not represent quantitatively its real motion, but qualitatively. Still, behavior of its semi-major axis suggests that chaos may be responsible for displacement of (778) Theobalda from the center of family in terms of $e_{p}$ and $I_{p}$. However, if this was the case, this asteroid probably spend some time residing in (5, -2, -2) resonance which is strong enough to increase its eccentricity from $e_{p}$$\approx$0.253 to $e_{p}$$\approx$0.259, on the time scale of several Myr.
Studying the distribution of family members shown in Fig. \[f:dis\_aei\], it can be noted that there are no family members located inside equivelocity ellipses, at $a_{p}$$\approx$3.165 au. Contrary to the gaps between the resonances, the absence of asteroids that belong to the family in this region cannot be explained by dynamical instability or by ”weakness” of the procedure of proper elements calculation. Although, a detail study of this problem is beyond the scope of our work, we believe that this may be related to the impact characteristics (cratering event), which ”forced” fragments to be symmetrically distributed around the semi-major axis of the largest fragment (778) Theobalda.
The age of Theobalda family {#s:age}
===========================
Backward integration {#ss:backwards}
--------------------
Backward integration of orbits is very accurate method for family age estimation, which works well with young families. It is based on the fact that due to the planetary perturbations the orientation of orbits in the space changes over time. Consequently, two angles that determine the orientation of orbits in space, the longitude of the ascending node ($\Omega$) and the longitude of perihelion ($\varpi$), evolve with different but nearly constant speeds for individual orbits. After some time this effect spreads out $\Omega$ and $\varpi$ uniformly around $360^{o}$. On the other hand, immediately after the disruption of the parent body, the orientations of the fragments’ orbits must have been nearly the same. Given that, the age of an asteroid family can be determined by integrating the orbits of the family members backwards, until the orbital orientation angles cluster around some value. This method was used by @nes2002 [@nes2003] to determine the ages of the Karin cluster (5.8$\pm$0.2 Myr) and Veritas family (8.3$\pm$0.5 Myr).
Here we applied Nesvorn' y et al.’s method to try to estimate the age of Theobalda family. By integrating the orbits of the Theobalda asteroid family back in time, hopefully, we can find a conjunction of orbital elements ($\Omega$ and $\varpi$), which occurred only immediately after the disruption of the parent body. This method is, however, limited to groups of objects moving on regular orbits, which, even in that case, can be accurately track up to 20 Myr in the past. Similarly as in the case of Veritas family [@nes2003; @menios07], only a fraction of Theobalda family members satisfies this condition and can be accurately integrated back in time. Also, as was pointed out by @nes2003 [see also @nes2008] the region around $a_{p}$=3.175 au is close enough to the 2/1 MMR with Jupiter to undergo fast differential evolution of the arguments of perihelion. This induces variability in the evolution histories and complicates any attempt to determine the age of the Theobalda family using arguments of perihelion. Thus, we selected 13 Theobalda family members which have Lyapounov times $T_{lyap}\geq$ $10^{5}$yr and propagated their orbits 20 Myr backwards. All these members are located at $a_{p}$$\geq$3.183 au. In Fig. \[f:clustering\] the average value of $\Delta\Omega$, for these 13 asteroids, is shown. Conjunction of nodal longitudes at $\approx$6.2 Myr suggests that the Theobalda family, or at least a part of the family located at $a_{p}$$\geq$3.183 au, was formed by a catastrophic collision at that time. The average $\Delta\Omega$, at $\approx$6.2 Myr, is $\approx$$58^{o}$, much smaller than at any other time. This suggests a statistical significance of the $\approx$6.2 Myr event. In this case, however, $\langle\Delta\Omega\rangle$ values are substantially more spread at $\approx$6.2 Myr than in the case of Karin cluster ($\langle\Delta\Omega\rangle$ is $\approx$$10^{o}$) or Veritas family ($\langle\Delta\Omega\rangle$ $\approx$$40^{o}$). This is primarily due to two reasons: (1) at least a few MMRs exist in the semi-major axis range from 3.18 to 3.19 au; thus, despite the present long Lyapounov times of the selected orbits, these orbits might have experienced periods of chaotic motion in the past; and (2) all regular bodies, whose orbits can be accurately tracked back in time, are small bodies ($\lesssim$5 km) and consequently subject to Yarkovsky thermal force, which, even on this relatively short time scale, can produce large enough changes in the semi-major axes, and consequently to affect the secular frequencies in a way that is difficult to reconstruct.
In order to estimate how sensitive this result is on the semi-major axis drift due to the Yarkovsky effect, an additional investigation should be carried out. As the Yarkovsky induced drift depends on several parameters, we had to decide the values of the parameters characterizing it. These are asteroid spin axis orientation ($\gamma$), rotational period (P), surface and bulk densities ($\rho$), surface thermal conductivity (K) and specific heat capacity (C). As Theobalda family members are most likely C-type asteroids, we have adopted the following values of parameters: $K=0.01-0.5~$ \[W(m K)$^{-1}$\], $C=680-1500~$\[J(Kkg)$^{-1}$\], and the same value for surface and bulk density $\rho=1300-1500~$\[kgm$^{-3}$\]. The rotational periods are chosen according to a Gaussian distribution peaked at $P=8~$h, while the distribution of spin axes orientation is assumed to be uniform. These values are consistent with C-type asteroids [@broz06; @broz08].
Next, we made 20 ”yarko” clones for each of the 13 regular members, by assigning random values of the parameters, from adopted intervals, to each clone. Then, we integrated[^14] the orbits of all clones (260 orbits in total), but accounting not only for gravitational perturbations, but also for Yarkovsky effect. The initial orbital elements of the asteroids and planets were the same as in the previous experiment. Finally, we checked how the value of average $\Delta\Omega$ change with different combinations of clones. We found that the result shown in Fig. \[f:clustering\] is very sensitive to the Yarkovsky induced drift, as expected. In a few cases any significant clustering even disappeared, but in most of the cases we obtained a deeper minimum. The deepest minimum that we found is related to the clustering within $\approx$$31^{o}$ at about 6.4 Myr ago (Fig. \[f:clustering\]), which is still within the error bars obtained from the integrations without Yarkovsky force. We would like to note here that this high sensitivity of the result on the Yarkovsky parameters could help us to estimate the rotational periods and spin axis orientations of these 13 asteroids. This can be achieved similarly as was done by @nesbot04 for the Karin cluster members, but we reserve this for a future work.
Although, the clustering at about 6.2 Myr within $\approx$$58^{o}$ is the most significant on the time scale of 20 Myr, there is another clustering at about 15.5 Myr within $\approx$$65^{o}$ (see Fig. \[f:clustering\]). As this clustering appears in the more distant past where we should expect less tight clustering, it is not possible to rule out its significance. Also, to use the argument of perihelions is impossible, because the changes in the semi-major axes, caused by Yarkovsky effect, coupled with large gradient[^15] of secular frequency ($dg/da$ $\approx$$0.3^{o}$ $yr^{-1}$ au, where $g$ is the longitude of perihelion frequency), erase evolution histories of these angles. Given that, we believe that, in the case of the Theobalda family, backward integration method is not enough to draw a firm conclusion about the age of the family.
Chaotic chronology {#ss:mch}
------------------
In this section we present results obtained by using MCC in order to estimate the age of Theobalda family. This model was successfully applied by @nov09 to estimate ages of Veritas and Lixiaohua asteroid families [see also @nov09b]. In order to apply MCC the family has to be located in the region of the main belt where diffusion takes place. Also, it is necessary that diffusion is fast enough to cause measurable effects, but slow enough so that most of the family members are still forming a robust family structure. As our results about diffusion speed suggest, Theobalda family is an excellent example in this respect (see Section \[sss:diffusion\]).
The basic steps and the model which we used in our Markov Chain Monte Carlo (MCMC) simulations are explained in @nov09, and thus we will describe these here only briefly. Our model simulate the evolution of the family in the 3-D space, i.e. proper semi-major axis $a_{p}$ and two actions $J_{1}$, $J_{2}$ (see Section \[sss:diffusion\] for definition of these actions). At the beginning of a simulation the *random walkers* are distributed in the region which was presumably occupied by the family members immediately after the impact event. Then, at each time step $dt$ the random walkers can change their positions in every direction, in the 3-D space. The length of the jump in $a_{p}$ is controlled by Yarkovsky thermal force [@farvok99], while the length of the jumps in $J_{1}$ and $J_{2}$ depend on diffusion speed, i.e. on the diffusion coefficients. At the time step when 0.3 per cent of random walkers leave an ellipse[^16] in the ($J_{1}$, $J_{2}$) plane, which corresponds to a 3$\sigma$ confidence interval of a two-dimensional Gaussian distribution, the simulation stops. The number of time steps multiplied by the time step $dt$ gives the age of the family.
### Diffusion coefficients {#sss:diffusion}
One of the most important information, which are needed as input for MCMC simulations, are the values of diffusion coefficients in the region of interest. As was shown by @nov09, to obtain good estimate of the family age by MCC, it is enough to determine diffusion coefficients as a function of proper semi-major axis $a_{p}$. This is our next step.
As well as MCC, the procedure of determination of diffusion coefficients, as the functions of proper semi-major axis, is described in @nov09. Let us mention here only its main features and numbers related to this work: the orbits of $\sim$5,000 fictitious bodies distributed randomly in the same ranges of osculating orbital elements as the real family members at present, are propagated for 10 Myr; then, the time series of mean elements [@MilKne98] for all of them are calculated; the mean elements are transformed to actions according to relations[^17] $J_{1}\approx\frac{1}{2}\sqrt{\frac{a_{p}}{a_{J}}}e^{2}_{m}$ and $J_{2}\approx\frac{1}{2}\sqrt{\frac{a_{p}}{a_{J}}}\sin^{2}I_{m}$; next, the family is split in the small cells, in terms of $a_{p}$, using a kind of moving-average technique with cell size of $\Delta a_{p} = 5 \times 10^{-4}$ au and step size of $\delta a_{p} = 2 \times 10^{-4}~$au; finally, the mean squared displacements $\langle(\Delta J)^{2}\rangle$, for both actions, are calculated, and the diffusion coefficients $D(J_{1})$, $D(J_{2})$ for each cell as the least-squares fit slope of the $\langle(\Delta J)^{2}\rangle$(t) curve, are determined.
The obtained values of diffusion coefficients $D(J_{1})$ and $D(J_{2})$, in the Theobalda family region are shown in Fig. \[f:diff\_coef\]. The fastest diffusion is associated to (5, -2, -2) three-body MMR, but the diffusion is very fast in (3, 3, -2) three-body MMR as well, and these two chaotic zones seem to be connected. The third chaotic zone, associated to (7, -7, -2) resonance, is connected to the first two in terms of diffusion in $J_{1}$, but not in terms of diffusion in $J_{2}$. This is in a relatively good agreement with results presented in Section \[ss:dynamics\]. The diffusion is somewhat faster in $J_{1}$ than in $J_{2}$, and while the local minimum near the center of (5, -2, -2) resonance exist for $D(J_{1})$, similarly as in the case of Veritas family [see @nov09], while there is no such a feature for $D(J_{2})$. In the region for $a_{p}\geq$3.18 au the values of $D(J_2)$ are practically zero, but there is some diffusion in the $J_{1}$. It should be noted here that this might affect age estimation using backward integration method, but not significantly, because the most of 13 asteroids that we used to apply backward integration method are located close to $a_{p}=$3.185 au, where both values, $D(J_{1})$ and $D(J_{2})$, are close to zero.
The important general conclusion can be drawn by comparing the values of diffusion coefficients obtained for the region occupied by Veritas family [@nov09] to the values obtained here. Two families stretch over the similar range of the semi-major axis, but members of Theobalda family have somewhat higher inclinations and significantly higher eccentricities. The estimated diffusion is about one order of magnitude faster in the region occupied by Theobalda family than in the region occupied by Veritas family. This confirms the fact that chaos is dominant at higher eccentricities.
### Monte Carlo simulations {#ss:mcmc}
Having obtained the values of diffusion coefficients, we are ready to apply MCC to estimate the age of the family. There are two separate parts of the Theobalda family suitable for application of MCC. These are bodies inside (5, -2, -2) and (3, 3, -2) three-body MMRs. Following @menios07 who deal with Veritas family, we called these bodies Group A (5, -2, -2) and Group B (3, 3, -2). As the results about diffusion coefficients confirmed, there exists significant diffusion in both groups. This gives an unique opportunity to apply MCC to both groups and to obtain two independent age estimates. A good agreement between these two estimates, as well as with the age derived using backward integration method, would suggest a reliable result.
As the present size of the chaotic zone is a critical parameter in our model, we start with the family as identified by applying HCM for velocity cutoff of $d_{cutoff}=65$ ms$^{-1}$. This is probably the lowest acceptable value of $d_{cutoff}$ in the case of Theobalda family. With this cutoff velocity we identified 30 bodies from Group A and 16 bodies from Group B. Corresponding sizes of these groups in $J_{1}$ and $J_{2}$ are: Group A ($10.67\pm1.03)\times 10^{-4}$ and ($3.82\pm0.47)\times 10^{-4}$; Group B ($10.32\pm1.05)\times 10^{-4}$ and ($4.00\pm1.06)\times 10^{-4}$.
Using these sizes of two chaotic groups, for each group, we performed 16 sets of MCMC simulations (each set consisting of 100 runs), by using different number of random walkers $n$ (2000 or 5000), time step $dt$ (from 100 yr to 2000 yr) and for two initial sizes of the family which correspond to velocities of $v=35$ ms$^{-1}$ and $v=40$ ms$^{-1}$ (see Fig. \[f:dis\_aei\]). From these simulations we derived the age of family to be 2.5 $\pm$ 1.1 Myr (using Group A bodies) and 7.2 $\pm$ 3.1 Myr (using Group B bodies).[^18] The obvious discrepancy between two results needs to be investigated further. The age obtained from Group B is in agreement with what we found using backward integration method, while the age obtained from Group A suggests that the family could be much younger. Also, the discrepancy between ages derived from two different groups, may be an indication that identification of family members has not been good, i.e. the velocity cut-off of 65 ms$^{-1}$ is too low.
Because of that, we repeated all simulations using our nominal family. For $d_{cutoff}=85$ ms$^{-1}$ we identified 40 bodies from Group A and 18 bodies from Group B. In this case, the corresponding sizes of these groups in $J_{1}$ and $J_{2}$ are: Group A ($14.86\pm1.24)\times 10^{-4}$ and ($6.03\pm0.65)\times 10^{-4}$; Group B ($11.87\pm1.29)\times 10^{-4}$ and ($3.80\pm0.95)\times 10^{-4}$. From these sizes and the same sets of MCMC simulations as in the previous case, we derived the age of Theobalda family to be 6.9 $\pm$ 1.8 Myr (using Group A bodies) and 7.2 $\pm$ 3.0 Myr (using Group B bodies).
Now, the agreement between two results is very good, and also, both results agree quite well with the age obtained by backward integration method.[^19] This, in our opinion, is a very strong indication that Theobalda family was formed about 7 Myr ago.
The fact, that four out of five, different age estimates, agree well, is the reason why we reject the age of 2.5 $\pm$ 1.1 Myr, derived using Group A ($d_{cutoff}=65$ ms$^{-1}$), as a possible solution. Thus, in order to obtain our final estimate of the age of Theobalda family, we use four results which are in a good agreement (see Fig. \[f:age\]). The values of the mean ($\mu$) and standard deviation ($\sigma$) of non-overlapping sub-samples, of the same size, can be calculated as:
$$\mu = \frac{\Sigma_{i=1}^{m}\mu_{i}}{m}
\label{eq:mean}$$
$$\sigma=\sqrt{\frac{\Sigma_{i=1}^{m} ((r-1)\sigma_{i}^{2}+r\mu_{i}^{2}) - mr\mu^{2}}{mr-1}}
\label{eq:std}$$
where $m$ is the number of sub-samples, $r$ is the size of each sub-sample (100 in our case), $\mu_{i}$ is the mean of $i$-th sub-sample, and $\sigma_{i}$ is the standard deviation of $i$-th sub-sample. Using Eqs. \[eq:mean\] and \[eq:std\] we obtain the final age estimate, Theobalda asteroid family is 6.9 $\pm$ 2.3 Myr old.
Summary, Discussions and Conclusions {#s:conclusions}
====================================
We have presented here a detailed study of Theobalda asteroid family. We found that family now consists of 128 members. By analyzing SFD of the identified family members we were able to infer diameter of the parent body to be $D_{PB}$ $\approx78\pm9$ km. However, this estimate is based on the assumption that all family members have the same albedo as the largest family member, asteroid (778) Theobalda. In order to obtain better estimate, the albedos of as many as possible family members are desirable. Ongoing projects, such as Wide-Field Infrared Survey Explorer (WISE) should improve situation significantly in this respect.
The most, but not all, of Theobalda family members move on chaotic orbits, thus, giving rise to the significant chaotic diffusion which has been changing the kinematical structure of the family over time. The study of dynamical characteristics, in the region occupied by the family, showed that three three-body MMRs are the most efficient in shaping the family. These are (3, 3, -2), (5, -2, -2) and (7, -7, -2) resonances, and they are connected in this high eccentricity region, allowing bodies to switch from one resonance to another.
The fact that some of the family members have stable orbits was the reason why we were able to apply backward integration method to estimate the age of the family. On the other hand, presence of the chaos in the region occupied by the family, allows to use MCC in order to estimate the age. Using both methods, and combining the results, we found the age of Theobalda family to be 6.9 $\pm$ 2.3 Myr. Given the very good agreement between results obtained with different methods as well as when applied to different groups, we believe this estimate is very robust. Thus, this is another family younger than 10 Myr. This result has several important implications, and some of them we mention bellow.
The young asteroid families are also known to be source of solar system dust bands [see e.g. @grun85; @nes2003]. The origin of three main dust bands is known, and they correspond to Karin, Veritas and Beagle family [@nes2006b; @nes2008]. Also, the very young Emilkowalski family is the most probably source of incomplete dust band at $17^{o}$ [@espy2009]. On the other hand, the origin of some less prominent bands, such as so called M/N dust band, is still not quite clear. The Theobalda family’s young age, and its proper inclination of $I_{p}$ $\approx14^{o}.3$ suggest that it might be a possible source of M/N dust band ($I_{p}$ $\approx15^{o}$) [@sykes90]. On the other hand, this dust band was linked to (170) Maria asteroid family [@reach97], and more recently to (1521) Seinajoki cluster [@nes2003]. In any case, dust band produced by such a young family, as Theobalda, should be observable. The size of its parent body also suggests that it should produce a prominent dust band. If this is not M/N dust band, then there must be another dust band which can be linked to this family. Alternatively, it should be explained why and how this dust band has disappeared.
Theobalda asteroid family is located very close to the region where three (out of four) so-called *main belt comets* (MBCs)[^20] have been discovered [see e.g. @jewitt2009]. As was suggested by @hsieh2009, it is possible that this kind of bodies can be found among the members of other young families, probably many of which waiting to be discovered. Being young and dominated by C-type asteroids, we believe Theobalda family is very good place to start.
Acknowledgements {#acknowledgements .unnumbered}
================
I am grateful to Zoran Kneževi' c and Rade Pavlovi' c for their useful suggestions on the manuscript. I also would like to thank David Nesvorn' y, the referee, for his useful comments and suggestions that help me to improve this article. This work has been supported by the Ministry of Science and Technological Development of the Republic of Serbia (Project No 146004 “Dynamics of Celestial Bodies, Systems and Populations”).
Bottke, W.F., Vokrouhlick' y, D., Brož, M., Nesvorn' y, D., Morbidelli, A., 2001, Science, 294, 1693
Bowell, E., Hapke, B., Domingue, D., Lumme, K., Peltoniemi, J., Harris, A.W., 1989, IN: Asteroids II; Proceedings of the Conference, Tucson, AZ, Mar. 8-11, 1988 (A90-27001 10-91). Tucson, AZ, University of Arizona Press, 1989, p. 524-556.
Brož, M., Vokrouhlick' y, D., Roig, F., Nesvorn' y, D., Bottke, W.F., Morbidelli, A., 2005, MNRAS, 359, 1437
Brož, M., 2006, Yarkovsky effect and the dynamics of Solar system. Ph.D. Thesis, Faculty of Mathematics and Physics, Charles University, Prague.
Bro[ž]{}, M., Vokrouhlický, D., 2008, MNRAS, 390, 715
Carruba, V., Burns, J.A., Bottke, W.F., Nesvorn' y, D., 2003, Icarus, 162, 308
Carruba, V., 2009, MNRAS, 395, 358
Dell’Oro, A., Bigongiari, G., Paolicchi, P., Cellino, A., 2004, Icarus, 169, 341
Dell’Oro, A., Cellino A., 2007, MNRAS, 380, 399
Durda, D.D., Bottke, W.F., Nesvorn' y, D., Enke, B.L., Merline, W.J., Asphaug, E., Richardson, D.C., 2007, Icarus, 186, 498
Espy A. J., Dermott S. F., Kehoe T. J. J., Jayaraman S., 2009, P&SS, 57, 235
Farinella, P., Vokrouhlick\` y, D., 1999, Science, 283, 1507
Gladman, B.J., Davis, D.R., Neese, C., Jedicke, R., Williams, G., Kavelaars, J.J., Petit, J-M., Scholl, H., Holman, M., Warrington, B., Esquerdo, G., Tricarico, P., 2009, Icarus, 202, 104
Grün, E., Fechtig, H., Zook, H.A., Giese, R.H., 1985, in R.H. Giese and P.L. Lamy (eds) *Properties and interactions of interplanetary dust; Proceedings of the 85 IAU Colloquium*, Dordrecht, D. Reidel Publishing Co., 411
Guzzo, M., Kneževi' c, Z., Milani, A., 2002, Cel. Mech. Dyn. Astron., 83, 121
Hsieh, H.H., Jewitt, D.C., Fern' andez, Y.R., 2004, Astron. J., 127, 2997
Hsieh, H.H., 2009, Astron. Astrophys., 505, 1297
Ivezić, Ž., and 32 colleagues, 2001, Astron. J., 122, 2749
Jewitt, D., Yang, B.,; Haghighipour, N., 2009, Astron. J., 137, 4313
Knežević, Z.; Milani, A.; Farinella, P.; Froeschl' e, Ch.; Froeschl' e, Cl., 1991, Icarus, 93, 316
Knežević, Z., Milani, A., 2000, Cel. Mech. Dyn. Astron., 78, 17
Knežević Z., Milani A., 2003, Astron. Astrophys., 403, 1165
Knežević, Z., Pavlović, R., 2002, EM&P, 88, 155
Laskar, J., 1994, Astron. Astrophys., 287, L9
Milani, A., Farinella, P., 1994, Nature 370, 40
Milani, A., Knežević, Z., 1990, Cel. Mech. Dyn. Astron., 49, 347
Milani, A., Knežević, Z., 1994, Icarus, 107, 219
Milani, A., Knežević, Z., 1998, Cel. Mech. Dyn. Astron., 71, 55
Moth' e-Diniz, T., Roig, F., Carvano, J.M., 2005, Icarus, 174, 54
Morbidelli, A., Zappal\` a, V., Moons,M., Cellino, A., Gonczi, R., 1995, Icarus 118, 132
Morbidelli, A., 2002, Modern Celestial Mechanics: Aspects of Solar System Dynamics. Taylor & Francis, London.
Morbidelli, A., Nesvorn' y, D., Bottke, W.F., Michel, P., Vokrouhlick\` y, D., Tanga, P., 2003, Icarus, 162, 328
Nesvorn' y, D., Morbidelli, A., 1998, Astron. J., 116, 3029
Nesvorn' y, D., Bottke, W.F., Dones, L., Levison, H.F., 2002, Nature, 417, 720
Nesvorn' y, D., Bottke, W.F., Levison, H.F., Dones, L., 2003, Astrophys. J., 591, 486
Nesvorn[ý]{}, D., Bottke W. F., 2004, Icarus, 170, 324
Nesvorn' y, D., Bottke, W.F., Vokrouhlick' y, D., Morbidelli, A., Jedicke, R., 2006, in: D. Lazzaro, S. Ferraz-Mello and J. Fernandez (eds.) *Proceedings of the 229th Symposium of the IAU, Asteroids, Comets, Meteors*, Cambridge University Press, Cambridge, 289
Nesvorn[ý]{} D., Vokrouhlick[ý]{} D., Bottke W. F., Sykes M., 2006b, Icarus, 181, 107
Nesvorn' y, D., Vokrouhlick' y, D., 2006, Astron. J., 132, 1950
Nesvorn' y, D., Bottke, W.F., Vokrouhlick' y, D., Sykes, M., Lien, D.J., Stansberry, J., 2008, Astrophys. J., 679, L143
Nobili, A.M., Milani, A., Carpino, M., 1989, Astron. Astrophys., 210, 313
Novakovi' c, B., Balaž, A., Kneževi' c, Z., Potočnik, M., 2009, Serb. Astron. J., 179, 75
Novakovi' c B., Tsiganis, K., Kne[ž]{}evi[' c]{}, Z., 2010a, MNRAS, 402, 1263
Novakovi' c, B., Tsiganis, K., Kne[ž]{}evi[' c]{}, Z., 2010b, Cel. Mech. Dyn. Astron., 107, 35
Parker, A., Ivezi' c, Ž., Juri' c, M., Lupton, R., Sekora, M.D., Kowalski, A., 2008, Icarus, 198, 138
Petit, J-M., Farinella, P., 1993, Cel. Mech. Dyn. Astron., 57, 1
Reach, W.T., Franz, B.A., Weiland, J.L., 1997, Icarus, 127, 461
Sykes, M. V., 1990, Icarus, 85, 267
Tedesco, E.F., Noah, P.V., Noah, M., Price, S.D., 2002, Astron. J., 123, 1056
Tedesco, E.F., Cellino, A., Zappalà, V., 2005, Astron. J., 129, 2869
Tsiganis, K., Knežević, Z., Varvoglis, H., 2007, Icarus, 186, 484
Vokrouhlick' y, D., Brož, M., Bottke, W.F., Nesvorn' y, D., Morbidelli, A., 2006, Icarus, 182, 118
Wiegert, P., Balam, D., Moss, A., Veillet, C., Connors, M., Shelton, I., 2007, Astron. J., 133, 1609
Zappalà, V., Cellino, A., Farinella, P., Kne[ž]{}evi[' c]{}, Z., 1990, Astron. J., 100, 2030
Zappalà, V., Cellino, A., Farinella, P., Milani, A., 1994, Astron. J., 107, 772
Yoshida, F., Nakamura, T., 2007, Planetary and Space Science, 55, 1113
[^1]: E-mail: [email protected]
[^2]: http://hamilton.dm.unipi.it/astdys/
[^3]: Note, that this is very similar to Veritas family, and situation with the largest member of this family (490) Veritas [@menios07].
[^4]: Instead of the index $\gamma$, the exponent of the cumulative distribution can be obtained in terms of diameters rather than absolute magnitudes [@delloro07]. However, as we do not know albedos for most of the asteroids, necessary to convert from absolute magnitudes to diameters, we chose to work with $\gamma$.
[^5]: Usually one adopts the value of $d_{cutoff}$ that corresponds to the center of the interval over which the index $\gamma$ is constant [@vok06b].
[^6]: The asteroid size distribution at diameters D$<$10 km is still poorly known. Various models and extrapolations yield very different estimates of the number of km-sized and smaller main belt asteroids. However, other possible estimates [e.g. @ivezic01; @ted05; @wiegert07; @yoshida07] would not affect our results significantly.
[^7]: Probably the better way to estimate the size of the parent body is the one proposed by @durda07. However, we were unable to find appropriate match with SFDs published in that paper using simple visual comparison of plots.
[^8]: Compensating for collective effects in the cloud of dispersing fragments, $V_{esc}$ = 1.64 $\times~GM/R$, where $GM$ is the product of the gravitational constant and the parent body mass, $R$ is radius of the parent body, while 1.64 is an empirical factor [@petfar1993].
[^9]: All three-body mean motion resonances [@nesmor98] discussed in this paper are among Jupiter, Saturn and asteroid.
[^10]: The secular frequencies of Jupiter ($g_{5}$) and Saturn ($g_{6}$,$s_{6}$) are taken from @nobili1989.
[^11]: The secular resonance $g+s-g_{6}-s_{6}$ has much more influence on dynamics of family members in the case of Padua family [see @carruba09].
[^12]: We did not use analytical proper elements, in our other analysis throughout the paper, because they are not enough accurate in this high eccentricity region. This can be appreciated comparing the distributions of Theobalda family members with $a_{p}$$\geq$3.183 au, shown for two different kinds of proper elements. Obvious grouping of the regular members, which is clearly visible in the space of synthetic proper elements (Fig. \[f:dis\_aei\]), disappears in the space of analytical proper elements (Fig. \[f:analytical\]).
[^13]: All integrations presented in this paper are performed using the public domain ORBIT9 integrator embedded in the multipurpose OrbFit package (http://hamilton.dm.unipi.it/astdys/), and dynamical model that includes the four major planets (from Jupiter to Neptune) as perturbing bodies. The indirect effect of the inner planets is accounted for by applying a barycentric correction to the initial conditions.
[^14]: These integrations were performed using ORBIT9 integrator in the Grid environment [@grid09].
[^15]: Caused by the proximity of the Theobalda family to the 2/1 resonance with Jupiter.
[^16]: The ellipse is determined by the present size of the family or, as in this case, by the present size of particular part of the family. It should not be confused with equivelocity ellipses shown e.g. in Fig. \[f:dis\_aei\].
[^17]: In these relations $a_{J}$ denotes Jupiter’s semi-major axis, $e_m$ the mean eccentricity and $I_m$ the mean inclination of the asteroids.
[^18]: The main source of the error is uncertainty in the determination of the present size of the Group B, due to the small number of members in this Group.
[^19]: The good agreement between ages obtained applying MCC to groups A i B, is one of the reasons why we adopted value of $d_{cutoff}=85$ ms$^{-1}$ to identify nominal family. It should be noted here that identification of resonant family members is not straightforward. Too small cut-off, on one hand, may prevents identification some of the real family members. On the other hand, the large cut-off could associate some interlopers with family. This may be one of the reasons for the variations of age estimates obtained by MCC for two different cut-off velocities.
[^20]: The MBCs are bodies with asteroid-like dynamical properties but comet like physical properties [@hsieh2004]. These are dynamically ordinary main-belt asteroids on which, probably, subsurface ice has recently been exposed e.g. because of a collision.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Quantitative assessment of a treatment progress in the Achilles tendon healing process - one of the most common musculoskeletal disorders in modern medical practice - is typically a long and complex process: multiple MRI protocols need to be acquired and analysed by radiology experts for proper assessment. In this paper, we propose to significantly reduce the complexity of this process by using a novel method based on a pre-trained convolutional neural network. We first train our neural network on over 500 000 2D axial cross-sections from over 3 000 3D MRI studies to classify MRI images as belonging to a healthy or injured class, depending on the patient’s condition. We then take the outputs of a modified pre-trained network and apply linear regression on the PCA-reduced space of the features to assess treatment progress. Our method allows to reduce up to 5-fold the amount of data needed to be registered during the MRI scan without any information loss. Furthermore, we are able to predict the healing process phase with equal accuracy to human experts in 3 out of 6 main criteria. Finally, contrary to the current approaches to healing assessment that rely on radiologist subjective opinion, our method allows to objectively compare different treatments methods which can lead to faster patient’s recovery.'
author:
- 'Norbert Kapinski, Jakub Zielinski, Bartosz A. Borucki, Tomasz Trzcinski, Beata Ciszkowska-Lyson, Krzysztof S. Nowinski'
bibliography:
- 'biblio.bib'
title: |
Estimating Achilles tendon healing progress\
with convolutional neural networks
---
Introduction
============
Injuries of the Achilles tendons are one of the most common musculoskeletal disorders in modern medical practice, with more than 18 case per 100,000 people per year [@Raikin14]. A ruptured tendon undergoes a surgical reconstruction followed by a rehabilitation process. A risk of the tendon re-rupture is between 20-40% and proper oversight of the healing process is needed to increase the chances that the tendon is not ruptured again. Furthermore, continuous monitoring of the healing process can be useful for perfecting treatment techniques and adjusting them to patient’s personal conditions.
Recently, several works showed how quantitative methods based on deep neural networks can successfully be used to monitor the healing of the Achilles tendon [@Kapinski2017; @Nowosielski17]. However, those methods are dedicated to simple classification tasks and do not fully take advantage of the mid-level neural network representation. As suggested by [@Oquab14], those representations can be useful, especially for medical image processing tasks [@Doe17; @Bar2015].
In this paper, we introduce a novel method for continuous evaluation of reconstructed Achilles tendon healing based on the responses of intermediate convolutional neural network layers. First, we train a neural network to classify MRI data as ‘healthy’ or ‘injured’. We then use the pre-trained network as a feature extractor and build the representation of the image data from the outputs of the first fully connected layer. After reducing the dimensionality of this representation with Principal Component Analysis (PCA), we use linear regression to fit the resulting representation to the scores assigned by human annotators that describe a healing phase. The obtained results indicate that the proposed approach allows to reduce the number of MRI protocols used for healing assessment from 10 to 2, without losing accuracy of the healing phase classification. Furthermore, using the responses of the proposed neural network, we are able to estimate the healing phase measured on a five-point scale with a mean square error of less then a half point, when compared with human experts.
Method
======
In this section, we present our method that relies on the outputs of the intermediate neural network layers to predict the healing process phase. We start by training the AlexNet architecture [@AlexNet] for the binary classification task of assigning ‘healthy’ [*vs*]{} ‘injured’ label to the input MRI images. Once trained, we use the truncated version of this architecture as a feature extractor. More precisely, we use the outputs of the first fully connected layer (fc6) and apply a Principal Component Analysis to reduce the dimensionality of the resulting representation. Finally, we introduce a metric $H$ which is a novel contribution of the paper, representing a score of the Achilles tendon condition, visible in a single 3D MRI study image of a given protocol: $$H = TM(PC1(x_1), PC1(x_2),..., PC1(x_n))$$ where TM is a truncated mean with 2.5 upper and lower hinges (a value we obtained in our initial set of experiments), $PC1(x_i)$ is the first principal component value for the network inference performed on the slice $x_i$ where $i$ is the index of the slice in the 3D MRI study.
Fig. \[fig:net\] shows the overview of our framework based on a neural network.
![The neural network topology used to generate tendon healing assessment.[]{data-label="fig:net"}](net.png){width="90.00000%"}
We take the feature extractor (FE) and fully connected layer (FC) from the pre-trained AlexNet model. The network uses features extracted for healthy tendons and those after rupture in different time of the healing. The FC layer performs initial weighting of the features, yet avoiding strong discretization that increases with the subsequent fully connected layers of the binary trained AlexNet model. The PC layer, comprising of Principal Components, reduces the number of output parameters while preserving the differentiation of the tendons in different healing state. Finally, $H$ reduces the number of outliers, as presented in details in Sec. 3.2, with the use of truncation and produces a single value output for the whole 3D MRI study of a given protocol.
Experiments
===========
In this section, we present experimental results obtained with the proposed method. First, we describe the dataset used in our experiments and show the results obtained using classification algorithms based on several neural network architectures. We then verify if the $H$ metric proposed in this work corresponds to the progress of the healing process. Finally, we evaluate our approach and present studies on minimal number of MRI protocols required for the healing process assessment.
Dataset
-------
We acquired MRI data of a lower limb of healthy volunteers and patients after the Achilles tendon rupture with the use of a GE Signa HDxt 1.5T scanner with Foot & Ankle array coil.
To monitor the progress of the healing, each of the individuals was scanned with 10 MRI protocols, typically used in orthopaedics (and containing the most significant visual information) i.e. four 3D FSPGR Ideal \[Fast Spoiled Gradient Echo\] (In Phase, Out Phase, Fat, Water), PD \[Proton Density\], T1, T2, T2 mapping, T2$^\ast$ GRE \[Gradient Echo\] and T2$^\ast$ GRE TE\_MIN \[Minimal Time Echo\]. The group of healthy patients was scanned only once, while the injured patients were scanned once before the tendon reconstruction surgery and 9 times afterwards (after 1, 3, 6, 9, 12, 20, 26, 40 and 52 weeks). As of March 2018, we collected 270 3D MRI scans of healthy individuals (including 27 volunteers, 10 protocols) and 2 772 of injured patients (including 60 patients, 10 protocols and up to 10 timesteps).
To overcome the possible artifacts of volumetric interpolation with spatially anisotropic data resolution and to represent the internal structure of tendon tissue, we performed an additional analysis only within axial slices of the 3D MRI scans. We augmented the healthy tendon slices number by mirroring and random rotations in the range of -10 to 10 degrees. We also mirrored the slices representing injured tendons. The resulting dataset contains 234,502 slices labeled as healthy and 277,208 slices labeled as injured.
To determine the ground truth for our dataset we worked with a group of expert radiologists. More precisely, we prepared a survey that includes 6 parameters describing the tendon healing process, visible in the MRI scans:
1. Structural changes within the tendon (SCT) - informs about the loss of cohesion within the tendon area.
2. Tendon thickening (TT) - informs about the maximum dimension in the sagittal direction.
3. Sharpness of the tendon edges (STE) - informs about the edge fractality.
4. Tendon edema (TE) - informs about an abnormal accumulation of fluid in the interstitium of the tendon.
5. Tendon uniformity (TU) - informs about the level of similarity of subsequent cross-sections of the tendon.
6. Tissue edema (TisE) - informs about an abnormal accumulation of fluid and enlarged size of the fascial compartment.
We asked experts to assess the healing condition of Achilles tendons based on the MRI images from our dataset. For each of the parameters, the experts could select a single score in a 5 point scale, where 1 describes a healthy tendon and 5 describes severely injured one.
Tendon healing process assessment
---------------------------------
### Binary classification task:
We first evaluate neural networks on the task of binary classification of the dataset. Tab. 1 presents the comparison of 5-fold cross-validation accuracy results of AlexNet, ResNet-18 [@ResNet] and GoogleNet [@GoogleNet] models used to classify the tendon condition (healthy vs. injured). 3 folds were used for training, one for validation and one for testing.
[ >l | >l | >l | >l | >l ]{} [ ]{} & Average & Min & Max & SD\
& 99.19 & 99.15 & 99.24 & 0.04\
& 95.98 & 92.78 & 99.04 & 2.5\
& 99.83 & 99.68 & 99.91 & 0.1\
The performance of all of the tested models on the classification task is satisfactory. Nevertheless, the average, minimum and maximum accuracy for both AlexNet and GoogleNet reaches over 99% and in case of ResNet-18 the results are scattered between 93 and 99%, giving an average accuracy of 96%.
The AlexNet is the least complex of the tested models and its training takes less than two hours, while in the case of ResNet-18 it takes over 20 hours to train the network, for GoogleNet this time increases to around 48h. All times were measured on a server station with NVIDIA V100 GPU. Taking into account the accuracy and the model complexity, we choose the AlexNet architecture for all the experiments in the remainder of this paper.
Most of the samples labeled as injured and misclassified by our network as healthy come from 3D MRI scans that focus on the area far away from the rupture. In this work, we eliminate those outliers by using a truncated mean of our $H$ metric, but in future work we plan to address this problem by incorporating fuzzy logic or soft classification methods.
### Principal Components Analysis:
For this task, we use the feature extractor part of the binary trained AlexNet model and its first fully connected layer (see Fig. 1). We evaluate whether the number of outputs can be decreased while preserving the differentiation of the stages of the healing process. For this purpose we use 48,225 slices derived from 10 patients that concluded a full year of the rehabilitation and were monitored in 10 timesteps. We perform PCA and compare the amount of variance preserved by 1, 10 and 200 most significant principal components. First 1, 10 and 200 principal components preserve 50.2, 90.8 and 98.8% amount of variance respectively. This means that 4096 outputs of our modified AlexNet topology can be reduced to a single output while preserving over 50% of variance. We therefore decide to test our approach using only the most significant principal component PC1 and this approach is used in the remaining experiments. Further study of consecutive components is a part of our future work.
### Healing monitoring:
In this task, we evaluate changes of the metric $H$ in different stages of the tendon healing. Fig. 2 presents the time curves of the average $H$ metric for 9 patients, monitored in 10 timesteps distributed over the year of the tendon recovery, for all of the 10 MRI protocols.
![Comparison of the curves of an average $H$ value over 9 patients monitored over the year of the tendon recovery for all of the 10 MRI protocols.[]{data-label="fig:H"}](H1.png){width="80.00000%"}
Except for the T2 mapping (T2MAP) with slices acquired from the raw signal data, all other protocols show a decreasing trend that can be interpreted as a negative difference between $H$ value before the reconstruction and at the end of the recovery period. According to the feedback provided by radiologists and medical professionals, all 9 curves correspond well with the real-life healing process assessment. In most cases the progress of healing is more pronounced at the beginning than in the following stages. However, one can also observe the fluctuations related to patient activity, diet and their obedience to the treatment prescription. Taking into account the above results, we select all 9 protocols in the following experiments.
### Correlation of $H$ with the ground truth:
Tab. 2 presents the values of Pearson correlation of the assessment done by the expert radiologist with our results obtained for 9 MRI protocols (excluding T2 mapping). The coefficients marked in bold are statistically significant with $p < 0.01, N=10$. For 4 protocols (i.e. PD, T1, T2, T2$^\ast$GRE) we can observe relatively high correlation with the ground truth in terms of the tendon and tissue edema size (TE and TisE), as well as the sharpness of the tendon edges (STE). Those parameters reach their maximum before the reconstruction and decrease with time. When they stabilize, the patients reach the end of the healing process. We also see that the parameter saturation coincides with plateauing of $H$ values for all 4 protocols.
PD IDELF IDELP IDELOP IDELW T1 T2 T2$^\ast$GRE T2$^\ast$GREMIN
------ --------------- ------- ------- -------- ------- --------------- --------------- --------------- -----------------
SCT 0.69 0.42 0.53 0.59 0.55 0.66 0.60 0.74 0.4
TT 0.52 0.71 0.46 0.30 0.37 0.55 0.40 0.38 0.45
STE [ **0.87**]{} 0.74 0.60 0.50 0.54 [ **0.84**]{} [ **0.81**]{} 0.72 0.64
TE [ **0.89**]{} 0.53 0.63 0.62 0.58 [ **0.81**]{} [ **0.82**]{} [ **0.82**]{} 0.60
TU 0.45 0.19 0.48 0.59 0.50 0.44 0.26 0.64 0.56
TisE [ **0.82**]{} 0.51 0.61 0.65 0.58 [ **0.84**]{} 0.67 [ **0.90**]{} 0.73
: Pearson correlation between the ground truth parameters and the $H$ measure. Coefficients with significance p $<$ 0.01 are shown in bold. []{data-label="correlation1"}
The TU parameter is assessed with the use of the sagittal slices. Thus, our neural network model is not able to successfully assess the healing phase using this parameter, as it is trained on axial cross-sections, not sagittal ones. We plan to further investigate the results of SCT and TT correlations and extend their analysis using more principal components.
### Inter-protocol correlation:
In this section, we analyse the correlation between results obtained with different MRI protocols to find a minimal subset of protocols that is needed by our method to provide satisfactory performance. We investigate 4 protocols that perform best in the previous task and present the inter-protocol correlations results in Tab. 3.
PD T1 T2 T2$^\ast$GRE
-------------- --------------- --------------- --------------- ---------------
PD 1.00 [ **0.96**]{} [ **0.90**]{} [ **0.89**]{}
T1 [ **0.96**]{} 1.00 [ **0.85**]{} [ **0.92**]{}
T2 [ **0.90**]{} [ **0.85**]{} 1.00 0.71
T2$^\ast$GRE [ **0.89**]{} [ **0.92**]{} 0.71 1.00
: Inter-protocol correlation of PD, T1, T2 and T2$^\ast$ GRE MRI protocols. The results marked in bold have a significance level p $<$ 0.01.[]{data-label="my-label"}
![Comparison of our model predictions with the ground truth. Our method predicts the correct outputs with error of less than a half of the score point.[]{data-label="fig:regression"}](image2){width="93.00000%"}
The coefficients marked in bold are statistically significant with $p < 0.01, N=10$. Due to the fact that the PD and T1 protocols correlate strongly with each other, it is sufficient to select only one of them for our experiments. We identify two least correlated protocols, namely PD and T2$^\ast$GRE, and choose this pair. Although alternative methods for protocol selection exist, [*e.g.*]{} one can use backward feature elimination strategy, however, our initial experiments indicate that the performance of the proposed approach is sufficient.
### Prediction of the healing parameters:
Here, we test how accurate predictions of the ground truth we can obtain using a combination of the results for PD and T2$^\ast$GRE-based inference. We use a linear regression of the results from both protocols and analyze the predictions of STE, TE and TisE parameters. Fig. 3 presents the results of our study. The prediction error is below 1 score point in a 5-point scale for every case. This result confirms the validity of our method and can justify using our method in the context of automatic assessment of the Achilles tendon healing.
Conclusions
===========
In this paper, we proposed to use convolutional neural networks to automatically assess the Achilles tendon healing phase. Currently, radiologists spend significant amount of time on analysing MRI images to manually evaluate tendond’s condition. The results presented in this paper prove that methods based on deep neural networks can provide automatic, quantitative analysis of the 3D MRI scans. This, in turn, can lead to significant time savings and increase the efficiency of healing assessment without losing its precision. More precisely, we proposed a novel method for the tendon healing process assessment based on the pre-trained convolutional neural network. Our method allows to significantly improve the clinical workflow by minimizing the number of MRI protocols used to assess the healing and by providing objective, single value assessment for the current condition of the tendon.
In this paper, we also presented the results confirming that our method allows to approximate 3 out of 6 assessment criteria with a human expert accuracy. At the same time, our method uses only 2 out of 10 typically used MRI protocols which can also prove the efficiency of the proposed method.
As future work, we plan to focus on approximating the remaining parameters. Our idea is to use more principal components to compute the $H$ metric, as well as test different types of regression methods and extend our neural network to process also the data from sagittal cross-sections.
Acknowledgments {#acknowledgments .unnumbered}
---------------
[ The following work was part of [*Novel Scaffold-based Tissue Engineering Approaches to Healing and Regeneration of Tendons and Ligaments (START)*]{} project, co-funded by The National Centre for Research and Development (Poland) within STRATEGMED programme (STRATEGMED1/233224/10/NCBR/2014). ]{}
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this work we propose an extension to the Standard Model in which we consider the model 2HDM type-III plus massive neutrinos and the horizontal flavor symmetry $S_{3}$ $(\nu$2HDM$\otimes S_3)$. In the above framework and with the explicit breaking of flavor symmetry $S_{3}$, the Yukawa matrices in the flavor adapted basis are represented by means of a matrix with two texture zeroes. Also, the active neutrinos are considered as Majorana particles and their masses are generated through type-I seesaw mechanism. The unitary matrices that diagonalize the mass matrices, as well as the flavor mixing matrices, are expressed in terms of fermion mass ratios. Consequently, in the mass basis the entries of the Yukawa matrices naturally acquire the form of the so-called [*Cheng-Sher ansatz*]{}. For the leptonic sector of $\nu$2HDM$\otimes S_3$, we compare, through a $\chi^{2}$ likelihood test, the theoretical expressions of the flavor mixing angles with the masses and flavor mixing leptons current experimental data. The results obtained in this $\chi^{2}$ analysis are in very good agreement with the current experimental data. We also obtained an allowed value ranges for the “Dirac-like” phase factor, as well as for the two Majorana phase factors. Furthermore, we study the phenomenological implications of these numerical values of the CP-violation phases on the neutrinoless double beta decay, and for Long Base-Line neutrino oscillation experiments such as T2K, NO$\nu$A, and DUNE.'
author:
- 'E. Barradas-Guevara'
- 'O. Félix-Beltrán'
- 'F. Gonzalez-Canales'
- 'M. Zeleny-Mora'
title: 'On the lepton CP violation in a $\nu$2HDM with flavor'
---
CIFFU-17-02
Introduction {#sec:introduction}
============
According to the most recent neutrino physics literature [@King:2017guk; @Capozzi:2016rtj], there are several issues still unresolved. Among others: whether the neutrinos are Dirac or Marjorana fermions; the absolute neutrino mass scale; the posible sources of Charge-Parity \[CP\] violation \[CPV\] in leptons. Nowadays, answers to these questions are being searched by means of the experimental results concerning KamLAND reactor neutrinos [@Decowski:2016axc; @Gando:2010aa; @Gando:2013nba], in each one of the current high-statistics Short Base-Line (SBL) reactor neutrino experiments RENO [@Seo:2014xei; @RENO:2015ksa], Double Chooz [@Abe:2014bwa] and Daya Bay [@An:2015nua]. Also, one of the most interesting effects related to the neutrino oscillation in matter is, that these periodic transformations of neutrinos from one flavor to another, can induce a fake CP violating effect. Therefore, the Long Base-Line (LBL) neutrino oscillation experiments are good candidates for determining the “Dirac-like” CP violation phase as well as resolving the mass hierarchy problem [@Feldman:2013vca]. The recent measures reported by T2K [@Abe:2017vif; @Batkiewicz:2017xoh; @Abe:2017uxa; @Abe:2015awa], NO$\nu$A [@Kolupaeva:2016fxw; @Bian:2015opa] and Super-Kamiokande experiments [@Cresst] suggest a nearly maximal CP violation. In these experiments the “Dirac-like” CP phase takes the value $\delta_{\text{CP}} \simeq 3 \pi /2$, with a statistical significance below 3$\sigma$ level. Moreover, the data obtained in the global fits of neutrino oscillations agree with a nonzero $\delta_{CP}$ phase, whereby the previous value is confirmed [@Forero:2014bxa; @Gonzalez-Garcia:2015qrr; @Capozzi:2016rtj; @Esteban:2016qun; @Ge:2017qqv].
The flavor and mass generation are two concepts strongly intertwined. In order to know the flavor dynamic in models beyond the Standard Model (SM), we need to understand the mechanism of flavor and mass generation arising in the standard theory. In the later, the Yukawa matrices are of great interest because its eigenvalues define the fermion masses. Moreover, for multi-Higgs models the Flavor Changing Neutral Current (FCNC) arises naturally from impediment to diagonalize simultaneously the mass and Yukawa matrices. In particular, models like the Two Higgs Doublet Model type-III (2HDM-III), in which the two Higgs doublets are coupled to all fermions, allow the presence of FCNC at tree level mediated by Higgs [@Atwood:1996vj; @DiazCruz:2009ek; @DiazCruz:2004pj; @Krawczyk:2007ne; @BarradasGuevara:2010xs; @Crivellin:2013wna]. The 2HDM predicts three neutral states $H^{0}_{1,2,3}$ and a pair of charged states denoted as $H^{\pm}_{1,2}$ [@Krawczyk:2005zy; @Branco:2011iw]. The Higgs-fermions couplings ($H^0 \bar{f}f$) in the 2HDM are given as [@Atwood:1996vj; @DiazCruz:2009ek; @Krawczyk:2007ne; @BarradasGuevara:2010xs; @Crivellin:2013wna]: $${\cal L} _Y = \bar{f}_i \left( S_{ij} + \gamma^5 P_{ij} \right) f_j H^0_a,
\qquad (a=1,2,3).$$ In 2HDM-III the FCNC’s are kept under control by imposing some texture zeroes in the Yukawa matrices. This fact reproduce the observed fermion masses and mixing angles [@Deppisch:2012vj]. Using texture shapes allow a direct relation between the Yukawa matrix entries and the parameters used to compute the decay widths and cross section, without losing the terms proportional to the light fermion masses. Specifically, considering a zero texture Yukawa matrix, one obtains the [*Cheng-Sher ansatz*]{} for flavor mix couplings, widely used in literature, where flavored couplings are considered proportional to the involved fermion masses [@Dorsner:2002wi].
The matter content in the 2HDM is divided among the quarks and leptons sectors. In turn these sectors are subdivided in two sectors, the up- and down-type for quarks sector, while charged leptons and neutrinos for the leptons sector. The fermions in each one of these subsectors are analogous each other, because they have completely identical couplings to all gauge bosons, although their mass values are not the same. Therefore, before of the spontaneous symmetry breaking (SSB), the Yukawa Lagrangian in the above subsectors is invariant under permutations of flavor indices. In other words, each one of these subsectors is invariant under the action of a $S_{3}$ symmetry group. This symmetry group has only three irreducible representations that correspond to two singlets and a doublet [@Georgi:1999wka; @Ishimori:2012zz].
On the other hand, obtained from the experimental data, the mass spectrum for Dirac fermions obeys the following strong hierarchy [@Olive:2016xmw]: $$\begin{array}{cc}
\widehat{m}_{e} \sim 10^{-6} , \;\; \widehat{m}_{\mu} \sim 10^{-3}, &
\quad \widehat{m}_{\tau} \sim 1 , \\
\widehat{m}_{u} \sim 10^{-5} , \;\; \widehat{m}_{c} \sim 10^{-3}, &
\quad \widehat{m}_{t} \sim 1 , \\
\underbrace{ \widehat{m}_{d} \sim 10^{-3} , \;\; \widehat{m}_{s} \sim 10^{-2},
}_{\bf 2} &
\quad \underbrace{ \widehat{m}_{b} \sim 1 . }_{\bf 1}
\end{array}$$ In the above expression $\widehat{m}_{l} = m_{l}/m_{\tau}$ ($l=e,\mu, \tau$) stands for charged leptons, $\widehat{m}_{U}= m_{U}/m_{t}$ ($U = u,c,t$) for up-type quarks, while $\widehat{m}_{D} = m_{D}/m_{b}$ ($D = d,s,b$) for down-type quarks. The behavior of these mass ratios in terms of irreducible representations of some symmetry group, can be interpreted as follows: the two lighter particles are associated with a doublet representation ${\bf 2}$, whereas for the heaviest particle it is assigned a singlet representation ${\bf 1}$. The smallest non-abelian group with irreducible representations of singlet and doublet, is the group of permutations of three objects. Hence, we expect the hierarchical nature of the Dirac fermion mass matrices to have its origin in the representation structure ${\bf 3} = {\bf 2} \oplus {\bf 1}$ of $S_{3}$. In the theoretical framework of SM, as well as for 2HDM, the neutrinos are massless particles, fact that is in disagree with the results obtained in the neutrino oscillation experiments.
Therefore, in this work we will study the flavor dynamics through Yukawa matrices in the specific scenario of 2HDM-III plus massive neutrinos and a horizontal flavor symmetry $S_3$ $(\nu$2HDM$\otimes S_3)$. In this context, under the action of $S_{3}$ flavor symmetry group the right-handed neutrinos as well as the two Higgs fields transform as singlets, while the active neutrinos are considered as Majorana particles and their masses are generated through type-I seesaw mechanism. Hence, it is necessary to consider the following hybrid mass term, which involves the Dirac and Majorana neutrino mass terms [@Xing:2010zzf], $${\cal L}_{\mathrm{M+D}} =
- \frac{1}{2} \bar{ \eta }_{L} \, {\bf M}_{\mathrm{M+D}} \, (\eta_{L})^{c} +
\textrm{h. c.},$$ where $\eta = \left(\, \nu_{L}, \, (N_{R})^{c} \, \right)^{\top} $ and $${\bf M}_{\mathrm{M+D}} =
\left( \begin{array}{cc}
{\bf 0} & {\bf M}_{\nu_{D}} \\
{\bf M}_{\nu_{D}}^{\top} & {\bf M}_{R}
\end{array} \right).$$ In the above expression ${\bf M}_{\nu_{D}}$ and ${\bf M}_{R}$ are the Dirac and right-handed neutrino mass matrix, respectively. In the special limit ${\bf M}_{R} \gg {\bf M}_{\nu_{D}}$, the effective mass matrix of left-handed neutrinos is given by the type-I seesaw mechanism whose expression is[^1]: $$\label{eq:seesaw}
{\bf M}_{ \nu } = {\bf M}_{\nu_{D}} {\bf M}_{R}^{-1} {\bf M}_{\nu_{D}}^{\top}.$$ If the fermion mass matrices do not have any element equal to zero, on one hand, the mass matrix of active neutrinos has twelve free parameters, since ${\bf M}_{\nu}$ is a complex symmetric matrix because this matrix comes from a Majorana mass term. On the other hand, the Dirac fermion mass matrices do not have any special feature, [*i.e.*]{}, these matrices are not Hermitian, nor symmetric. This is mainly due to the fact that the Yukawa matrices are represented through a $3 \times 3$ complex matrix. Hence, for Dirac mass matrices we have eighteen free parameters.
After the explicit sequential breaking of flavor symmetry according to the chain $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} } \supset
S_{3}^{\mathrm{diag}} \supset S_{2}^{\mathrm{diag}}$, all Yukawa matrices in the flavor adapted basis are represented by means of a matrix with two texture zeroes. Therefore, all fermion mass matrices in the model have the same generic form with two texture zeroes.
The difference between 2HDM and $\nu$2HDM$\otimes S_{3}$ lays on the Yukawa structure, on the symmetries of the Higgs sector and on the possible appearance of new CPV sources. This CPV can arise from the same phase appearing in the Cabibbo-Kobayashi matrix, as in the SM, or some extra phase which arises from the Yukawa field or from the Higgs potential, either explicitly or spontaneously. The Higgs potential preserves CP symmetry, whereby CPV comes from the Yukawa matrices.
In order to validate our hypothesis where the $S_{3}$ horizontal flavor symmetry is explicitly breaking, hence all fermion mass matrices are represented through a matrix with two texture zeroes, we make a likelihood test where the $\chi^{2}$ function is defined in terms of leptonic flavor mixing angles. Afterwards, we shall investigate the phenomenological implications of these results on the neutrinoless double beta decay and the CPV in neutrino oscillations in matter.
The organization of this work is as follows. In section \[sec:Yukawa\_Lag\] we present the Yukawa Lagrangian in the $\nu$2HDM$\otimes S_3$, the form of the Dirac and Majorana fermion mass matrices in terms of its eigenvalues. In this way, we derive explicit and analytical expressions for the leptonic flavor mixing angles and Higgs-fermions couplings. In section \[sec:numericalanal\] we present a detailed likelihood test where the $\chi^{2}$ function is defined in terms of leptonic mixing angles. Also, in section \[sec:PenImp\] we explore the phenomenological implications of the numerical values obtained for the CP violating phase factors, on the neutrinoless double beta decay and the neutrino oscillations in matter. Finally, in the section \[sec:conclusions\] we present the conclusions and remarks of the present work.
The Yukawa Lagrangian in the $\nu$2HDM$\otimes S_{3}$ {#sec:Yukawa_Lag}
=====================================================
In the fermion matter content of the SM, which is the same for the 2HDM, there are no right-handed neutrinos, consequently in both models a neutrino mass term is not allowed. This latter fact gainsay the results obtained in the neutrino oscillation experiments which requires neutrinos to have nonzero masses [@Olive:2016xmw]. In order to include a Majorana neutrino mass term to the 2HDM, we need to increase its matter content. For this reason, we consider six neutrino fields: three left-handed $\nu_{L} = \left( \nu_{e L}, \nu_{\mu L}, \nu_{\tau L} \right)^{\top}$ and three right-handed $N_{R} = \left( N_{1 R}, N_{2 R}, N_{3 R} \right)^{\top}$. The right-handed neutrinos must be uncharged under the weak and electromagnetic interactions, which means that this kind of neutrinos are singlets under $G_{\mathrm{EW}} \equiv SU(2)_L \otimes U(1)_Y$. In other words, only the left-handed neutrinos take part in the electroweak interaction. In this theoretical framework, we have all SM matter content plus massive neutrinos and an extra Higgs boson, whereby it is called as $\nu$2HDM. In the weak basis, the Yukawa interaction Lagrangian for Dirac fermions in the $\nu$2HDM is given by [@Branco:2011iw; @Barradas-Guevara:2016gda; @Wang:2016vfj]: $$\label{2HDMlagrangian}
{\cal L}_{\mathrm{Y}}^{\mathrm{w}} =
\sum_{k = 1}^{2}
\left(
{\bf Y}_{k}^{\mathrm{w}, u} \, \bar{Q} \tilde{\Phi}_{k} u_{R}
+ {\bf Y}_{k}^{\mathrm{w}, d} \, \bar{Q} \Phi_{k} d_{R}
+ {\bf Y}_{k}^{\mathrm{w}, \nu_{D}} \, \bar{L} \tilde{\Phi}_{k} N_{R}
+ {\bf Y}_{k}^{\mathrm{w}, l} \, \bar{L} \Phi_{k} l_{R} \right)
+ \textrm{h. c.,}$$ where $Q = \left( \, u , \, d \, \right)^{\top}_{L}$ and $L = \left( \, \nu_{l} , \, l \, \right)^{\top}_{L}$ are the left-handed doublets of $SU(2)_{L}$; $u_{R}$, $d_{R}$ and $l_{R}$ are the right-handed singlets of the electroweak gauge group. In this expression, the $\mathrm{w}$ superscript indicates that we are working in the weak basis, while the indices $l$, $u$ and $d$ represent the charged leptons, $u$- and $d$-type quarks, respectively. Also, $\Phi_{k} = \left( \phi^{+}_{k}, \, \phi^{0}_{k} \right)^{\top}$ denotes the two Higgs fields which are doublets of $SU(2)_{L}$, with $\tilde{\Phi}_{k} = i \sigma_2 \Phi_{k}^{*}$. Finally, the ${\bf Y}_{k}^{\mathrm{w},\texttt{j}}$ are the Yukawa matrices in the weak basis, where the $\texttt{j}$ superscript denotes the Dirac fermions, ($\texttt{j} = u,d,l,\nu_{D}$) [@Branco:2011iw]. In general, after the SSB and in the context of $\nu$2HDM, the Dirac fermion mass matrix in the weak basis can be written as [@Branco:2011iw; @Felix-Beltran:2013tra; @Barradas-Guevara:2016gda]: $$\label{masa-fermiones}
{\bf M}_{ \texttt{j} }^{\mathrm{w}} =
\frac{1}{ \sqrt{2} }
\sum_{k=1}^{2} v_{k} \, {\bf Y}_{k}^{\mathrm{w},\texttt{j}} ,
$$ where $v_{k}$ are the vacuum expectation values (vev’s) of the two Higgs bosons $\Phi_{k}$, with $k=1,2$.
Mass matrices from the $S_{3}$ flavor symmetry
----------------------------------------------
To reduce the free parameters in the fermion mass matrices, we will consider a horizontal symmetry which correlates the particle flavor indices each one with the other, thus the Yukawa matrices could be represented by means of a $3\times3$ Hermitian matrix. Consequently, for three families or generations of quarks and leptons, we propose[^2] that after the SSB, the Yukawa Lagrangian in the $\nu$2HDM presents the permutations group $S_{3}$ as horizontal flavor symmetry. In this context, the right-handed neutrinos, as well as the two Higgs bosons, transform as singlets under the action of $S_{3}$ flavor symmetry. In other words, the right-handed neutrinos and the two scalar fields $\Phi_{k}$ are flavorless particles, whereby these fields are treated as scalars with respect to the $S_{3}$ symmetry transformations. The general way to implement the flavor symmetry is considering that under the action of $S_{3}$ symmetry, the left- and right-handed spinors transform as [@Canales:2013cga]: $$\begin{array}{l}
\psi^{ \mathrm{\,s} }_{ \texttt{j} L } = {\bf g}_{\mathrm{a}}^{ \texttt{j} } \,
\psi_{\texttt{j} L}
\quad \textrm{and} \quad
\psi^{ \mathrm{\,s} }_{ \texttt{j} R } = \widetilde{\bf g}_{\mathrm{b}}^{ \texttt{j} } \,
\psi_{ \texttt{j} R },
\qquad \mathrm{a,b} = 1, \ldots, 6.
\end{array}$$ At this point, the proposed flavor symmetry for the Yukawa Lagrangian is the $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} }$ group, whose elements are the pairs $\left( {\bf g}_{\mathrm{a}}^{ \texttt{j} }, \widetilde{\bf g}_{\mathrm{b}}^{ \texttt{j} }
\right)$, where ${\bf g}_{\mathrm{a}}^{ \texttt{j} } \in S_{3L}^{ \texttt{j} }$ and $\widetilde{\bf g}_{\mathrm{b}}^{ \texttt{j} } \in S_{ 3R }^{ \texttt{j} }$, while the superscript “$\mathrm{s}$” means that fields are in the flavor symmetry adapted basis. However, the mass terms ${\cal L}_{m} \sim \bar{ \psi }_{\texttt{j} L} \, {\bf M}_{\texttt{j}} \,
\psi_{\texttt{j} R} $ and charged currents ${\cal J}_{\mu} \sim \bar{ \psi }_{\texttt{j} L} \ \gamma_{\mu} \psi_{\texttt{j} L} W^{\mu}$ in the Lagrangian are not invariant under $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} }$ group. In order to make ${\cal L}_{m}$ and ${\cal J}_{\mu}$ invariant under the flavor group transformations, the elements of the $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} }$ group must satisfy the relation ${\bf g}_{\mathrm{a}} \equiv {\bf g}_{\mathrm{a}}^{\texttt{j}}
= \widetilde{\bf g}_{\mathrm{b}}^{\texttt{j}}$. The latter condition implies that the flavor group is reduced according to the chain: $S_{ 3L }^{ \mathrm{j} } \otimes S_{ 3R }^{ \mathrm{j} } \supset
S_{3}^{\mathrm{diag}}$ [@Canales:2013cga]. This flavor symmetry breaking chain should be interpreted as: all fermions in the model must be transformed with the same flavor group and the same element thereof. In the above, the flavor group is called $S_{3}^{\mathrm{diag}}$ because its elements are the pairs $\left({\bf g}_{\mathrm{a}} , {\bf g}_{\mathrm{a}} \right)$, where ${\bf g}_{\mathrm{a}} \in S_{3L}$ [@Canales:2013cga].
Finally, it is easy conclude that $S_{3}^{\mathrm{diag}}$ is the horizontal flavor symmetry which conserves the invariance of ${\cal L}_{\mathrm{Y}}^{\mathrm{w}}$ under the action of electroweak gauge group. Also, from the invariance of ${\cal L}_{m}$ Lagrangian under the action of $S_{3}^{\mathrm{diag}}$ group, we obtain that fermion mass matrices commute with all elements of the flavor group.
In the weak basis, the two Yukawa matrices in Eq. (\[masa-fermiones\]) are represented by means of a matrix with the exact $S_{3}^{\mathrm{diag}}$ symmetry. Therefore, these Yukawa matrices are expressed as $$\label{eq:Ykj0}
{\bf Y}_{k}^{\mathrm{w},\texttt{j}} \equiv {\bf Y}_{k}^{\texttt{j} 3}
= \alpha^{\texttt{j}}_{k} \, {\bf P}_{\bf 1} ,$$ where $\alpha^{\texttt{j}}_{k}$ are real constants associated with the flavor symmetry. The explicit form of ${\bf P}_{\bf 1}$ matrix is given by Eq. (\[eq:A-P1\]), and corresponds to the projector associated with the symmetric singlet representation of $S_{3}$. Hence, the Dirac fermion mass matrix is $${\bf M}_{ \texttt{j} }^{\mathrm{w}} =
\frac{1}{\sqrt{2}} \sum_{k=1}^{2} v_{k} {\bf Y}_{k}^{\texttt{j} 3} =
m_{\texttt{j}3} {\bf P}_{\bf 1} ,$$ where $m_{\texttt{j}3} = \frac{1}{\sqrt{2}}\left( v_{1} \, \alpha^{\texttt{j}}_{1}
+ v_{2} \, \alpha^{\texttt{j}}_{2} \right)$. In the flavor adapted basis, the Dirac fermion mass matrices are [@Felix-Beltran:2013tra] $$\label{eq:MS3_exact}
{\bf M}_{ \texttt{j} }^{\mathrm{s}} =
{\bf U}_{\mathrm{s}}^{\dagger} {\bf M}_{ \texttt{j} }^{\mathrm{w}} {\bf U}_{\mathrm{s}} =
m_{\texttt{j}3} {\bf U}_{\mathrm{s}}^{\dagger} {\bf P}_{\bf 1} {\bf U}_{\mathrm{s}} =
\textrm{diag} \left( 0, 0, m_{\texttt{j}3} \right) ,$$ where $$\label{eq:Us}
{\bf U}_{\mathrm{s}} =
\frac{1}{\sqrt{6}}
\left( \begin{array}{ccc}
\sqrt{3} & 1 & \sqrt{2} \\
- \sqrt{3} & 1 & \sqrt{2} \\
0 & -2 & \sqrt{2}
\end{array} \right) .$$ The interpretation of Eq. (\[eq:MS3\_exact\]) is that under an exact $S_{3}^{\mathrm{diag}}$ symmetry, the mass spectrum for Dirac fermions consists of a massive particle and two massless particles [@PhysRevD.59.093009]. The only massive particle in each of fermion mass spectrum corresponds to the heaviest fermion. However, this result disagrees with the experimental data on quarks and leptons masses [@Olive:2016xmw]. Since the two Higgs fields are invariant under flavor symmetry transformations, these are naturally assigned to $S_{3}$ flavor singlets. If the Yukawa Lagrangian is exactly invariant under $S_{3}$ flavor transformations, the two scalar fields can only couple with the $S_{3}$-singlet component of fermion fields. Consequently, only the $S_{3}$-singlet component of fermion fields acquires a mass non zero. As the third family is the heaviest, here we assign the fermion fields in the third family to singlet irreducible representation of $S_{3}$.
So, with the aim to generate a non zero mass for all fermions in the model, here we will break the flavor symmetry in an explicit sequential way, according to the chain $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} }\supset S_{3}^{\mathrm{diag}} \supset S_{2}^{\mathrm{diag}}$. Respectively, the first two fermion families and third one are assigned to the doublet and singlet irreducible representations of $S_{3}^{\mathrm{diag}}$. The mass of the second fermion family is generated when the $S_{3}^{\mathrm{diag}}$ flavor symmetry is explicitly breaks into the $S_{2}^{\mathrm{diag}}$ group. This symmetry breaking is carried out when we add the following term to the ${\bf Y}_{k}^{\texttt{j3}}$ matrix in Eq. (\[eq:Ykj0\]) : $$\label{eq:Ykj2}
{\bf Y}_{k}^{\texttt{j2} } = \beta_{k}^{\texttt{j}} \, {\bf T}_{z1}^{+} +
\gamma_{k}^{\texttt{j}} \, {\bf T}_{z2}^{+},$$ where $\beta_{k}^{\texttt{j}}$ and $\gamma_{k}^{\texttt{j}}$ are real constant parameters. The explicit form of the tensors ${\bf T}_{z1}^{+}$ and ${\bf T}_{z2}^{+}$ is given by Eq. (\[eq:Tz12+\]). The ${\bf Y}_{k}^{\texttt{j2} }$ matrix mixes the symmetric component of the doublet with the singlet. Finally, the first fermion family’s mass is generated by adding the term $$\label{eq:Ykj1}
{\bf Y}_{k}^{\texttt{j1} } =
\epsilon_{k}^{\texttt{j}} \, {\bf T}_{x}^{+} +
\rho_{k}^{\texttt{j}} \, {\bf T}_{x}^{-}$$ to matrices in Eqs. (\[eq:Ykj0\]) and (\[eq:Ykj1\]), where $\epsilon_{k}^{\texttt{j}}$ and $\rho_{k}^{\texttt{j}}$ are real constant parameters. Thus, the explicit form of the tensors ${\bf T}_{x}^{+}$ and ${\bf T}_{x}^{-}$ is given by Eq. (\[eq:Ty+-\]). The ${\bf Y}_{k}^{\texttt{j1} }$ matrix mixes the components of the doublet representation between each other in the weak basis. So, in the weak basis and under the explicit sequential breaking of flavor symmetry according to the chain; $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} } \supset
S_{3}^{\mathrm{diag}} \supset S_{2}^{\mathrm{diag}}$, we obtain the Yukawa matrices which produce three massive fermions. These Yukawa matrices are the sum of the three expressions given by Eqs. (\[eq:Ykj0\]), (\[eq:Ykj2\]), and (\[eq:Ykj1\]). Then, $$\label{eq:Yuka-w}
\begin{array}{l}\vspace{2mm}
{\bf Y}_{k}^{\mathrm{w}, \texttt{j}} =
{\bf Y}_{k}^{\texttt{j3} } +
{\bf Y}_{k}^{\texttt{j2} } +
{\bf Y}_{k}^{\texttt{j1} } =
\alpha^{\texttt{j}}_{k} \, {\bf P}_{\bf 1} +
\beta_{k}^{\texttt{j}} \, {\bf T}_{z1}^{+} +
\gamma_{k}^{\texttt{j}} \, {\bf T}_{z2}^{+} +
\epsilon_{k}^{\texttt{j}} \, {\bf T}_{x}^{+} +
\rho_{k}^{\texttt{j}} \, {\bf T}_{x}^{-} , \\
{\bf Y}_{k}^{\mathrm{w}, \texttt{j} } =
\left( \begin{array}{ccc}
e^{\mathrm{w} , \texttt{j} }_{k} & a^{\mathrm{w} , \texttt{j} }_{k} &
f^{\mathrm{w} , \texttt{j} }_{k} \\
a^{\mathrm{w} , \texttt{j} \,* }_{k} & b^{\mathrm{w} , \texttt{j} }_{k} &
c^{\mathrm{w} , \texttt{j} }_{k} \\
f^{\mathrm{w} , \texttt{j} \,* }_{k} & c^{\mathrm{w} , \texttt{j} \,* }_{k} &
d^{\mathrm{w} , \texttt{j} }_{k}
\end{array} \right) ,
\end{array}$$ where $$\begin{array}{lll} \vspace{2mm}
a^{\mathrm{w} , \texttt{j} }_{k} =
\frac{
\alpha^{\texttt{j}}_{k} + 3 \left( \beta_{k}^{\texttt{j}} + i \rho_{k}^{\texttt{j}}
\right) }{ 3 }, &
b^{\mathrm{w} , \texttt{j} }_{k} =
\frac{
\alpha^{\texttt{j}}_{k} + 3 \left( \beta_{k}^{\texttt{j}} + \epsilon_{k}^{\texttt{j}}
\right) }{ 3 } , &
c^{\mathrm{w} , \texttt{j} }_{k} =
\frac{
\alpha^{\texttt{j}}_{k} + 3 \left( \gamma_{k}^{\texttt{j}} - \epsilon_{k}^{\texttt{j}}
+ i \rho_{k}^{\texttt{j}}
\right) }{ 3 } , \\ \vspace{2mm}
d^{\mathrm{w} , \texttt{j} }_{k} =
\frac{
\alpha^{\texttt{j}}_{k} - 6 \beta_{k}^{\texttt{j}}
}{
3
}, &
e^{\mathrm{w} , \texttt{j} }_{k} =
\frac{
\alpha^{\texttt{j}}_{k} + 3 \left( \beta_{k}^{\texttt{j}} - \epsilon_{k}^{\texttt{j}}
\right) }{ 3 } , &
f^{\mathrm{w} , \texttt{j} }_{k} =
\frac{
\alpha^{\texttt{j}}_{k} + 3 \left( \gamma_{k}^{\texttt{j}} + \epsilon_{k}^{\texttt{j}}
- i \rho_{k}^{\texttt{j}}
\right) }{ 3 } .
\end{array}$$ In this same basis, now the fermion mass matrices ${\bf M}_{\texttt{j}}^{\mathrm{w}}$ take the form: $$\label{eq:Mwj}
\begin{array}{l}
{\bf M}_{ \texttt{j} }^{\mathrm{w}} =
{\displaystyle \frac{ 1 }{ \sqrt{2} } \sum_{k=1}^{2} } \, v_{k}
{\bf Y}_{k}^{\mathrm{w} , \texttt{j}}
= {\displaystyle \frac{ 1 }{ \sqrt{2} } \sum_{k=1}^{2} } \, v_{k}
\left(
\alpha^{\texttt{j}}_{k} \, {\bf P}_{\bf 1} +
\beta_{k}^{\texttt{j}} \, {\bf T}_{z1}^{+} +
\gamma_{k}^{\texttt{j}} \, {\bf T}_{z2}^{+} +
\epsilon_{k}^{\texttt{j}} \, {\bf T}_{x}^{+} +
\rho_{k}^{\texttt{j}} \, {\bf T}_{x}^{-}
\right).
\end{array}$$ Thus, with help of previous expression and Eq. (\[eq:Yuka-w\]), it is easy to conclude that the Dirac fermion mass matrices are Hermitian matrices without any of their elements equal to zero. However, in the flavor adapted basis the mass matrices in Eq. (\[eq:Mwj\]) acquire the following form: $$\label{MMM}
\begin{array}{l}\vspace{2mm}
{\bf M}_{ \texttt{j} }^{\mathrm{s}} =
{\bf U}_{\mathrm{s}}^{\dagger} {\bf M}_{ \texttt{j} }^{\mathrm{w}} {\bf U}_{\mathrm{s}} =
{\displaystyle \frac{1}{ \sqrt{2} } \sum_{k=1}^{2} } \,
v_{k} {\bf U}_{\mathrm{s}}^{\dagger} {\bf Y}_{k}^{\mathrm{w} , \texttt{j}}
{\bf U}_{\mathrm{s}} , \\
{\bf M}_{\texttt{j}}^{\mathrm{s}} =
{\bf P}_{ \texttt{j} }
\left( \begin{array}{ccc}
0 & \left| A_{ \texttt{j} } \right| & 0 \\
\left| A_{ \texttt{j} } \right| & B_{ \texttt{j} } & C_{ \texttt{j} } \\
0 & C_{ \texttt{j} } & D_{ \texttt{j} }
\end{array}\right)
{\bf P}_{ \texttt{j} }^{\dagger}=
\frac{ v \cos \beta }{\sqrt{2}} \times
\left[ \left( \begin{array}{ccc}
0 & A^{\texttt{j} }_{1} & 0 \\
A^{\texttt{j} \,* }_{1} & B^{\texttt{j} }_{1} & C^{\texttt{j} }_{1} \\
0 & C^{\texttt{j} }_{1} & D^{\texttt{j} }_{1}
\end{array} \right)
+ \tan \beta \left( \begin{array}{ccc}
0 & A^{\texttt{j} }_{2} & 0 \\
A^{\texttt{j} \,* }_{2} & B^{\texttt{j} }_{2} & C^{\texttt{j} }_{2} \\
0 & C^{\texttt{j} }_{2} & D^{\texttt{j} }_{2}
\end{array} \right) \right],
\end{array}$$ where ${{\bf P}_{ \texttt{j} } } = \textrm{diag}
\left( 1 , e^{ - i \phi_{ \texttt{j} } }, e^{ - i \phi_{ \texttt{j} } } \right)$ with $\phi_{ \texttt{j} } = \arg \left\{ A_{ \texttt{j} } \right\}$, and $$\label{Eq:Yuk:FB}
\begin{array}{llll}
A^{\texttt{j} }_{k} =
- \sqrt{3} \left( \epsilon_{k}^{\texttt{j}} - i \rho_{k}^{\texttt{j}} \right) , &
B^{\texttt{j} }_{k} =
- \frac{ 2 }{ 3 } \left( \beta_{k}^{\texttt{j}} + 2 \gamma_{k}^{\texttt{j}} \right) , &
C^{\texttt{j} }_{k} =
\frac{ \sqrt{2} }{ 3 }
\left( 4 \beta_{k}^{\texttt{j}} - \gamma_{k}^{\texttt{j}} \right) , &
D^{\texttt{j} }_{k} =
\alpha^{\texttt{j}}_{k}
+ \frac{ 2 }{ 3 } \left( \beta_{k}^{\texttt{j}} + 2 \gamma_{k}^{\texttt{j}} \right),
\end{array}$$ with $k=1,2$. The parameters in Eq. (\[Eq:Yuk:FB\]) are the Yukawa matrices elements expressed in the flavor adapted basis. Finally, in the Higgs sector $\tan \beta = \dfrac{ v_{2} }{ v_{1} }$ with $v^{2} = v_{1}^{2} + v_{2}^{2} = \left( 246.22~\textrm{GeV}\right)^{2}$.
Here, we consider that the active neutrinos acquire their mass through the type-I seesaw mechanism, Eq. (\[eq:seesaw\]), where the Dirac fermion mass matrix is given by Eq. (\[MMM\]), while we suppose that in the flavor adapted basis the right-handed neutrino mass matrix has the form: $${\bf M}_{R}^{\mathrm{s}} = \textrm{diag} \left( A_{R}, A_{R}, D_{R} \right) {\bf D}^{(3)}
\left( A_{1} \right).$$ In the latter expression $A_{R}$ and $D_{R}$ are real parameters, and the form of ${\bf D}^{(3)}\left( A_{1} \right)$ matrix is given by Eq. (\[eq:A-1\]). So, in the flavor adapted basis the active neutrinos mass matrix is $$\label{eq:Mnu}
\begin{array}{ccc}
{\bf M}_{\nu}^{\mathrm{s}} =
{\bf P}_{\nu}^{\dagger}
\left( \begin{array}{ccc}
0 & a_{\nu} & 0 \\
a_{\nu} & \left| b_{\nu} \right| & \left| c_{\nu} \right| \\
0 & \left| c_{\nu} \right| & d_{\nu}
\end{array} \right)
{\bf P}_{\nu}^{\dagger} ,
\end{array}$$ where ${\bf P}_{\nu} = e^{ i \phi_{ \nu } } \textrm{diag}
\left( 1, e^{ - i 2 \phi_{ \nu } }, e^{ -i \phi_{ \nu } } \right)$ with $\phi_{ \nu } = \arg \left\{ C_{ \nu } \right\}$, and $\arg \left\{ C_{ \nu } \right\} = 2 \arg \left\{ B_{ \nu } \right\}$, $$\begin{array}{l}
a_{\nu} = \frac{ \left| A_{ \nu_{D} } \right|^{2} }{ A_{R} }, \quad
b_{\nu} = \frac{ C_{ \nu_{D} }^{2} }{ D_{R} } +
\frac{ 2 B_{ \nu_{D} } A_{ \nu_{D} }^{*} }{ A_{R} } , \quad
c_{\nu} = \frac{ C_{ \nu_{D} } D_{ \nu_{D} } }{ D_{R} } +
\frac{ C_{ \nu_{D} } A_{ \nu_{D} }^{*} }{ A_{R} } ,
\quad \textrm{and} \quad
d_{\nu} = \frac{ D_{ \nu_{D} }^{2} }{ D_{R} }.
\end{array}$$ In this work, we study the flavor dynamics through the Yukawa matrices in the 2HDM-III plus massive neutrinos and a horizontal flavor symmetry $S_3$. This theoretical framework is called $\nu$2HDM$\otimes S_3$.
The mass and mixing matrices as function of fermion masses {#sec:PMNS_Masses}
----------------------------------------------------------
For a normal \[inverted\] hierarchy[^3] in the mass spectrum, the real symmetric matrices in Eqs. (\[MMM\]) and (\[eq:Mnu\]), which are associated with the fermion mass matrices, can be reconstructed through the orthogonal transformation[^4] $$\label{eq:DMSym}
\bar{\bf M}_{f}^{ \mathrm{n[i]} } = m_{f 3[2] } \,
{\bf O}_{f}^{ \mathrm{n[i]} } {\bf \Delta}_{f}^{ \mathrm{n[i]} }
\left( { \bf O}_{f}^{ \mathrm{n[i]} } \right)^{\top}, \qquad f = u, d, l, \nu,$$ where ${\bf \Delta}_{f}^{ \mathrm{n[i]} } = \textrm{diag}
\left( \widehat{m}_{f 1[3]}, -\widehat{m}_{f 2[1] }, 1 \right) $, in this expression $\widehat{m}_{ f 1[3] } = m_{ f 1[3] } / m_{ f 3[2] }$ and $\widehat{m}_{ f 2[1] } = | m_{ f 2[1] } | / m_{ f 3[2] }$. Here, $m_{f 2[1] } = -| m_{f 2[1] } |$ and the $m_{f}$’s are the eigenvalues of fermion mass matrices, [*i.e.*]{}, the particle masses. From algebraic invariants of the expression in Eq. (\[eq:DMSym\]) we have $$\begin{array}{ll}\vspace{2mm}
a_{f}^{ \mathrm{n[i]} } \equiv
\dfrac{ \left( {\bf M}_{f} \right)_{12} }{ m_{f 3[2] } }
= \sqrt{ \frac{ \widehat{m}_{f 1[3]} \widehat{m}_{f 2[1]} }{ 1 - \delta_{f} } } , &
b_{f}^{ \mathrm{n[i]} } \equiv
\dfrac{ \left( {\bf M}_{f} \right)_{22} }{ m_{f 3[2] } } =
\widehat{m}_{f 1[3]} - \widehat{m}_{f 2[1]} + \delta_{f}, \\
c_{f}^{ \mathrm{n[i]} } \equiv
\dfrac{ \left( {\bf M}_{f} \right)_{23} }{ m_{f 3[2] } } =
\sqrt{ \frac{ \delta_{f} }{ 1 - \delta_{f} } \xi_{ f 1[3] } \xi_{ f 2[1] } } ,&
d_{f}^{ \mathrm{n[i]} } \equiv
\dfrac{ \left( {\bf M}_{f} \right)_{33} }{ m_{f 3[2] } } =
1 - \delta_{f} ,
\end{array}$$ where $ \xi_{ f 1[3] } = 1 - \widehat{m}_{ f 1[3] } - \delta_{f}$ and $\xi_{ f 2[1] } = 1 + \widehat{m}_{ f 2[1] } - \delta_{f}$. Also, the free parameter $\delta_{f}$ must satisfy the relation $1 - \widehat{m}_{ f 1[3] } > \delta_{f} > 0$ [@Canales:2012dr]. The real orthogonal matrix given in Eq. (\[eq:DMSym\]) written in terms of fermion masses has the shape [@Barradas-Guevara:2016gda; @Barranco:2010we] $$\label{eq:Real-O}
{\bf O}_{f}^{ \mathrm{n[i]} } =
\left( \begin{array}{ccc} \vspace{2mm}
\sqrt{ \frac{ \widehat{m}_{ f 2 [1] }
\xi_{ f 1[3] } }{ {\cal D}_{ f 1[3] } } } &
-\sqrt{ \frac{ \widehat{m}_{ f 1 [3] }
\xi_{ f 2[1] } }{ {\cal D}_{ f 2[1] } } } &
\sqrt{ \frac{ \widehat{m}_{ f 1 [3] } \widehat{m}_{ f 2[1] }
\delta_{f} }{ {\cal D}_{ f 3[2] } } } \\ \vspace{2mm}
\sqrt{ \frac{ \widehat{m}_{ f 1 [3] } \left( 1 - \delta_{f} \right)
\xi_{ f 1[3] } }{ {\cal D}_{ f 1[3] } } } &
\sqrt{ \frac{ \widehat{m}_{ f 2 [1] }
\left( 1 - \delta_{f} \right) \xi_{ f 2[1] } }{ {\cal D}_{ f 2[1] } } } &
\sqrt{ \frac{ \delta_{f} \left( 1 - \delta_{f} \right)
}{ {\cal D}_{ f 3[2] } } } \\ \vspace{2mm}
-\sqrt{ \frac{ \widehat{m}_{ f 1[3] } \delta_{f} \xi_{ f 2[1] } }{
{\cal D}_{ f 1[3] } } } &
-\sqrt{ \frac{ \widehat{m}_{ f 2[1] } \delta_{f} \xi_{ f 1[3] } }{
{\cal D}_{ f 2[1] } } } &
\sqrt{ \frac{ \xi_{ f 1[3] } \xi_{ f 2[1] } }{ {\cal D}_{ f 3[2] } } }
\end{array} \right) .$$ In this orthogonal matrix we have $$\begin{array}{l}
{\cal D}_{f 1[3]} =
\left( 1 - \delta_{f} \right)
\left( \widehat{m}_{ f 1[3] } + \widehat{m}_{ f 2[1] } \right)
\left( 1 - \widehat{m}_{ f 1[3] } \right), \\
{\cal D}_{f 2[1]} =
\left( 1 - \delta_{f} \right)
\left( \widehat{m}_{ f 1[3] } + \widehat{m}_{ f 2[1] } \right)
\left( 1 + \widehat{m}_{ f 2[1] } \right), \\
{\cal D}_{f 3[2]} =
\left( 1 - \delta_{f} \right) \left( 1 - \widehat{m}_{f 1 [3]} \right)
\left( 1 + \widehat{m}_{f 2 [1]} \right).
\end{array}$$ In the theoretical framework of $\nu$2HDM$\otimes S_{3}$, the unitary matrices that diagonalize the mass matrices of charged leptons and active neutrinos are defined as $${\bf U}_{\ell}^{\mathrm{n[i]}} = {\bf U}_{\mathrm{s}} {\bf P}_{\ell}
{\bf O}_{\ell}^{\mathrm{n[i]}}, \qquad \ell = l, \nu.$$ The matrix ${\bf U}_{\mathrm{s}}$ is given by Eq. (\[eq:Us\]), and the diagonal phase matrices can be founded in Eqs. (\[MMM\]) and (\[eq:Mnu\]). Finally, the real orthogonal matrix ${\bf O}_{\ell}^{ \mathrm{n[i]} }$ is given in Eq. (\[eq:Real-O\]). From the previous unitary matrix, in the mass states basis, the Dirac fermion mass matrices take the shape $$\begin{array}{l}\vspace{2mm}
{\bf \Delta}_{ \texttt{j} } =
{\bf U}_{\texttt{j} }^{\dagger}
{\bf M}_{ \texttt{j} }^{\mathrm{w}}
{\bf U}_{\texttt{j} } =
{\displaystyle \frac{ 1 }{ \sqrt{2} }\sum_{k=1}^{2} } \, v_{k} \,
{\bf U}_{\texttt{j} }^{\dagger}
{\bf Y}_{k}^{\mathrm{w} , \texttt{j} }
{\bf U}_{\texttt{j} } =
{\displaystyle \frac{ 1 }{ \sqrt{2} } \sum_{k=1}^{2} } \, v_{k} \,
\widetilde{\bf Y}_{k}^{ \texttt{j} } , \qquad \texttt{j} = u,d,l,\nu_{D},
\end{array}$$ where $\widetilde{\bf Y}_{k}^{ \texttt{j} } =
{\bf U}_{\texttt{j} }^{\dagger} {\bf Y}_{k}^{\mathrm{w}, \texttt{j} } {\bf U}_{\texttt{j} }$ are the Yukawa matrices in the mass states basis. Now, with the help of the orthogonal matrix given in Eq. (\[eq:Real-O\]) and, as the mass spectrum of Dirac particles only has the normal hierarchy, it is easy to obtain that the elements of Yukawa matrices $\widetilde{\bf Y}_{k}^{\texttt{j}}$ can be expressed in terms of geometric mean of Dirac fermion masses normalized with respect of electroweak scale, [*i.e.*]{}, $$\label{eq:Cheng-Sher}
\begin{array}{l}
\left( \widetilde{\bf Y}_{k}^{ \texttt{j} } \right)_{ \texttt{rt} } =
\dfrac{ \sqrt{ m_{ \texttt{jr} } m_{ \texttt{jt} } } }{ v }
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ \texttt{rt} } \qquad
(\texttt{r,t} = 1, 2, 3).
\end{array}$$ Here, $\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ \texttt{rt} }$ are complex parameters, whose the explicit form is given in Appendix \[App:B\]. In the literature, this last relation is called [*Cheng-Sher ansatz*]{} [@Cheng:1987rs], which is associated with Higgs-fermions couplings. Also, the result obtained in Eq. (\[eq:Cheng-Sher\]) can be extended to any Hermitian mass matrix that may be brought to a two zero texture matrix by means of an unitary transformation.
As one knowns, the 2HDM-III is a generic description of particle physics at a higher energy scale ($\gtrsim \textrm{TeV}$), and its imprint at low energies is reflected in Yukawa couplings structure. A detailed study of Yukawa Lagrangian within the 2HDM-III is given in [@DiazCruz:2009ek; @DiazCruz:2004pj; @DiazCruz:2004tr; @HernandezSanchez:2013xj; @Felix-Beltran:2013tra], while the phenomenological implications of this model in scalar sector including Lepton Violation (LV) and/or FCNC’s are presented in [@GomezBock:2009xz; @Martinez:2008hu; @Martinez:2005sya]. In these works, the FCNC’s are under control because the authors have been assuming that Yukawa matrices in the weak and mass basis, are represented by means of an Hermitian matrix with two texture zeroes and the [*Cheng-Sher ansatz*]{}, respectively. In our model $\nu\textrm{2HDM}\otimes S_3$, the two texture zeroes shape for the Yukawa matrices is obtained by imposing a flavor symmetry $S_{3}$ and from its explicit sequential breaking according to the chain $S_{3L}^{ \texttt{j} } \otimes S_{3R }^{ \texttt{j} } \supset S_{3}^{\mathrm{diag}} \supset
S_{2}^{\mathrm{diag}}$. An immediate consequence from the above is that Yukawa matrices in the mass basis naturally take the form of so-called [*Cheng-Sher ansatz*]{}. Therefore, we can say that the FCNC’s are under control in the $\nu$2HDM$\otimes S_{3}$.
The lepton flavor mixing matrix is defined as ${\bf U}_{\mathrm{PMNS}} = {\bf U}_{l}^{\dagger} {\bf U}_{\nu}$ [@Hochmuth:2007wq]. In this context the $\textrm{PMNS}$ matrix takes the expression $$\label{Eq:PMNS_th}
{\bf U}_{\mathrm{PMNS}} =
{\bf O}_{l}^{ \top } {\bf P}^{(\nu - l)} {\bf O}_{\nu}^{ \mathrm{n[i]} } ,$$ where ${\bf P}^{(\nu - l)} =
\textrm{diag} \left( 1, e^{i \phi_{\ell 1} }, e^{i \phi_{\ell2} } \right)$ with $\phi_{\ell 1} = \phi_{l} - 2 \phi_{\nu}$ and $\phi_{\ell 2} = \phi_{l} - \phi_{\nu}$. The explicit form of the entries of the PMNS matrix are given in Appendix \[App:C\]. The first conclusion of this $\textrm{PMNS}$ matrix is that ${\bf U}_{\mathrm{s}}$ unitary matrix, the one that allows us to pass from the weak basis to the flavor symmetry adapted basis, is unobservable in the lepton flavor mixing matrix.
The mixing angle and CP violation phases {#Sec:Angle-Phase}
----------------------------------------
In the symmetric parametrization of lepton flavor mixing matrix, the relation between mixing angles and the entries of $\textrm{PMNS}$ matrix is [@Hochmuth:2007wq; @Bilenky:2010zza] $$\label{Eq:senoscuadrados}
\sin^{2} \theta_{13} \equiv
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{13} \right|^{2}, \quad
\sin^{2} \theta_{12} \equiv
\frac{
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{12} \right|^{2}
}{
1 - \left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{13} \right|^{2} }, \quad
\sin^{2} \theta_{23} \equiv
\frac{
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{23} \right|^{2}
}{
1 - \left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{13} \right|^{2}
}.$$ On the one hand, the lepton Jarlskog invariant which appears in conventional neutrino oscillations is defined as: $${\cal J}_{\mathrm CP} =
{\cal I}m \left \{
\left( {\bf U}_{\mathrm{PMNS}} \right)_{11}^{*}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{23}^{*}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{13}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{21}
\right \},$$ and in the symmetric parametrization it has the form $$\label{INV:JCP}
\begin{array}{l}
{\cal J}_{\mathrm CP} =
\frac{1}{8}
\sin 2 \theta_{12} \, \sin 2 \theta_{23} \, \sin 2 \theta_{13} \, \cos \theta_{13} \,
\sin \delta_{\mathrm CP} ,
\end{array}$$ where $\delta_{\mathrm CP } = \phi_{13} - \phi_{23} - \phi_{12}$. Moreover, the invariants $$\begin{array}{l}
{\cal I}_{1} = {\cal I}m \left \{
\left( {\bf U}_{\mathrm{PMNS}} \right)_{12}^{2}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{11}^{*2}
\right \}
\quad \textrm{and} \quad
{\cal I}_{2} = {\cal I}m \left \{
\left( {\bf U}_{\mathrm{PMNS}} \right)_{13}^{2}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{11}^{*2}
\right \},
\end{array}$$ associated with the Majorana phases [@Branco:1986gr; @Branco:2011zb; @Jenkins:2007ip] take the expressions $$\label{INV:I1-I2}
\begin{array}{l}
{\cal I}_{1} = \frac{1}{4} \sin^{2} 2 \theta_{12} \cos^{4} \theta_{13} \sin ( - 2 \phi_{12} )
\quad \textrm{and} \quad
{\cal I}_{2} = \frac{1}{4} \sin^{2} 2 \theta_{13} \cos^{2} \theta_{12} \sin ( - 2 \phi_{13} ).
\end{array}$$ Then, the phase factors associated with the CP violation can be written as: $$\label{Eq:CP_Phases}
\begin{array}{l}\vspace{2mm}
\sin ( \delta_{\mathrm CP } ) =
\frac{ {\cal J}_{\rm CP}
\left( 1 - \left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{13} \right|^{2} \right)
}{
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{11} \right|
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{12} \right|
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{13} \right|
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{23} \right|
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{33} \right| } , \\
\sin \left( - 2 \phi_{12} \right) =
\frac{
{\cal I}_{1}
}{
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{11} \right|^{2}
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{12} \right|^{2}
} , \quad
\sin \left( - 2 \phi_{13} \right) =
\frac{
{\cal I}_{2}
}{
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{11} \right|^{2}
\left| \left( {\bf U}_{\mathrm{PMNS}} \right)_{13} \right|^{2}
} .
\end{array}$$
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![In the left-panel we show the sum of neutrino masses. The right-panel shows the neutrino masses. The neutrino oscillation parameters $\Delta m_{i j}^{2}$ are taken within the currently allowed $3\sigma$ range [@Esteban:2016qun]. The upper bound on the mass of the lightest neutrino is obtained from the Planck result where $\sum m_{\nu_{i}} < 0.23$ eV at 95% level [@Ade:2015xua]. []{data-label="Fig:Masa_neutrinos"}](suma.pdf "fig:"){width="8.7cm" height="6.3cm"} ![In the left-panel we show the sum of neutrino masses. The right-panel shows the neutrino masses. The neutrino oscillation parameters $\Delta m_{i j}^{2}$ are taken within the currently allowed $3\sigma$ range [@Esteban:2016qun]. The upper bound on the mass of the lightest neutrino is obtained from the Planck result where $\sum m_{\nu_{i}} < 0.23$ eV at 95% level [@Ade:2015xua]. []{data-label="Fig:Masa_neutrinos"}](masas.pdf "fig:"){width="8.7cm" height="6.3cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The equivalence between the PDG and symmetric parameterization may be expressed as ${\bf U}_{ \mathrm{PDG} } = {\bf K} {\bf U}_{ \mathrm{Sym} }$, where ${\bf K} = \textrm{diag} \left( 1, e^{i \frac{ \alpha_{21} }{2} }, e^{i \frac{ \alpha_{31} }{2} } \right)$ with $\delta_{CP} = \phi_{13} - \phi_{23} - \phi_{12}$, $\alpha_{21} = - 2 \phi_{12}$ and $\alpha_{31} = - 2 ( \phi_{12} + \phi_{23} )$.
Numerical analysis\[sec:numericalanal\]
=======================================
In the three flavor neutrino scheme there are six independent parameters which rule the behavior of neutrino oscillation phenomena: flavor mixing angles, the “Dirac-like” CP-violating phase and squared-mass splitting. The latest neutrino oscillation parameter is defined as $\Delta m_{i j}^{2} \equiv m^{2}_{ \nu_{i} } - m^{2}_{ \nu_{j} }$, in agreement with the results obtained in the global fit reported in Ref. [@Esteban:2016qun]. These neutrino oscillation parameters have the following numerical values (at BFP$\pm 1 \sigma$ and $3\sigma$)[^5] $$\label{Eq:Val_DMij}
\begin{array}{rl}\vspace{2mm}
\Delta m_{21}^{2} \; \left( 10^{-5}~\textrm{eV}^{2} \right) = &
7.50_{-0.17}^{+0.19}, \; 7.03-8.09, \\ \vspace{2mm}
\Delta m_{31}^{2} \; \left( 10^{-3} \textrm{eV}^{2} \right) = &
2.524_{-0.040}^{+0.039}, \; 2.407-2.643, \; \textrm{for NH},\\
\Delta m_{23}^{2} \; \left( 10^{-3} \textrm{eV}^{2} \right) = &
2.514_{-0.041}^{+0.038}, \; 2.399-2.635, \; \textrm{for IH}.
\end{array}$$ From the definition of squared-mass splitting $\Delta m_{i j}^{2}$ two of neutrino masses can be written as: $$\label{Eq:For_mnu}
\begin{array}{l}\vspace{2mm}
m_{ \nu_{3 [2]} } = \sqrt{ m_{\nu_{1 [3]} }^{2} + \Delta m_{31 [23]}^{2} },
\quad \textrm{and} \quad
m_{ \nu_{2 [1]} } = \sqrt{ m_{\nu_{1 [3]} }^{2} + \Delta m_{21 [31]}^{2} }.
\end{array}$$ where $m_{\nu_{1 [3]} }$ is the lightest neutrino mass[^6]. Also, this neutrino mass is considered as the only one free parameter in the above expressions, since the mass-squared differences $\Delta m_{i j}^{2}$ are determined by experimental means.
From the results reported by Planck Collaboration for cosmological parameters, the upper limit on the active neutrino masses sum is $\sum m_{\nu_{i}} < 0.23$ eV, for an active neutrinos number equal to $N_{\textrm eff} = 3.15 \pm 0.23$ [@Ade:2015xua]. This upper bound is independent of hierachy in the neutrino mass spectrum. So, with all the above experimental information and considering the expressions in Eq. (\[Eq:For\_mnu\]), we can obtain the following value range for neutrino masses $$\begin{array}{lll}
m_{ \nu_{1} } \left( 10^{-2}~\textrm{eV} \right) =
\left \{ \begin{array}{l}
\left[0.00 , 7.10 \right], \\
\left[4.90 , 8.25 \right],
\end{array} \right. &
m_{ \nu_{2} } \left( 10^{-2}~\textrm{eV} \right) =
\left \{ \begin{array}{l}
\left[0.84 , 7.13 \right], \\
\left[4.97 , 8.30 \right],
\end{array} \right. &
m_{ \nu_{3} } \left( 10^{-2}~\textrm{eV} \right) =
\left \{ \begin{array}{l}
\left[4.80 , 8.75 \right], \\
\left[0 , 6.45 \right].
\end{array} \right.
\end{array}$$ Here, the oscillation parameters $\Delta m_{ij}^{2}$ are taken within the currently allowed $3\sigma$ range [@Esteban:2016qun]. The values in the first and second row correspond to a normal and inverted hierarchy in the neutrino mass spectrum, respectively. For both hierarchies there is the possibility that the lightest neutrino mass could be zero. Namely in this case, the lightest neutrino is a massless particle. In the Fig. \[Fig:Masa\_neutrinos\], the behaviour of neutrino masses (right panel), as well as of the neutrino masses sum as function of ligntest neutrino mass mnu (left panel), are shown.
The likelihood test $\chi^{2}$
------------------------------
$\Delta m_{ij}^{2}$ at $\Phi_{\ell 1}~[^{\circ}]$ $\Phi_{\ell 2}~[^{\circ}]$ $m_{\nu_{\mathrm{lightest} } }$ \[eV\] $\delta_{\ell}$ $\delta_{\nu}$ $\theta_{12}^{ ^{\mathrm{th}} }~[^{\circ}]$ $\theta_{23}^{ ^{\mathrm{th}} }~[^{\circ}]$ $\theta_{13}^{ ^{\mathrm{th}} }~[^{\circ}]$ $\delta_{\mathrm{CP}}~[^{\circ}]$ $\phi_{12}~[^{\circ}]$ $\phi_{13}~[^{\circ}]$ $\chi^{2}_{\mathrm{min}}$
---- ------------------------ ---------------------------- ---------------------------- ---------------------------------------- ----------------- ---------------- --------------------------------------------- --------------------------------------------- --------------------------------------------- ----------------------------------- ------------------------ ------------------------ ---------------------------
3$\sigma$ $270$ $195$ $2.57 \times 10^{-3} $ $0.20460$ $0.63519$ $33.58$ $41.60$ $8.47$ $-68.65$ $-5.86$ $14.77$ $4.63 \times 10^{-4}$
NH BFP$\pm 1\sigma$ $270$ $195$ $2.57 \times 10^{-3} $ $0.22256$ $0.64507$ $33.59$ $41.61$ $8.46$ $-70.74$ $-5.79$ $14.67$ $3.13 \times 10^{-4}$
BFP $270$ $195$ $2.57 \times 10^{-3} $ $0.21492$ $0.64008$ $33.80$ $41.63$ $8.45$ $-69.85$ $-5.80$ $14.73$ $8.75 \times 10^{-2}$
3$\sigma$ $290$ $187$ $2.49 \times 10^{-2}$ $0.59943$ $0.01999$ $33.67$ $50.08$ $8.48$ $-80.90$ $-5.25$ $-2.18$ $2.30 \times 10^{-2}$
IH BFP$\pm 1\sigma$ $290$ $187$ $2.49 \times 10^{-2}$ $0.59888$ $0.01995$ $33.74$ $50.05$ $8.49$ $-80.88$ $-5.25$ $-2.18$ $4.56 \times 10^{-2}$
BFP $290$ $187$ $2.49 \times 10^{-2}$ $0.59798$ $0.01971$ $33.83$ $49.99$ $8.49$ $-80.83$ $-5.25$ $-2.19$ $1.08 \times 10^{-1}$
: Numerical values obtained in the BFP for the five parameters, the lepton mixing angles and the phase factors associated with the CPV. These results were obtained considering to $\Delta m_{ij}^{2}$ at BFP, BFP$\pm \sigma$ and 3$\sigma$ range [@Esteban:2016qun], and simultaneously $\Phi_{\ell 1}$, $\Phi_{\ell 2}$, $\delta_{\ell}$, $\delta_{\nu}$, and $m_{\nu_{1 [3] } }$ are free parameters in the $\chi^{2}$ function. []{data-label="tab:table1"}
[cc]{}
To validate our hypothesis where the $S_{3}$ horizontal flavor symmetry is explicitly sequential breaking according to the chain $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} }
\supset S_{3}^{\mathrm{diag}} \supset S_{2}^{\mathrm{diag}}$, hence all fermion mass matrices are represented through a matrix with two texture zeroes, we make a likelihood test where the $\chi^{2}$ function is defined as: $$\label{chicuadradadef}
\chi^{2} =
\sum_{i < j}^{3}
\frac{
\left(
\sin^{2} \theta_{ij}^{ ^{\mathrm{exp} } }
- \sin^{2} \theta_{ij}^{ ^{ \mathrm{th} } } \right)^2
}{
\sigma_{ \theta ij }^{2}
}.$$ In this expression, the “$\mathrm{th}$” superscript is used to denote the theoretical expressions of lepton mixing angles, while the terms with superscript “$\mathrm{exp}$” denote to the experimental data with uncertainty $\sigma_{\theta_{ij}}$ for lepton mixing angles. For these latter we consider the following values, at BFP$\pm 1 \sigma$ [@Esteban:2016qun]: $$\label{Eq:Exp:Thetas}
\begin{array}{l}
\sin^{2} \theta_{12}^{ ^{\mathrm{exp}} } (10^{-1}) =
3.06 \pm 0.12 , \quad
\sin^{2} \theta_{23}^{ ^{\mathrm{exp}} } (10^{-1}) =
\left \{ \begin{array}{l}
4.41_{-0.21}^{+0.27} , \\
5.87_{-0.24}^{+0.20} ,
\end{array} \right. \quad
\sin^{2} \theta_{13}^{ ^{\mathrm{exp}} } (10^{-2}) =
\left \{ \begin{array}{l}
2.166 \pm 0.0075 , \\
2.179 \pm 0.0076 ,
\end{array} \right.
\end{array}$$ the values in the first and second row correspond to a normal and inverted hierarchy in the neutrino mass spectrum, respectively. From expressions in Eqs. (\[eq:Real-O\]), (\[Eq:PMNS\_th\]), (\[Eq:senoscuadrados\]), and (\[Eq:For\_mnu\]), it is easy to conclude that the $\chi^{2}$ function depends of five free parameters $\chi^{2} =
\chi^{2} \left( \Phi_{\ell 1}, \Phi_{\ell 2}, \delta_{\ell}, \delta_{\nu},
m_{\nu_{1 [3] } }
\right)$. However, the $\chi^{2}$ function depends only on three experimental data which correspond to the leptonic flavor mixing angles. Therefore, if simultaneously we consider $\Phi_{\ell 1}$, $\Phi_{\ell 2}$, $\delta_{\ell}$, $\delta_{\nu}$, and $m_{\nu_{1 [3] } }$ as free parameters in the likelihood test, we can only determine the values of these parameters in the best fit point (BFP). In accordance with the above, we first seek the BFP by means of a likelihood test where the $\chi^{2}$ function have all those the five free parameters $\Phi_{\ell 1}$, $\Phi_{\ell 2}$, $\delta_{\ell}$, $\delta_{\nu}$, and $m_{\nu_{1 [3] } }$. To minimize the $\chi^{2}$ function we have done a scanning of the parameter space where we considered the following values for the charged lepton masses [@Olive:2016xmw] $$\begin{array}{l}
m_{e}=0.5109998928 \pm 0.000000011, \quad
m_{\mu}=105.6583715\pm 0.0000035, \quad
m_{\tau}=1776.82\pm 0.16,
\end{array}$$ while in Eq. (\[Eq:Exp:Thetas\]) are given the experimental values for leptonic mixing angles.
[cc]{}\
In the Table \[tab:table1\] we show the numerical values obtained in the BFP for the five parameters, the lepton mixing angles and the phase factors associated with the CPV. All of these results were obtained considering to $\Delta m_{ij}^{2}$ at BFP, BFP$\pm \sigma$ and 3$\sigma$ range, and simultaneously $\Phi_{\ell 1}$, $\Phi_{\ell 2}$, $\delta_{\ell}$, $\delta_{\nu}$, and $m_{\nu_{1 [3] } }$ are free parameters in the $\chi^{2}$ function of the likelihood test.
Now, as we know the numerical values of the five free parameters at BFP, we perform a new $\chi^{2}$ analysis for the case when the oscillation parameters $\Delta m_{ij}^{2}$ take the values at BFP and where we fix $m_{\nu_{1 [3] } }$, $\Phi_{\ell 1}$ and $\Phi_{\ell 2}$ parameters to the values given in Table \[tab:table1\].
So, $\chi^{2} = \chi^{2}(\delta_l,\delta_\nu)$ function implies one degree of freedom. In the Fig. \[Fig:calorchi\], we show the allowed regions in the parameter space for $\delta_l$ and $\delta_\nu$ at 3$\sigma$ C.L., the red point (${\color{red}}\bullet$) represent the BFP. The left-panel is for normal hierarchy and we can see that $\delta_\nu$ and $\delta_l$ parameters are of the order of $10^{-1}$. The right-panel is for the inverted hierarchy and here $\delta_\nu$ parameter is $\sim10^{-1}$, while the $\delta_l$ parameter is of the order of $10^{-2}$.
[cc]{}\
Associated to the parameter regions of $\delta_{\nu}$ and $\delta_{l}$ given in Fig. \[Fig:calorchi\] for both hierarchies on neutrino mass spectrum, and based on Eq. (\[Eq:CP\_Phases\]), we found the predicted regions by the $\nu\textrm{2HDM}\otimes S_{3}$ for “Dirac-like” phase $\delta_{CP}$. These regions are shown in Fig. \[Fig:deltacp\]. In concordance with experimental data, plots as function of $\delta_{l}$ are more restricted than the other one as function of $\delta_{\nu}$. These results correspond closely with allowed regions obtained in the global fit reported in Ref. [@Esteban:2016qun].
In the same way, for both hierarchies, we analyze the three leptonic flavor mixing angles, but in Fig. \[Fig:senos\] we just show the allowed regions for the atmospheric mixing angle $\theta_{23}$, at BFP, $\pm 1\sigma$ and $3\sigma$ C.L. In order to round the above results, from our analysis we obtain the following values for the three mixing angles, at BFP$\pm 1\sigma$ C.L.: $$\begin{array}{c}
\sin^{2} \theta_{12}^{ ^{\mathrm{th}} } (10^{-1}) =
\left \{ \begin{array}{l}
3.09_{-0.065}^{+0.066} , \\
3.10_{-0.011}^{+0.011} ,
\end{array} \right. \quad
\sin^{2} \theta_{23}^{ ^{\mathrm{th}} } (10^{-1}) =
\left \{ \begin{array}{l}
4.41_{-0.14}^{+0.10} , \\
5.87_{-0.223}^{+0.224} ,
\end{array} \right. \quad
\sin^{2} \theta_{13}^{ ^{\mathrm{th}} } (10^{-2}) =
\left \{ \begin{array}{l}
2.160 \pm 0.14 ,\\
2.177 \pm 0.12 .
\end{array} \right.
\end{array}$$ We also obtained the following allowed value ranges at BFP$\pm 1\sigma$ for the “Dirac-like” phase $\delta_{\mathrm{CP}}$, as well as for the two Majorana phase factors $\phi_{12}$ and $\phi_{13}$: $$\label{Eq:CP_phases-Fit}
\begin{array}{c}
\delta_{CP} (^{\circ}) =
\left \{ \begin{array}{l}
-69.8_{-6.110}^{+5.508} , \\
-80.83_{-0.709}^{+0.652} ,
\end{array} \right. \quad
\phi_{12} (^{\circ}) =
\left \{ \begin{array}{l}
-5.800_{-0.150}^{+0.170} , \\
-5.24_{-0.148}^{+0.153} ,
\end{array} \right. \quad
\phi_{13} (^{\circ}) =
\left \{ \begin{array}{l}
14.744_{-1.366}^{+1.266} ,\\
-2.190_{-0.0005}^{+0.0030} .
\end{array} \right.
\end{array}$$ From Eq. (\[Eq:CP\_phases-Fit\]) and the Table \[tab:table1\] we can conclude that values for the $\delta_{\text{CP}}$ phase obtained in our scheme are consistent with a maximal CP violation.
Finally, as a immediate result of the above likelihood analysis, the entries magnitude of ${\mathbf U}_{\textrm{PMNS}}$ mixing matrix can numerically computed. So, at $3\sigma$ C.L., we have that ${\mathbf U}_{\textrm{PMNS}}$ matrix takes the form: $$\begin{pmatrix}
0.822^{+0.0044}_{-0.0045} & 0.550^{+0.0055}_{-0.0054} & 0.147^{+0.0047}_{-0.0048} \\
0.395^{+0.0181}_{-0.0154} & 0.642^{+0.0008}_{-0.0001} & 0.657^{+0.0082}_{-0.0111} \\
0.410^{+0.0056}_{-0.0089} & 0.534^{+0.0045}_{-0.0056} & 0.739^{+0.0088}_{-0.0064} \\
\end{pmatrix}, \qquad {\bf {\textrm{Normal Hierarchy}}},$$ $$\begin{pmatrix}
0.822^{+0.0012}_{-0.0012} & 0.551^{+0.0007}_{-0.0006} & 0.147^{+0.0041}_{-0.0041} \\
0.355^{+0.0072}_{-0.0071}&
0.547^{+0.0144}_{-0.0149} &
0.758^{+0.0138}_{-0.0141} \\
0.446^{+0.0077}_{-0.0080} & 0.630^{+0.0120}_{-0.0122} &
0.636^{+0.0174}_{-0.0178} \\
\end{pmatrix}, \qquad {\bf {\textrm{Inverted Hierarchy}}}.$$
Phenomenological implications {#sec:PenImp}
=============================
In the above section we have seen that in our theoretical framework, where the $S_{3}$ flavor symmetry sets up that the fermion mass matrices should have two texture zeroes, we can reproduce the values of oscillation parameters in very good agreement with the last experimental data. In the following, we shall investigate the phenomenological implications of these results for the neutrinoless double beta decay ($0\nu \beta \beta $) and the CP violation in neutrino oscillations in matter.
Neutrinoless double beta decay\[subsec:neutrinoless\]
-----------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![In the upper panel we show the effective mass $\left| m_{ee} \right|$ which is involved in the $0\nu \beta \beta$ decay. The red and blue bands are obtained with the current experimental data on neutrino oscillations, at 3$\sigma$ [@Esteban:2016qun], for an inverted and normal neutrino mass hierarchy, respectively. On the one hand, from the combination of EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] results we have the follow upper bound for $\left| m_{ee} \right| < 0.120$eV. On the other hand, from the results reported by Planck Collaboration we have that $\sum_{i} m_{i} < 0.230$eV at $95\%$ level [@Ade:2015xua], thus an upper bound on the lightest neutrino mass is established. In the left- and right-lower panels we show a zoom in of the allowed regions for $\left| m_{ee} \right|$ obtained at 95% C.L. in the context of $\nu$2HDM$\otimes S_{3}$ for a normal and inverted hierarchy, respectively. []{data-label="Fig:zbbb"}](Mee_nu2HDMS3.pdf "fig:"){width="0.45\linewidth"}
![In the upper panel we show the effective mass $\left| m_{ee} \right|$ which is involved in the $0\nu \beta \beta$ decay. The red and blue bands are obtained with the current experimental data on neutrino oscillations, at 3$\sigma$ [@Esteban:2016qun], for an inverted and normal neutrino mass hierarchy, respectively. On the one hand, from the combination of EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] results we have the follow upper bound for $\left| m_{ee} \right| < 0.120$eV. On the other hand, from the results reported by Planck Collaboration we have that $\sum_{i} m_{i} < 0.230$eV at $95\%$ level [@Ade:2015xua], thus an upper bound on the lightest neutrino mass is established. In the left- and right-lower panels we show a zoom in of the allowed regions for $\left| m_{ee} \right|$ obtained at 95% C.L. in the context of $\nu$2HDM$\otimes S_{3}$ for a normal and inverted hierarchy, respectively. []{data-label="Fig:zbbb"}](Mee_nu2HDMS3_NH_zoom.pdf "fig:"){width="0.4\linewidth"} ![In the upper panel we show the effective mass $\left| m_{ee} \right|$ which is involved in the $0\nu \beta \beta$ decay. The red and blue bands are obtained with the current experimental data on neutrino oscillations, at 3$\sigma$ [@Esteban:2016qun], for an inverted and normal neutrino mass hierarchy, respectively. On the one hand, from the combination of EXO-200 [@Auger:2012ar; @Albert:2014awa] and KamLAND-ZEN [@Gando:2012zm] results we have the follow upper bound for $\left| m_{ee} \right| < 0.120$eV. On the other hand, from the results reported by Planck Collaboration we have that $\sum_{i} m_{i} < 0.230$eV at $95\%$ level [@Ade:2015xua], thus an upper bound on the lightest neutrino mass is established. In the left- and right-lower panels we show a zoom in of the allowed regions for $\left| m_{ee} \right|$ obtained at 95% C.L. in the context of $\nu$2HDM$\otimes S_{3}$ for a normal and inverted hierarchy, respectively. []{data-label="Fig:zbbb"}](Mee_nu2HDMS3_IH_zoom.pdf "fig:"){width="0.4\linewidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The $0\nu \beta \beta$ is a rare second-order weak process where a nucleus $(A,Z)$ decays into another one by the emission of two electrons, whose mode decay is $(A,Z) \rightarrow
(A, Z + 2) + e^{-} + e^{-}$. The observation of this process would establish that neutrino are Majorana particles and that total lepton number is not a conserved symmetry in nature [@Schechter:1981bd; @Duerr:2011zd]. In the most simple version of the process, the amplitude for the decay is proportional to a quantity called the effective mass $m_{ee}$ [@Beringer:1900zz; @Barger:2003vs; @King:2013psa]. In the symmetric parametrization of lepton mixing matrix the effective mass parameter have the shape [@Schechter:1980gr; @Chen:2015siy] $$\left| m_{ee} \right| =
\left| m_{ \nu_{1} } \cos^{2} \theta_{12} \cos^{2} \theta_{13}
+ m_{ \nu_{2} } \sin^{2} \theta_{12} \cos^{2} \theta_{13} e^{ - i 2 \phi_{12} }
+ m_{ \nu_{3} } \sin^{2} \theta_{13} e^{ - i 2 \phi_{13} } \right|\, ,$$ where $\phi_{12}$ and $\phi_{13}$ are the Majorana phases given in Eq. (\[Eq:CP\_Phases\]). In the Fig. \[Fig:zbbb\] we show the allowed regions for the magnitude of effective mass parameter $m_{ee}$, which were obtained in the context of $\nu$2HDM$\otimes S_{3}$. Each one of these regions was obtained by setting the values of some of the five free parameters in the $\chi^{2}$ function, Eq. (\[chicuadradadef\]), to the values given in the Table \[tab:table1\] for $\Delta m_{ij}^{2}$ at BFP. Then, for both hierarchies in the lower panels of Fig. \[Fig:zbbb\], the blue lines were obtained by means a likelihood test where the values of $\phi_{\ell 1}$, $\phi_{\ell 2}$, $\delta_{e}$ and $\delta_{\nu}$ are fixed, while $m_{\nu_{\mathrm{lightest} } }$ is free parameter. The orange bands were obtained by means a likelihood test where the values of $\phi_{\ell 2}$, $\delta_{e}$ and $\delta_{\nu}$ are fixed, while $m_{\nu_{\mathrm{lightest} } }$ and $\phi_{\ell 1}$ are free parameters. The yellow bands were obtained by means of a likelihood test where the values of $\phi_{\ell 1}$, $\delta_{e}$ and $\delta_{\nu}$ are fixed, while $m_{\nu_{\mathrm{lightest} } }$ and $\phi_{\ell 2}$ are free parameters. The sky blue bands were obtained through a likelihood test where the values of $\phi_{\ell 1}$, $\phi_{\ell 2}$ and $\delta_{\nu}$ are fixed, while $m_{\nu_{\mathrm{lightest} } }$ and $\delta_{e}$ are free parameters. Finally, the turquoise bands were obtained via a likelihood test where the values of $\phi_{\ell 1}$, $\phi_{\ell 2}$ and $\delta_{e}$ are fixed, while $m_{\nu_{\mathrm{lightest} } }$ and $\delta_{\nu}$ are free parameters.
To round the previous results, in the Table \[tab:table2\] we show the allowed numerical ranges, at 95% C.L., for the magnitude of effective mass parameter $m_{ee}$ and the lightest neutrino mass $m_{\nu_{\mathrm{lightest} } }$. From the results in the Table \[tab:table2\] it is easy conclude that for the normal hierarchy $m_{ \nu_{1} } \sim 2 \times 10^{-3}$ eV and $|m_{ee}| \sim 3 \times 10^{-3}$ eV, while for the inverted hierarchy $m_{ \nu_{3} } \sim 2 \times 10^{-2}$ eV and $|m_{ee}| \sim 3 \times 10^{-2}$ eV.
Fixed parameters $m_{\nu_{\mathrm{lightest} } }~[ 10^{-2}$ eV\] $\left| m_{ee} \right|~[ 10^{-2}$ eV\]
---- ---------------------------------------------------------------- ------------------------------------------------ ----------------------------------------
$\phi_{\ell 1}$, $\phi_{\ell 2}$, $\delta_{e}$, $\delta_{\nu}$ \[ 0.2360 , 0.2768 \] \[ 0.3204 , 0.3608 \]
$\phi_{\ell 2}$, $\delta_{e}$, $\delta_{\nu}$ \[ 0.2374 , 0.2735 \] \[ 0.3215 , 0.3583 \]
NH $\phi_{\ell 1}$, $\delta_{e}$, $\delta_{\nu}$ \[ 0.2404 , 0.2711 \] \[ 0.3251 , 0.3563 \]
$\phi_{\ell 1}$, $\phi_{\ell 2}$, $\delta_{\nu}$ \[ 0.2349 , 0.2761 \] \[ 0.3229 , 0.3577 \]
$\phi_{\ell 1}$, $\phi_{\ell 2}$, $\delta_{e}$ \[ 0.2164 , 0.2908 \] \[ 0.3121 , 0.3659 \]
$\phi_{\ell 1}$, $\phi_{\ell 2}$, $\delta_{e}$, $\delta_{\nu}$ \[ 2.268 , 2.685 \] \[ 3.491 , 3.717 \]
$\phi_{\ell 2}$, $\delta_{e}$, $\delta_{\nu}$ \[ 2.317 , 2.635 \] \[ 3.511 , 3.694 \]
IH $\phi_{\ell 1}$, $\delta_{e}$, $\delta_{\nu}$ \[ 2.311 , 2.650 \] \[ 3.515 , 3.696 \]
$\phi_{\ell 1}$, $\phi_{\ell 2}$, $\delta_{\nu}$ \[ 2.301 , 2.648 \] \[ 3.512 , 3.695 \]
$\phi_{\ell 1}$, $\phi_{\ell 2}$, $\delta_{e}$ \[ 2.123 , 2.787 \] \[ 3.416 , 3.767 \]
: The allowed numerical ranges, at 95% C.L., for the effective mass parameter magnitude $m_{ee}$ and the lightest neutrino mass $m_{\nu_{\mathrm{lightest} } }$.[]{data-label="tab:table2"}
CP violation in neutrino oscillations in matter\[subsec:CPV\]
-------------------------------------------------------------
[cc]{}\
\
[cc]{}\
\
In the recent years, we have entered into a precision era in the determination of flavor leptonic mixing angles. However, it is not the same situation for the CP violation in this sector, since has yet to be determined experimentally the numerical value of CP violation phase. But we have a hunch of where to look: the neutrino oscillations with matter effects [@Diwan:2016gmz]. One of the aims of the LBL neutrino experiments such as T2K [@Abe:2011ks] and NO$\nu$A [@Adamson:2016tbq], as well as the proposed experiment DUNE [@Acciarri:2015uup], it is determination of the “Dirac-like” CP violation phase and other parameters that rule the neutrino oscillations $\nu_{\mu} \to \nu_{e}$ and $\bar{\nu}_{\mu} \to \bar{\nu}_{e}$. The transition probability in matter for the oscillation between electron and muon neutrinos, as well as for the oscillation between electron and muon antineutrinos, have the form [@Nunokawa2008338; @Chen:2015siy; @Gonzalez-Canales:2016xia]: $$\label{Eq:P_mu-e}
\begin{array}{l} \vspace{2mm}
P \left( \nu_{\mu} \rightarrow \nu_{e} \right) \simeq
P_{ \textrm{atm} } + P_{ \textrm{sol} }
+ 2 \sqrt{ P_{ \textrm{atm} } } \sqrt{ P_{ \textrm{sol} } }
\cos \left( \Delta_{32} + \delta_{\textrm{CP}} \right) , \\
P \left( \bar{\nu}_{\mu} \rightarrow \bar{\nu}_{e} \right) \simeq
{\cal P}_{ \textrm{atm} } + P_{ \textrm{sol} }
+ 2 \sqrt{ {\cal P}_{ \textrm{atm} } } \sqrt{ P_{ \textrm{sol} } }
\cos \left( \Delta_{32} - \delta_{\textrm{CP}} \right) ,
\end{array}$$ where $$\begin{array}{l} \vspace{2mm}
\sqrt{ P_{ \textrm{sol} } } =
\cos \theta_{23} \sin 2 \theta_{12} \frac{ \sin a L }{ a L } \Delta_{21},
\\ \vspace{2mm}
\sqrt{ P_{ \textrm{atm} } } =
\sin \theta_{23} \sin 2 \theta_{13} \frac{ \sin \left( \Delta_{31} - a L \right) }{
\left( \Delta_{31} - a L \right) } \Delta_{31} , \\ \vspace{2mm}
\sqrt{ {\cal P}_{ \textrm{atm} } } =
\sin \theta_{23} \sin 2 \theta_{13} \frac{ \sin \left( \Delta_{31} + a L \right) }{
\left( \Delta_{31} + a L \right) } \Delta_{31} .
\end{array}$$ In the above expressions, $L$ is the Base-Line, $$\begin{array}{l}
\Delta_{ij} = \frac{ \Delta m_{ij}^{2} L }{ 4E }, \quad
\Delta m_{ij}^{2} = m_{i}^{2} - m_{j}^{2}
\quad \textrm{and} \quad
a = \frac{ G_{F} N_{e} }{ \sqrt{2}}.
\end{array}$$ Here, $E$ is the energy of neutrino beam, $G_{F}$ is the Fermi constant and $N_{e}$ is the density of electrons. The $a$ parameter is $a \approx ( 3500~\textrm{km})^{-1}$ for the Earth crust [@Nunokawa2008338] . The asymmetry between $P \left( \nu_{\mu} \rightarrow \nu_{e} \right)$ and $P \left( \bar{\nu}_{\mu} \rightarrow \bar{\nu}_{e} \right)$ in matter is [@Gonzalez-Canales:2016xia] $$\label{Eq:A_mu-e}
\begin{array}{rl}\vspace{2mm}
{\cal A}_{\mu e} &
=
\frac{
P \left( \nu_{\mu} \rightarrow \nu_{e} \right)
-
P \left( \bar{\nu}_{\mu} \rightarrow \bar{\nu}_{e} \right)
}{
P \left( \nu_{\mu} \rightarrow \nu_{e} \right)
+
P \left( \bar{\mu}_{\nu} \rightarrow \bar{\nu}_{e} \right)
} \\
&
=
\frac{
\left( P_{ \textrm{atm} } -{\cal P}_{ \textrm{atm} } \right)
+ 2 \sqrt{ P_{ \textrm{sol} } }
\left(
\sqrt{ P_{ \textrm{atm} } }
\cos \left( \Delta_{32} + \delta_{\textrm{CP}} \right)
-
\sqrt{ {\cal P}_{ \textrm{atm} } }
\cos \left( \Delta_{32} - \delta_{\textrm{CP}}
\right)
\right)
}{
\left( P_{ \textrm{atm} } + {\cal P}_{ \textrm{atm} } \right)
+ 2 \sqrt{ P_{ \textrm{sol} } }
\left(
\sqrt{ P_{ \textrm{atm} } }
\cos \left( \Delta_{32} + \delta_{\textrm{CP}} \right)
+
\sqrt{ {\cal P}_{ \textrm{atm} } }
\cos \left( \Delta_{32} - \delta_{\textrm{CP}}
\right)
\right)
+ 2 P_{ \textrm{sol} }
}.
\end{array}$$ The above asymmetry ${\cal A}_{\mu e}$ is basically due to the absence of positrons in the journey of neutrino (anti-neutrino) through the earth. Hence, a neutrino experiment with a LBL would be more sensitive to measure this asymmetry.
The T2K neutrino oscillation experiment has a LBL of 295 km, while the energy of its neutrino beam has a peak around to 0.6 GeV and width of $\sim$0.3 GeV [@Abe:2011ks]. In Fig. \[Fig:probabilitiesNH\] we show the transition probability $\nu_{\mu} ( \bar{\nu}_{\mu} ) \rightarrow \nu_{e} (\bar{\nu}_{e})$, as well as the asymmetry ${\cal A}_{\mu e}$ for T2K experiment.
The NO$\nu$A neutrino oscillation experiment has a LBL of 810 km, while the energy of its neutrino beam has a peak around to 2 GeV [@Adamson:2016tbq]. In Fig. \[Fig:probabilitiesNH\] we show the transition probability $\nu_{\mu} ( \bar{\nu}_{\mu} ) \rightarrow \nu_{e} (\bar{\nu}_{e})$, as well as the asymmetry ${\cal A}_{\mu e}$ for NO$\nu$A experiment.
Finally, the future neutrino oscillation experiment DUNE will have a LBL of $\sim1300$ km, while the energy of its neutrino beam will have a peak around to $2.5-3.0$ GeV [@Acciarri:2015uup]. In Fig. \[Fig:probabilitiesNH\] the transition probability $\nu_{\mu} ( \bar{\nu}_{\mu} ) \rightarrow \nu_{e} (\bar{\nu}_{e})$, and the asymmetry ${\cal A}_{\mu e}$ for DUNE experiment are shown.
Conclusions\[sec:conclusions\]
==============================
We have studied the theory of neutrino masses, mixings and CPV as the realization of an $S_{3}$ flavor symmetry in the framework of the Two Higgs Doublet Model type-III. In this $\nu$2HDM$\otimes S_{3}$ extension of Standard Model, on the one hand, the active neutrinos acquire their little masses via the type-I seesaw mechanism. On the other hand, the explicit sequential breaking of flavor symmetry according the chain $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} } \supset
S_{3}^{\mathrm{diag}} \supset S_{2}^{\mathrm{diag}}$, allow us to represent in the flavor basis the Yukawa matrices with an Hermitian matrix with two texture zeroes. Consequently, we obtained an unified treatment for all fermion mass matrices in the model, which are represented through of a matrix with two texture zeroes.
The unitary matrices that diagonalize the mass matrices are expressed in terms of fermion mass ratios. Then, the entries of the Yukawa matrices in the mass basis naturally acquire the form of the so-called [*Cheng-Sher ansatz*]{}. Also, the lepton flavor mixing matrix PMNS is expressed as function of the masses of charged leptons and neutrinos, two phases associated with the CP-violation, and two parameters associated with the flavor symmetry breaking. The unitary matrix that allows us to pass from the weak basis to the flavor symmetry adapted basis, is unobservable in the Higgs-fermions couplings and lepton flavor mixing matrix.
To validate our hypothesis where the $S_{3}$ horizontal flavor symmetry is explicitly breaking according to the chain $S_{ 3L }^{ \texttt{j} } \otimes S_{ 3R }^{ \texttt{j} }
\supset S_{3}^{\mathrm{diag}} \supset S_{2}^{\mathrm{diag}}$, all fermion mass matrices are represented through a matrix with two texture zeroes. Furthermore, we make a likelihood test where we compare the theoretical expressions of the flavor mixing angles with the current experimental data on masses and flavor mixing of leptons. The results obtained in this $\chi^{2}$ analysis are in very good agreement with the current experimental data.
We also obtained the following allowed value ranges, at BFP$\pm 1\sigma$ for the “Dirac-like” phase factor, as well as for the two Majorana phase factors: $$\begin{array}{c}
\delta_{CP} (^{\circ}) =
\left \{ \begin{array}{l}
-69.8_{-6.110}^{+5.508} ,\\
-80.83_{-0.709}^{+0.652} ,
\end{array} \right. \quad
\phi_{12} (^{\circ}) =
\left \{ \begin{array}{l}
-5.800_{-0.150}^{+0.170} ,\\
-5.24_{-0.148}^{+0.153} ,
\end{array} \right. \quad
\phi_{13} (^{\circ}) =
\left \{ \begin{array}{l}
14.744_{-1.366}^{+1.266} ,\\
-2.190_{-0.0005}^{+0.0030} .
\end{array} \right.
\end{array}$$ The upper (lower) row corresponds to the normal (inverted) hierarchy in the neutrino mass spectrum. These values of the phase factors are in agreement with a maximum CPV in the neutrino oscillation in matter. Finally, we also analyzed the phenomenological implications of the above numerical values of the CP-violation phases on the neutrinoless double beta decay, as well as for LBL neutrino oscillation experiments such as T2K, NO$\nu$A, and DUNE.
Three dimensional representation of $S_{3}$
===========================================
The permutations of symmetry group $S_{3}$ can be represented on the reducible triplet as [@Georgi:1999wka; @Ishimori:2012zz]: $$\label{eq:A-1}
\begin{array}{ccc} \vspace{2mm}
{\bf D}^{(3)} \left( E \right) =
\left( \begin{array}{ccc}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1
\end{array} \right), &
{\bf D}^{(3)} \left( A_{1} \right) =
\left( \begin{array}{ccc}
0 & 1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1
\end{array} \right), &
{\bf D}^{(3)} \left( A_{2} \right) =
\left( \begin{array}{ccc}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array} \right) , \\
{\bf D}^{(3)} \left( A_{3} \right) =
\left( \begin{array}{ccc}
1 & 0 & 0 \\
0 & 0 & 1 \\
0 & 1 & 0
\end{array} \right) , &
{\bf D}^{(3)} \left( A_{4} \right) =
\left( \begin{array}{ccc}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0
\end{array} \right) , &
{\bf D}^{(3)} \left( A_{5} \right) =
\left( \begin{array}{ccc}
0 & 1 & 0 \\
0 & 0 & 1 \\
1 & 0 & 0
\end{array} \right) .
\end{array}$$ In this representation the projection operators take the form: $$\label{eq:A-P1}
\begin{array}{ll}
\textrm{Symmetric singlet}, &
{\bf P}_{\bf 1 } =
\frac{1}{3}
\left( \begin{array}{ccc}
1 & 1 & 1 \\
1 & 1 & 1 \\
1 & 1 & 1
\end{array} \right) =
| v_{\bf 1} \rangle \langle v_{\bf 1} | . \\
\textrm{Anti-symmetric singlet}, &
{\bf P}_{\bf 1' } = 0. \\
\textrm{Doublet}, &
{\bf P}_{\bf 2 } =
\frac{1}{3}
\left( \begin{array}{ccc}
2 & -1 & -1 \\
-1 & 2 & -1 \\
-1 & -1 & 2
\end{array} \right) =
| v_{\bf 2A} \rangle \langle v_{\bf 2A} | +
| v_{\bf 2S} \rangle \langle v_{\bf 2S} |. \\
\end{array}$$ Here, the vector $| v_{\bf 1} \rangle = \frac{1}{\sqrt{3}} \left( 1, 1, 1 \right)^{\top}$ is associated with the symmetric singlet. In the projection operator ${\bf P}_{\bf 2 }$, we have the vectors $| v_{\bf 2A} \rangle = \frac{1}{ \sqrt{2} } \left( -1, 1, 0 \right)^{\top}$ and $| v_{\bf 2S} \rangle = \frac{1}{ \sqrt{2} } \left( 1, 1, -2 \right)^{\top}$, which are associated with the doublet.
Correspondingly, the vectors $| v_{\bf 2A} \rangle$ and $| v_{\bf 2S} \rangle$ are antisymmetric and symmetric, under the permutation of first two elements. With the previous three vectors we can construct some tensors that can be helpful. Then, $$\begin{array}{ccc}
| v_{\bf 2S} \rangle \langle v_{\bf 2A} | =
\frac{ 1 }{ \sqrt{12} }
\left( \begin{array}{ccc}
-1 & 1 & 0 \\
-1 & 1 & 0 \\
2 & -2 & 0
\end{array} \right) &
\textrm{and} &
| v_{\bf 2A} \rangle \langle v_{\bf 2S} | =
\frac{ 1 }{ \sqrt{12} }
\left( \begin{array}{ccc}
-1 & -1 & 2 \\
1 & 1 & -2 \\
0 & 0 & 0
\end{array} \right).
\end{array}$$ If we define the tensors ${\bf T}_{x}^{+} = | v_{\bf 2S} \rangle \langle v_{\bf 2A} | + | v_{\bf 2A} \rangle
\langle v_{\bf 2S} |$ and ${\bf T}_{x}^{-} = i \left( | v_{\bf 2A} \rangle \langle v_{\bf 2S} | -
| v_{\bf 2S} \rangle \langle v_{\bf 2A} | \right) $, we obtain $$\label{eq:Ty+-}
\begin{array}{ccc}
{\bf T}_{x}^{+} = \frac{1}{\sqrt{3}}
\left( \begin{array}{ccc}
-1 & 0 & 1 \\
0 & 1 & -1 \\
1 & -1 & 0
\end{array} \right) &
\textrm{and} &
{\bf T}_{x}^{-} = \frac{1}{\sqrt{3}}
\left( \begin{array}{ccc}
0 & i & -i \\
-i & 0 & i \\
i & -i & 0
\end{array} \right).
\end{array}$$ The terms proportional to tensors ${\bf T}_{x}^{+}$ and ${\bf T}_{x}^{-}$ mix the components of the doublet representation each other. Now, $$\begin{array}{ccc}
| v_{\bf 1} \rangle \langle v_{\bf 2A} | =
\frac{ 1 }{ \sqrt{6} }
\left( \begin{array}{ccc}
-1 & 1 & 0 \\
-1 & 1 & 0 \\
-1 & 1 & 0
\end{array} \right) &
\textrm{and} &
| v_{\bf 2A} \rangle \langle v_{\bf 1} | =
\frac{ 1 }{ \sqrt{6} }
\left( \begin{array}{ccc}
-1 & -1 & -1 \\
1 & 1 & 1 \\
0 & 0 & 0
\end{array} \right).
\end{array}$$ If we define the tensors ${\bf T}_{y}^{+} = | v_{\bf 1} \rangle \langle v_{\bf 2A} | + | v_{\bf 2A} \rangle
\langle v_{\bf 1} |$ and ${\bf T}_{y}^{-} = i \left( | v_{\bf 2A} \rangle \langle v_{\bf 2A} | -
| v_{\bf 2A} \rangle \langle v_{\bf 2S} | \right) $, we obtain $$\begin{array}{ccc}
{\bf T}_{y}^{+} = \frac{1}{\sqrt{6}}
\left( \begin{array}{ccc}
-2 & 0 & -1 \\
0 & 2 & 1 \\
-1 & 1 & 0
\end{array} \right) &
\textrm{and} &
{\bf T}_{y}^{-} = \frac{1}{\sqrt{6}}
\left( \begin{array}{ccc}
0 & 2i & i \\
-2i & 0 & -i \\
-i & i & 0
\end{array} \right).
\end{array}$$ The terms proportional to tensors ${\bf T}_{y}^{+}$ and ${\bf T}_{y}^{-}$ mix the antisymmetric component of doublet with the singlet. Finally, $$\begin{array}{ccc}
| v_{\bf 1} \rangle \langle v_{\bf 2S} | =
\frac{ 1 }{ 3\sqrt{2} }
\left( \begin{array}{ccc}
1 & 1 & -2 \\
1 & 1 & -2 \\
1 & 1 & -2
\end{array} \right) &
\textrm{and} &
| v_{\bf 2S} \rangle \langle v_{\bf 1} | =
\frac{ 1 }{ 3\sqrt{2} }
\left( \begin{array}{ccc}
1 & 1 & 1 \\
1 & 1 & 1 \\
-2 & -2 & -2
\end{array} \right) .
\end{array}$$ If we define the tensors ${\bf T}_{z}^{+} = | v_{\bf 1} \rangle \langle v_{\bf 2S} | + | v_{\bf 2S} \rangle
\langle v_{\bf 1} |$ and ${\bf T}_{y}^{-} = i \left( | v_{\bf 2A} \rangle \langle v_{\bf 2A} | -
| v_{\bf 2A} \rangle \langle v_{\bf 2S} | \right) $, then $$\begin{array}{ccc}
{\bf T}_{z}^{+} = \frac{1}{3 \sqrt{2}}
\left( \begin{array}{ccc}
2 & 2 & -1 \\
2 & 2 & -1 \\
-1 & -1 & -4
\end{array} \right) &
\textrm{and} &
{\bf T}_{z}^{-} = \frac{1}{\sqrt{2}}
\left( \begin{array}{ccc}
0 & 0 & -i \\
0 & 0 & -i \\
i & i & 0
\end{array} \right).
\end{array}$$ The terms proportional to tensors ${\bf T}_{z}^{+}$ and ${\bf T}_{z}^{-}$ mix the symmetric component of doublet with the singlet. The tensor ${\bf T}_{z}^{+}$ can be written as a linear combination of two independent matrices, $${\bf T}_{z}^{+} =
\sqrt{ \frac{2}{3} } {\bf T}_{z1}^{+} - \frac{1}{3 \sqrt{2}} {\bf T}_{z2}^{+},$$ where $$\label{eq:Tz12+}
\begin{array}{ccc}
{\bf T}_{z1}^{+} =
\left( \begin{array}{ccc}
1 & 1 & 0 \\
1 & 1 & 0 \\
0 & 0 & -2
\end{array} \right) &
\textrm{and} &
{\bf T}_{z2}^{+} =
\left( \begin{array}{ccc}
0 & 0 & 1 \\
0 & 0 & 1 \\
1 & 1 & 0
\end{array} \right).
\end{array}$$
Cheng-Sher Parameters {#App:B}
=====================
$$\begin{array}{l}\vspace{2mm}
\widetilde{\bf Y}_{k}^{ \texttt{j} } =
{\bf U}_{ \texttt{j} }^{\dagger}
{\bf Y}_{k}^{ \mathrm{w}, \texttt{j} } {\bf U}_{ \texttt{j} } =
{\bf O}_{ \texttt{j} }^{ \top } {\bf P}_{ \texttt{j} }^{ \dagger }
{\bf U}_{ \mathrm{s} }^{\dagger}
{\bf Y}_{k}^{ \mathrm{w}, \texttt{j} }
{\bf U}_{ \mathrm{s} } {\bf P}_{ \texttt{j} } {\bf O}_{ \texttt{j} } , \\
\left( \widetilde{\bf Y}_{k}^{ \texttt{j} } \right)_{ \texttt{rt} } =
\dfrac{ \sqrt{ m_{ \texttt{jr} } m_{ \texttt{js} } } }{ v }
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ \texttt{rt} } ,
\quad \texttt{r,t} = 1,2,3
\end{array}$$
where $$\begin{array}{ll}\vspace{2mm}
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ 11 } = &
2 \frac{
\xi_{ \texttt{J} 1}
}{
D_{\texttt{j}1 }
}
\sqrt{ \frac{
\widehat{m}_{ \texttt{j} 2 }
}{
\widehat{m}_{ \texttt{j} 1 }
}
\left( 1 - \delta_{ \texttt{j} } \right)
}
\cos \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right)
\widetilde{a}_{k}^{ \texttt{j} }
+
\frac{
\left( 1 - \delta_{\texttt{j}} \right) \xi_{ \texttt{j} 1}
}{
D_{ \texttt{j} 1}
}
\widetilde{b}_{k}^{ \texttt{j} }
-
2 \frac{
\sqrt{ \delta_{ \texttt{j} } \left( 1 - \delta_{ \texttt{j} } \right)
\xi_{ \texttt{j} 1 } \xi_{ \texttt{j} 2 } }
}{
D_{ \texttt{j} 1 }
}
\widetilde{c}_{k}^{ \texttt{j} }
+
\frac{
\delta_{ \texttt{j} } \xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 1 }
}
\widetilde{d}_{k}^{ \texttt{j} } , \\ \vspace{2mm}
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ 12 } = &
\sqrt{
\frac{
\left( 1 - \delta_{ \texttt{j} } \right) \xi_{ \texttt{j}1 }
\xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 1 } D_{ \texttt{j} 2 }
}
}
\left(
\sqrt{
\frac{
\widehat{m}_{ \texttt{j} 2 }
}{
\widehat{m}_{ \texttt{j} 1 }
}
}
e^{ i \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right) }
-
\sqrt{
\frac{
\widehat{m}_{ \texttt{j} 1 }
}{
\widehat{m}_{ \texttt{j} 2 }
}
}
e^{-i \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right) }
\right) \widetilde{a}_{k}^{ \texttt{j} }
+
\left( 1 - \delta{ \texttt{j} } \right)
\sqrt{
\frac{
\xi_{ \texttt{j} 1 } \xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 1 } D_{ \texttt{j} 2 }
}
}
\widetilde{b}_{k}^{ \texttt{j} } \\ \vspace{2mm}
&
-
\left( \xi_{ \texttt{j}1 } + \xi_{ \texttt{j} 2 } \right)
\sqrt{
\frac{
\delta_{ \texttt{j} } \left( 1 - \delta_{ \texttt{j} } \right)
}{
D_{ \texttt{j} 1 } D_{ \texttt{j} 2 }
}
} \widetilde{c}_{k}^{ \texttt{j} }
+
\delta_{ \texttt{j} }
\sqrt{
\frac{
\xi_{ \texttt{j} 1 } \xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 1 } D_{ \texttt{j} 2}
} }
\widetilde{d}_{k}^{ \texttt{j} }, \\ \vspace{2mm}
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ 13 } = &
\sqrt{
\frac{ \widehat{m}_{ \texttt{j} 2 } }{ \widehat{m}_{ \texttt{j} 1 } }
\frac{
\left( 1- \delta_{ \texttt{j} } \right) \delta_{ \texttt{j} } \xi _{ \texttt{j} 1 }
}{
D_{ \text{j} 1 } D_{ \texttt{j} 3 }
}
}
\left(
\widehat{m}_{ \texttt{j} 1}
e^{ - i \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right) }
+ e^{ i \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right) }
\right)
\widetilde{a}_{k}^{ \texttt{j} }
+
\left( 1 - \delta_{ \texttt{j} } \right)
\sqrt{
\frac{
\delta_{ \texttt{j} } \xi_{ \texttt{j} 1 }
}{
D_{ \texttt{j} 1 } D_{ \texttt{j} 3 }
}
} \widetilde{b}_{k}^{ \texttt{j} } \\ \vspace{2mm}
& +
\left( \xi _{ \texttt{j} 1 } - \delta_{ \texttt{j} } \right)
\sqrt{
\frac{
\left( 1 - \delta_{ \texttt{j} } \right) \xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 1} D_{ \texttt{j} 3 }
}
}
\widetilde{c}_{k}^{ \texttt{j} }
-
\xi _{ \texttt{j} 2 }
\sqrt{
\frac{
\delta_{\texttt{j}} \xi_{ \texttt{j} 1}
}{
D_{ \texttt{j} 1 } D_{ \texttt{j} 3 }
}
}
\widetilde{d}_{k}^{ \texttt{j} } , \\ \vspace{2mm}
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ 22 } = &
- 2
\sqrt{
\frac{
\widehat{m}_{ \texttt{j} 1 }
}{
\hat{m}_{ \texttt{j} 2 }
}
\left( 1 - \delta_{ \texttt{j} } \right)
}
\frac{
\xi_{ \texttt{j} 2 }
}{
D_{\texttt{j}2 }
}
\cos \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right)
\widetilde{a}_{k}^{ \texttt{j} }
+
\frac{
\left( 1 - \delta_{ \texttt{j} } \right) \xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 2}
}
\widetilde{b}_{k}^{ \texttt{j} }
- 2
\frac{
\sqrt{
\delta_{ \texttt{j} } \left( 1 - \delta_{ \texttt{j} } \right)
\xi_{ \texttt{j} 1 } \xi_{ \texttt{j} 2 }
}
}{
D_{ \texttt{j} 2 }
}
\widetilde{c}_{k}^{ \texttt{j} }
+
\frac{
\delta _j \xi_{ \texttt{j} }
}{
D_{ \texttt{j} 2 }
}
\widetilde{d}_{k}^{ \texttt{j} } , \\ \vspace{2mm}
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ 23 } = &
- \sqrt{
\frac{ \widehat{m}_{ \texttt{j} 1 } }{ \widehat{m}_{ \texttt{j} 2} }
\frac{ \delta_{ \texttt{j} } \left( 1 - \delta_{ \texttt{j} } \right)
\xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 2 } D_{ \texttt{j} 3 }
}
}
\left(
e^{ i \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right) }
- \widehat{m}_{ \texttt{j} 2 }
e^{ - i \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right) }
\right)
\widetilde{a}_{k}^{ \texttt{j} }
+
\left( 1 - \delta_{ \texttt{j} } \right)
\sqrt{
\frac{
\delta_{ \texttt{j} } \xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 2 } D_{ \texttt{j} 3 }
}
}
\widetilde{b}_{k}^{ \texttt{j} } \\ \vspace{2mm}
&
+
\left( \xi_{ \texttt{j} 2 } - \delta_{ \texttt{j} } \right)
\sqrt{
\frac{
\left( 1 - \delta_{ \texttt{j} } \right)
\xi_{ \texttt{j} 1 }
}{
D_{ \texttt{j} 2 } D_{ \texttt{j} 3 }
}
}
\widetilde{c}_{k}^{ \texttt{j} }
-
\xi_{ \texttt{j} 1 }
\sqrt{
\frac{
\delta_{ \texttt{j} } \xi_{ \texttt{j} 2 }
}{
D_{ \texttt{j} 2 } D_{ \texttt{j} 3 }
}
}
\widetilde{d}_{k}^{ \texttt{j} } , \\ \vspace{2mm}
\left( \widetilde{\bf \chi}_{k}^{ \texttt{j} } \right)_{ 33 } = &
2
\sqrt{
\left( 1 - \delta_{ \texttt{j} } \right)
\widehat{m}_{ \texttt{j} 1 } \widehat{m}_{ \texttt{j} 2 }
}
\frac{ \delta_{ \texttt{j} } }{ D_{ \texttt{j} 3 } }
\cos \left( \phi_{k}^{ \texttt{j} } - \phi_{ \texttt{j} } \right)
\widetilde{a}_{k}^{ \texttt{j} }
+
\frac{
\delta_{ \texttt{j} } \left( 1 - \delta_{ \texttt{j} } \right)
}{
D_{ \texttt{j} 3 }
}
\widetilde{b}_{k}^{ \texttt{j} }
+
2
\frac{
\sqrt{
\delta_{ \texttt{j} } \left( 1 - \delta_{ \texttt{j} } \right)
\xi_{ \texttt{j} 1 } \xi_{ \texttt{j} 2 }
}
}{
D_{ \texttt{j} 3 }
}
\widetilde{c}_{k}^{ \texttt{j} }
+
\frac{d_{\text{jk}}
\xi _{\text{j1}} \xi _{\text{j2}}}{D_{\text{j3}}}
\widetilde{d}_{k}^{ \texttt{j} } ,
\\ \vspace{2mm}
\end{array}$$ with $$\begin{array}{l}
\widetilde{a}_{k}^{ \texttt{j} } = \frac{ v }{ m_{ \texttt{j} 3 } }
\left | A_{k}^{ \texttt{j} } \right|, \quad
\widetilde{b}_{k}^{ \texttt{j} } = \frac{ v }{ m_{ \texttt{j} 3 } }
B_{k}^{ \texttt{j} }, \quad
\widetilde{c}_{k}^{ \texttt{j} } = \frac{ v }{ m_{ \texttt{j} 3 } }
C_{k}^{ \texttt{j} },
\quad \textrm{and} \quad
\widetilde{d}_{k}^{ \texttt{j} } = \frac{ v }{ m_{ \texttt{j} 3 } }
D_{k}^{ \texttt{j} }.
\end{array}$$
Mixing Matrix {#App:C}
=============
The lepton flavor mixing matrix is $$\begin{array}{l}
{\bf U}_{\mathrm{PMNS}} = {\bf U}_{l}^{\dagger} {\bf U}_{\nu} =
{\bf O}_{l}^{\top}
{\bf P}_{l}^{ \dagger }
{\bf U}_{\mathrm{s}}^{ \dagger }
{\bf U}_{\mathrm{s}} {\bf P}_{\nu}
{\bf O}_{\nu}^{\mathrm{n[i]}}
= {\bf O}_{l}^{\top} {\bf P}^{ ( \nu - l ) } {\bf O}_{\nu}^{\mathrm{n[i]}}.
\end{array}$$ The explicit form of entries of previous matrix are: $$\begin{array}{l} \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{11} =
\sqrt{
\frac{
\hat{m}_{\mu} \hat{m}_{\nu 2[1]} \xi_{l1} \xi _{\nu 1[3]}
}{
D_{l1} D_{ \nu 1[3] }
}
}
+
\sqrt{
\frac{
\hat{m}_{e} \hat{m}_{\nu 1[3]}
}{
D_{l1} D_{ \nu 1[3] }
}
}
\left(
\sqrt{
\left( 1 - \delta_{\nu} \right) \left( 1 - \delta_{l} \right)
\xi_{l1} \xi _{\nu 1[3]} } e^{ i \phi_{ l1 } }
+
\sqrt{
\delta_{\nu} \delta_{l} \xi_{l2} \xi_{\nu 2[1] } }
e^{ i \phi_{ l2 } }
\right) , \\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{12} =
- \sqrt{
\frac{
\hat{m}_{\mu} \hat{m}_{\nu 1[3]} \xi_{l1} \xi _{\nu 2[1] }
}{
D_{l1} D_{\nu 2[1]}
}
}
+
\sqrt{
\frac{
\hat{m}_{e} \hat{m}_{\nu 2[1]}
}{
D_{l1} D_{\nu 2[1]}
}
}
\left(
\sqrt{
\left( 1 - \delta_{\nu} \right) \left( 1 - \delta_{l} \right) \xi_{l1}
\xi _{\nu 2[1] } } e^{ i \phi_{ l1 } }
+
\sqrt{
\delta_{\nu} \delta_{l} \xi_{l2} \xi _{\nu 1[3]} } e^{ i \phi_{ l2 } }
\right) , \\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{13} =
\sqrt{
\frac{
\widehat{m}_{\mu} \widehat{m}_{\nu 1[3]} \widehat{m}_{\nu 2[1]}
\delta_{\nu} \xi_{l1}
}{
D_{l1} D_{\nu 3[2]}
}
}
+
\sqrt{
\frac{
\widehat{m}_{e}
}{
D_{l1} D_{\nu 3[2]}
}
}
\left(
\sqrt{
\left( 1 -\delta_{\nu} \right) \delta_{\nu} \left( 1 - \delta_{l} \right)
\xi_{l1} } e^{ i \phi_{ l1 } }
-
\sqrt{ \delta_{l} \xi_{l2} \xi_{\nu 1[3]} \xi _{\nu 2[1] } }
e^{ i \phi_{ l2 } }
\right) , \\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{21} =
- \sqrt{
\frac{
\widehat{m}_{e} \widehat{m}_{\nu 2[1]} \xi_{l2} \xi _{\nu 1[3]}
}{
D_{l2} D_{ \nu 1[3] }
}
}
+
\sqrt{
\frac{
\widehat{m}_{\mu} \widehat{m}_{\nu 1[3]}
}{
D_{l2} D_{ \nu 1[3] }
}
}
\left(
\sqrt{ \left( 1 - \delta_{\nu} \right) \left( 1 - \delta_{l} \right) \xi_{l2}
\xi _{\nu 1[3]} } e^{ i \phi_{l1} }
+
\sqrt{ \delta_{\nu} \delta_{l} \xi_{l1} \xi _{\nu 2[1] } }
e^{ i \phi_{ l2 } }
\right), \\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{22} =
\sqrt{
\frac{
\widetilde{m}_{e} \widetilde{m}_{\nu 1 [3]}
\xi_{l2} \xi_{\nu 2 [1]}
}{
D_{l2} D_{ \nu 2[1] }
}
}
+
\sqrt{
\frac{
\widetilde{m}_{\mu} \widetilde{m}_{\nu 2 [1]}
}{
D_{l2} \, D_{\nu 2 [1]}
}
}
\left(
\sqrt{ \left( 1 - \delta_{\nu} \right) \left( 1 - \delta_{l} \right)
\xi_{l2} \xi_{\nu 2 [1]} } e^{i \phi_{l1} }
+
\sqrt{ \delta_{\nu} \delta_{l} \xi_{l1} \xi_{\nu 1 [3]} }
e^{i \phi_{ l2 } }
\right), \\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{23} =
- \sqrt{
\frac{
\widetilde{m}_{e} \widetilde{m}_{\nu 1[3]} \widetilde{m}_{\nu 2[1]} \delta_{\nu}
\xi_{l2}
}{
D_{l2} D_{\nu 3[2]}
}
}
+
\sqrt{
\frac{
\widetilde{m}_{\mu}
}{
D_{l2} D_{\nu 3[2]}
}
}
\left(
\sqrt{ \delta_{\nu} \left( 1 - \delta_{\nu } \right)
\left( 1 - \delta_{l} \right) \xi_{l2} } e^{i \phi_{l1} }
-
\sqrt{ \delta_{l} \xi_{l1} \xi_{\nu 1[3]} \xi_{\nu 2[1]} }
e^{i \phi_{ l2 } }
\right) , \\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{31} =
\sqrt{
\frac{
\widetilde{m}_{e} \widetilde{m}_{\mu} \widetilde{m}_{\nu 2[1]} \delta_{l}
\xi_{\nu 1[3]}
}{
D_{l3} D_{\nu 1[3]}
}
}
+
\sqrt{
\frac{
\widetilde{m}_{\nu 1[3]}
}{
D_{l3} D_{\nu 1[3]}
}
}
\left(
\sqrt{ \delta_{l} \left( 1 - \delta_{\nu } \right)
\left( 1 - \delta_{l} \right) \xi_{\nu 1[3]} } e^{i \phi_{l1} }
-
\sqrt{ \delta_{\nu} \, \xi_{l1} \, \xi_{l2} \, \xi_{\nu 2[1]} }
e^{i \phi_{ l2 } }
\right) ,\\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{32} =
-
\sqrt{
\frac{
\widetilde{m}_{e} \widetilde{m}_{\mu} \widetilde{m}_{\nu 1[3]} \delta_{l}
\xi_{\nu 2[1]}
}{
D_{l3} D_{\nu 2[1]}
}
}
+
\sqrt{
\frac{
\widetilde{m}_{\nu 2[1]}
}{
D_{l3} D_{\nu 2[1]}
}
}
\left(
\sqrt{ \delta_{l} \left( 1 - \delta_{\nu } \right)
\left( 1 - \delta_{l} \right) \xi_{\nu 2[1]} } e^{i \phi_{l1} }
-
\sqrt{ \delta_{\nu} \xi_{l1} \xi_{l2} \xi_{\nu 1[3]} }
e^{i \phi_{ l2 } }
\right) , \\ \vspace{2mm}
\left( {\bf U}_{\mathrm{PMNS}} \right)_{33} =
\sqrt{
\frac{
\widetilde{m}_{e} \widetilde{m}_{\mu} \widetilde{m}_{\nu 1[3]}
\widetilde{m}_{\nu 2[1]} \delta_{l} \delta_{\nu}
}{
D_{l3} \, D_{\nu 3[2]}
}
}
+
\sqrt{ \frac{ 1 }{ D_{l3} \, D_{\nu 3[2]} } }
\left(
\sqrt{ \delta_{l} \delta_{\nu } \left( 1 - \delta_{\nu } \right)
\left( 1 - \delta_{l} \right) } e^{i \phi_{l1} }
- \sqrt{ \xi_{l1} \xi_{l2} \xi_{\nu 1[3]} \xi_{\nu 2[1]} }
e^{i \phi_{ l2 } }
\right) .
\end{array}$$
This work has been partially supported by *CONACYT-SNI (México)*. The authors thankfully acknowledge the computer resources, technical expertise and support provided by the Laboratorio Nacional de Supercómputo del Sureste de México CONACYT network of national laboratories." FGC acknowledges the financial support from [*CONACYT*]{} and [*PRODEP*]{} under Grant No. 511-6/17-8017.
[10]{} \[1\][`#1`]{} \[2\]\[\][[\#2](#2)]{}
S. F. King, *[Unified Models of Neutrinos, Flavour and CP Violation]{}*, [[](http://dx.doi.org/10.1016/j.ppnp.2017.01.003)]{}[[](http://dx.doi.org/10.1016/j.ppnp.2017.01.003)]{}, [[](http://arxiv.org/abs/1701.04413)]{}.
F. Capozzi et al., *[Neutrino masses and mixings: Status of known and unknown $3\nu$ parameters]{}*, [[](http://dx.doi.org/10.1016/j.nuclphysb.2016.02.016)]{}[[](http://dx.doi.org/10.1016/j.nuclphysb.2016.02.016)]{}, [[](http://arxiv.org/abs/1601.07777)]{}.
M. P. Decowski (KamLAND), *[KamLAND’s precision neutrino oscillation measurements]{}*, [[](http://dx.doi.org/10.1016/j.nuclphysb.2016.04.014)]{}[[](http://dx.doi.org/10.1016/j.nuclphysb.2016.04.014)]{}.
A. Gando et al. (KamLAND), *[Constraints on $\theta_{13}$ from A Three-Flavor Oscillation Analysis of Reactor Antineutrinos at KamLAND]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.83.052002)]{}[[](http://dx.doi.org/10.1103/PhysRevD.83.052002)]{}, [[](http://arxiv.org/abs/1009.4771)]{}.
A. Gando et al. (KamLAND), *[Reactor On-Off Antineutrino Measurement with KamLAND]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.88.033001)]{}[[](http://dx.doi.org/10.1103/PhysRevD.88.033001)]{}, [[](http://arxiv.org/abs/1303.4667)]{}.
S.-H. Seo (RENO), *[New Results from RENO and The 5 MeV Excess]{}*, [[](http://dx.doi.org/10.1063/1.4915563)]{}[[](http://dx.doi.org/10.1063/1.4915563)]{}, [[](http://arxiv.org/abs/1410.7987)]{}.
J. H. Choi et al. (RENO), *[Observation of Energy and Baseline Dependent Reactor Antineutrino Disappearance in the RENO Experiment]{}*, [[](http://dx.doi.org/10.1103/PhysRevLett.116.211801)]{}[[](http://dx.doi.org/10.1103/PhysRevLett.116.211801)]{}, [[](http://arxiv.org/abs/1511.05849)]{}.
Y. Abe et al. (Double Chooz), *[Improved measurements of the neutrino mixing angle $\theta_{13}$ with the Double Chooz detector]{}*, [[](http://dx.doi.org/10.1007/JHEP02(2015)074,
10.1007/JHEP10(2014)086)]{}[[](http://dx.doi.org/10.1007/JHEP02(2015)074,
10.1007/JHEP10(2014)086)]{}, \[Erratum: JHEP02,074(2015)\], [[](http://arxiv.org/abs/1406.7763)]{}.
F. P. An et al. (Daya Bay), *[Measurement of the Reactor Antineutrino Flux and Spectrum at Daya Bay]{}*, [[](http://dx.doi.org/10.1103/PhysRevLett.116.061801,
10.1103/PhysRevLett.118.099902)]{}[[](http://dx.doi.org/10.1103/PhysRevLett.116.061801,
10.1103/PhysRevLett.118.099902)]{}, \[Erratum: Phys. Rev. Lett.118,no.9,099902(2017)\], [[](http://arxiv.org/abs/1508.04233)]{}.
G. J. Feldman, J. Hartnell and T. Kobayashi, *[Long-baseline neutrino oscillation experiments]{}*, [[](http://dx.doi.org/10.1155/2013/475749)]{}[[](http://dx.doi.org/10.1155/2013/475749)]{}, [[](http://arxiv.org/abs/1210.1778)]{}.
K. Abe et al. (T2K), *[Measurement of neutrino and antineutrino oscillations by the T2K experiment including a new additional sample of $\nu_e$ interactions at the far detector]{}* (2017), [[](http://arxiv.org/abs/1707.01048)]{}.
M. Batkiewicz (T2K), *[Latest Results from T2K]{}*, in *[Proceedings, 17th Lomonosov Conference on Elementary Particle Physics: Moscow, Russia, August 20-26, 2015]{}* (2017) pages 66–72, [[](http://arxiv.org/abs/1705.04277)]{}, <http://inspirehep.net/record/1599081/files/arXiv:1705.04277.pdf>.
K. Abe et al. (T2K), *[Combined Analysis of Neutrino and Antineutrino Oscillations at T2K]{}*, [[](http://dx.doi.org/10.1103/PhysRevLett.118.151801)]{}[[](http://dx.doi.org/10.1103/PhysRevLett.118.151801)]{}, [[](http://arxiv.org/abs/1701.00432)]{}.
K. Abe et al. (T2K), *[Measurements of neutrino oscillation in appearance and disappearance channels by the T2K experiment with 6.6×10$^{20}$ protons on target]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.91.072010)]{}[[](http://dx.doi.org/10.1103/PhysRevD.91.072010)]{}, [[](http://arxiv.org/abs/1502.01550)]{}.
L. Kolupaeva (NO$\nu$A), *[Current results of the NO$\nu$A experiment]{}*, [[](http://dx.doi.org/10.1051/epjconf/201612501002)]{}[[](http://dx.doi.org/10.1051/epjconf/201612501002)]{}.
J. Bian (NO$\nu$A), *[First Results of $\nu_e$ Appearance Analysis and Electron Neutrino Identification at NO$\nu$A]{}*, in *[Meeting of the APS Division of Particles and Fields (DPF 2015) Ann Arbor, Michigan, USA, August 4-8, 2015]{}* (2015) [[](http://arxiv.org/abs/1510.05708)]{}, <http://inspirehep.net/record/1399048/files/arXiv:1510.05708.pdf>.
W. Seidel, *Talk on behalf of the cresst collaboration at idm montpellier.* (2010), <http://indico.in2p3.fr/conferenceDisplay.py?confId=1565/>.
D. Forero, M. Tortola and J. Valle, *[Neutrino oscillations refitted]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.90.093006)]{}[[](http://dx.doi.org/10.1103/PhysRevD.90.093006)]{}, [[](http://arxiv.org/abs/1405.7540)]{}.
M. C. Gonzalez-Garcia, M. Maltoni and T. Schwetz, *[Global Analyses of Neutrino Oscillation Experiments]{}*, [[](http://dx.doi.org/10.1016/j.nuclphysb.2016.02.033)]{}[[](http://dx.doi.org/10.1016/j.nuclphysb.2016.02.033)]{}, [[](http://arxiv.org/abs/1512.06856)]{}.
I. Esteban et al., *[Updated fit to three neutrino mixing: exploring the accelerator-reactor complementarity]{}*, [[](http://dx.doi.org/10.1007/JHEP01(2017)087)]{}[[](http://dx.doi.org/10.1007/JHEP01(2017)087)]{}, [[](http://arxiv.org/abs/1611.01514)]{}.
S.-F. Ge, *[Measuring the Leptonic Dirac CP Phase with TNT2K]{}*, in *[Prospects in Neutrino Physics (NuPhys2016) London, London, United Kingdom, December 12-14, 2016]{}* (2017) [[](http://arxiv.org/abs/1704.08518)]{}, <http://inspirehep.net/record/1597125/files/arXiv:1704.08518.pdf>.
D. Atwood, L. Reina and A. Soni, *[Phenomenology of two Higgs doublet models with flavor changing neutral currents]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.55.3156)]{}[[](http://dx.doi.org/10.1103/PhysRevD.55.3156)]{}, [[](http://arxiv.org/abs/hep-ph/9609279)]{}.
J. L. Diaz-Cruz et al., *[Yukawa Textures and Charged Higgs Boson Phenomenology in the 2HDM-III]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.79.095025)]{}[[](http://dx.doi.org/10.1103/PhysRevD.79.095025)]{}, [[](http://arxiv.org/abs/0902.4490)]{}.
J. L. Diaz-Cruz, R. Noriega-Papaqui and A. Rosado, *[Measuring the fermionic couplings of the Higgs boson at future colliders as a probe of a non-minimal flavor structure]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.71.015014)]{}[[](http://dx.doi.org/10.1103/PhysRevD.71.015014)]{}, [[](http://arxiv.org/abs/hep-ph/0410391)]{}.
M. Krawczyk and D. Sokolowska, *[The Charged Higgs boson mass in the 2HDM: Decoupling and CP violation]{}*, eConf **C0705302** (2007) HIG09, \[,141(2007)\], [[](http://arxiv.org/abs/0711.4900)]{}.
J. E. [Barradas Guevara]{} et al., *[Implications of Yukawa Textures in the decay $H^+ \to W^+ \gamma$ within the 2HDM-III]{}*, [[](http://dx.doi.org/10.1088/0954-3899/37/11/115008)]{}[[](http://dx.doi.org/10.1088/0954-3899/37/11/115008)]{}, [[](http://arxiv.org/abs/1002.2626)]{}.
A. Crivellin, A. Kokulu and C. Greub, *[Flavor-phenomenology of two-Higgs-doublet models with generic Yukawa structure]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.87.094031)]{}[[](http://dx.doi.org/10.1103/PhysRevD.87.094031)]{}, [[](http://arxiv.org/abs/1303.5877)]{}.
M. Krawczyk, *[Testing Higgs sector of 2HDM]{}*, PoS **HEP2005** (2006) 335, [[](http://arxiv.org/abs/hep-ph/0512371)]{}.
G. C. Branco et al., *[Theory and phenomenology of two-Higgs-doublet models]{}*, [[](http://dx.doi.org/10.1016/j.physrep.2012.02.002)]{}[[](http://dx.doi.org/10.1016/j.physrep.2012.02.002)]{}, [[](http://arxiv.org/abs/1106.0034)]{}.
F. F. Deppisch, *[Lepton Flavour Violation and Flavour Symmetries]{}*, [[](http://dx.doi.org/10.1002/prop.201200126)]{}[[](http://dx.doi.org/10.1002/prop.201200126)]{}, [[](http://arxiv.org/abs/1206.5212)]{}.
I. Dorsner and S. M. Barr, *[Flavor exchange effects in models with Abelian flavor symmetry]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.65.095004)]{}[[](http://dx.doi.org/10.1103/PhysRevD.65.095004)]{}, [[](http://arxiv.org/abs/hep-ph/0201207)]{}.
H. Georgi, *[Lie algebras in particle physics]{}*, Front. Phys. **54** (1999) 1–320.
H. Ishimori et al., *[An introduction to non-Abelian discrete symmetries for particle physicists]{}*, [[](http://dx.doi.org/10.1007/978-3-642-30805-5)]{}[[](http://dx.doi.org/10.1007/978-3-642-30805-5)]{}.
C. Patrignani et al. (Particle Data Group), *[Review of Particle Physics]{}*, [[](http://dx.doi.org/10.1088/1674-1137/40/10/100001)]{}[[](http://dx.doi.org/10.1088/1674-1137/40/10/100001)]{}.
Z.-z. Xing, *[Neutrino masses and flavor mixing]{}*, [[](http://dx.doi.org/10.1016/j.nuclphysbps.2010.08.006)]{}[[](http://dx.doi.org/10.1016/j.nuclphysbps.2010.08.006)]{}.
A. Das and N. Okada, *[Inverse seesaw neutrino signatures at the LHC and ILC]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.88.113001)]{}[[](http://dx.doi.org/10.1103/PhysRevD.88.113001)]{}, [[](http://arxiv.org/abs/1207.3734)]{}.
A. Das and N. Okada, *[Bounds on heavy Majorana neutrinos in type-I seesaw and implications for collider searches]{}* (2017), [[](http://arxiv.org/abs/1702.04668)]{}.
E. Barradas-Guevara et al., *[Analysis of the Lepton Mixing Matrix in the Two Higgs Doublet Model]{}*, [[](http://dx.doi.org/10.15415/jnp.2016.41021)]{}[[](http://dx.doi.org/10.15415/jnp.2016.41021)]{}, [[](http://arxiv.org/abs/1606.05388)]{}.
W. Wang and Z.-L. Han, *[Global $U(1)_{L}$ Breaking in Neutrinophilic 2HDM: From LHC Signatures to X-Ray Line]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.94.053015)]{}[[](http://dx.doi.org/10.1103/PhysRevD.94.053015)]{}, [[](http://arxiv.org/abs/1605.00239)]{}.
O. Felix-Beltran et al., *[Analysis of the quark sector in the 2HDM with a four-zero Yukawa texture using the most recent data on the CKM matrix]{}*, [[](http://dx.doi.org/10.1016/j.physletb.2015.02.003)]{}[[](http://dx.doi.org/10.1016/j.physletb.2015.02.003)]{}, [[](http://arxiv.org/abs/1311.5210)]{}.
F. Gonzalez Canales, A. Mondragon and M. Mondragon, *[The $S_3$ Flavour Symmetry: Neutrino Masses and Mixings]{}*, [[](http://dx.doi.org/10.1002/prop.201200121)]{}[[](http://dx.doi.org/10.1002/prop.201200121)]{}, [[](http://arxiv.org/abs/1205.4755)]{}.
F. González Canales et al., *[Quark sector of $S_{3}$ models: classification and comparison with experimental data]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.88.096004)]{}[[](http://dx.doi.org/10.1103/PhysRevD.88.096004)]{}, [[](http://arxiv.org/abs/1304.6644)]{}.
J. Barranco, F. Gonzalez Canales and A. Mondragon, *[Universal Mass Texture, CP violation and Quark-Lepton Complementarity]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.82.073010)]{}[[](http://dx.doi.org/10.1103/PhysRevD.82.073010)]{}, [[](http://arxiv.org/abs/1004.3781)]{}.
A. Mondragón and E. Rodríguez-Jáuregui, *Breaking of the flavor permutational symmetry: Mass textures and the ckm matrix*, [[](http://dx.doi.org/10.1103/PhysRevD.59.093009)]{}[[](http://dx.doi.org/10.1103/PhysRevD.59.093009)]{}, <http://link.aps.org/doi/10.1103/PhysRevD.59.093009>.
T. Cheng and M. Sher, *[Mass Matrix Ansatz and Flavor Nonconservation in Models with Multiple Higgs Doublets]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.35.3484)]{}[[](http://dx.doi.org/10.1103/PhysRevD.35.3484)]{}.
J. L. Diaz-Cruz, R. Noriega-Papaqui and A. Rosado, *[Mass matrix ansatz and lepton flavor violation in the THDM-III]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.69.095002)]{}[[](http://dx.doi.org/10.1103/PhysRevD.69.095002)]{}, [[](http://arxiv.org/abs/hep-ph/0401194)]{}.
J. Hernandez-Sanchez, S. Moretti, R. Noriega-Papaqui and A. Rosado, *[Update of the 2HDM-III with a four-zero texture in the Yukawa matrices and phenomenology of the charged Higgs Boson]{}*, PoS **CHARGED2012** (2012) 029, [[](http://arxiv.org/abs/1302.0083)]{}.
M. Gomez-Bock, G. Lopez Castro, L. Lopez-Lozano and A. Rosado, *[Flavor-changing neutral current in production and decay of pseudoscalar mesons in a type III two-Higgs-doublet-model with four-texture Yukawa couplings]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.80.055017)]{}[[](http://dx.doi.org/10.1103/PhysRevD.80.055017)]{}, [[](http://arxiv.org/abs/0905.3351)]{}.
R. Martinez, J. A. Rodriguez and S. Sanchez, *[Charged Higgs production at photon colliders in 2HDM-III]{}*, [[](http://dx.doi.org/10.1590/S0103-97332008000400025)]{}[[](http://dx.doi.org/10.1590/S0103-97332008000400025)]{}, [[](http://arxiv.org/abs/0810.4303)]{}.
R. Martinez, J. A. Rodriquez and D. A. Milanes, *[The Lightest Higgs boson production at photon colliders in the 2HDM-III]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.72.035017)]{}[[](http://dx.doi.org/10.1103/PhysRevD.72.035017)]{}, [[](http://arxiv.org/abs/hep-ph/0502087)]{}.
K. A. Hochmuth, S. T. Petcov and W. Rodejohann, *$u_{\mathrm{pmns}} =
u^{\dagger}_{l} u_{\nu}$*, [[](http://dx.doi.org/10.1016/j.physletb.2007.08.072)]{}[[](http://dx.doi.org/10.1016/j.physletb.2007.08.072)]{}, [[](http://arxiv.org/abs/0706.2975)]{}.
S. Bilenky, *[Introduction to the physics of massive and mixed neutrinos]{}*, [[](http://dx.doi.org/10.1007/978-3-642-14043-3)]{}[[](http://dx.doi.org/10.1007/978-3-642-14043-3)]{}.
G. C. Branco, L. Lavoura and M. N. Rebelo, *[Majorana Neutrinos and [CP]{} Violation in the Leptonic Sector]{}*, [[](http://dx.doi.org/10.1016/0370-2693(86)90307-2)]{}[[](http://dx.doi.org/10.1016/0370-2693(86)90307-2)]{}.
G. C. Branco, R. G. Felipe and F. R. Joaquim, *[Leptonic CP Violation]{}*, [[](http://dx.doi.org/10.1103/RevModPhys.84.515)]{}[[](http://dx.doi.org/10.1103/RevModPhys.84.515)]{}, [[](http://arxiv.org/abs/1111.5332)]{}.
E. E. Jenkins and A. V. Manohar, *[Rephasing Invariants of Quark and Lepton Mixing Matrices]{}*, [[](http://dx.doi.org/10.1016/j.nuclphysb.2007.09.031)]{}[[](http://dx.doi.org/10.1016/j.nuclphysb.2007.09.031)]{}, [[](http://arxiv.org/abs/0706.4313)]{}.
P. A. R. Ade et al. (Planck), *[Planck 2015 results. XIII. Cosmological parameters]{}*, [[](http://dx.doi.org/10.1051/0004-6361/201525830)]{}[[](http://dx.doi.org/10.1051/0004-6361/201525830)]{}, [[](http://arxiv.org/abs/1502.01589)]{}.
M. Auger et al. (EXO-200), *[Search for Neutrinoless Double-Beta Decay in $^{136}$Xe with EXO-200]{}*, [[](http://dx.doi.org/10.1103/PhysRevLett.109.032505)]{}[[](http://dx.doi.org/10.1103/PhysRevLett.109.032505)]{}, [[](http://arxiv.org/abs/1205.5608)]{}.
J. B. Albert et al. (EXO-200), *[Search for Majorana neutrinos with the first two years of EXO-200 data]{}*, [[](http://dx.doi.org/10.1038/nature13432)]{}[[](http://dx.doi.org/10.1038/nature13432)]{}, [[](http://arxiv.org/abs/1402.6956)]{}.
A. Gando et al. (KamLAND-Zen), *Limit on neutrinoless betabeta decay of xe-136 from the first phase of kamland-zen and comparison with the positive claim in ge-76*, Phys. Rev. Lett. **110** (2013) 062502, [[](http://arxiv.org/abs/1211.3863)]{}.
J. Schechter and J. W. F. Valle, *[Neutrinoless Double beta Decay in $SU(2)\times U(1)$ Theories]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.25.2951)]{}[[](http://dx.doi.org/10.1103/PhysRevD.25.2951)]{}.
M. Duerr, M. Lindner and A. Merle, *[On the Quantitative Impact of the Schechter-Valle Theorem]{}*, [[](http://dx.doi.org/10.1007/JHEP06(2011)091)]{}[[](http://dx.doi.org/10.1007/JHEP06(2011)091)]{}, [[](http://arxiv.org/abs/1105.0901)]{}.
J. Beringer et al. (Particle Data Group), *[Review of Particle Physics (RPP)]{}*, [[](http://dx.doi.org/10.1103/PhysRevD.86.010001)]{}[[](http://dx.doi.org/10.1103/PhysRevD.86.010001)]{}.
V. Barger, D. Marfatia and A. Tregre, *[Neutrino mass limits from SDSS, 2dFGRS and WMAP]{}*, [[](http://dx.doi.org/10.1016/j.physletb.2004.06.049)]{}[[](http://dx.doi.org/10.1016/j.physletb.2004.06.049)]{}, [[](http://arxiv.org/abs/hep-ph/0312065)]{}.
S. F. King, A. Merle and A. J. Stuart, *[The Power of Neutrino Mass Sum Rules for Neutrinoless Double Beta Decay Experiments]{}*, [[](http://dx.doi.org/10.1007/JHEP12(2013)005)]{}[[](http://dx.doi.org/10.1007/JHEP12(2013)005)]{}, [[](http://arxiv.org/abs/1307.2901)]{}.
J. Schechter and J. W. F. Valle, *Neutrino masses in $su(2)\times u(1)$ theories*, Phys. Rev. **D22** (1980) 2227.
P. Chen, G.-J. Ding, F. Gonzalez-Canales and J. W. F. Valle, *[Generalized $\mu-\tau$ reflection symmetry and leptonic CP violation]{}*, [[](http://dx.doi.org/10.1016/j.physletb.2015.12.069)]{}[[](http://dx.doi.org/10.1016/j.physletb.2015.12.069)]{}, [[](http://arxiv.org/abs/1512.01551)]{}.
M. V. Diwan, V. Galymov, X. Qian and A. Rubbia, *[Long-Baseline Neutrino Experiments]{}*, [[](http://dx.doi.org/10.1146/annurev-nucl-102014-021939)]{}[[](http://dx.doi.org/10.1146/annurev-nucl-102014-021939)]{}, [[](http://arxiv.org/abs/1608.06237)]{}.
K. Abe et al. (T2K), *[The T2K Experiment]{}*, [[](http://dx.doi.org/10.1016/j.nima.2011.06.067)]{}[[](http://dx.doi.org/10.1016/j.nima.2011.06.067)]{}, [[](http://arxiv.org/abs/1106.1238)]{}.
P. Adamson et al. (NO$\nu$A), *[First measurement of electron neutrino appearance in NO$\nu$A]{}*, [[](http://dx.doi.org/10.1103/PhysRevLett.116.151806)]{}[[](http://dx.doi.org/10.1103/PhysRevLett.116.151806)]{}, [[](http://arxiv.org/abs/1601.05022)]{}.
R. Acciarri et al. (DUNE), *[Long-Baseline Neutrino Facility (LBNF) and Deep Underground Neutrino Experiment (DUNE)]{}* (2015), [[](http://arxiv.org/abs/1512.06148)]{}.
H. Nunokawa, S. Parke and J. W. Valle, *Cp violation and neutrino oscillations*, [[](http://dx.doi.org/http://dx.doi.org/10.1016/j.ppnp.2007.10.001)]{}[[](http://dx.doi.org/http://dx.doi.org/10.1016/j.ppnp.2007.10.001)]{}, ISSN 0146-6410.
F. Gonzalez-Canales, *[The remnant CP transformation and its implications]{}*, [[](http://dx.doi.org/10.1088/1742-6596/761/1/012046)]{}[[](http://dx.doi.org/10.1088/1742-6596/761/1/012046)]{}, [[](http://arxiv.org/abs/1609.01346)]{}.
[^1]: Analysis of heavy Majorana neutrinos implications in LHC is dealt in Refs. [@Das:2012ze; @Das:2017nvm].
[^2]: As other authors have done in the SM [@Canales:2012dr; @Canales:2013cga; @Barranco:2010we] (and references in there).
[^3]: The inverted hierarchy is only valid for neutrinos.
[^4]: The superscript $\mathrm{n[i]}$ denote the normal \[inverted\] hierarchy in the neutrino mass spectrum.
[^5]: Here, NH and IH denote the normal and inverted hierarchy in neutrino mass spectrum, respectively.
[^6]: The subscript $i[j]$ denote to normal \[inverted\] hierarchy in the neutrino masses spectrum.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We analyse the inter-scale transfer of energy for two types of plane Poiseuille flow: the P4U exact coherent state of @Park15 and turbulent flow in a minimal channel. For both flows, the dominant energy-producing modes are streamwise-constant streaks with a spanwise spacing of approximately 100 wall units. Since the viscous dissipation for these scales is not sufficient to balance production, the nonlinear terms redistribute the excess energy to other scales. Spanwise-constant scales (that is, Tollmien-Schlichting-like modes with zero spanwise wavenumber), in particular, account for a significant amount of net energy gain from the nonlinear terms. We compare the energy balance to predictions from resolvent analysis and we show that it does not model energy transfer well. Nevertheless, we find that the energy transferred from the streamwise-constant streaks can be predicted reasonably well by a Cess eddy viscosity profile. As such, eddy viscosity is an effective model for the nonlinear terms in resolvent analysis and explains good predictions for the most energetic streamwise-constant streaks. It also improves resolvent modes as a basis for structures whose streamwise lengths are greater than their spanwise widths by counteracting non-normality of the resolvent operator. Eddy viscosity does not respect the conservative nature of the nonlinear energy transfer which must sum to zero over all scales. It is less effective, consequently, for scales which receive energy from the nonlinear terms.'
author:
- 'Sean Symon , Simon J. Illingworth'
- Ivan Marusic
bibliography:
- 'energy\_transfer.bib'
title: Energy transfer in turbulent channel flows and implications for resolvent modelling
---
Introduction
============
Energy transfer plays a key role in the organisation and evolution of turbulent flows. It is responsible for the multi-scale nature of turbulence through the Richardson-Kolmogorov turbulent energy cascade [@Kolmogorov41] and lends insight into the self-sustaining process [@Hamilton95]. Energy transfer for an individual scale is described by the spectral turbulent kinetic energy (TKE) equation, which contains a nonlinear term sometimes referred to as turbulent transport. As noted by [@Domaradzki94], the nonlinearity poses considerable theoretical difficulties by permitting inter-scale energy exchange. It is not possible, for example, to study a scale in isolation without a closure model and, in the context of a large eddy simulation (LES), subgrid models need to account for the influence of small scales on the large scales of interest. An improved understanding of nonlinear interactions in turbulent flows, therefore, is essential to improve turbulence modelling and simulation.
It is also known that linear mechanisms are important in energy transfer. These are described well by the linear operator obtained after linearising the Navier-Stokes equations around a suitable base flow [@Schmid01]. This operator is highly non-normal due to the mean shear found in wall-bounded flows [@Trefethen93]. As a result, infinitesimal disturbances may experience significant transient growth by extracting energy from the mean shear [@Butler92; @Reddy93]. Linear mechanisms have also been identified in mean (time-averaged) flows by the resolvent analysis of [@McKeon10]. In this framework, the equations are linearised around the mean flow to obtain the resolvent operator that maps the nonlinear terms, treated as an intrinsic forcing, to the velocity in the frequency domain. Energy production by the linear system is excited by the nonlinear forcing which, when large enough, also excites dissipative modes to dissipate energy [@Sharma09]. The nonlinear forcing is itself composed of quadratic interactions between various outputs of the linear amplification process to complete a feedback loop [@McKeon13]. Unless the flow is dominated by a single Fourier mode [@Rosenberg19b], it is not tractable to isolate the principal interactions that comprise the nonlinear forcing.
An objective of this paper, therefore, is to investigate the extent to which energy transfer is correctly modelled by resolvent analysis. To address this question, we first examine how energy is produced, dissipated, and transferred among various scales in turbulent channel flow at low Reynolds numbers. Similar to other studies [@Mizuno16; @Cho18; @Lee19], we calculate these terms in spectral space and integrate them over the wall-normal direction. The true energy transfer is compared to predictions from the optimal resolvent mode, which is often representative of the true velocity field observed in DNS or experiments [@McKeon17]. The agreement can be improved by adding the [@Cess58] eddy viscosity profile to the resolvent operator as done in many studies [@Hwang10; @Morra19; @Symon20]. It has been likened to a crude model for the energy cascade by [@Hwang16], suggesting that it assumes the role of turbulent transport in resolvent analysis. To provide insight into the matter, we quantify the contribution of eddy viscosity to the energy balance for each scale and compare it to the true nonlinear transfer. Finally, we aim to shed insight into the role of non-normality in the energy balance and the attenuation of non-normality by eddy viscosity to improve predictions from resolvent analysis.
The flows selected for this study are the P4U exact coherent state (ECS) of [@Park15] and turbulent flow in a minimal channel at low Reynolds number. The former is a nonlinear travelling wave whose mean properties resemble those of near-wall turbulence. It is a particularly appealing choice to study energy transfer since it is low-dimensional [@Sharma16; @Rosenberg19] and travels at a fixed convection velocity. As such, all computations can be performed on a standard personal computer and no integration in time is necessary to obtain each term in the energy budget. To verify that the transfer mechanisms are similar in a time-evolving flow, we compare the results for the P4U ECS to those of more standard turbulence in a “minimal flow unit” [@Jimenez91].
The paper is organised as follows. In §\[sec:methods\], the relevant equations for resolvent analysis, energy transfer and the eddy viscosity model are derived. The simulation parameters for the P4U ECS and minimal channel flows are described in §\[sec:description\]. The energy balances computed from DNS and resolvent analysis are compared for the ECS in §\[sec:ECS\] and the minimal channel in §\[sec:channel\]. In §\[sec:discussion\], we examine the influence of non-normality on the ability of the first resolvent mode to describe energy transfer processes. We also analyse the role of eddy viscosity on the efficiency of resolvent modes as a basis for the velocity fluctuations. This leads to a discussion on the typical scales for which an eddy viscosity leads to an improvement before we conclude in §\[sec:conclusions\].
Methods {#sec:methods}
=======
In §\[sec:equations\], we describe the governing equations for plane Poiseuille flow and their non-dimensionalisation. A brief overview of resolvent analysis is provided in §\[sec:resolvent analysis\]. The energy balance for each scale is then derived from the fluctuation equations in §\[sec:energy\] and we show that this balance is maintained for each resolvent mode. Finally, we describe the eddy viscosity model in §\[sec:eddy viscosity\].
Plane Poiseuille flow equations {#sec:equations}
-------------------------------
The non-dimensional Navier-Stokes equations for statistically steady, turbulent plane Poiseuille flow are
\[eq:NSE\] $$\label{eq:Momentum}
\frac{\partial \boldsymbol{u}}{\partial t} + \boldsymbol{u} \cdot \boldsymbol{\nabla} \boldsymbol{u} = -\boldsymbol{\nabla} p + \frac{1}{Re_{\tau}}\boldsymbol{\nabla}^2\boldsymbol{u},$$ $$\boldsymbol{\nabla} \cdot \boldsymbol{u} = 0,$$
where $\boldsymbol{u} = [u,v,w]^T$ is the velocity in the $x$ (streamwise), $y$ (spanwise) and $z$ (wall-normal) directions and $p$ is the pressure. The friction Reynolds number $Re_{\tau} = u_\tau h /\nu$ is defined in terms of the friction velocity $u_{\tau}$, channel half height $h$, and kinematic viscosity $\nu$. No-slip boundary conditions are applied at the walls and periodic boundary conditions are imposed in the streamwise and spanwise directions. The density of the fluid is $\rho$ and the velocities are non-dimensionalized by $u_\tau$, the spatial variables by $h$ and the pressure by $\rho u_{\tau}^2$. A ‘$+$’ superscript denotes spatial variables that have been normalized by the viscous length scale $\nu/u_{\tau}$.
Resolvent analysis {#sec:resolvent analysis}
------------------
We begin by Reynolds-decomposing (\[eq:Momentum\]), which leads to the following equations for the fluctuations: $$\label{eq:input-output}
\frac{\partial \boldsymbol{u}'}{\partial t} + \boldsymbol{U} \cdot \boldsymbol{\nabla} \boldsymbol{u}' + \boldsymbol{u}' \cdot \boldsymbol{\nabla} \boldsymbol{U} + \boldsymbol{\nabla} p' - \frac{1}{Re_{\tau}}\boldsymbol{\nabla}^2 \boldsymbol{u}' = -\boldsymbol{u}'\cdot \boldsymbol{\nabla} \boldsymbol{u}' + \overline{\boldsymbol{u}' \cdot \boldsymbol{\nabla} \boldsymbol{u}'} = \boldsymbol{f}',$$ where $(\overline{\cdot})$ and $(\cdot)'$ denote a time-average and fluctuation, respectively, and $\boldsymbol{U} = [U(y),0,0]^T$ is the mean velocity. Equation (\[eq:input-output\]) is written such that all linear terms appear on the left-hand side while all nonlinear terms appear on the right-hand side. Equation (\[eq:input-output\]) is then Fourier-transformed in time and in the homogeneous directions $x$ and $y$ $$\label{eq:Fourier transform}
\hat{\boldsymbol{u}}(k_x,k_y,z,\omega) = \int_{\infty}^{\infty}\int_{-\infty}^{\infty} \int_{-\infty}^{\infty} \boldsymbol{u}'(x,y,z,t) e^{-i(k_xx+k_yy-\omega t)}dxdydt,$$ where $(\hat{\cdot})$ denotes the Fourier-transformed coefficient, $k_x$ is the streamwise wavenumber, $k_y$ is the spanwise wavenumber and $\omega$ the temporal frequency. The equivalent wavelengths in the streamwise and spanwise directions are $\lambda_x = 2\pi/k_x$ and $\lambda_y = 2\pi/k_y$. The equations are arranged into state-space form [@Jovanovic05] after substituting (\[eq:Fourier transform\]) into (\[eq:input-output\])
\[eq:OSSQ\] $$i\omega \hat{\boldsymbol{q}}(\boldsymbol{k}) = \boldsymbol{A}(k_x,k_y)\hat{\boldsymbol{q}}(\boldsymbol{k}) + \boldsymbol{B}(k_x,k_y)\hat{\boldsymbol{f}}(\boldsymbol{k}) ,$$ $$\hat{\boldsymbol{u}}(\boldsymbol{k}) = \boldsymbol{C}(k_x,k_y)\hat{\boldsymbol{q}}(\boldsymbol{k}) ,$$
where $\boldsymbol{k} = (k_x,k_y,\omega)$ is the wavenumber triplet and $\hat{\boldsymbol{q}}$ consists of the wall-normal velocity and vorticity $\hat{\eta} = ik_y\hat{u} - ik_x\hat{v}$. The operators $\boldsymbol{A}$, $\boldsymbol{B}$ and $\boldsymbol{C}$ represent the linear Navier-Stokes operator, the input matrix and the output matrix, respectively, and are defined in appendix \[sec:operators\]. These operators are independent of $\omega$ but are functions of the wavenumber pair $(k_x,k_y)$ under consideration. In the interest of readability, this dependence is omitted for the rest of the paper.
Once (\[eq:OSSQ\]) is recast into input-output form, i.e, $$\hat{\boldsymbol{u}}(\boldsymbol{k}) = \boldsymbol{C}(i\omega \boldsymbol{I}- \boldsymbol{A})^{-1} \hat{\boldsymbol{f}}(\boldsymbol{k}) = \mathcal{H}(\boldsymbol{k}) \hat{\boldsymbol{f}}(\boldsymbol{k}),$$ a linear operator called the resolvent $\mathcal{H}(\boldsymbol{k})$ relates the input forcing $\hat{\boldsymbol{f}}(\boldsymbol{k})$ to the output velocity $\hat{\boldsymbol{u}}(\boldsymbol{k})$. Even if the nonlinear forcing is unknown, the resolvent identifies structures due to linear mechanisms. These structures can be obtained from a singular value decomposition of the resolvent operator: $$\label{eq:SVD}
\mathcal{H}(\boldsymbol{k}) = \hat{\boldsymbol{\Psi}} (\boldsymbol{k}) \boldsymbol{\Sigma}(\boldsymbol{k}) \hat{\boldsymbol{\Phi}}^*(\boldsymbol{k}),$$ where $\hat{\boldsymbol{\Psi}}(\boldsymbol{k}) = [\hat{\boldsymbol{\psi}}_1(\boldsymbol{k}), \hat{\boldsymbol{\psi}}_2(\boldsymbol{k}), \cdots, \hat{\boldsymbol{\psi}}_p(\boldsymbol{k})]$ and $\hat{\boldsymbol{\Phi}}(\boldsymbol{k}) = [\hat{\boldsymbol{\phi}}_1(\boldsymbol{k}),\hat{\boldsymbol{\phi}}_2(\boldsymbol{k}), \cdots, \hat{\boldsymbol{\phi}}_p(\boldsymbol{k})]$ are orthogonal basis functions for the velocity and nonlinear forcing, respectively. The diagonal matrix $\boldsymbol{\Sigma}(\boldsymbol{k})$ ranks the $p$th structure by its gain $\sigma_p(\boldsymbol{k})$ using an inner product that is proportional to its kinetic energy. Consequently, the structure $\hat{\boldsymbol{\psi}}_1(\boldsymbol{k})$, referred to as the optimal or first resolvent mode, is the most amplified response by the linear dynamics contained in the operator. The true velocity field is the weighted sum of resolvent modes, i.e., $$\label{eq:weighted sum}
\hat{\boldsymbol{u}}(\boldsymbol{k}) = \sum_{p = 1}^N \hat{\boldsymbol{\psi}}_p(\boldsymbol{k}) \sigma_p(\boldsymbol{k}) \chi_p(\boldsymbol{k}),$$ where $\chi_p(\boldsymbol{k})$ is the projection of $\hat{\boldsymbol{\phi}}_p(\boldsymbol{k})$ onto $\hat{\boldsymbol{f}}(\boldsymbol{k})$.
Energy balance {#sec:energy}
--------------
We now derive the energy balance that must be satisfied by the velocity field and individual resolvent modes. Equation (\[eq:input-output\]) is rewritten in index notation $$\label{eq:fluctuating NSE}
\frac{\partial u'_i}{\partial t} + U_j\frac{\partial u'_i}{\partial x_j} + u_j'\frac{\partial U_i}{\partial x_j} + \frac{\partial p'}{\partial x_i} - \frac{1}{Re}\frac{\partial^2u_i'}{\partial x_j \partial x_j} = - u'_j\frac{\partial u_i'}{\partial x_j} + \overline{u'_j\frac{\partial u_i'}{\partial x_j}} .$$ Similar to (\[eq:input-output\]), all nonlinear terms appear on the right-hand side although they are not treated as an unknown forcing. The indices $i,j =1,2,3$ and $U_i = (U(z),0,0)$ is the mean velocity, which is a function of the wall-normal direction only. It can be noted, therefore, that $U_1 = U$, $U_j = 0$ if $j=2,3$ and $\partial U_i/\partial x_j \neq 0$ for $i = 1$ and $j = 3$ only. The kinetic energy of the full system is characterised by the inner product between (\[eq:fluctuating NSE\]) and $u_i'$ integrated over the volume $V$: $$\begin{aligned}
\label{eq:Reynolds-Orr}
\underbrace{ \frac{1}{2} \int_V \frac{\partial u_i^{'2}}{\partial t} dV }_{\dot{E}(t)} = \underbrace{- \int_V u'_i u_j' \frac{\partial U_i}{\partial x_j} dV }_{P(t)} \underbrace{ - \frac{1}{Re} \int_V \frac{\partial u_i'}{\partial x_j} \frac{\partial u_i'}{\partial x_j} dV}_{D(t)},\\
\int_V u_i'f_i'dV = \int_V u_i'u_j'\frac{\partial u_i'}{\partial x_j}dV = 0. \label{eq:conservative}\end{aligned}$$ Equation (\[eq:Reynolds-Orr\]) is the Reynolds-Orr equation [@Schmid01] where the evolution of kinetic energy in the system is a balance between production and dissipation, which must be negative. Due to the conservative nature of the nonlinear terms, their contribution to the Reynolds-Orr equation sums to zero when integrated over the volume as expressed in (\[eq:conservative\]). For a statistically stationary flow, a time average of (\[eq:Reynolds-Orr\]) implies that production balances dissipation since $\overline{dE/dt} = 0$.
The kinetic energy for a specific spatial scale is obtained after multiplying (\[eq:fluctuating NSE\]) by $u_i^{'*}$ and Fourier-transforming in $x$ and $y$. The result is integrated over the wall-normal direction and time-averaged to arrive at the spectral turbulent kinetic energy (TKE) equation: $$\begin{aligned}
\label{eq:budget}
\nonumber \overline{\frac{\partial \hat{E}(k_x,k_y)}{\partial t}} = & \underbrace{- \int_{-h}^h\frac{dU}{dz}\overline{\hat{u}^*(k_x,k_y)\hat{v}(k_x,k_y)}dz}_{\hat{P}(k_x,k_y)} \underbrace{-\frac{1}{Re}\int_{-h}^h \overline{ \frac{\partial \hat{u}_i(k_x,k_y)}{\partial x_j} \frac{\partial \hat{u}^*_i(k_x,k_y)}{\partial x_j}} dz}_{\hat{D}(k_x,k_y)} \\ &
\underbrace{ - \int_{-h}^h \overline{\hat{u}_i^*(k_x,k_y)\frac{\partial}{\partial x_j} \widehat{u_iu_j}(k_x,k_y) } dz}_{\hat{N}(k_x,k_y)} = 0.\end{aligned}$$ The pressure terms vanish in (\[eq:budget\]) after integrating over the channel height [@Aubry88]. Following [@Muralidhar19], we consider the real part of (\[eq:budget\]), which consists of three terms: production, viscous dissipation and nonlinear transfer. In general, production $\hat{P}$ is positive for a given scale as perturbations extract energy from the mean flow. Viscous dissipation $\hat{D}$, on the other hand, is guaranteed to be real and negative according to (\[eq:budget\]) as it is the mechanism through which kinetic energy is removed from the system and converted into heat. Nonlinear transfer $\hat{N}$ may be positive or negative depending on the scale selected. If $\hat{P} > \hat{D}$, for example, then $\hat{N} < 0$ in order to achieve a balance. In a similar fashion, if $\hat{P} < \hat{D}$, then $\hat{N} > 0$. The integral of $\hat{N}$ over all $k_x$ and $k_y$, nevertheless, is zero as stated in (\[eq:conservative\]).
To obtain the energy balance for resolvent modes, which are defined for a wavenumber triplet $\boldsymbol{k}$, (\[eq:fluctuating NSE\]) is multiplied by $u_i^{'*}$ and Fourier-transformed in $x$, $y$ and $t$. The result is integrated over the wall-normal direction $$\label{eq:budget k}
\int_{-h}^h\frac{dU}{dz} \hat{u}^*(\boldsymbol{k})\hat{v} (\boldsymbol{k})dz
-\frac{1}{Re} \int_{-h}^h \frac{\partial \hat{u}_i(\boldsymbol{k})}{\partial x_j} \frac{\partial \hat{u}^*_i(\boldsymbol{k})}{\partial x_j} dz
- \int_{-h}^h \hat{u}_i^*(\boldsymbol{k})\hat{f}_i(\boldsymbol{k}) dz = 0.$$ In this form, the nonlinear forcing $\hat{\boldsymbol{f}}(\boldsymbol{k})$ appears explicitly in the energy balance. We can now express the velocity field in terms of resolvent modes. In the special case where $\hat{\boldsymbol{f}}(\boldsymbol{k})$ is white noise, the velocity field can be written as $$\label{eq:white noise resolvent}
\hat{\boldsymbol{u}}(\boldsymbol{k}) = \sum_{p = 1}^N \hat{\boldsymbol{\psi}}_p(\boldsymbol{k}) \sigma_p(\boldsymbol{k}).$$ Substituting (\[eq:white noise resolvent\]) into (\[eq:budget k\]) yields $$\begin{aligned}
\label{eq:res mode balance}
\nonumber \sum_p \sigma_p(\boldsymbol{k}) \left( \int_{-h}^h \frac{dU}{dz} \hat{\boldsymbol{\psi}}_p^{*,i=1}(\boldsymbol{k}) \hat{\boldsymbol{\psi}}_p^{j=2} (\boldsymbol{k}) dz + \frac{1}{Re} \int_{-h}^h \frac{\partial \hat{\boldsymbol{\psi}}_p^{*,i}(\boldsymbol{k})} {\partial \boldsymbol{x}_j} \frac{\partial \hat{\boldsymbol{\psi}}_p^i (\boldsymbol{k})}{\partial \boldsymbol{x}_j} dz \right) + \\
\sum_p \int_{-h}^h \boldsymbol{\psi}_p^*(\boldsymbol{k})\hat{\boldsymbol{\phi}}_p(\boldsymbol{k})dz = 0.\end{aligned}$$ Each term in the sum can be decoupled since the basis functions $\hat{\boldsymbol{\psi}}_p$ are orthogonal. This means that production, dissipation and nonlinear transfer must be balanced across each resolvent mode. If $\sigma_1(\boldsymbol{k}) \gg \sigma_2(\boldsymbol{k})$ then it would imply that the sum over all $p$ is dominated by the first resolvent mode, or $$\begin{aligned}
\label{eq:budget rank 1}
\nonumber \sigma_1(\boldsymbol{k}) \left( \int_{-h}^h \frac{dU}{dz} \hat{\boldsymbol{\psi}}_1^{*,i=1}(\boldsymbol{k}) \hat{\boldsymbol{\psi}}_1^{j=2} (\boldsymbol{k}) dz + \frac{1}{Re} \int_{-h}^h \frac{\partial \hat{\boldsymbol{\psi}}_1^{*,i}(\boldsymbol{k})} {\partial \boldsymbol{x}_j} \frac{\partial \hat{\boldsymbol{\psi}}_1^i (\boldsymbol{k})}{\partial \boldsymbol{x}_j} dz \right) + \\
\int_{-h}^h \boldsymbol{\psi}_1^*(\boldsymbol{k})\hat{\boldsymbol{\phi}}_1(\boldsymbol{k})dz = 0.\end{aligned}$$ The bulk of production, dissipation and nonlinear transfer for a particular scale $\boldsymbol{k}$, therefore, could also be accounted for by the first resolvent mode.
Eddy viscosity model {#sec:eddy viscosity}
--------------------
If the nonlinear forcing is not white noise, then (\[eq:res mode balance\]) is not applicable since it does not take into account the complex amplitude of each mode. One method to model the nonlinear forcing is to add an eddy viscosity to the linearised equations after performing a triple decomposition of the velocity field $\tilde{\boldsymbol{u}}$ into a mean component $\boldsymbol{U}$, coherent motions $\boldsymbol{u}$ and incoherent turbulent fluctuations $\boldsymbol{u}'$ [@Reynolds72]. The equations governing the coherent velocity and pressure are $$\label{eq:eddy LNSE}
\frac{\partial \boldsymbol{u}}{\partial t} + \boldsymbol{U} \cdot \boldsymbol{\nabla} \boldsymbol{u} + \boldsymbol{u} \cdot \boldsymbol{\nabla} \boldsymbol{U} + \boldsymbol{\nabla} p + \boldsymbol{\nabla} \cdot \left[ \nu_T ( \boldsymbol{\nabla} \boldsymbol{\boldsymbol{u}} + \boldsymbol{\nabla} \boldsymbol{\boldsymbol{u}}^T) \right] = \boldsymbol{d},$$ where $\nu_T(z)$ is the total effective viscosity and $\boldsymbol{d} = -\boldsymbol{u} \cdot \boldsymbol{\nabla} \boldsymbol{u} + \overline{\boldsymbol{u} \cdot \boldsymbol{\nabla} \boldsymbol{u}}$ is the forcing term. It should be noted that $\boldsymbol{d}$ is different from $\boldsymbol{f}'$ in (\[eq:input-output\]). Following [@Reynolds67] and [@Hwang10], we use the [@Cess58] eddy viscosity profile of the form $$\nu_T(z) = \frac{\nu}{2}\left(1 + \left[ \frac{\kappa}{3}(1-z^2)(1+2z^2)\left(1-e^{|z-1| \frac{Re_{\tau}}{A}}\right) \right]^2 \right)^{1/2} + \frac{\nu}{2},$$ where $\kappa = 0.426$ and $A = 25.4$ are chosen based on a least-squares fit to experimentally obtained mean velocity profiles at $Re_{\tau} = 2000$ [@delAlamo06].
Fourier-transforming (\[eq:eddy LNSE\]) in time and the homogeneous directions and rearranging it into the following input-output form yields $$\hat{\boldsymbol{u}}(\boldsymbol{k}) = \mathcal{H}^e(\boldsymbol{k})\hat{\boldsymbol{d}}(\boldsymbol{k}),$$ where $\mathcal{H}^e(\boldsymbol{k})$ is a new linear operator that relates the forcing $\hat{\boldsymbol{d}} (\boldsymbol{k})$ to the velocity field $\hat{\boldsymbol{u}}(\boldsymbol{k})$. Similar to the resolvent operator, we can analyse structures that are preferentially amplified by performing a singular value decomposition $$\mathcal{H}^e(\boldsymbol{k}) = \hat{\boldsymbol{\Psi}}^e (\boldsymbol{k}) \boldsymbol{\Sigma}^e (\boldsymbol{k}) \hat{\boldsymbol{\Phi}}^{*,e}(\boldsymbol{k}),$$ although the individual modes $\hat{\boldsymbol{\psi}}_p^e(\boldsymbol{k})$ do not satisfy the energy balance in (\[eq:res mode balance\]). Instead, the addition of eddy viscosity in (\[eq:eddy LNSE\]) introduces two terms, the first of which is $$\label{eq:eddy dissipation}
\hat{V}(k_x,k_y) = -\int_{-h}^h (\nu_T(z)-\nu) \overline{ \frac{\partial \hat{u}_i(k_x,k_y)}{\partial x_j} \frac{\partial \hat{u}^*_i(k_x,k_y)}{\partial x_j}} dz,$$ where the kinematic viscosity $\nu$ has been subtracted in order to remove the contribution of viscous dissipation $\hat{D}(k_x,k_y)$. The remainder $\hat{V}(k_x,k_y)$ represents, therefore, the additional dissipation provided by the wall-normal varying portion of $\nu_T$. Similar to $\hat{D}(k_x,k_y)$, this term is real and negative, signifying that it removes energy. The second term is related to the wall-normal gradient of $\nu_T$ $$\hat{G}(k_x,k_y) = -\int_{-h}^h \frac{d \nu_T}{dz} \overline{ \left( \hat{u}^*_i(k_x,k_y) \frac{\partial \hat{u}_i(k_x,k_y)}{\partial y} + \hat{u}_i^*(k_x,k_y) \frac{\partial \hat{v}(k_x,k_y)}{\partial x_i} \right) } dz.$$ Unlike $\hat{V}(k_x,k_y)$, the sign of $\hat{G}(k_x,k_y)$ cannot be determined *a priori*.
The combined effect of $\hat{V}(k_x,k_y)$ and $\hat{G}(k_x,k_y)$ is referred to as eddy dissipation $\widehat{Edd}(k_x,k_y)$, i.e., $$\widehat{Edd}(k_x,k_y) = \hat{V}(k_x,k_y) + \hat{G}(k_x,k_y).$$ Eddy dissipation is computed in §§\[sec:ECS\] and \[sec:channel\] using the true velocity field to determine its accuracy in modelling the effect of nonlinear transfer in (\[eq:budget\]). If $\widehat{Edd}(k_x,k_y) \approx \hat{N}(k_x,k_y)$, then it is expected that eddy viscosity will lead to an improvement in the structures predicted by resolvent analysis.
Flow descriptions {#sec:description}
=================
In this section, we describe the two flows that are analysed from an energy transfer perspective. These are the P4U ECS computed by [@Park15] and turbulent channel flow in the minimal unit [@Jimenez91] which are discussed in §§\[sec:ECS description\] and \[sec:min chan description\], respectively.
P4U ECS {#sec:ECS description}
-------
----------------- ------------- ----------------- ------- --------- ------- ------- -------
$Re_{\tau}$ $ c^+$ $L_x$ $L_y$ $N_x$ $N_y$ $N_z$
\[3pt\] P4U 85 14.2 $\pi$ $\pi/2$ 24 24 81
Minimal Channel 180 $ \in [0,19.4]$ $\pi$ $\pi/4$ 96 48 128
----------------- ------------- ----------------- ------- --------- ------- ------- -------
: Relevant parameters for the flows under consideration.[]{data-label="tab:parameters"}
The P4U ECS, henceforth referred to as P4U, is a nonlinear travelling wave with a friction Reynolds number of $Re_{\tau} = 85$ and fixed wave speed of $c^+ = 14.2$. As seen in table \[tab:parameters\], P4U is solved in a computational domain with 24 equally spaced grid points in the streamwise and spanwise directions, which have lengths of $\pi$ and $\pi/2$, respectively. There are 81 points in the wall-normal direction on a Chebyshev grid. The spatial structure of P4U is in the form of low-speed streaks, which are wavy in the streamwise direction, straddled by counter-rotating vortices. As mentioned by [@Park15], its structure is qualitatively similar to near-wall turbulence and its mean velocity profile closely resembles a standard turbulent mean. This is seen more clearly in figure \[fig:means\](a) where the mean profile for P4U in blue is compared to the Cess model in red at $Re_{\tau} = 85$. Despite good overall agreement, the P4U ECS profile has a more wavy nature since the structure has a single convection velocity.
Even though the simulation size is small, there are still many wavenumber pairs which may participate in the transfer of energy. We begin by computing the kinetic energy of each wavenumber pair $$\hat{E}(k_x,k_y) = \frac{1}{2}\left(\overline{\hat{u}^2(k_x,k_y) + \hat{v}^2(k_x,k_y) + \hat{w}^2(k_x,k_y)}\right),$$ and plot the most energetic pairs in figure \[fig:energy\](a). The area and colour intensity of the square marker at the centre of each tile are directly proportional to the kinetic energy. The most energetic scale is streamwise-constant with a spanwise width of approximately 100 wall units, which matches the near-wall streak spacing of [@Smith83]. Most of the kinetic energy, furthermore, is concentrated in structures with small streamwise wavenumbers.
Minimal channel {#sec:min chan description}
---------------
The minimal channel flow is computed for $Re_{\tau} = 180$ using an unstructured finite difference solver (see [@Chung14] for details) on a domain with dimensions $\pi \times \pi/4 \times 2h$ in the streamwise, spanwise and wall-normal directions. There are 96 and 48 equally spaced points in the streamwise and spanwise directions, respectively, and 128 points in the wall-normal direction on a Chebyshev grid. The mean profile for the minimal channel in figure \[fig:means\](b) is more smooth than the P4U mean profile since there exists a distribution of wave speeds which range between $0 < c^+ < 19.4$ as seen in table \[tab:parameters\]. The minimal channel mean profile is in good agreement with the mean profile of [@Lee15] for most areas of the flow other than the wake region, where the minimal channel mean profile overshoots the one from [@Lee15]. This phenomenon has been observed by [@Jimenez91] and stems from the fact that the minimal domain is too small to accommodate the largest structures which reside in the outer region. Despite this disagreement, there is no impact on near-wall turbulence in the buffer and viscous regions where the bulk of energy resides [@Jimenez91; @Jimenez99].
Similar to P4U, the kinetic energy for the most energetic wavenumber pairs is plotted in figure \[fig:energy\](b). Although there are more energetic scales in the minimal channel since the friction Reynolds number is higher, the relative distribution of energy among the scales is quite similar to P4U. The most energetic scale is also streamwise-constant with a spanwise width of approximately 100 wall units. This supports the notion that P4U is a simple model for turbulent channel flow at very low Reynolds number. Therefore, to facilitate visualisation later in the paper, we choose to plot only those wavenumber pairs that appear in figure \[fig:energy\] although the energy balance will be computed across all of them.
Results: P4U ECS {#sec:ECS}
================
In this section, we analyse energy transfer for P4U. We begin with a comparison of production, dissipation and nonlinear transfer across the most energetic scales in §\[sec:P4 balance\]. These results are compared to the resolvent predictions in §\[sec:P4 resolvent\]. Finally, the additional dissipation introduced by eddy viscosity is quantified for each scale and compared to nonlinear transfer in §\[sec:P4 eddy\].
Energy balance {#sec:P4 balance}
--------------
Production, dissipation and nonlinear transfer are computed for P4U and are illustrated in figure \[fig:ecs balance\] for the subset of wavenumber pairs discussed in the previous section. A square marker appears at the centre of each tile. Both colour intensity and area of the square marker indicate each term’s magnitude. The colours red and blue denote positive and negative quantities, respectively. In order to satisfy (\[eq:budget\]), the sum across tiles which appear in the same position in each of the three figure panels must be zero. Additionally, the sum over all tiles in figure \[fig:ecs balance\](c) is approximately zero since the nonlinear terms are conservative when summed over all scales (this sum would be exactly zero if all wavenumber pairs were displayed in the figure).
Half of the production terms in figure \[fig:ecs balance\](a) are positive with the largest energy-producing modes being the streamwise-constant modes. The maximum production occurs at $(k_x,k_y) = (0,4)$, which is also the most energetic mode in the flow (see figure \[fig:energy\]a). Production is negative for some scales. Of particular note is that production is negative for all of the spanwise-constant modes. Even though $\hat{P} \approx 0$ for most of these spanwise-constant modes, the same cannot be said for $(2,0)$ for which the production is negative and of large amplitude. In fact, its magnitude is comparable to that of $(0,8)$ even though it is less energetic, i.e. $|\hat{P}(2,0)| \approx |\hat{P}(0,8)|$ even though $\hat{E}(2,0) < \hat{E}(0,8)$. As expected, all dissipation terms in figure \[fig:ecs balance\](b) are negative.
The nonlinear transfer in figure \[fig:ecs balance\](c) contains both positive and negative terms as the sum over all scales must equal zero. Consistent with the turbulent cascade, most values are positive, indicating that they are receiving energy from nonlinear transfer. The most notable exception is the $(0,4)$ mode, which must redistribute energy to other scales since dissipation offsets less than half of production. The additional scales that lose energy due to nonlinear transfer all have low streamwise wavenumbers. The (0,8) mode is one that receives energy from nonlinear transfer since $\hat{D}(0,8) > \hat{P}(0,8)$. Perhaps surprisingly, the spanwise-constant modes receive a considerable share of the nonlinearly-transferred energy. In particular the (2,0) mode receives more energy than any other mode. The (4,0) and (6,0) modes also receive rather than donate energy. Therefore, in addition to a cascade of energy from large scales to small scales, there is also a significant transfer from scales that are streamwise-constant to scales that are spanwise-constant. Indeed, the (2,0) mode (the largest recipient) is in fact larger in scale than the (0,4) mode (the largest donor). Thus in addition to a cascade, there also exists a transfer to scales of a similar scale but with a different orientation of their wavenumber vector.
Resolvent predictions {#sec:P4 resolvent}
---------------------
Having considered the true energy balance from (\[eq:budget\]), we now focus on its counterpart for the first resolvent mode in (\[eq:budget rank 1\]). To do so, it is necessary to set $\omega = c^+k_x$ since the wave speed is fixed at $c^+ = 14.2$. Figure \[fig:ecs resolvent 1\] illustrates the production, dissipation and nonlinear transfer in a manner analogous to that of figure \[fig:ecs balance\]. The resolvent prediction for production in figure \[fig:ecs resolvent 1\](a) is positive for every scale and the largest value occurs when $(k_x,k_y) = (0,4)$. The predictions for the largest scales are similar to the true values of production in figure \[fig:ecs balance\](a) and reflect the resolvent operator’s ability to identify linear amplification mechanisms.
The dissipation and nonlinear transfer from the first resolvent mode in figures \[fig:ecs resolvent 1\](b) and (c), respectively, are less similar to the true values in figures \[fig:ecs balance\](b) and (c). For all scales, the dissipation is nearly equal and opposite to production resulting in very small values for nonlinear transfer. A similar phenomenon is observed by [@Jin20] for the first resolvent mode in low Reynolds number cylinder flow. It can therefore be concluded that suboptimal resolvent modes are necessary to correctly model nonlinear transfer.
Eddy dissipation {#sec:P4 eddy}
----------------
As discussed in §\[sec:eddy viscosity\], one way to model nonlinear transfer is through the use of an eddy viscosity. Figure \[fig:ecs eddy\](a) presents the eddy dissipation from (\[eq:eddy dissipation\]), which is negative for all wavenumber pairs considered. Unlike nonlinear transfer, therefore, eddy dissipation is not conservative and contributes net energy loss to every scale. Ideally the eddy dissipation would resemble $\hat{N}(k_x,k_y)$ in figure \[fig:ecs balance\](c), so its error $\epsilon(k_x,k_y)$ with respect to nonlinear transfer is computed in figure \[fig:ecs eddy\](b) using the expression $$\label{eq:epsilon}
\epsilon(k_x,k_y) = \frac{\widehat{Edd}(k_x,k_y) - \hat{N}(k_x,k_y)}{|\hat{N}(k_x,k_y)|}.$$ The error for all wavenumber pairs exceeds 1 other than $(0,4)$ where $\epsilon \approx 0.28$. The size of the square marker in figure \[fig:ecs eddy\](b) reflects the magnitude of the error and the smallest marker coincides with the tile belonging to $(0,4)$. The fact that $\epsilon$ is lowest for this scale indicates that the eddy viscosity is most effective for highly amplified linear mechanisms. In other words, the eddy viscosity works best for scales where viscous dissipation is not sufficient to balance production.
Results: Minimal channel {#sec:channel}
========================
Having analysed the energy transfer for the P4U ECS, this section examines the same quantities for the minimal channel. Since each wavenumber pair has a distribution of temporal frequencies, all terms in the energy balance are time-averaged.
DNS and resolvent energy balances
---------------------------------
Production, dissipation and nonlinear transfer for the minimal channel are illustrated in figure \[fig:minchan balance\]. For almost all wavenumber pairs shown, production is positive as seen in figure \[fig:minchan balance\](a) with the maximum occurring for $(0,8)$. The only scales where production is negative are spanwise-constant, i.e. $k_y = 0$. The dissipation in figure \[fig:minchan balance\](b) is negative for all scales, as expected. Even though the largest dissipation occurs for $(0,8)$, its value is comparable to that for other wavenumbers.
The nonlinear transfer in figure \[fig:minchan balance\](c) illustrates that the surfeit of energy not dissipated by viscosity from $(0,8)$ is redistributed to other scales. Moreover, all scales which lose energy due to nonlinear transfer are clustered around low streamwise wavenumbers. This is consistent with the turbulence cascade in that energy from the large-scales trickles down to smaller scales which are more effective at dissipating energy. All but five scales in figure \[fig:minchan balance\](c) receive energy from nonlinear transfer with the largest amounts going to spanwise-constant structures. This is interesting given that these Tollmien-Schlichting-type waves are the first to become unstable [@Tollmien29; @Schlichting33] yet they play a damping role in the minimal channel.
The energy balance for the first resolvent mode is presented in figure \[fig:minchan resolvent 1\]. Since each wavenumber pair has a distribution of energetic temporal frequencies, we compute the singular values across a discretisation of $\omega$ and choose the $\omega$ that results in the largest amplification. As one example, $\omega = 0$ leads to the largest amplification for $(0,8)$. The production in figure \[fig:minchan resolvent 1\] is so large for this scale that the choice of $\omega$ for other scales has little impact on the result. It can be remarked that viscous dissipation in figure \[fig:minchan resolvent 1\](b) is sufficient to completely counteract production for the majority of scales considered. Since the sum of all three terms must be zero for each resolvent mode, it follows that nonlinear transfer is negligible for nearly all scales as seen in figure \[fig:minchan resolvent 1\](c).
To identify for which scales eddy viscosity can model nonlinear transfer, the eddy dissipation is computed and displayed in figure \[fig:minchan eddy\](a). As expected, it is negative for all scales even though nonlinear transfer tends to be positive outside the cluster around $(0,8)$. The error $\epsilon$, as defined in (\[eq:epsilon\]), is thus large for the majority of scales as seen in figure \[fig:minchan eddy\](b). The only scale where $\epsilon < 1$ is $(0,8)$. Although $\epsilon > 1$ for every other scale, those where nonlinear transfer is negative such as $(0,16)$ or $(2,8)$ have lower values of $\epsilon$ than scales where nonlinear transfer is positive.
Comparison of P4U and minimal channel
-------------------------------------
The energy transfer processes in P4U are similar to those of the minimal channel. Production is positive for the majority of scales and its maximum occurs for $(\lambda_x^+,\lambda_y^+) \approx (\infty,100)$, which corresponds to $(k_x,k_y) = (0,4)$ in P4U and $(k_x,k_y) = (0,8)$ in the minimal channel. The production for spanwise-constant scales, on the other hand, is mostly negative. Dissipation is always negative but it is insufficiently large to counterbalance production for the largest streamwise-constant scales. Nonlinear transfer contains both positive and negative terms as the sum over all scales must be zero according to (\[eq:conservative\]). Energy is primarily removed from the largest structures and redistributed to smaller ones. Interestingly, there is also a polarisation effect where nonlinear transfer reallocates energy from streamwise-constant to spanwise-constant modes. Finally, our choice to integrate over the wall-normal domain and analyse energy transfer in $(k_x,k_y)$ space permits quantification of energy loss due to eddy dissipation. For both flows, the only scale where eddy dissipation can quantitatively predict energy loss due to nonlinear transfer is the most energetic scale $(\lambda_x^+,\lambda_y^+) \approx (\infty,100)$.
The only notable difference between the two flows is that nonlinear transfer has a clearer pattern for the minimal channel. Scales where this term is negative are localised in one cluster of large structures. For P4U, there is more scatter primarily due to the $(0,8)$ mode which is able to dissipate more energy than it produces.
Non-normality in the energy balance {#sec:discussion}
===================================
The results in §§\[sec:ECS\] and \[sec:channel\] indicate that the optimal resolvent mode does not accurately account for energy transfer between scales. As we will explain in §\[sec:nonnormality\], the root of this discrepancy is non-normality induced by the mean shear. In §\[sec:counteract\], we show that its influence can be weakened by eddy viscosity. The efficiency of resolvent modes, therefore, is lower than that for eddy modes when reconstructing the energy balance in §\[sec:projection\] using correctly weighted resolvent and eddy modes. Finally, we highlight in §\[sec:aspect ratio\] that the eddy viscosity is most effective for high aspect ratio modes where the influence of non-normality is most pronounced.
Competition between production and nonlinear transfer {#sec:nonnormality}
-----------------------------------------------------
In this section, we demonstrate that non-normality leads to a competition between the production and nonlinear transfer terms in the energy balance. To simplify the discussion, we will only consider the first resolvent mode although similar arguments can be made for suboptimal modes. We begin by rewriting the resolvent norm, or the first singular value $\sigma_1(\boldsymbol{k})$, as $$\label{eq:sigma1}
\sigma_1(\boldsymbol{k}) \approx \frac{1}{(i\omega - \lambda_{min}(\boldsymbol{k}))} \cdot \frac{1}{\int_{-h}^{h} \hat{\boldsymbol{\psi}}_1^*(\boldsymbol{k})\hat{\boldsymbol{\phi}}_1(\boldsymbol{k})dz}.$$ Equation (\[eq:sigma1\]) is a result from [@Symon18] and is applicable for resonant mechanisms such as $k_x = 0$ modes in channel flow or the shedding mode in cylinder flow. In (\[eq:sigma1\]), the resolvent norm is rewritten as the product of two terms. The first is the inverse distance between the imaginary axis and the least stable eigenvalue $\lambda_{min}(\boldsymbol{k})$ of the linear Navier-Stokes operator $\boldsymbol{A}$. The second is a metric of non-normality, originally proposed by [@Chomaz05], which is equivalent to the inverse of nonlinear transfer in (\[eq:res mode balance\]). In parallel shear flows, the lift-up mechanism [@Landahl80] results in component-type non-normality [@Marquet09] so all the energy for $\hat{\boldsymbol{\psi}}_1$ is concentrated in $\hat{u}$ while all the energy for $\hat{\boldsymbol{\phi}}_1$ is concentrated in $\hat{v}$ and $\hat{w}$. Consequently, non-normality results in $\sigma_1$ being very large.
We proceed by designating $C_1 = (i\omega - \lambda_{min})$ and substituting (\[eq:sigma1\]) into (\[eq:budget rank 1\]) to arrive at the following $$\label{eq:non-normality}
\frac{1}{C_1 \int_{-h}^{h} \hat{\boldsymbol{\psi}}_1^*(\boldsymbol{k})\hat{\boldsymbol{\phi}}_1(\boldsymbol{k})} \left( \check{P}(\boldsymbol{k})) + \check{D}(\boldsymbol{k}) \right) +
\underbrace{\int_{-h}^h \boldsymbol{\psi}_1^*(\boldsymbol{k}) \hat{\boldsymbol{\phi}}_1(\boldsymbol{k})dz}_{\hat{N}(\boldsymbol{k})} = 0,$$ where $\check{(\cdot)}$ denotes a quantity that has been normalised by $\sigma_1(\boldsymbol{k})$. Equation (\[eq:non-normality\]) illustrates that as non-normality increases (thus leading to higher amplification), so does the disparity between production and nonlinear transfer. This behaviour is problematic because even for the low Reynolds number flows considered here, we find in the DNS that dissipation is not sufficient to counteract production. Therefore, the nonlinear transfer term cannot be small because it needs to remove a considerable amount of energy for the scale to reach equilibrium. Despite identifying amplification mechanisms, non-normality therefore hinders the efficiency of the optimal resolvent mode in representing the true velocity fluctuations by hindering energy transfer between scales.
As an aside, we note that the same arguments are applicable to cylinder flow for which [@Jin20] found that nonlinear transfer for the first resolvent mode was nearly zero. The primary difference for spatially-developing flows is that convective non-normality [@Chomaz05; @Marquet09; @Symon18] is responsible for minimising $\hat{\boldsymbol{\psi}}_1^* \hat{\boldsymbol{\phi}}_1$. Mean advection in the resolvent operator localises $\hat{\boldsymbol{\psi}}_1$ downstream of the cylinder and $\hat{\boldsymbol{\phi}}_1$ upstream of the cylinder. The non-normality increases amplification by at least one order of magnitude [@Symon18] but leads to a drastic underestimation of nonlinear transfer by the first resolvent mode.
Counteracting non-normality with eddy viscosity {#sec:counteract}
-----------------------------------------------
Returning to the channel flow case, the root of non-normality is the coupling term $-ik_yU'$ of the resolvent operator. Since only the spanwise wavenumber appears in this term, higher aspect ratio structures where $k_x < k_y$ are more prone to amplification. The simplest way to explain the weakening of non-normality by eddy viscosity is that it damps the linear operator. In other words, the effect of viscosity is increased to counteract the mean shear. As a representative example, we plot the first resolvent and eddy mode in figure \[fig:mode comparison\] for $(k_x,k_y) = (0,4)$ and compare it to the DNS, i.e. the true Fourier mode for P4U.
In terms of the streamwise velocity component, the first eddy mode is in better agreement with the DNS than the first resolvent mode. In terms of the spanwise and wall-normal components, however, neither the eddy nor the resolvent mode are in close agreement with the DNS. Nevertheless, the eddy mode has more energy in these velocity components than its resolvent counterpart. Despite having a more complicated structure, the DNS also has more energy in $v$ and $w$, suggesting that the eddy viscosity has sufficiently dampened non-normality to provide a better basis for the flow.
It is worth mentioning that eddy viscosity is not the only way to counteract mean shear. In [@Rosenberg19], a componentwise analysis of the resolvent operator yielded two distinct families of modes which, when correctly weighted, destructively interfere to reduce bias towards the streamwise velocity component. The approach has been applied in [@McMullen20] to higher Reynolds number flows where the destructive interference is more pronounced due to stronger non-normality and higher mean shear.
Resolvent and eddy reconstructions of the energy budget {#sec:projection}
-------------------------------------------------------
Since the first resolvent mode cannot provide the nonlinear transfer that was observed in the DNS, we hypothesise that the role of suboptimal modes in energy transfer is important. We test this hypothesis by reconstructing the energy budget from resolvent modes for P4U since there is a unique wave speed for every wavenumber pair (see appendix \[sec:SPOD\] for the minimal channel case). The weights of resolvent modes $\tilde{\chi}_p(\boldsymbol{k})$ are determined by projecting them onto the velocity field as done in [@Sharma16] $$\label{eq:projection}
\tilde{\chi}_p(\boldsymbol{k}) = \sigma_p(\boldsymbol{k}) \chi_p (\boldsymbol{k})= \hat{\boldsymbol{\psi}}^*_p (\boldsymbol{k}) \hat{\boldsymbol{u}}(\boldsymbol{k}).$$ The approximate velocity field $\hat{\boldsymbol{u}}_a(\boldsymbol{k})$ can be written as $$\label{eq:weights}
\hat{\boldsymbol{u}}_a(\boldsymbol{k}) = \sum_{p=1}^n \tilde{\chi}_p(\boldsymbol{k}) \hat{\boldsymbol{\psi}}_p(\boldsymbol{k}),$$ where $n$ is the number of resolvent modes used in the approximation. The reconstructed energy budget is evaluated as a function of $n$ by replacing $\hat{\boldsymbol{u}}(\boldsymbol{k})$ with $\hat{\boldsymbol{u}}_a(\boldsymbol{k})$ in (\[eq:budget k\]). Since eddy viscosity can lead to improvements, we will also do this for eddy modes by replacing $\hat{\boldsymbol{\psi}}_p(\boldsymbol{k})$ with $\boldsymbol{\psi}^e_p(\boldsymbol{k})$ in (\[eq:weights\]).
In figure \[fig:projection resolvent\](a), the reconstructions are plotted for $(k_x,k_y) = (0,4)$, which is both the most energetic structure in the flow and most amplified by the resolvent. Solid and dotted lines denote resolvent and eddy modes, respectively. Production, which appears as red, requires almost 60 resolvent modes to be adequately captured. Moreover, modes 10 to 30 contribute almost zero net production and therefore modes 30 to 60 are needed. This might seem at odds with the success of the resolvent in identifying sources of production but it is actually consistent. Because the first resolvent mode is strongly biased towards production, the suboptimal modes cannot balance it with nonlinear transfer unless $\tilde{\chi}_1/\sigma_1 \ll 1$, i.e. the projection of $\hat{\boldsymbol{f}}$ onto $\hat{\boldsymbol{\phi}}_1$ is small. Therefore, the nonlinear forcing has to be heavily biased towards suboptimal modes in order to maintain an energy balance. The reconstruction using eddy modes, on the other hand, performs reasonably well with a single mode. Although 30 eddy modes are required to converge to the true value, this is considerably less than the 60 resolvent modes needed. Dissipation (blue) and nonlinear transfer (black) converge more quickly to their true values for both sets of modes.
Resolvent and eddy modes perform equally well as each other for $(k_x,k_y) = (2,0)$ as seen in figure \[fig:projection resolvent\](b). In contrast to $(0,4)$, the spanwise-constant structure has negative production and positive nonlinear transfer. The first eddy mode, which was nearly able to capture all terms in the energy budget, is no longer sufficient to reconstruct any term. Although they converge slightly faster than the resolvent modes, both sets require $n \approx 20$ in order to converge to within 5% of their true values.
We finally consider the oblique wave $(k_x,k_y) = (2,4)$ in figure \[fig:projection resolvent\](c). This structure is less straightforward than the previous two cases. The eddy modes reconstruct production with just 10 modes whereas almost 30 resolvent modes are needed. Dissipation and nonlinear transfer, on the other hand, require more than 30 eddy modes to converge to the true values while they are reconstructed with roughly the same number of resolvent modes needed to reconstruct production ($n \approx 28$). It can be concluded from figure \[fig:projection resolvent\] that the primary benefit of adding eddy viscosity to the operator is in the reconstruction of production. For some scales, however, this might slow down reconstruction of dissipation and nonlinear transfer.
Role of aspect ratio {#sec:aspect ratio}
--------------------
For a more rigorous comparison between eddy and resolvent modes, we choose an error threshold $\mathcal{T}$. The number of modes $n$ is gradually increased until every term in the reconstructed energy balance is within $\mathcal{T}\%$ of the true value. We then compute the difference $\Delta n = n_{res}$ - $n_{edd}$ between the number of resolvent $n_{res}$ and eddy modes $n_{edd}$ required and present the results in figure \[fig:reconstruction error\]. The colour red indicates that eddy modes are more efficient, or $\Delta n > 0$, and thus more resolvent modes must be included to reconstruct the energy budget. Alternatively, the colour blue designates scales for which resolvent modes are more efficient, or $\Delta n < 0$. We also assess the impact of $\mathcal{T}$ on $\Delta n$ by setting the threshold to $\mathcal{T} = 25\%$ in figure \[fig:reconstruction error\](a) and reducing it to $\mathcal{T} = 1\%$ in figure \[fig:reconstruction error\](b).
The first observation that can be made regarding figure \[fig:reconstruction error\] is that the choice of $\mathcal{T}$ has a relatively minor influence on the results. When $\mathcal{T}$ is reduced from 25% to 1%, only five tiles change sign. If we recall that most of the kinetic energy and transfer processes are among $k_x = 0$ and $k_x = 2$ scales, then the only important scale impacted by $\mathcal{T}$ is $(k_x,k_y) = (2,4)$. The reason it is positive for $\mathcal{T} = 25\%$ is that eddy modes more quickly reconstruct the dominant production term as seen in figure \[fig:projection resolvent\](c). It is negative for $\mathcal{T} = 1\%$ because the eddy modes take longer to reconstruct dissipation and nonlinear transfer.
The most striking trend for both values of $\mathcal{T}$ is that the eddy modes are a more efficient basis when $k_x < k_y$. Streamwise-constant structures, in particular, are represented with more than 20 fewer modes if the velocity field is projected onto eddy modes. When we define aspect ratio as $\AR = k_y/k_x$, then $\AR = 3$ is a conservative estimate for when eddy modes are a superior basis to resolvent modes. The pattern is less clear for structures with $\AR = 2$ since the sign of $\Delta n$ is sensitive to the choice of $\mathcal{T}$. Excluding spanwise-constant structures, those with $\AR < 1$ are slightly better represented with resolvent modes although they tend to be weak energetically.
While we only consider P4U in this section, parallel analyses for the minimal channel in appendix \[sec:SPOD\] are consistent with the above. We opt to focus only on P4U for simplicity since the Fourier modes for a particular wavenumber pair are deterministic and travel at a single convection velocity. The details for handling statistical variability and multiple convection velocities are discussed in appendix \[sec:SPOD\].
Conclusions {#sec:conclusions}
===========
We have investigated energy transfer for the P4U ECS and low Reynolds number turbulent flow in a minimal channel. For every wavenumber pair, a balance must be achieved across production, dissipation and nonlinear transfer. For both flows, production for the energetic scales is generally positive and the largest contribution is generated by the near-wall streaks with a spanwise spacing of $\lambda_y^+ \approx 100$. Production is negative for some scales such as those that are spanwise-constant. Dissipation is negative for every scale although it is not sufficiently large to counteract production produced by the most energetic structures. As such, nonlinear transfer redistributes energy to smaller scales through the turbulent cascade and ensures each scale achieves an energy balance. The net effect of nonlinear transfer across all scales is zero but it is negative for the largest scales and positive for the rest. Spanwise-constant structures are among the largest recipients of energy through nonlinear transfer since both their production and dissipation are negative. It is interesting that they play a damping role in the minimal channel since these Tollmien-Schlichting-type waves are the first to become unstable.
Energy transfer in the DNS was compared to predictions from resolvent analysis. Similar to DNS, each mode has to satisfy a balance across production, dissipation and nonlinear transfer. Since the first resolvent mode is often representative of the true velocity field, we computed its energy balance and compared it to DNS. The first resolvent mode was successful in identifying the main production mechanisms in the flow. These are the most amplified structures by resolvent analysis and highlight the role of linear mechanisms in the sustenance of wall-bounded turbulence. For nearly all scales, production was counteracted primarily by dissipation. The nonlinear transfer, consequently, was nearly zero for every scale for resolvent analysis even though it played a major role in redistributing energy in the DNS.
We demonstrated that nonlinear transfer could be modelled by the addition of eddy viscosity, which introduced additional dissipation into the energy balance. Its quantitative accuracy, however, was limited to the most energetic mode and we noted that it can only remove energy, suggesting it less applicable for the many scales that receive energy through nonlinear transfer. The addition of eddy viscosity, nevertheless, had an impact on the number of resolvent modes required to reconstruct the energy budget. To explore this in greater detail, each term in the energy balance was reconstructed as a function of the number of correctly resolvent weighted modes. We only considered the P4U ECS since there was a unique wave speed and deterministic structure for each wavenumber pair to simplify the analysis. We determined that the eddy basis performed better for high aspect ratio structures, particularly with respect to reconstructing production. In some cases, as many as 20 fewer eddy modes than resolvent modes were needed to properly reconstruct all terms in the energy balance. The resolvent basis was slightly better for high aspect ratio structures although these were energetically less significant. A threshold of $\AR = 3$ was a conservative estimate for when eddy modes were more efficient.
We showed that eddy viscosity improved the basis for high aspect ratio structures by counteracting non-normality. This term resulted in a trade-off between production and nonlinear transfer in the energy balance equation for the first resolvent mode. Higher non-normality resulted in more production but less nonlinear transfer. The most amplified mechanisms, which tended to be non-normal, were therefore the most poorly represented by the first resolvent mode. The damping introduced by eddy viscosity mitigated this trade-off and resulted in a better basis for the velocity field. Its refinement for higher Reynolds number flows could improve the potential of linear models for estimation and control.
Acknowledgements
================
The authors wish to thank J. S. Park and M. D. Graham, for providing the P4U solution analysed in this article. The authors are also very grateful to M. Xie and D. Chung for providing the DNS data in the case of the minimal channel. Finally, the authors acknowledge the financial support of the Australian Research Council.
Linear operators {#sec:operators}
================
After elimination of the pressure, the linearized Navier-Stokes equations can be rewritten for the wall-normal velocity $\hat{w}$ and wall-normal vorticity $\hat{\eta} = ik_y\hat{u}-ik_x\hat{v}$). The matrices $\boldsymbol{A}$, $\boldsymbol{B}$, and $\boldsymbol{C}$ that appear in (\[eq:OSSQ\]) are
$$\boldsymbol{A} = \boldsymbol{M}\left[\begin{array}{cc} \mathcal{L}_{OS} & 0 \\ -ik_y U' & \mathcal{L}_{SQ} \end{array} \right],$$
$$\boldsymbol{B} = \boldsymbol{M} \left[\begin{array}{ccc} -i k_x \mathcal{D} & -i k_y \mathcal{D} & -k^2 \\ ik_y & -ik_x & 0 \end{array} \right],$$
$$\boldsymbol{C} = \frac{1}{k^2}\left[\begin{array}{cc} ik_x \mathcal{D} & -ik_y \\ ik_y \mathcal{D} & ik_x \\ k^2 & 0 \end{array} \right].$$
Both $\mathcal{D}$ and $'$ represent differentiation in the wall-normal direction and $k^2 = k_x^2+ k_y^2$. The mass matrix $\boldsymbol{M}$ is defined as $$\boldsymbol{M}(k_x,k_y) = \left[\begin{array}{cc} \Delta^{-1} & 0 \\ 0 & \boldsymbol{I} \end{array} \right],$$ where $\Delta = \mathcal{D}^2-k^2$ and $\boldsymbol{I}$ is the identity matrix. The Orr-Sommerfeld $\mathcal{L}_{OS}$ and Squire $\mathcal{L}_{SQ}$ operators are
$$\mathcal{L}_{OS} = -ik_xU\Delta + ik_xU'' + (1/Re_{\tau})\Delta^2,$$
$$\mathcal{L}_{SQ} = -ik_xU + (1/Re_{\tau}) \Delta.$$
With the addition of eddy viscosity, they become
$$\mathcal{L}_{OS} = -ik_xU\Delta + ik_xU'' + \nu_T\Delta^2 + 2 \nu_T'\mathcal{D}\Delta + \nu_T''(\mathcal{D}^2 + k^2),$$
$$\mathcal{L}_{SQ} = -ik_xU + \nu_T \Delta + \nu_T'\mathcal{D}.$$
Non-normality in the minimal channel case {#sec:SPOD}
=========================================
As alluded to in §\[sec:aspect ratio\], analysis of the minimal channel is complicated by the fact that each wavenumber pair has a distribution of energetic temporal frequencies. For the sake of simplicity, we choose to analyse the most energetic temporal frequency for each wavenumber pair only. The second complication is that, unlike the P4U case, there is no deterministic mode to describe the velocity field or nonlinear forcing at each $\boldsymbol{k}$ in the minimal channel. As discussed by [@Towne18], this stems from the statistical variability in turbulent flows. Instead, we can replace $\hat{\boldsymbol{u}}(\boldsymbol{k})$ in (\[eq:projection\]) with the most energetic mode from spectral proper orthogonal decomposition (SPOD) [@Lumley70; @Picard00] which optimally represents the space-time flow statistics for any $\boldsymbol{k}$.
The SPOD modes are computed using the procedure described in [@Towne18] and [@Muralidhar19]; a brief summary is provided here. Following Welch’s method [@Welch67], the DNS data for a particular $(k_x,k_y)$ are split into overlapping segments containing 128 snapshots with 50% overlap and are Fourier-transformed in time. The data are arranged into the matrix $$\hat{\boldsymbol{Q}}(\boldsymbol{k}) = \left[\begin{array}{cccc} \hat{\boldsymbol{q}}_{\omega}^{(1)} & \hat{\boldsymbol{q}}_{\omega}^{(2)} & \cdots & \hat{\boldsymbol{q}}_{\omega}^{(s)} \end{array} \right] ,$$ where each row represents a different temporal frequency contained in the $s$th segment. The cross-spectral density matrix for a specific wavenumber triplet $\hat{\boldsymbol{S}}(\boldsymbol{k})$ is $$\hat{\boldsymbol{S}}(\boldsymbol{k}) = \hat{\boldsymbol{Q}}(\boldsymbol{k})\hat{\boldsymbol{Q}}^*(\boldsymbol{k}) .$$ The SPOD eigenvectors $\hat{\boldsymbol{V}}(\boldsymbol{k})$ and eigenvalues $\boldsymbol{\Lambda}(\boldsymbol{k})$ can be obtained via an eigenvalue decomposition of the cross-spectral density matrix $$\hat{\boldsymbol{S}}(\boldsymbol{k})\hat{\boldsymbol{V}}(\boldsymbol{k}) = \hat{\boldsymbol{V}}(\boldsymbol{k})\boldsymbol{\Lambda}(\boldsymbol{k}).$$
In figure \[fig:modes minchan\], we compare the most energetic SPOD, resolvent and eddy modes for $\boldsymbol{k} = (0,8,0)$. The streamwise component of SPOD and the eddy mode are nearly identical whereas the resolvent mode predicts a much wider structure. The key improvement, however, is observed for the spanwise and wall-normal components. While the SPOD and eddy modes are not equivalent, the eddy mode is much closer than the resolvent mode in terms of magnitude. Without the presence of eddy viscosity, the first resolvent mode is more biased towards the streamwise velocity component.
Figure \[fig:projection minchan\] shows the resolvent reconstructions for production and dissipation only. Nonlinear transfer is not included since it would require performing SPOD for the nonlinear forcing in addition to the velocity field. Production is the term of interest since it is slowest to converge for P4U. The wavenumber pairs considered in figure \[fig:projection minchan\] reinforce that this observation is not specific to P4U. The resolvent reconstruction of the near wall streaks in figure \[fig:projection minchan\](a) is particularly slow for production, requiring over 100 modes. Similar to the P4U results, a plateau region emerges around $20 < n < 50$ where the addition of resolvent modes has virtually no impact on production. Dissipation, on the other hand, converges around $n = 50$. Only 40 eddy modes, meanwhile, are needed to reconstruct the energy budget. In fact, a substantial portion is captured by the first pair of eddy modes alone, which is consistent with their close resemblance to the SPOD modes in figure \[fig:modes minchan\].
The reconstructions for $(k_x,k_y) = (2,0)$ and $(k_x,k_y) = (2,8)$ in figures \[fig:projection minchan\](b) and (c), respectively, reinforce the efficiency of eddy modes in reconstructing the energy budget. Nevertheless, the first pairs of eddy modes for these two scales are much less effective than they were for $(k_x,k_y) = (0,8)$.
Finally, we consider $\Delta n$ for the minimal channel case in figure \[fig:minchan compare\]. Eddy modes outperform resolvent modes for nearly every wavenumber pair regardless of the threshold $\mathcal{T}$ chosen. Due to the dimensions of the computational domain, however, $\AR > 2$ for nearly every scale that appears in figure \[fig:minchan compare\]. Based on the P4U results, therefore, it is expected that eddy modes will be more efficient.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In this work, we report the excitation of inertial waves in a librating sphere even for libration frequencies where these waves are not directly forced. This spontaneous generation comes from the localized turbulence induced by the centrifugal instabilities in the Ekman boundary layer near the equator and does not depend on the libration frequency. We characterize the key features of these inertial waves in analogy with previous studies of the generation of internal waves in stratified flows from localized turbulent patterns. In particular, the temporal spectrum exhibits preferred values of excited frequency. This first-order phenomenon is generic to any rotating flow in the presence of localized turbulence and is fully relevant for planetary applications.'
bibliography:
- 'biblio.bib'
title: Spontaneous generation of inertial waves from boundary turbulence in a librating sphere
---
geophysical and geological flows, rotating flows
Introduction
============
Rotating fluids support so-called inertial waves, which are associated with the Coriolis force [@kelvin1880]. These waves can be directly excited by various harmonic forcings whose frequencies range between plus and minus twice the spin frequency [@greenspanbook]. For instance, [@aldridge1969] showed the direct forcing of inertial waves by small periodic oscillations of the spinning rate of a rotating sphere (i.e. “librations”) for libration frequencies non-dimensionalized by the mean rotation rate $|\omega_{lib}| \leq 2$ [see also @zhang2013]. The nonlinear self-interaction of the excited inertial waves can lead to strong axisymmetric jets, as observed experimentally for tidal forcing [@morize2010]. Outside this range of frequencies, inertial waves cannot be excited by direct forcing and were never observed except very recently in an axially librating cylinder in the presence of instabilities near the outer boundary [@lopez2011; @sauret2012_pof]. The corresponding mechanism is still controversial and is the subject of the present study. We focus here on [[longitudinal]{}]{} libration in a spherical geometry, which has recently received renewed interest mainly because of planetary applications [e.g. @rambaux2011].
Libration leads to rich dynamics in the contained fluid. At sufficiently large libration amplitudes, centrifugal instabilities are induced near the equator of a sphere where they generate turbulence [@noir2009; @calkins2010]. This instability is generic to any librating container, as for instance in a cylinder [@noir2010; @sauret2012_pof]. Libration also induces a mean zonal flow due to nonlinear interactions in the boundary layers even in the absence of inertial waves [@busse2010a; @sauret2010; @noir2012; @sauret2012_jfm]. For libration frequency $|\omega_{lib}| \leq 2$, the dynamic of the fluid becomes more complicated as the contribution of the inertial waves to the mean flow is non negligible. The energy fed to inertial waves is then localized near Ekman-layer eruptions [@calkins2010; @sauret2012_pof; @koch2013]. Finally, elliptical instability can also be excited by the libration forcing in ellipsoidal containers [@cebronldei].
In this work, we characterize for the first time the inertial waves excited in a sphere by the boundary flow induced by a longitudinal libration for $\omega_{lib}>2$. In section 2, we present the governing equations and the numerical methods used to tackle this problem. Then section 3 is devoted to the description of the main numerical results. Taking advantage of the well-known similarities between rotating and stratified flows [e.g. @veronis1970], those results are finally explained in section 4 in the form of an analytical model extended from closely related studies of internal waves generation from turbulence [@townsend1966; @dohan2003; @dohan2005; @taylor2007].
Governing equations and numerical methods
=========================================
Consider the flow in a sphere of radius $R$, filled with an incompressible, homogeneous and Newtonian fluid of kinematic viscosity $\nu$ and density $\rho$, rotating about the $z$-axis at the mean angular velocity $\Omega_0$. In addition to this mean rotation, the sphere oscillates with an angular frequency $\omega_{wall}$ and amplitude $\Delta \Omega$. Using ${\Omega_0}^{-1}$ and $R$ as time and length scales respectively, the instantaneous angular velocity is given by $$\boldsymbol{\Omega}=[1+\epsilon\,\cos(\omega_{lib}\,t)]\,\boldsymbol{e_z}$$ where $\epsilon=\Delta \Omega/\Omega_0$ and $\omega_{lib}=\omega_{wall}/{\Omega_0}$ are, respectively, the dimensionless libration amplitude and frequency. A schematic of the system is shown in figure \[schematic\_sphere\](a). The dimensionless equations of motion, written in the rotating frame of reference at the mean angular velocity $\Omega_0$, are $$\begin{aligned}
\frac{\partial \boldsymbol{u}}{\partial t}+(\boldsymbol{u}\cdot\boldsymbol{\nabla})\,\boldsymbol{u}+2\,\boldsymbol{e_z}\times \boldsymbol{u}&=&-\boldsymbol{\nabla}p+E\,\boldsymbol{\nabla}^2\,\boldsymbol{u}, \\
\boldsymbol{\nabla}\cdot\boldsymbol{u}&=&0,\end{aligned}$$ where $\boldsymbol{u}$ and $p$ are, respectively, the velocity field in the rotating frame of reference and the reduced pressure which takes into account the centrifugal force. There are three governing parameters in our problem: the Ekman number, $E=\nu/(\Omega_0\,R^2)$, describing the ratio of the viscous effects and the Coriolis force, $\omega_{lib}$ and $\epsilon$. Throughout this study we use the cylindrical polar unit vectors ($\boldsymbol{e_r}$,$\boldsymbol{e_\phi}$,$\boldsymbol{e_z}$). A no-slip boundary condition is used on the outer boundary where $\boldsymbol{u}=\epsilon\,\cos(\omega_{lib}\,t)\,\boldsymbol{e_z} \times \boldsymbol{r}$.
![(a) Schematic of the upper quarter of the sphere with the cylindrical polar vectors ($\boldsymbol{e_r}$,$\boldsymbol{e_\phi}$,$\boldsymbol{e_z}$). (b) Mesh grid used in the numerical model showing two zones: a bulk zone with triangular elements and a boundary-layer domain with quadrangular elements. [[The inset is a zoom of the mesh grid near the outer boundary]{}]{}.[]{data-label="schematic_sphere"}](Figure1.jpg)
An analogous problem of internal wave generation from turbulence in a stratified flow has shown, by comparison with laboratory experiments, that 2D numerical simulations capture the physical mechanism and key features of internal wave generation [@dohan2005]. Therefore, to be able to reach sufficiently small Ekman numbers, we base our study on axisymmetric simulations. We use a commercial finite element code, Comsol Multiphysics. This numerical model has already been successfully used to study similar problems of libration-driven flows in spherical geometries [@sauret2010; @sauret2012_jfm] and cylindrical geometries [@sauret2012_pof]. For more details about the numerical procedure and the validation of the numerical code, we refer the reader to [@sauret2010]. To ensure a high accuracy, all the simulations used in this work are performed with standard Lagrange elements of $P2-P3$ type (i.e. quadratic for the pressure field and cubic for the velocity field), and the number of degrees of freedom (DoF) used in the simulations is typically of order $350\,000$. A typical mesh grid is shown in figure \[schematic\_sphere\](b). [[The mesh grid is composed of $17\,794$ elements, of which $6\,500$ are quadrilateral elements devoted to the mesh in the boundary-layer domain and the others are triangular elements for the bulk. More precisely, the boundary-layer domain has a thickness of $0.034$ along the outer boundary and is discretized in the direction normal to the boundary into $25$ quadrilateral elements with an initial thickness of $5.9\,\times \,10^{-5}$ and a stretching factor of $1.2$. This ensures that we have a sufficient number of mesh elements in the viscous layer, typically around $13$ for the parameters considered in this paper. In addition, the adaptive Backward Differentiation Formula (BDF) order is between $1$ and $5$ and the time-step leads to more than $500$ points per period of libration.]{}]{}
Numerical results
=================
From stable regime to boundary turbulence
-----------------------------------------
In all the following we consider the small Ekman number (${E}\ll1$) and no spin-up regime ($\omega_{lib} \gg \sqrt{E}$), i.e. no spin-up occurs in the bulk at each libration cycle [see e.g. @greenspanbook; @busse2010a]. This situation is satisfied in planetary fluid layers where the Ekman number is typically smaller than $10^{-12}$ and where the main libration components have frequencies of order $1$ [see e.g. @rambaux2011]. As shown in figure \[diagramme\_omega=3\], three flow regimes are observed in the sphere depending on the libration amplitude $\epsilon$ [see also @noir2009; @calkins2010]. For a libration amplitude $\epsilon$ lower than a critical value, called $\epsilon_{TG}$, the flow remains stable and only the viscous layer is visible near the outer boundary; the flow remains laminar everywhere. For $\epsilon > \epsilon_{TG}$ but lower than a critical value, noted $\epsilon_{turb}$, longitudinal rolls, called Taylor-Görtler vortices, develop near the equator along the outer boundary. Then, for an amplitude of libration $\epsilon$ larger than the critical amplitude $\epsilon_{turb}$, the longitudinal rolls turn into a turbulent patch localized around the equator. In this last regime, even if the frequency of libration is such that no inertial waves are directly forced (i.e. $\omega_{lib} >2$), inertial waves are excited in the bulk, emitted from the turbulent patch.
![Radial velocity $u_r$ in a librating sphere ($\omega_{lib}=3$, $E=5\times 10^{-5}$) for different values of the libration amplitude $\epsilon$ (from left to right, $\epsilon=0.3$, $0.55$, $0.85$); $\epsilon_{TG}$ and $\epsilon_{turb}$ are, respectively, the critical value for the appearance of longitude rolls (see the zoom) and boundary turbulence. Spontaneous generation of inertial waves is observed for $\epsilon > \epsilon_{turb}$.[]{data-label="diagramme_omega=3"}](Figure2.jpg)
Characteristics of the generated inertial waves field
-----------------------------------------------------
To study the generated inertial waves, we performed series of simulations during $30$ libration periods (once the permanent regime is reached). An example is shown in figure \[spatio\_temporelle\](a) for $\omega_{lib}=2.1$, $\epsilon=0.8$ and $E=4\times10^{-5}$, together with the space-time diagram of the radial velocity $u_r$ taken at $r=0.2$ in figure \[spatio\_temporelle\](b). The norm of the two-dimensional Fourier transform of this time series is shown in figure \[spatio\_temporelle\](c). It exhibits a patch of large amplitude located at $\omega = 2.1$ and small values of $k_z$. This patch is directly related to the libration forcing at the frequency $\omega_{lib} = 2.1$. It is due to the velocity variations associated with the Ekman pumping at each libration cycle. In addition to this main frequency, another patch is observed around $\omega \sim 1.6$ and $k_z \sim 2 - 4$, which is the signature of the inertial waves emitted in the bulk. The norm of the Fourier transform averaged over $z$ for $r=0.2$ is shown in figure \[spatio\_temporelle\](d). Again, it clearly exhibits a peak located at the libration frequency $\omega=2.1$. In addition, it shows a secondary peak around $\omega=1.4-1.6$ as already observed in the cylinder [see @sauret2012_pof].
 along the dashed line. Apart from the oscillations of the sphere, the propagation of inertial waves is also visible. (c) Corresponding norm of the Fourier transform $P(\omega,k_z)$ showing the axial wavelength and frequency of the generated inertial waves. The colour bar represents $P(\omega,k_z)$. (d) Norm of the Fourier transform of the radial velocity $u_r$ taken at $r=0.2$ and averaged over $z$.[]{data-label="spatio_temporelle"}](Figure3.jpg)
We have performed the same analysis for other libration frequencies: $\omega_{lib}=0.1$ (figures \[fig:fft\_various\]a,b) and $\omega_{lib}=3$ (figures \[fig:fft\_various\]c,d). In both cases, the norm of the Fourier transform taken at $r=0.2$ and averaged over $z$ shows a large peak at the libration frequency $\omega= 0.1$ and $\omega=3$, respectively. For $\omega_{lib}=3$, the situation is close to the previous case with a peak around $\omega \sim 1.5$. For $\omega_{lib}=0.1$ a wider range of frequencies is excited in the bulk that may be related to the extended size of the turbulent patch in the Ekman layer. [[Note also that, in this case, the frequency of libration as well as its higher harmonics $2\,\omega_{lib}$, $3\,\omega_{lib}$, ... allow direct forcing of inertial waves. However, the mechanism described above remains fully generic and is superimposed onto these direct forcings, the only difference being that the spectrum is less localized around $\omega \sim 1.4 - 1.6$]{}]{}.
![Snapshots of the radial velocity field $u_r$ for $E=4\times 10^{-5}$, $\epsilon=0.8$ at (a) $\omega_{lib}=0.1$ and (c) $\omega_{lib}=3$. Corresponding norms of the Fourier transform of the radial velocity averaged over $z$ for $r=0.2$ at (b) $\omega_{lib}=0.1$ and (d) $\omega_{lib}=3$.[]{data-label="fig:fft_various"}](Figure4.jpg)
Comparison with an idealized analytical model
=============================================
Following the seminal study of [@townsend1966] and the more recent work of [@taylor2007] for internal waves, the inertial waves propagation in the bulk of the sphere can be understood by a linear model. For the sake of simplicity, we consider here a two-dimensional model and neglect the curvature of the boundary of the sphere. In his study of internal waves generated at a horizontal interface, [@townsend1966] considered a gaussian perturbation in both space and time along the horizontal plane $z=0$ of the vertical displacement, which is directly related to the vertical velocity. Following the analogy between stratified and rotating flows, it is natural to consider here a perturbation of the radial velocity along the plane $r=1$ and localized around the equator $(r=1, z=0)$. To perform a tractable analysis, we assume this perturbation to be separable in space and time, i.e. $$\label{TT1}
u_r^{p}(z,t)=f(z)\,g(t),$$ where the spatial function $f(z)$ is maximum at $z=0$. In the following, owing to the symmetry of the system with respect to the equator of the sphere, we only consider the half-domain $z>0$. The Fourier transform of the expression (\[TT1\]) leads to $$\begin{aligned}
\hat{u}_r^{p}(k_z,{\omega})& =&\int \int u_r^{p}(z,t)\,\exp\left[-\text{i}\,k_z\,z+\text{i}\,{\omega}\,t\right] \text{d}t \,\text{d}z\, = \hat{f}(k_z)\,\hat{g}({\omega}) \label{eq:cotcot}\end{aligned}$$ where $\mathbf{k}=(k_r,k_z)$ is the wavevector and ${\omega}$ is the frequency. The propagation of this perturbation along the r-direction can be written as : $$u_r(r,z,t)=\frac{1}{4\,\pi^2}\,\int \int \hat{f}(k_z)\,\hat{g}({\omega})\,\text{exp}\left[\text{i}\,(k_r\,(1-r)+k_z\,z-{\omega}\,t)\right] \text{d}{\omega}\, \text{d}k_z
\label{eq:cotcotback1}$$ The dispersion relation for inertial waves links the axial and radial components of the wavevector: $${k_r}=\pm {k_z}\,\sqrt{\frac{4}{{\omega}^2}-1} \label{gen_omega}$$ The radial velocity field then can be written as $$u_r(r,z,t)=\frac{1}{4\,\pi^2}\,\int_{-2}^2\,\hat{g}({\omega})\,\text{e}^{-\text{i}{\omega} t}\,\text{d}{\omega}\,\int \hat{f}(k_z)\,\text{exp}\left[\text{i}\,k_z\left(z\pm (1-r)\,\sqrt{\frac{4}{{\omega}^2}-1}\right)\right] \text{d}k_z,
\label{eq:cotcotback}$$ neglecting the contribution for $|{\omega}|>2$, which corresponds to evanescent waves [@greenspanbook]. The relation (\[eq:cotcotback\]) can be rewritten as $$u_r(r,z,t)=\frac{1}{2\,\pi}\,\int_{-2}^2\,\hat{g}({\omega})\,f\!\left(z\pm (1-r)\,\sqrt{\frac{4}{{\omega}^2}-1}\right)\,\text{e}^{-\text{i}{\omega} t}\,\text{d}{\omega}
\label{eq:cotcotback2}$$ The temporal Fourier transform of the radial velocity at a position ($r$, $z$) is thus $$\hat{u}_r(r,z,{\omega})=\hat{g}({\omega})\,f\!\left(z\pm (1-r)\,\sqrt{\frac{4}{{\omega}^2}-1}\right)
\label{eq:cotcotback3}$$ for $|{\omega}| \leq 2$, with a negligible contribution from $|{\omega}|>2$. This expression shows that the signal in the sphere, at first order and neglecting the viscosity, is the product of the temporal signal of the initial excitation $\hat{g}({\omega})$ and the spatial function $f$. Remembering that this spatial function is localized around $0$, the energy in the $z>0$ space is mainly focused around trajectories $$\label{IW_pos_theo}
z-(1-r)\,\sqrt{\frac{4}{{\omega}^2}-1}=0, \mbox{\quad i.e \quad} {\omega}=\frac{2\,(1-r)}{\sqrt{(1-r)^2+z^2}}.$$ This is readily interpreted geometrically. Starting from a point source of fluctuations, waves of all frequencies are emitted with a given amplitude, which depends on the initial excitation in the turbulent patch. Waves with $|{\omega}| > 2$ are evanescent, while waves with $|{\omega}| \leq 2$ propagate following the dispersion relation ${\omega}=2\,\cos\theta$, where $\theta$ is the angle of propagation [see e.g. @greenspanbook]. This means that a given point $(r,z)$ can only be reached by a wave having the specific frequency given by (\[IW\_pos\_theo\]). This geometrical path is plotted in figure \[fig:fft\_rrrr\](a)-(d) for different radial positions and shows good agreement with the numerical results.
![Norm of the Fourier transform $P({\omega})$ as a function of $z$ for slices at different radial positions: (a) $r=0.2$, (b) $r=0.4$, (c) $r=0.6$ and (d) $r=0.8$. The parameters used in the simulations are $\omega_{lib}=2.1$, $\epsilon=0.7$ and $E=4\times10^{-5}$. The continuous cyan lines represent the geometrical path given by relation (\[IW\_pos\_theo\]), and the dashed cyan lines take into account the axial extension of the turbulent patch following (\[relation\_ff\]). []{data-label="fig:fft_rrrr"}](Figure5.jpg)
To go further, we need the two functions $f$ and $g$, which can be obtained from simulations. The axial extension of the source, $f(z)$, is well modelled by a gaussian fit, as illustrated in figure \[fig:exict\_initial\](a), where we have plotted the maximum amplitude of the fluctuation of the velocity normal to the outer boundary, i.e. the spherical radial velocity. [[Note that considering the time-averaged velocity normal to the outer boundary leads to a similar lateral extension of the turbulent patch. ]{}]{} Taking into account this extension of the source allows one to estimate the extension of the propagating signal around the geometrical path (\[IW\_pos\_theo\]), taking for instance $$\label{relation_ff}
z-(1-r)\,\sqrt{\frac{4}{{\omega}^2}-1}= \pm \sigma,$$ where $\sigma$ is the standard deviation of the gaussian fit. Results are shown in figure \[fig:fft\_rrrr\](a)-(d) and show good agreement with the numerical results.
![(a) Maximum fluctuation of the velocity normal to the outer boundary as a function of $z$. Parameters are: $\epsilon=0.8$, $E=4\times10^{-5}$ and $\omega_{lib}=2.1$. Red squares are the numerical results and the black dashed line is a gaussian fit centered at the equator and of standard deviation $\sigma = 0.2$. (b) Norm of the Fourier transform $P({\omega})$ of the fluctuation of the radial velocity $u_r$ close to the outer boundary ($r=0.98,\,z=0$). Parameters are the same as panel (a). The dashed red line shows the slope ${\omega}^{-3}$. []{data-label="fig:exict_initial"}](Figure6.jpg)
A typical temporal Fourier transform of the excitation is shown in figure \[fig:exict\_initial\](b). At large frequencies, the turbulent patch is well fitted by a slope ${\omega}^{-3}$ characteristic of a two-dimensional turbulence [see e.g. @smith; @bofetta] For ${\omega} \in [0,2]$ corresponding to the propagative waves, the norm of the Fourier transform is nearly constant; hence according to (\[eq:cotcotback3\]), no frequency is a priori preferred in the bulk. Yet, as seen before, the numerical estimation of the norm of the Fourier transform averaged over $z$ exhibits a maximum, whose location depends on the radial position (see fig. \[cotcotback3982\]). This focalization was also observed in a librating cylinder [@sauret2012_pof] for a turbulent patch localized all along the outer boundary and for the generation of internal waves from a turbulent layer [@dohan2005; @taylor2007]. In this latter case, the observed frequency selection was attributed to the influence of viscosity. Adapting their study to inertial waves, the viscous attenuation of a given wave with axial wavenumber $k_z$ and frequency ${\omega}$ propagating in the $r$-direction, is given by$$\label{ohayoo}
\text{exp}\left[ -\frac{16\,E\,|{k_z}|^3}{{\omega}^4\,\sqrt{4-{\omega}^2}}\,(1-r)\right],$$ which, according to the typical values considered here (see e.g. figure \[spatio\_temporelle\](c)), induces a change in the wave amplitude of less than $0.01\%$. Viscous attenuation can clearly not account for the frequency selection shown here, and this will be even more true in planetary applications where $E < 10^{-12}$. In the present case, the finite size and the shape of the spherical container are actually responsible for the selection of preferred frequencies. Indeed, using the numerically determined values of the source functions $f$ and $g$, we can average over $z$ the signal given by the relation (\[eq:cotcotback3\]) at a constant $r$ between the boundaries of the sphere $z=\pm \sqrt{1-r^2}$, which gives $$\begin{aligned}
<\hat{u}_r(r,{\omega})>_z & \propto & \left(\text{erf}\left[\frac{(1-r)\, |k_r/k_z|}{\sigma\,\sqrt{2}}\right]-\text{erf}\left[\frac{(1-r)\,|k_r/k_z|-\sqrt{1-r^2}}{\sigma\,\sqrt{2}}\right]\right)\,\hat{g}({\omega}), \nonumber
\label{eq:cotcotback398} \\\end{aligned}$$ where $\text{erf}$ is the Gauss error function and $|k_r/k_z|=\sqrt{{4}/{{\omega}^2}-1}$ (see eq. \[gen\_omega\]). Figure \[cotcotback3982\] shows the comparison between numerical and analytical results at different radial positions. It exhibits reasonable agreement, taking into account the strong assumptions used in the analytical model. [[Starting from an excitation localized at the equator and with a flat spectrum between $[0,\,2]$, energy is emitted in all directions. But because of geometrical constraints, energy propagation along or close to the direction of the axis of rotation, corresponding to low-frequency waves, is more prevented than propagation along the equator, corresponding to frequencies close to 2. This geometrical frequency selection has a growing importance while going further away from the source, i.e. while $r$ decreases. Hence, while $r$ decreases, selected frequencies are more and more localized close to 2.]{}]{} [[Note also that the mismatch between the numerical results and the analytical model close to $\omega \sim 2.0$ can be explained by our assumptions. The analytical model is based on a flat excited spectrum in the range $\omega\in[0,\,2]$ only. In the numerical simulation, a peak around the frequency $\omega=\omega_{lib}=2.1$ is also present, associated with the spin-up and spin-down close to the outer boundary at each libration cycle (see also Fig. 3).]{}]{}
![Rescaled norm of the Fourier transform $P({\omega})/\text{max}\left(P({\omega})\right)$ for ${\omega} \in[0,\,2]$ of the fluctuation of $u_r$ at different radial locations averaged over $z$. (a) Numerical results [[for $\omega_{lib}=2.1$]{}]{} and (b) analytical model plotted with the relation (\[eq:cotcotback398\]). We have reported the norm of the Fourier transform at $r=0.8$ (blue thick line), $r=0.6$ (green dash-dotted line), $r=0.4$ (red dashed line) and $r=0.2$ (black continuous line).[]{data-label="cotcotback3982"}](Figure7.jpg)
Conclusion
==========
In conclusion, this work presents the first evidence of spontaneous generation of inertial waves from a localized patch of turbulence in a rotating container and of the related mechanism of frequency selection by geometrical constraints. These processes are illustrated here by a simple two-dimensional analytical model and by simple axisymmetric simulations of a librating sphere. This configuration is especially interesting for planetary applications: celestial bodies such as Io and Europa, are indeed thought to present turbulence around their equator driven by libration [@noir2009]. The present mechanism may then participate in the explanation of their preferred eigenmodes of vibration. More generally, we expect the processes presented here to be fully generic to any source of turbulence and to any type of container. A difference must be made between localized and extended patches, as illustrated by the libration of a sphere and of a cylinder for inertial waves. Following the study of [@townsend1966] for internal waves [see also @taylor2007], the resulting signal for an extended source is equal to the superposition of the effects of random localized sources in space and time. In all cases, a focusing of energy at a given frequency is thus to be expected.
Acknowledgments {#acknowledgments .unnumbered}
===============
D.C. acknowledges support from the ETH Zürich Postdoctoral Fellowship Progam and from the Marie Curie Actions for People COFUND Program. M.L.B. acknowledges support from the Marie Curie Actions of the European Commission (FP7-PEOPLE-2011-IOF).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate the influence of confinement on the positional order of a quasi-1D electron system trapped on the surface of liquid helium. We find evidence that the melting of the Wigner solid (WS) depends on the confinement strength, as well as electron density and temperature. A reentrant solid-liquid-solid transition is observed for increasing electron density under constant electrostatic confinement. As the electron row number $N_y$ changes, varying commensurability results in a modulation of the WS order, even when $N_y$ is large (several tens). This is confirmed by Monte Carlo simulations.'
author:
- 'David G. Rees'
- 'Niyaz R. Beysengulov'
- Yoshiaki Teranishi
- 'Chun-Shuo Tsao'
- 'Sheng-Shiuan Yeh'
- 'Shao-Pin Chiu'
- 'Yong-Han Lin'
- 'Dmitrii A. Tayurskii'
- 'Juhn-Jong Lin'
- Kimitoshi Kono
title: 'Structural Order and Melting of a Quasi-One-Dimensional Electron System'
---
I. Introduction
===============
For interacting particles in (quasi-)1D channels, the competing influences of temperature, interaction energy and confinement give rise to many complex phenomena. For Fermi degenerate electrons (or holes), 1D confinement results in Luttinger liquid behaviour[@deshpande2010electron], Wigner crystallisation[@deshpande2008one] and anomalous transport close to the first quantised conductance plateau[@PhysRevLett.77.135]. As the confinement weakens, ‘zig-zag’ transitions from 1 to 2 electron rows can occur[@Hewetal; @chaplik1980instability]. Similar structural transitions can be observed directly in trapped ion experiments[@birkl1992multiple]. For a growing number of particle chains the quasi-1D order depends critically on commensurability, as demonstrated in experiments with colloids[@LeidererLayering], dusty plasmas[@PhysRevLett.90.245004], vortices in superconducting films[@VorticesInChannels] and electrons on liquid helium substrates[@PhysRevLett.109.236802; @*rees2013reentrant], as well as numerical simulations[@Peeters1DCrystal; @PackingAndMelting; @PhysRevB.84.024117]. However, the parameter range explored in such experiments is typically quite narrow, whilst simulations are limited by processor speed. Here, we gain new insights into the ordering and melting of a quasi-1D system of electrons on He by using a multigated microchannel device to tune the particle density and confinement over a wide range. In particular, we find that the strength of the lateral confinement, rather than simply the reduced width, can play a key role in determining the melting behaviour of the quasi-1D electron lattice.
Surface-state electron (SSE) systems on liquid He substrates are ideally suited to the study of strongly correlated electron behaviour[@Andrei]. The typical surface density ($n_s\approx10^{13}$ m$^{-2}$) is low, ensuring that electron-electron interactions are purely Coulombic. In 2D, the electrons form a triangular lattice, the Wigner solid (WS), with increasing $n_s$ or decreasing temperature $T$[@GrimesAdamsWignerCrystal]. Once the electrons become localised, the electrostatic pressure from each electron on the He beneath results in the formation of surface ‘dimples’ that increase the system effective mass[@MonarkhaShikinDimple]. For the moving electron lattice, resonant Bragg-Cherenkov (BC) scattering with surface excitations (ripplons) deepens the dimple lattice and the resistive force exerted on the electron system increases[@DykmanRubo]. The electron velocity is then limited to that of the phase velocity of ripplons whose wavevector is commensurate with one of the reciprocal lattice vectors of the electron lattice (usually the shortest). Hence the Wigner solid transport is strongly nonlinear. When the driving force reaches a critical value, the electrons decouple from the dimple lattice and ‘slide’ along the He surface with high velocity[@PhysRevLett.74.781]. This transport anomaly is a sensitive alert to the Wigner solid formation[@DahmMobillity]. The onset of nonlinearity in the electric conductivity is a more reliable experimental criterion for determining the Wigner solid formation than the conductivity change, which can appear smeared[@rees2013reentrant].
Microchannels filled with liquid He can provide quasi-1D confinement for SSE systems[@PhysRevLett.87.176802; @ReesPRL]. Using such devices, the melting of the Wigner solid was found to be suppressed for small electron row number $N_y$ whilst, separately, reentrant ordering of the quasi-1D lattice was observed with increasing $N_y$[@PhysRevB.82.201104; @*Ikegami2015; @PhysRevLett.109.236802; @*rees2013reentrant]. Here we map structural and phase diagrams for a quasi-1D SSE system, for $1\leq N_y\lesssim 30$. The phase boundary is determined by a unique method, namely measuring the third harmonic component of the SSE current to find the onset of nonlinear transport. We find evidence of a scaled relationship between Coulomb energy, confinement strength and temperature at the Wigner solid melting point, and observe a novel solid-liquid-solid melting behaviour as $n_s$ increases under constant electrostatic confinement. Our experiment demonstrates a uniquely sensitive electrostatic control over the positional order of a quasi-1D electron system, and so is an important step towards utilising SSE in quantum information and quantum optics applications[@PlatzmanandDykmanscience; @LyonQubits; @SchusterProposal].
![(a) False-color scanning electron micrograph of the central microchannel in the device. (b) Schematic diagram of the microchannel cross-section. Gold layers, separated by an insulating layer, are fabricated on a silicon wafer section. (c) $\phi_y$ and $n_s$ in the central microchannel calculated by FEM for $V_{sg}=-0.2$ V and $V_{bg}=+1$ V. \[Fig:2\]](Fig1asmall.eps "fig:"){width="23.50000%"} ![(a) False-color scanning electron micrograph of the central microchannel in the device. (b) Schematic diagram of the microchannel cross-section. Gold layers, separated by an insulating layer, are fabricated on a silicon wafer section. (c) $\phi_y$ and $n_s$ in the central microchannel calculated by FEM for $V_{sg}=-0.2$ V and $V_{bg}=+1$ V. \[Fig:2\]](Fig1bc.eps "fig:"){width="23.50000%"}
II. Results and Discussion
==========================
A. Sample and Method
--------------------
The device, shown in Fig. 1(a), has been described in detail elsewhere[@Beysengulov]. Two metal layers are separated by an insulating layer approximately 2 $\SI{}{\micro\meter}$ thick formed by hard-baked photoresist. The Guard electrode (upper layer) and Reservoir electrodes (lower layer) define two large arrays of microchannels that act as electron reservoirs. The reservoirs are connected by a smaller central microchannel, 100 $\SI{}{\micro\meter}$ long and 7.5 $\SI{}{\micro\meter}$ wide, that is formed by the Split Gate electrode (upper layer) and Bottom Gate electrode (lower layer).
Dc voltages $V_{gu}=-0.2$ V, $V_{res}=0$ V, $V_{bg}$ and $V_{sg}$ were applied to the Guard, Reservoir, Bottom Gate and Split Gate electrodes, respectively. Transport measurements were made by applying an ac voltage $V_{in}$, of frequency $f=20.2$ kHz, to the Left Reservoir electrode and measuring the ac current $I$ induced in the Right Reservoir electrode. (Note that all ac parameters are given in peak-to-peak units.) The circuit was well-described by a lumped-element RC model[@Iye], which was used to extract the SSE resistance $R$. The area of the reservoirs greatly exceeds that of the central microchannel. Therefore, the number of electrons in the reservoirs, and so the electrostatic potential of the electron system $\phi_e$, can be assumed to remain constant whilst the lateral confinement potential $\phi_y$ in the central microchannel is controlled by changing $V_{bg}$ and $V_{sg}$ (Fig. 1(c)). The electron density in the reservoirs was kept low to prevent Wigner solid formation, a crucial advantage over previous experiments[@PhysRevLett.109.236802; @*rees2013reentrant]. Finite element modelling (FEM) was used to calculate $\phi_y$, the average $n_s$ and the effective width of the electron system $w_e$ in the central microchannel for all values of $V_{bg}$ and $V_{sg}$, using $\phi_e$ and the channel depth $h$ as fitting parameters[@Beysengulov; @hecht2012new]. $N_y$ was then estimated as $N_y=w_e\sqrt{n_s}$. Values of the angular frequency $\omega$, which describes strength of the lateral confinement, were found by fitting the parabolic function $\phi_y e=\frac{1}{2}m_e\omega^2y^2$ to the FEM results in the central region of the central microchannel, where $m_e$ is the bare electron mass.
B. Experimental Results
-----------------------
The magnitude of $I$ for varying $V_{bg}$ and $V_{sg}$ is shown in Fig. 2(a), for $V_{in}=3$ mV and $T=0.6$ K. The threshold for conductance through the central microchannel depends on both electrode voltages. As in other similar devices[@ReesPRL; @Beysengulov], electrons can enter the central microchannel when the maximum of the potential at its centre (or minimum, for electrons) $\phi_y^{max}=\alpha V_{bg}+\beta V_{sg}$ exceeds $\phi_e$. Examination of the conductance threshold in Fig. 2(a) yields the values $\alpha=0.60$ and $\beta=0.40$. These values are reproduced by the FEM for $h=2.20$ $\SI{}{\micro\meter}$.
![(a) Magnitude of $I$ against $V_{bg}$ and $V_{sg}$, at $T=0.6$ K. Here $V_{in}=3$ mV. (b) Magnitude of $I_{3f}$ measured under the same conditions, but with $V_{in}=10$ mV. The measurement noise floor was 5 pA. The solid line indicates the boundary between the electron liquid (EL) and WS phases. Regions corresponding to the EL phase and the sliding and dynamically pinned WS phases are labelled accordingly, as are lines representing the conditions $\Gamma=130$ and $\xi^+=3w_e$. \[Fig:4\]](Fig2a.EPS "fig:"){width="45.00000%"} ![(a) Magnitude of $I$ against $V_{bg}$ and $V_{sg}$, at $T=0.6$ K. Here $V_{in}=3$ mV. (b) Magnitude of $I_{3f}$ measured under the same conditions, but with $V_{in}=10$ mV. The measurement noise floor was 5 pA. The solid line indicates the boundary between the electron liquid (EL) and WS phases. Regions corresponding to the EL phase and the sliding and dynamically pinned WS phases are labelled accordingly, as are lines representing the conditions $\Gamma=130$ and $\xi^+=3w_e$. \[Fig:4\]](Fig2b.EPS "fig:"){width="45.00000%"}
From the current measurement shown in Fig. 2(a), for each point in the $V_{bg}$-$V_{sg}$ plane, it is not straightforward to determine whether the SSE system is in the electron liquid or Wigner solid phase. This is because there is no sharp contrast between high and low current regions. Although the current should drop when the electron system becomes dynamically pinned to the dimple lattice, the sinusoidal driving voltage can induce the decoupling the dimple lattice during each ac cycle, if the pinning effect is not sufficiently strong[@StickSlip]. In this case the measured current magnitude may not be greatly changed from that measured in the electron liquid phase. However, the nonlinear response in this transport regime gives rise to higher odd harmonics in the ac current signal[@ThirdHarmonic]. Therefore, in order to precisely determine the boundary between the liquid and solid phases, the third harmonic component of the SSE current, $I_{3f}$, was measured for varying $V_{bg}$ and $V_{sg}$. The result is shown in Fig. 2(b), for $V_{in}=10$ mV and $T=0.6$ K. In the upper left-hand corner of the plot, where $n_s$ is generally low, the ac response is linear signifying the electron liquid regime. In the Wigner solid regime, distinct regions in which $I_{3f}$ is high and low are evident. These correspond to the sliding and dynamically pinned transport regimes, respectively. In the sliding regime, the decoupling of the electron system from the dimple lattice during each ac current cycle leads to large $I_{3f}$. The decoupling occurs more readily close to the conduction threshold, where $n_s$ is low. In the dynamically pinned regime, in which $n_s$ is higher, the driving force cannot induce the decoupling and BC scattering limits the electron velocity during each ac cycle. This also results in a nonlinear SSE response, but with a reduced current magnitude and so smaller $I_{3f}$.
In 2D, the Wigner solid melts when the value of $\Gamma$, which is defined as the ratio of the electron Coulomb energy $U_C=e^2\sqrt{\pi n_s}/4\pi\varepsilon_0$ to kinetic energy $k_B T$, falls below a critical value of $\Gamma^{2\text{D}}\approx130$. Here $e$, $\varepsilon_0$ and $k_B$ are the elementary charge, vacuum permittivity and the Boltzmann constant, respectively. The melting is a Kosterlitz-Thouless (KT)-type transition, occurring due to the unbinding of lattice dislocation pairs[@KosterlitzThouless1973; @HalperinNelson1978; @Young1979; @Morf1979]. In a small temperature range above the melting temperature there exists a ‘hexatic’ phase, in which the system exhibits nearest-neighbour bond-orientational order. This order is finally destroyed at a higher temperature by the emergence of a second type of defect, lattice disclinations, and the system enters the isotropic liquid phase. In this work we define melting as the loss of long-range positional order due to the appearance of free dislocations because, for electrons on helium, the dimple lattice formation and associated transport properties depend on this positional ordering. The distance over which positional ordering can be expected is described by the correlation length $\xi^+=ae^{b/t^{\nu}}$ where $a\approx n_s^{-0.5}$ is the dislocation core size, $b=1.8$ is the ratio of the core energy to the thermal energy, $\nu=0.37$ is a constant and $t=\frac{\Gamma^{2D}}{\Gamma}-1$. The correlation length is finite in the electron liquid phase and diverges as $\Gamma$ approaches $\Gamma^{2\text{D}}$. However, for quasi-1D systems, an ordered state might be expected when $\xi^+$ exceeds the system width $w_e$, which can occur for $\Gamma<\Gamma^{2\text{D}}$. Recently, the melting of a quasi-1D SSE system was found to be described by the empirical relation $\xi^+=3w_e$[@Ikegami2015]. We note that a strictly quantitative comparison between experiment and theory is difficult in this case because the values of $b$ and $\nu$ are valid only when $t\leq0.07$.
![(a) Phase boundaries determined by measuring $I_{3f}$, at $T=0.6$, 0.8 and 1.0 K. The solid line represents the threshold for conductance through the central microchannel. Along the dashed line $\Delta V$ is constant. (b) $U_C/\hbar\omega$ against $\Gamma$ for the data shown in (a). The dotted line represents the 2D melting criterion. The dashed line is equivalent to the dashed line in (a), for $T=1$ K.\[Fig:4\]](Fig3a.EPS "fig:"){width="45.00000%"} ![(a) Phase boundaries determined by measuring $I_{3f}$, at $T=0.6$, 0.8 and 1.0 K. The solid line represents the threshold for conductance through the central microchannel. Along the dashed line $\Delta V$ is constant. (b) $U_C/\hbar\omega$ against $\Gamma$ for the data shown in (a). The dotted line represents the 2D melting criterion. The dashed line is equivalent to the dashed line in (a), for $T=1$ K.\[Fig:4\]](Fig3b.eps "fig:"){width="45.00000%"}
Lines representing the 2D and quasi-1D melting criteria are shown in Fig. 2(b). (In our calculations the screening of the Coulomb interaction by the underlying electrode, which modifies $U_C$ by a few per cent, is taken into account[@PhysRevB.82.201104; @*Ikegami2015].) When $V_{sg}$ is close to $V_{gu}$ the boundary between the electron liquid and solid regimes is close to the 2D melting criterion, confirming the KT-type melting. However, for more negative $V_{sg}$, the electron system is solid within (to the left of) both the $\Gamma=130$ and $\xi^+=3w_e$ boundaries. We conclude that electron ordering in quasi-1D generally occurs for values of $\Gamma$ below 130 and, in our sample, is found to be more strongly enhanced than in previous experiments where the phase boundary followed the condition $\xi^+=3w_e$[@Ikegami2015].
Phase boundaries for $T=0.6$, 0.8 and 1.0 K are shown in Fig. 3(a). The boundaries were determined by finding the values of $V_{bg}$ and $V_{sg}$ for which $I_{3f}$ exceeds the measurement noise floor, after adjacent-point averaging over a 50 mV window. The liquid region expands as $T$ increases, as expected. To elucidate the relationship between $U_C$, $\omega$ and $T$ at the phase boundaries, we show in Fig. 3(b) the dimensionless ratio $U_C/\hbar\omega$ against $\Gamma$. (It is convenient to express the confinement in units of energy although no quantum mechanical effects are considered here). The data points fall close to a single curve, suggesting that $U_C$, $\omega$ and $T$ are interrelated at the Wigner solid melting point. When $U_C/\hbar\omega$ is large the melting occurs close to the 2D criterion $\Gamma=130$. As $U_C/\hbar\omega$ decreases, due to increasing confinement strength or decreasing electron density, the Wigner solid melts at values of $\Gamma$ much lower than 130. This observation is in agreement with numerical simulations, which have shown that strong electrostatic confinement restricts lateral particle motion and thereby suppresses the melting of the quasi-1D Wigner solid[@Peeters1DCrystal; @PackingAndMelting]. The dependence of melting on confinement strength is qualitatively different from finite-size effects considered in other studies[@PhysRevB.82.201104; @*Ikegami2015]. We conclude that the KT-type melting of the Wigner solid can be significantly modified by strong lateral confinement, as it promotes the positional order of the electron system. To our knowledge, this is the first evidence of a scaled relationship between interaction energy, confinement strength and temperature that describes the melting of a quasi-1D system.
The interplay between $U_C$, $\omega$ and $T$ at the Wigner solid melting point results in a curvature of the phase boundaries shown in Figs. 3(a) and 3(b). Lines drawn on these plots for constant confinement voltage $\Delta V=V_{bg}-V_{sg}$, and therefore constant $\omega$, can intersect the phase boundary twice, for a given temperature. Moving along these lines, by increasing $n_s$ under the constant electrostatic confinement, therefore results in a reentrant solid-liquid-solid transition. We are aware of no previous demonstration of such behaviour, but consider it likely to be observed in other strongly correlated quasi-1D systems subjected to parabolic-like confinement.
![d$R$/d$V_{bg}$ against $V_{bg}$ and $V_{sg}$ at $T=0.6$ K and $V_{in}=3$ mV. In the hatched region $I$ becomes too small to measure. The dashed lines represent constant values of $N_y$ as given by the FEM analysis. The solid line represents the liquid-solid boundary determined by the measurement of $I_{3f}$, as also shown in Fig. 2(b).\[Fig:2\]](Fig4.EPS){width="45.00000%"}
In Fig. 2, both $I$ and $I_{3f}$ exhibit fringe-like features that lie close to parallel with the conductance threshold. Previous studies have shown that such oscillatory transport behaviour arises due to the modulation of the electron lattice structural order with changing $N_y$[@PhysRevLett.109.236802; @*rees2013reentrant]. Close to the structural transitions between adjacent $N_y$, lattice defects and fluctuations between lattice configurations of similar energy should occur[@Peeters1DCrystal]. The reduced positional order weakens the Bragg ripplon scattering, and so the effective electron mass and the SSE resistivity. A plot of d$R$/d$V_{bg}$ for varying $V_{bg}$ and $V_{sg}$ is shown in Fig. 4. Oscillatory resistance features follow arc-like paths in the $V_{bg}$-$V_{gt}$ plane. No resistance oscillations are recorded in the electron liquid region. Several lines corresponding to constant $N_y$ values, as calculated by the FEM for $\phi_e=-0.151$ V, are also shown. These lines closely follow the $R$ maxima, confirming that the resistance oscillations are related to the changing $N_y$. However, this effect was previously only observed close to the melting point of the Wigner solid and for small $N_y$[@PhysRevLett.109.236802; @*rees2013reentrant]. Here we demonstrate that structural transitions strongly influence the electron positional order up to large $N_y$ and for $\Gamma \gg 130$.
The solid-liquid-solid transition that occurs when $n_s$ increases for certain values of $\Delta V$ results in the loss of the resistance oscillations for intermediate values of $N_y$. In Fig. 5 we show both $R$ and $I_{3f}$ recorded along the line in the $V_{bg}$-$V_{sg}$ plane for which $\Delta V=0.55$ V, at $T=0.6$ K. In both measurements, signatures of electron ordering are exhibited for small and large $N_y$ but not for intermediate values $5 \lesssim N_y \lesssim 13$, confirming the reentrant behaviour. The close agreement between the two independent measurements, performed for different $V_{in}$, confirms that the phase boundary is accurately determined in our transport measurements.
![$R$ and the magnitude of $I_{3f}$ recorded along the line $\Delta V=V_{bg}-V_{sg} = 0.55$ V, at $T=0.6$ K. The electron liquid and Wigner solid regimes, and resistance peaks corresponding to several $N_y$, are labelled. \[Fig:2\]](Fig5.EPS){width="45.00000%"}
C. Monte Carlo Simulations
--------------------------
To better understand the nature of the structural transitions that occur as $N_y$ increases, we performed Monte Carlo simulations of electrons interacting via a screened Coulomb potential under parabolic confinement. Techniques similar to those described in Ref.[@Peeters1DCrystal] were used, but extended to larger $N_y$. We consider $N$ classical particles in 2 dimensional space interacting with each other through a Yukawa potential. The particles are confined in the $y$ direction by a parabolic potential, and there is no confinement along the $x$ axis. Periodic boundary conditions are inserted in the $x$ direction to simulate an infinite length in $x$. We solved the Langevin equation with a friction force proportional to the velocity and a temperature dependent random force. The ground state structures at zero temperature were searched for by the annealing method. The number of particles required to achieve the smallest energy per particle were found by repeating the annealing simulation for various $N$, with the box length adjusted to keep the linear density constant. In this way the proper number of particles to obtain the lowest-energy structures could be determined. We used the scaled Hamiltonian used in Ref.[@Peeters1DCrystal], with $\kappa =7.25$. This corresponds to a screening length $\lambda=1$ $\SI{}{\micro\meter}$ for an electron system under parabolic confinement $\hbar\omega/k_B=0.6$ K. These values are comparable with the experimental conditions.
![(a) $y_i$ against $n_l$ as calculated by Monte Carlo simulations. (b) Electron lattice configuration for $n_l=13$ $\SI{}{\micro\meter}^{-1}$. The simulation cell contains 84 particles. (c) Electron lattice configuration for $n_l=23.5$ $\SI{}{\micro\meter}^{-1}$. The simulation cell contains 141 particles. Electrons exhibiting 5 nearest neighbours are marked by triangles.\[Fig:3\]](Fig6a.EPS "fig:"){width="50.00000%"} ![(a) $y_i$ against $n_l$ as calculated by Monte Carlo simulations. (b) Electron lattice configuration for $n_l=13$ $\SI{}{\micro\meter}^{-1}$. The simulation cell contains 84 particles. (c) Electron lattice configuration for $n_l=23.5$ $\SI{}{\micro\meter}^{-1}$. The simulation cell contains 141 particles. Electrons exhibiting 5 nearest neighbours are marked by triangles.\[Fig:3\]](Fig6b.EPS "fig:"){width="18.97500%"} ![(a) $y_i$ against $n_l$ as calculated by Monte Carlo simulations. (b) Electron lattice configuration for $n_l=13$ $\SI{}{\micro\meter}^{-1}$. The simulation cell contains 84 particles. (c) Electron lattice configuration for $n_l=23.5$ $\SI{}{\micro\meter}^{-1}$. The simulation cell contains 141 particles. Electrons exhibiting 5 nearest neighbours are marked by triangles.\[Fig:3\]](Fig6c.EPS "fig:"){width="25.30000%"}
Lateral particle positions $y_i$ against linear electron density $n_l$ are shown in Fig. 6(a). Generally, the electrons are arranged in rows and $N_y$ increases sequentially. An example of a highly ordered row structure, found for $n_l=13$ $\SI{}{\micro\meter}^{-1}$, is shown in Fig. 6(b). However, close to values of $n_l$ at which $N_y$ changes, an increased scatter in $y_i$ reflects reduced positional order. An example is shown in Fig. 6(c) for $n_l=15.4$ $\SI{}{\micro\meter}^{-1}$; the electron lattice becomes distorted and domains containing 6 or 7 rows appear, along with structural defects that break the six-fold symmetry of the electron lattice. This behaviour is in agreement with other similar studies made for small values of $N_y$[@PackingAndMelting]. The increased disorder observed here occurs at each transition and up to the largest $N_y$ values, supporting our explanation that the resistance oscillations observed in the experiment reflect changes in $N_y$ as $V_{bg}$ or $V_{sg}$ is varied.
III. CONCLUSION
===============
We have used a multigated microchannel device to map the first structural and phase diagrams for a quasi-1D electron system on a liquid He substrate. The KT-type Wigner solid melting is strongly modified by the lateral electrostatic confinement. We have demonstrated for the first time that the melting of the quasi-1D Wigner solid is determined by a scaled relationship between Coulomb energy, temperature and the confinement strength. In addition we have shown that the positional order of electrons in the quasi-1D Wigner solid depends strongly on the commensurability with the confinement geometry, even when the number of electron rows is large. This observation was confirmed using Monte Carlo simulations. Because electrons on helium are a model system, our results are relevant to a wide variety of micro- and macroscopic many body systems.
ACKNOWLEDGEMENTS
================
We thank A. D. Chepelianskii and Yu. Lysogorskiy for helpful discussions. This work was supported by JSPS KAKENHI Grant No. JP24000007, and by the Taiwan Ministry of Science and Technology (MOST) through Grant Nos. MOST 103-2112-M-009-001, MOST 103-2112-M-009-017 and MOST 102-2112-M-009-014-MY2, and by the MOE ATU Program. This work was performed according to the Russian Government Program of Competitive Growth of Kazan Federal University.
[37]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.77.135) [****, ()](\doibase
10.1103/PhysRevLett.102.056804) @noop [**** ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.97.208302) [****, ()](\doibase
10.1103/PhysRevLett.90.245004) [****, ()](\doibase 10.1103/PhysRevLett.88.247004) [****, ()](\doibase
10.1103/PhysRevLett.109.236802) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.69.045324) [****, ()](\doibase 10.1103/PhysRevE.82.041504) [****, ()](\doibase 10.1103/PhysRevB.84.024117) , ed., @noop [**]{} (, ) [****, ()](\doibase 10.1103/PhysRevLett.42.795) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.78.4813) [****, ()](\doibase 10.1103/PhysRevLett.74.781) [****, ()](\doibase 10.1103/PhysRevLett.48.641) [****, ()](\doibase
10.1103/PhysRevLett.87.176802) [****, ()](\doibase
10.1103/PhysRevLett.106.026803) [****, ()](\doibase 10.1103/PhysRevB.82.201104) [****, ()](\doibase 10.1007/s10909-014-1272-8) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.74.052338) [****, ()](\doibase 10.1103/PhysRevLett.105.040503) [****, ()](\doibase
10.1007/s10909-015-1344-4) [****, ()](\doibase 10.1007/BF00119515) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.116.206801) [****, ()](\doibase 10.1103/PhysRevLett.28.1129) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.41.121) [****, ()](\doibase 10.1103/PhysRevB.19.1855) [****, ()](\doibase 10.1103/PhysRevLett.43.931)
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'R. P. Mignani'
- 'C. Motch'
- 'F. Haberl'
- 'S. Zane'
- 'R. Turolla'
- 'A. Schwope'
date: 'Received ...; accepted ...'
title: |
VLT optical observations of the isolated neutron star\
[^1]
---
Introduction
============
X-ray observations performed with the [[*Röntgen Satellite*]{}]{} ([[*ROSAT*]{}]{}) yielded to the identification of a group of seven radio-silent (Kondriatev et al. 2008)[^2] Isolated Neutron Stars (INSs). Their relatively dim X-ray emission ($L_X\approx 10^{30}$–$10^{31}$ erg$\,{\rm s}^{-1}$) originally earned them the nickname of X-ray Dim INSs, or XDINSs (see Haberl 2007; van Kerkwijk & Kaplan 2007, and Kaplan 2008 for recent reviews). Recently, a new XDINS candidate has been identified in archival [[*XMM-Newton*]{}]{} observations (Pires et al. 2009). XDINSs have purely thermal X-ray spectra which are best represented by a blackbody ($kT\approx 50$–100 eV), as expected for middle-aged ($\sim 1$ Myr) cooling INSs, whose emission radius is consistent with a sizable fraction of the neutron star surface. The derived hydrogen column densities $N_{\rm H} \approx 10^{20}$ cm$^{-2}$ suggest distances $<$ 500 pc (Posselt et al. 2007), as confirmed in two cases by their optical parallaxes (e.g. van Kerkwijk & Kaplan 2007). X-ray pulsations ($P=3$–12 s) have been detected for all of them (Haberl et al. 1997, 1999; Haberl & Zavlin 2002; Hambaryan et al. 2002; Zane et al. 2005; Tiengo & Mereghetti 2007) but , although with different pulsed fractions. The measurement of the period derivative $\dot{P}$ (Cropper et al. 2004; Kaplan & van Kerkwijk 2005a; Kaplan & van Kerkwijk 2005b; van Kerkwijk & Kaplan 2008) yielded spin-down ages of $\sim 1.5-3.8$ Myrs and rotational energy losses $\dot {E} \sim (3-5) \times 10^{30}$ erg s$^{-1}$. Broad absorption features ($E_{\mathrm{line}}\approx 0.2$–0.7 keV) have been observed in all XDINSs but (Haberl et al. 2003, 2004; van Kerkvijk et al. 2004; Zane et al. 2005), superimposed to the thermal continuum. These features are likely due to proton cyclotron and/or bound-free, bound-bound transitions in H, H-like and He-like atoms. The inferred magnetic fields of $\sim 10^{13}-10^{14}$ G are consistent with the values derived from the neutron star spin down and suggest that XDINSs might be (evolutionary) linked to other class of INSs, the magnetar candidates (see Mereghetti 2008 for a recent review) and the Rotating Radio Transients (e.g. Popov, Turolla, Possenti 2006).\
In the optical, only (Walter & Matthews 1997; Walter 2001), (Motch & Haberl 1998; Kulkarni & van Kerkwijk; 1998; Motch et al. 2003) and (Kaplan et al. 2003; Motch et al. 2005; Zane et al. 2006) have counterparts certified by their proper motion measurements, while likely candidates have been proposed for (Kaplan et al. 2002) and (Zane et al. 2008; Schwope et al. 2009) based on their coincidence with the X-ray positions. Apart from providing a clear evidence of the optical identification, proper motion measurements are important to obtain an estimate of the kinematic age of the neutron star, to be compared with the characteristic age derived from the spin-down. The XDINS optical fluxes usually exceed by a factor of $\sim 5$ (or more) the extrapolation of the X-ray blackbody, and their optical spectra, when measured, seem to follow a Rayleigh-Jeans distribution (e.g., Kaplan 2008). The XDINS optical emission has been interpreted either in terms of a non-homegeneous surface temperature distribution, with the cooler part emitting the optical (e.g., Pons et al. 2002), or of reprocessing of the surface radiation by a thin H atmosphere around a bare neutron star (Zane et al. 2004; Ho 2007), or of non-thermal emission from particles in the star magnetosphere (Motch et al. 2003). However, for the measured $\dot{E}$, magnetospheric emission would not be detectable, at least if an average optical emission efficiency of rotation–powered neutron stars (e.g. Zharikov et al. 2006) is assumed. Alternatively, like for the magnetars, optical magnetospheric emission might be powered by the neutron star magnetic field, as proposed for (Zane et al. 2008).\
One of the XDINSs without a certified optical counterpart is . The first optical observations of the field performed with the [[*New Technology Telescope*]{}]{} ([[*NTT*]{}]{}) soon after the discovery of the X-ray source (Haberl et al. 1999) did not reveal any candidate counterpart brighter than B$\sim25.2$ and R$\sim25.2$. More recently, thanks to the updated [[*Chandra*]{}]{} position, a possible optical identification was proposed by Haberl et al. (2004) with a faint object (B$=26.6 \pm 0.3$, V$\ge 25.5$) tentatively detected on archival [[*Very Large Telescope*]{}]{} ([[*VLT*]{}]{}) images. However, the identification has not been confirmed so far. The field of was also observed in the near-infrared (NIR) with the [[*VLT*]{}]{} but no candidate counterpart was detected down to H$\sim 21.7$ (Mignani et al. 2007; Lo Curto et al. 2007; Posselt et al. 2009) and K$_s \sim 21.5$ (Mignani et al. 2008).\
In this paper we re-analyze the original [[*VLT*]{}]{} observations of presented by Haberl et al. (2004) and we report on follow-up, longer optical observations of the candidate counterpart, performed by our team with the [[*VLT*]{}]{}. Observations and data analysis are described in Sect. 2, while results are presented and discussed in Sect. 3 and Sect. 4, respectively.
Observations
============
Observation description
-----------------------
Optical observations of were performed in service mode with the [[*VLT*]{}]{} at the ESO Paranal observatory on November 21st 2000, on November 25th 2006, January 16th and 22nd, and February 11th 2007 (see Tab. 1 for a summary). The 2000 observations were performed with [[*FOcal Reducer/low dispersion Spectrograph*]{}]{} ([[*FORS1*]{}]{}), a multi-mode camera for imaging and long-slit/multi-object spectroscopy, as part of the ESO guaranteed time programme. At the epoch of the observations [[*FORS1*]{}]{} was equipped with the original four port 2048$\times$2084 CCD detector and it was mounted at the [[*VLT*]{}]{} Antu telescope. The observations were performed in standard resolution mode, with a 02 pixel size and a field of view of 6$\farcm8 \times 6\farcm8$. The low gain, fast read-out, single port mode was chosen. A sequence of three 1200 s exposures was obtained through the Bessel B filter, with an airmass of $\sim 1.14$, an image quality of $\sim 0\farcs7$, and dark time conditions. Since the seeing values measured by the differential image motion monitor (DIMM) are relative to the zenith and not to the pointing direction of the telescope they are not necessarily indicative of the actual image quality. We thus computed the actual image quality from the measured point spread function (PSF), derived by fitting the full width half maximum (FWHM) of a number of well-suited field stars using the [*Sextractor*]{} tool (Bertin & Arnouts 1996), as documented in the [[*FORS1*]{}]{} data quality control pages[^3].\
Two 600 s additional exposures were obtained in the V filter but since the proposed candidate counterpart was not detected we focus our analysis on the B-band data only. Sky conditions were reported to be photometric (see Haberl et al. 2004 for a more detailed observations description). A very bright star, CD-50 1353 ($B=9.9$, as listed in [*Simbad*]{}), located at $\approx 45\arcsec$ from the position of was partially masked using the [[*FORS1*]{}]{} occulting bars. Bias, twilight flat–fields frames, and images of the standard star fields SA 92 and Rubin 149 (Landolt 1992) were obtained as part of the [[*FORS1*]{}]{} science calibration plan.
Date N T (s) IQ ($\arcsec$) Airmass
------- ------------ ---- ------- ---------------- ---------
FORS1 2000-11-21 3 3600 0.74 (0.08) 1.14
FORS2 2006-11-25 5 2915 0.71 (0.10) 1.24
2007-01-16 10 5830 0.93 (0.18) 1.18
2007-01-22 5 2915 0.82 (0.11) 1.19
2007-02-11 5 2915 0.82 (0.12) 1.30
: Log of the [[*VLT*]{}]{} [[*FORS1*]{}]{} and [[*FORS2*]{}]{} B-band observations of ). Columns report the observing date (yyyy-mm-dd), the number of exposures (N) and the total integration time per night (T), the image quality (IQ) and rms (in parentheses), as computed on the image, and the airmass. Values are the average computed over the exposure sequence.
\[data\]
The 2006/2007 observations were performed with [[*FORS2*]{}]{} as part of the ESO open time programme. At the epoch of the observations, [[*FORS2*]{}]{} had swapped with [[*FORS1*]{}]{} at the [[*VLT*]{}]{} Antu telescope. [[*FORS2*]{}]{} is equipped with two 2k$\times$4k MIT CCD detectors. Due to vignetting, the effective sky coverage of the two detectors is smaller than the projected detector field of view, and it is larger for the upper CCD chip. Observations were performed in high resolution mode, with a 2$\times$2 binning and a pixel size of 0$\farcs$125. The low gain, fast read-out mode was chosen. The telescope pointing was set in order to position in the upper CCD chip to include a larger number of reference stars for a precise image astrometry thanks to its larger effective sky coverage ($3\farcm5 \times 2\arcmin$). Sequences of 580 s exposures were obtained through the Bessel B filter. The bright star CD-50 1353 was more efficiently masked both by positioning it at the centre of the gap between the two chips and by using the [[*FORS2*]{}]{} occulting bars. Unfortunately, the distance of from the gap ($\sim 17\arcsec$) and the width of the occulting bars ($\sim 25\arcsec$), which can move along one CCD direction only, made it impossible to mask completely the star halo. Exposures were taken in dark time and under mostly clear but not perfectly photometric sky conditions. In particular, the night of February 11th was affected by the presence of thin variable cirri. Atmospheric conditions were not optimal either. The first two nights were affected by a strong wind, close to the telescope pointing limit, while the nights of January 16th and February 11th were affected by 40 % humidity. Unfortunately, although foreseen by the instrument science calibration plan, only for some nights both day and night time calibration frames were taken. In particular, no twilight flat–fields were taken for the night of November 25th, while B-band standard star images (of the Rubin 152 field) were taken on the night of February 11th only.
Data reduction and calibration
------------------------------
We retrieved the [[*FORS1*]{}]{} science images from the public ESO archive[^4] and we reduced them using tools available in [[*MIDAS*]{}]{} for bias subtraction, and flat–field correction. The same reduction steps were applied to the [[*FORS2*]{}]{} science images through the ESO [[*FORS2*]{}]{} data reduction pipeline[^5]. We searched the archive for suitable twilight flat–fields to reduce the [[*FORS2*]{}]{} November 25th science images but the closest in time were those associated with our January 16th images. We evaluated the possibility of using lamp flat–fields as backup calibration frames, with the caveat that they are affected by reflections produced by the instrument atmospheric dispersion correctors. However, since lamp flat–fields are only taken for trending purposes, and the high resolution mode is not the standard one for [[*FORS2*]{}]{}, no suitable data was found in the archive. Thus, since the November 25th science images can not be calibrated with twilight flat–fields taken on the same night, initially we do not use them in the subsequent analysis. For both the [[*FORS1*]{}]{} and [[*FORS2*]{}]{} data sets the photometric calibration was applied using the available, extinction corrected, night zero points available through the instrument data quality control database[^6]. For the January 16th and 22nd [[*FORS2*]{}]{} observations, for which no standard star images were taken, we assumed as a zero point the value extrapolated from the night zero point trend. Since none of the [[*FORS2*]{}]{} observations was taken in perfectly photometric conditions, we estimated that a relative photometry calibration, with the February 11th observations taken as a reference, would introduce an uncertainty comparable to that associated to the extrapolation of the zero point trend. We converted the trended [[*FORS1*]{}]{} and [[*FORS2*]{}]{} zero points, computed in units of electrons/s, to units of ADU/s by applying the corresponding electrons–to–ADU conversion factors. For each of the two data sets, we then used the [*MIDAS*]{} task [average/window]{} to cosmic rays filter and stack single exposures.
{height="5.9cm"} {height="5.9cm"} {height="5.9cm"}
Astrometry
----------
As a reference for the astrometric calibration we used the [[*GSC-2*]{}]{} version 2.3 (Lasker et al. 2008). Approximately 70 [[*GSC-2*]{}]{} objects are identified in the 2000 [[*FORS1*]{}]{} image. From this list we filtered out extended objects, stars that are either saturated or too faint to be used as reliable astrometric calibrators or too close to the CCD edges. We finally performed our astrometric calibration using 30 well-suited [[*GSC-2*]{}]{} reference stars, evenly distributed in the [[*FORS1*]{}]{} field of view. The pixel coordinates of the selected [[*GSC-2*]{}]{} stars were measured by fitting their intensity profiles with a Gaussian function using the dedicated tool of the [*Graphical Astronomy and Image Analysis*]{} ([*GAIA*]{}) interface[^7]. The coordinate transformation between the detector and the celestial reference frame was then computed using the [*Starlink*]{} package [ASTROM]{}[^8] using higher order polynomials to accounts for the CCD distortions. The rms of the astrometric solution turned out to be $\approx$ 02, accounting for the rms of the fit in the right ascension and declination components. Following Lattanzi et al. (1997), we estimated the overall uncertainty of our astrometry by adding in quadrature the rms of the astrometric fit and the precision with which we can register our field on the [[*GSC-2*]{}]{} reference frame. This is estimated as $\sqrt 3 \times \sigma_{GSC} / \sqrt N_{s}$, where the $\sqrt 3$ term accounts for the free parameters (x-scale, y-scale, and rotation angle) in the astrometric fit, $\sigma_{GSC}$ is the mean positional error of the [[*GSC-2*]{}]{} coordinates (03, Lasker et al. 2008) and $N_{s}$ is the number of stars used for the astrometric calibration. The uncertainty on the reference stars centroids is below 001 and was neglected. We also added in quadrature the 015 uncertainty (Lasker et al. 2008) on the tie of the [[*GSC-2*]{}]{} to the International Celestial Reference Frame (ICRF). Thus, the overall accuracy of the [[*FORS1*]{}]{} astrometry is 027 ($1\sigma$). The astrometric calibration of the [[*FORS2*]{}]{} image was computed in the same way but with a lower number of reference stars due to the smaller field of view of the [[*FORS2*]{}]{} chip. The rms of the astrometric fit then turned out to be 036. Again, after accounting for systematic uncertainties (see above) the overall accuracy of the [[*FORS2*]{}]{} astrometry is 043 ($1\sigma$).\
As a reference to compute the position we considered X-ray coordinates derived from [[*Chandra*]{}]{} observations which are closest in time to our [[*VLT*]{}]{} observations. In particular, for the [[*FORS1*]{}]{} observations (epoch 2000.89) we used the [[*Chandra*]{}]{} coordinates (epoch 2002.86) published in Haberl et al. (2004), i.e. $\alpha_{J2000}=04^h 20^m 01.95^s$, $\delta_{J2000}= -50^\circ 22\arcmin 48\farcs1$ which have a nominal error of 06 (90% confidence level). For the [[*FORS2*]{}]{} observations (epoch 2007.04) we reanalysed a more recent [[*Chandra*]{}]{} observations (epoch 2005.85). Like in Haberl et al. (2004), we determined the source position with the [*CIAO*]{} task [celldetect]{} and we obtained $\alpha_{J2000}=04^h 20^m 01.94^s$, $\delta_{J2000}= -50^\circ 22\arcmin 48\farcs2$ (06; 90% confidence level). As shown in Haberl et al. (2004) a match between the coordinates of the X-ray sources detected in the [[*Chandra*]{}]{} field with those of their possible [*USNOB.10*]{} counterpart did not reveal any significant systematic shift. No significant shift is found between the coordinates of the same X-ray sources between the 2002 and 2007 [[*Chandra*]{}]{} observations either. Thus, no boresight correction was applied to our reference coordinates. Recently, an upper limit on the proper motion (123 mas yr$^{-1}$, $2 \sigma$) was obtained with [[*Chandra*]{}]{} (Motch et al. 2009). We accounted for the proper motion uncertainty when we registered the reference [[*Chandra*]{}]{} coordinates on the [[*FORS1*]{}]{} and on the [[*FORS2*]{}]{} images. This yields an additional position uncertainty due to the unknown proper motion of $\sim 0\farcs123$ and $\sim 0\farcs073$ ($1 \sigma$) for each of the two images, respectively. The overall uncertainty to be attached to the position at the epoch of the [[*FORS1*]{}]{} and [[*FORS2*]{}]{} observations was finally obtained by adding in quadrature the error on the [[*Chandra*]{}]{} coordinates ($1 \sigma$), the coordinate uncertainty due to the proper motion, and the overall error of the astrometric calibration. This yields to uncertainties of 087 and 111 (90% confidence level) on the position on the [[*FORS1*]{}]{} and on the [[*FORS2*]{}]{} image, respectively.
{height="5.9cm"} {height="5.9cm"} {height="5.9cm"}
Results
=======
The [[*FORS1*]{}]{} observations
--------------------------------
We first re–analyzed the [[*VLT*]{}]{} observations taken in 2000 to better assess the confidence of the optical identification of proposed in Haberl et al. (2004). Fig. 1 (left) shows the computed [[*Chandra*]{}]{} position of overlaied on a cutout of the 2000 [[*FORS1*]{}]{} co–added B-band image. As seen from Fig. 1 (left), only the bright PSF core of star CD-50 1353 is masked, while its halo extends close to the target position. This increases the local sky background as well as the background noise, which results in a larger number of spurious detections. In order to enhance the detection significance for fainter objects we tried to minimise the effects of the halo of star CD-50 1353 on the local sky background. Firstly, we fitted the sky background in an area of $\sim 15\arcsec \times 15\arcsec$ around the target position using a second order polynomial and we subtracted the fitted value from the co–added B-band image using the [*MIDAS*]{} task [fit/flat\_sky]{}. We warn here that the fit to the sky background is biased by the choice of the sampling areas. This can yield to more or less evident feature enhancements when the sky background subtraction is applied to the image. We thus carefully choose the sampling areas not to introduce systematic effects in our procedure. Fig. 1 (middle) shows a zoom of the sky-subtracted image. As already shown by Haberl et al. (2004), four objects are clearly detected close to the [[*Chandra*]{}]{} position. In addition, a very faint feature is possibly recognised within the [[*Chandra*]{}]{} error circle. We identify this feature with object $I$ of Haberl et al. (2004), which they tentatively proposed as a candidate counterpart to . However, the excess of counts at the feature position is comparable to the rms of the local sky background, which corresponds to a very low detection significance of $\approx 2 \sigma$. We re-computed the magnitude of the feature through PSF photometry. We derived the PSF parameters from a set of several non saturated objects selected for their stellar like profiles, located close to the , and spanning a large range of magnitudes. The airmass correction was applied using the Paranal extinction coefficients measured with [[*FORS1*]{}]{}[^9]. We found $B=27.52 \pm 0.61$. This is fainter than the value of $B=26.57 \pm 0.30$ reported in Haberl et al. (2004) but it is still compatible at the $1 \sigma$ level when systematic uncertainties in their photometry are taken into account (see section 3 of Haberl et al.). For a better visualisation, we smoothed the image using a Gaussian filter over cells of $3\times3$ pixels i.e. of size comparable to that of the image PSF. Since the image smoothing enhances the detection of very faint objects but also that of fluctuations of the noisy sky background we fine-tuned the smoothing parameters not to produce an over-enhancement of background features. The result is shown in Fig. 1 (right). However, the image processing (sky subtraction and smoothing) does not single out object $I$ against the many, similarly significant, background features recognised around the [[*Chandra*]{}]{} position. Thus, we can not rule out that object $I$ was a spurious detection due to the high background noise induced by the halo of star CD-50 1353.
The [[*FORS2*]{}]{} observations
--------------------------------
We used our follow-up [[*FORS2*]{}]{} observations to search for a higher confidence candidate counterpart to . In order to minimise the effects of the halo of star CD-50 1353, we first co–added only the exposures taken with an image quality better than $1\arcsec$. In first place, we used the co-addition of the best image quality exposures of all nights with the exception of those taken on November 25th 2006, which were calibrated using twilight flat–fields taken about 40 days apart (see Sect 2.2).\
Fig. 2 (left) shows the computed [[*Chandra*]{}]{} position of overlaid on a cutout of the [[*FORS2*]{}]{} B-band image (6960 s) obtained from the co-addition of the twelve best image quality (08-09) January/February 2007 exposures. Indeed, although the more efficient masking reduced the contamination from the halo of star CD-50 1353, the sky background at the [[*Chandra*]{}]{} position remained significantly affected by scattered light. As we did in Sect. 3.1, we fitted and subtracted the sky background from the co–added image. In order to increase the S/N ratio per pixel we then rebinned the sky-subtracted image by a factor of 2, ending up with a pixel size of 025 which well matches that of [[*FORS1*]{}]{} (02). We did not find evidence for object $I$, the feature tentatively proposed by Haberl et al. (2004) as a candidate counterpart to . However, we possibly recognised a second feature within the [[*Chandra*]{}]{} error circle, $\approx 0\farcs5$ north of the expected position of object $I$. Unfortunately, the low number of counts only yields to a marginal detection significance ($\sim 3 \sigma$). As a test, and being aware of possible issues related to the non optimal flat–fielding, we decided to use the five November 2006 exposures (2915 s) which happen to have the best image quality (07) in the [[*FORS2*]{}]{} data set. As done for the January/February data set, we fitted and subtracted the sky background from the co–added image and we rebinned the sky-subtracted image by a factor of 2. Interestingly, a feature appears right at the same position of that seen in the co–addition of the January/February best image quality exposures, although with only a $\sim 2.5 \sigma$ detection significance. While we do not claim that this is a strong detection evidence, it is quite unusual that a background feature appears at the same position in images taken weeks apart. To increase the S/N ratio, we both co–added all the twenty 2007 exposures (11680 s) and all the available exposures (14575 s), again applying sky-subtraction and rebinning, and we obtained a detection significance of $\sim 3.5 \sigma$ and $\approx 3.9 \sigma$, respectively. A zoom of the longest integration time, co–added image is shown in Fig. 2 (middle), where the feature detected in the [[*Chandra*]{}]{} error circle is labelled $X$. For a better visualisation, we smoothed the image using a Gaussian filter over cells of $5\times5$ pixels (Fig. 2, right).\
[llll]{}\
ID & $\alpha_{J2000} ^{(hms)}$ & $\delta_{J2000}^{(\circ ~'~")}$ & $B$\
\
X & 04 20 01.94 & -50 22 47.75 & 27.5 $\pm$ 0.3\
A & 04 20 02.10 &-50 22 42.60 & 24.35$\pm$0.05\
D & 04 20 01.59 &-50 22 43.17 & 24.95$\pm$0.05\
E & 04 20 01.57 &-50 22 48.75 & 25.37$\pm$0.07\
H & 04 20 02.28 &-50 22 46.30 & 26.49$\pm$0.10\
\[data\]
As done in Sect. 3.1, we measured the flux of object $X$ through PSF photometry. The airmass correction was applied using the Paranal extinction coefficients measured with [[*FORS2*]{}]{}[^10]. Due to the still low S/N in the aperture and to the noisy sky background the flux measurement is obviously affected by a large error. Our best estimate gives $B=27.5 \pm 0.3$, where the statistical error obviously dominates over the uncertainty of our absolute photometry (Sect. 2.2). The object magnitude and coordinates are listed in Table 2 together with those of the other objects identified in Fig. 2, as a reference. Interestingly, the flux of object $X$ coincides with that of object $I$ ($B=27.52 \pm 0.61$), the candidate counterpart tentatively proposed by Haberl et al. (2004), which we re-computed in Sect. 3.1. One may thus speculate whether we detected the same feature both in the [[*FORS2*]{}]{} and in the [[*FORS1*]{}]{} images, although at slightly different positions. The measured angular separation between object $X$ and object $I$ is $0\farcs5 \pm 0\farcs3$, accounting for an estimated uncertainty of one pixel on the object centroid in both the [[*FORS1*]{}]{} and [[*FORS2*]{}]{} images. This would imply a yearly displacement of $80 \pm 50$ mas yr$^{-1}$, consistent with the upper limit on the proper motion (Motch et al. 2009). The yearly displacement would thus imply a transverse velocity of $\approx$ 140 $d_{350}$ km s$^{-1}$, where $d_{350}$ is the neutron star distance in units of 350 pc (Posselt et al. 2007), i.e. within the range of the tangential velocities inferred for neutron stars. The actual proper motion measurement, to be eventually obtained with [[*Chandra*]{}]{}, will unambiguously address this speculation.
Discussion
==========
![ Upper panel: blackbody temperature $T_{o}$ as a function of the emission radius $r_{o}$ for different values of the optical excess $f$. The red, blue, and green lines correspond to an optical excess of $f = 7, 4$ and 2, respectively. For each value of $f$, the three curves are drawn for different values of the neutron star distance, 550, 350 and 200 pc (dashed, dot-dashed, and solid lines, respectively). Lower panel: relative contribution $R$ to the total 0.1-1keV X-ray flux of a blackbody with temperature $T_{o}$ versus the radius of the emitting region $r_{o}$ for different values of the optical excess $f$ and of the source distance. The line style and colour coding is the same as in the upper panel. The horizontal dotted line corresponds to the threshold $R=0.1$. The allowed region in the parameters space lies below this line. []{data-label="spec"}](12600fig8.ps "fig:"){height="6cm"} ![ Upper panel: blackbody temperature $T_{o}$ as a function of the emission radius $r_{o}$ for different values of the optical excess $f$. The red, blue, and green lines correspond to an optical excess of $f = 7, 4$ and 2, respectively. For each value of $f$, the three curves are drawn for different values of the neutron star distance, 550, 350 and 200 pc (dashed, dot-dashed, and solid lines, respectively). Lower panel: relative contribution $R$ to the total 0.1-1keV X-ray flux of a blackbody with temperature $T_{o}$ versus the radius of the emitting region $r_{o}$ for different values of the optical excess $f$ and of the source distance. The line style and colour coding is the same as in the upper panel. The horizontal dotted line corresponds to the threshold $R=0.1$. The allowed region in the parameters space lies below this line. []{data-label="spec"}](12600fig9.ps "fig:"){height="6cm"}
The very marginal detection significance ($\sim 3.9 \sigma$) of object $X$ against the number of local spurious detections makes it difficult to determine whether or not it is real and, thus, whether or not we have detected a candidate optical counterpart to . We compared the flux of the putative candidate counterpart with the extrapolation in the optical domain of the models which best fit the [[*XMM-Newton*]{}]{} [*EPIC-pn*]{} spectrum of . To this aim, we have re-analysed the original data of Haberl et al. (2004) using updated calibration files. The spectrum can be fit by a single blackbody with temperature $kT_{X} = 46.2 \pm 1.4$ eV and N$_{\rm H}$=$(0.73 \pm 0.21) \times 10^{20}$ cm$^{-2}$ (reduced $\chi^{2}=2.08$, 64 d.o.f.), corresponding to an emission radius $r_{X} = 5.11 ~ d_{350}$ km, where $r_{X}$ is the X-ray emission as seen from infinity and $d_{350}$ is the neutron star distance in units of 350 pc (Posselt et al. 2007). However, a blackbody with $kT_{X} = 47.8 \pm 2.2$ eV and N$_{\rm H}$=$1.19^{+0.45}_{-0.31} \times 10^{20}$ cm$^{-2}$ ($r_{X} = 5.44 ~ d_{350}$ km) plus an absorption line with centroid energy $E_{line} = 337 \pm 24$ eV and equivalent width $EW_{line} = 47 \pm 5$ eV gives a better fit (reduced $\chi^{2}=1.33$, 62 d.o.f.). The line width $\sigma_{line}$) was fixed at 70 eV, as in Haberl et al. (2004). For the spectral fits we used element abundances both from Anders & Grevesse (1989) and Wilms et al. (2000), obtaining virtually the same results. The best-fit, absorption-corrected X-ray spectra of are shown in Fig. 3 together with the optical flux of its putative counterpart. We corrected for the absorption in the B band using as a reference the N$_{\rm H}$ derived from the best-fit X-ray spectral model (blackbody plus absorption line) and applying the relation of Predehl & Schmitt (1995) with the extinction coefficients of Fitzpatrick (1999). From the flux of the putative counterpart we can rule out a $> 7$ optical excess with respect to the extrapolation of the [[*XMM-Newton*]{}]{} spectrum. We note that an optical excess of $\sim 5$ is usually observed in other optically identified XDINSs with the exception of and , where it is as large as $\approx 15$ (Motch et al. 2005) and $\approx 30-40$ (Zane et al. 2008; Schwope et al. 2009), respectively.\
As a limit case, we checked whether an optical excess of $\sim 7$ would be compatible with either rotation-powered emission from the neutron star magnetosphere or with thermal emission from a fraction of the neutron star surface, colder and larger than that responsible for the X-ray emission. In the first case, the value of the X-ray period and the upper limit on the period derivative of ($P=3.45$s; $\dot {P} < 92 \times 10^{-13}$s s$^{-1}$; see Haberl 2007) only yield a rotational energy loss $\dot {E} < 8.8 \times 10^{33}$ erg s$^{-1}$. The flux of the putative counterpart would imply an optical luminosity $L_{B} \sim 1.2 \times 10^{27}$ erg s$^{-1} ~ d_{350}^2$. This would correspond to an emission efficiency $\eta_{B} \equiv L_{B}/\dot{E} > 1.3 \times 10^{-7}$, which could still be compatible with the values expected for $10^{6}-10^{7}$ years old neutron stars (Zharikov et al. 2006). However, a period derivative $\dot{P} \sim 10^{-13}$s s$^{-1}$, comparable to that of other XDINSs, would imply a factor of 100 lower $\dot{E}$ and would make it less likely that the optical emission is powered by the rotational energy loss. In the second case, we can constrain both the blackbody temperature $T_{o}$ and the emission radius $r_{o}$, as seen from infinity. Since the fit to the [[*XMM-Newton*]{}]{} spectrum does not require the presence of a second blackbody component at lower temperature, we can impose that its relative contribution $R$ to the total X-ray flux in the 0.1-1 keV band (see Sect. 3 of Zane et al. 2008) must be $<<1$. We chose $R=0.1$ as a reasonable threshold. We first computed the values of $T_o$ for a grid of values of $r_{o}$ and for different values of the optical excess $f=\frac {r_{o}^{2} ~T_{o} }{r_{X}^{2} ~T_{X}}$ and of the source distance (Fig. 4, upper panel), where $T_{X} = 47.8 $ eV and $r_{X} = 5.44 ~ d_{350}$ km are derived from the best X-ray spectral fit (blackbody plus absoption line, see above). We then computed $R$ from the values of $r_{o}$ and $T_{o}$ (Fig. 4, lower panel). As it is seen, for a neutron star distance of 350 pc an optical excess of $\sim 7$ would be compatible with a blackbody with $kT_{o} \leq 25$ eV and an an implausibly large emitting radius of $r_{o} \geq$ 23 km. Thus, would our putative counterpart be confirmed, an optical excess of $\sim 7$, for a neutron star distance of $\sim 350$ pc, might rather point towards a non-thermal origin for the optical emission, as proposed for RBS\~ 1774 (Zane et al. 2008). Actually, as Fig. 4 shows, unphysical large radii are required even if the actual counterpart is dimmer, $f\sim 2$–4, unless the neutron star is at $\la 200$ pc.
Conclusions
===========
We carefully re-analysed archival [[*VLT*]{}]{}/[[*FORS1*]{}]{} observations of the field of the XDINS , taken in 2000, and we performed deeper follow-up observations with [[*FORS2*]{}]{} in 2006 and in 2007. With a measured detection significance of $\sim 2 \sigma$ and a re-computed flux of $B=27.52 \pm 0.61$, we can not rule out that the candidate counterpart tentatively detected in the [[*FORS1*]{}]{} images by Haberl et al. (2004) was a feature of the noisy sky background, produced by the very bright nearby star CD-50 1353. While we could not confirm this detection in our deeper [[*FORS2*]{}]{} images, we detected an apparently new feature ($B=27.5 \pm 0.3$) within the updated [[*Chandra*]{}]{} error circle of , $\approx 0\farcs5$ north of the expected position of that detected in the [[*FORS1*]{}]{} images. Interestingly, both their similar flux and their angular separation, compatible with the upper limit on the proper motion, suggest that we might have actually detected the same feature both in the [[*FORS1*]{}]{} and in the [[*FORS2*]{}]{} images. However, its still marginal detection significance ($\sim 3.9 \sigma$) makes it difficult to determine whether the latter feature is associated with a real object, and thus it is the candidate counterpart, or it is also a possible background feature. From the flux of the putative counterpart we can rule out a $> 7$ optical excess with respect to the extrapolation of the [[*XMM-Newton*]{}]{} spectrum. An optical excess of $\sim 7$ (or lower) could be compatible either with rotation-powered emission from the neutron star magnetosphere or with thermal emission from the neutron star surface for a distance $\la 200$ pc , i.e. much lower than the current best estimate of $\sim$ 350 pc (Posselt et al. 2007).\
More observations are required to confirm the detection of the putative candidate counterpart and to validate its identification with . Unfortunately, the presence of star CD-50 1353 severely hampers ground-based follow-up observations, even if performed under sub-arcsec seeing conditions and using a very careful masking. High spatial resolution observations with the refurbished [[*Hubble Space Telescope*]{}]{} ([[*HST*]{}]{}), possibly to be performed in the ultraviolet, are the only way to settle the identification issue.
RPM acknowledges STFC for support through a Rolling Grant and thanks S. Moheler (ESO) for reducing our observations through the ESO data reduction pipeline and C. Izzo (ESO) for technical details. SZ acknowledges STFC for support through an Advanced Fellowship. We thank the anonymous referee for his/her comments to the manuscript.
[99]{}
Bertin, E. & Arnouts, S., 1996, A. & A. Suppl., 117, 393
Cropper, M., Haberl, F., Zane, S., Zavlin, V. E., 2004, MNRAS, 351, 1099
Fitzpatrick, E.L., 1999, PASP, 111, 63
Grevesse, N. & Anders, E., 1989, AIPC, 183, 1
Haberl, F., Motch, C., Buckley, D. A. H., Zickgraf, F.-J., Pietsch, W., 1997, A&A, 326, 662
Haberl, F., Pietsch, W., Motch, C., 1999, A&A, 351, L53
Haberl, F., Zavlin, V.E., 2002, A&A, 391, 571
Haberl, F., Schwope, A. D., Hambaryan, V., Hasinger, G., Motch, C., 2003, A&A 403, L19
Haberl, F., Zavlin, V. E., Trümper, J., Burwitz, V., 2004, A&A 419, 1077
Haberl, F., Motch, C., Zavlin, V. E., et al., 2004, A&A, 424, 635
Haberl, F., 2007, Ap&SS, 308, 181
Hambarayan, V., Hasinger, G., Schwope, A.D., Schulz, N.S., 2002, A&A, 381, 98
Ho, W.C.G., et al. 2007, MNRAS, 375, 821
Kaplan, D.L., Kulkarni, S.R., van Kerkwijk, M.H., 2002, ApJ, 579, L29
Kaplan, D.L., Kulkarni, S.R., van Kerkwijk, M.H., 2003, ApJ, 588, L33
Kaplan, D.L., van Kerkwijk, M.H., 2005a, ApJ, 628, L45
Kaplan, D.L., van Kerkwijk, M.H., 2005b, ApJ, 635, L65
Kaplan, D. L., 2008, Proc. of “40 Years of Pulsars: Millisecond Pulsars, Magnetars, and More”, AIP, 983, 331
Kondratiev, V. I., Burgay, M., Possenti, A., et al., 2008, Proc. of “40 Years of Pulsars: Millisecond Pulsars, Magnetars, and More”, AIP, 983, 348
Kulkarni, S. R. & van Kerkwijk, M. H., 1998, ApJ, 507, L49
Lasker, B., Lattanzi, M. G., McLean, B. J., Bucciarelli, B., Drimmel, R., et al., 2008, AJ, 136, 735
Lattanzi, M. G., Capetti, A., Macchetto, F. D. 1997, A&A, 318, 997
Lo Curto, G., Mignani, R. P., Perna, R., et al. 2007, A&A 473, 539
Malofeev, V. M., Malov, O. I., Teplykh, D. A., 2007, Ap&SS, 308, 211
McLaughlin, M.A., Lyne, A.G., Lorimer, D.R. et al., 2006, Nature, 439, 817
Mereghetti, S., 2008, A&A Rev., 15, 225
Mignani, R., Bagnulo, S., de Luca, A., et al., 2007a, AP&SS, 308, 203
Mignani, R.P., Falomo, R., Moretti, A., Treves, A., Turolla, R., et al., 2008, A&A, 488, 267
Motch, C., Haberl, F., 1998, A&A, 333, L59
Motch, C., Zavlin, V.e., Haberl, F., 2003, A&A, 408, 323
Motch, C., Sekiguchi, K., Haberl, F., Zavlin, V. E., et al., 2005, A&A, 429, 257
Motch, C., Pires, A. M., Haberl, F., Schwope, A., Zavlin, V.E., 2009, A&A, 497, 423
Perna, R., Hernquist, L., Narayan, R., 2000, ApJ, 541, 344
Pires, A. M., Motch, C., Turolla, R., Treves, A., Popov, S. B., 2009, A&A, 498, 233
Pons, J.A., Walter, F.M., Lattimer, J.M., Prakash, M., Neuäuser, R., An, P. 2002, ApJ, 564, 981
Popov, S. B., Turolla, R. & Possenti, A. 2006, MNRAS, 369, L23
Posselt, B., Popov, S. B., Haberl, F., et al., 2007, Ap&SS, 308, 171
Posselt, B., Neuhäuser, R., Haberl, F., 2009, A&A, 496, 533
Predehl, P. & Schmitt, J.H.M.M. 1995, A&A 293, 889
Schwope, A., Erben, T., Kohnert, J., 2009, A&A, 499, 267
Tiengo, A., Mereghetti, S., 2007, ApJ, 657, L101
Walter, F.M., & Matthews, L.D., 1997, Nature, 389, 358
Walter, F.M., 2001, ApJ, 549, 433
van Kerkwijk, M.H., Kulkarni, S.R., 2001, A&A, 378, 986
van Kerkwijk, M. H., et al., 2004, 608, 432
van Kerkwijk, M. H., Kaplan, D. L., 2007, Ap&SS, 308, 191
van Kerkwijk, M. H. & Kaplan, D. L., 2008, ApJ, 673, L163
Wilms, J., Allen, A., McCray, R., 2000, ApJ, 542, 914
Zane, S.,Turolla, R., Drake, J.J., et al, 2004, AdSpR, 33, pp. 531-536.
Zane, S., Cropper, M., Turolla, R., et al., 2005, ApJ, 627, 397
Zane, S., De Luca, A., Mignani, R.P., Turolla, R., 2006, A&A, 457, 619
Zane, S., Mignani, R.P., Turolla, R., et al., 2008, ApJ, 682, 487
Zharikov, S., Shibanov, Yu., Komarova, V., 2006, AdSpR, 37, 1979
[^1]: Based on observations collected at ESO, Paranal, under Programmes 66.D-0128(A), 078.D-0162(A)
[^2]: The claimed low-frequency pulsed emission from two of them (Malofeev et al. 2007) has not been confirmed yet.
[^3]: http://www.eso.org/observing/dfo/quality/FORS1/qc/qc1.html
[^4]: http://archive.eso.org
[^5]: http://www.eso.org/observing/dfo/quality/FORS2/pipeline
[^6]: http://www.eso.org/observing/dfo/quality/FORS2/qc/qc1.html
[^7]: star-www.dur.ac.uk/ pdraper/gaia/gaia.html
[^8]: http://star-www.rl.ac.uk/Software/software.htm
[^9]: http://www.eso.org/observing/dfo/quality/FORS1/qc/qc1.html
[^10]: http://www.eso.org/observing/dfo/quality/FORS2/qc/qc1.html
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Renormalization factors for three- and four-quark operators, which appear in the low energy effective Lagrangian of the proton decay and the weak interactions, are perturbatively calculated in domain-wall QCD. We find that the operators are multiplicatively renormalizable up to one-loop level without mixing with any other operators that have different chiral structures. As an application, we evaluate a renormalization factor for $B_K$ at the parameters where previous simulations have been performed, and find one-loop corrections to $B_K$ are 1-5% in these cases.'
address: |
$^1$Institute of Physics, University of Tsukuba, Tsukuba, Ibaraki 305-8571, Japan\
$^2$Department of Physics, Washington University, St. Louis, Missouri 63130, USA\
author:
- '$^1$Sinya Aoki, $^1$Taku Izubuchi, $^{2}$Yoshinobu Kuramashi [^1] and $^1$Yusuke Taniguchi'
title: |
[hep-th/9902008]{}\
[UTHEP-398]{}\
Perturbative renormalization factors of three- and four-quark operators for domain-wall QCD
---
\#1\#2
Introduction
============
Calculation of hadron matrix elements of phenomenological interest represents an inevitable application of lattice QCD. In the past decade much efforts have been devoted for the calculation of three- and four-quark hadron matrix elements relevant to the proton decay amplitude and the weak interaction ones using the Wilson and the Kogut-Susskind(KS) quark actions. However, the satisfactorily precise measurement of the matrix elements has not been achieved so far because of the inherent defects in these quark actions: the explicit chiral symmetry breaking in the Wilson quark action causes the non-trivial operator mixing between different chiralities and for the KS quark it is hard to treat the heavy-light cases due to the flavor symmetry breaking.
The domain-wall quark formulation in lattice QCD, which is based on the introduction of many heavy regulator fields, was proposed by Shamir[@Shamir93; @Shamir95] anticipating superior features over other quark formulations: no need of the fine tuning to realize the chiral limit and no restriction for the number of flavors. Recent simulation results seem to support the former feature non-perturbatively[@Blum-Soni; @Wingate; @Blum]. It is also perturbatively shown that the massless mode at the tree level still remains stable against the quantum correction[@Aoki-Taniguchi]. These advantageous features fascinate us to the application of the domain-wall quark for calculation of the three- and four-quark hadron matrix elements.
In order to convert the matrix elements obtained by lattice simulations to those defined in some continuum renormalization scheme(${\it e.g.}, {{\overline {\rm MS}}}$), we must know the renormalization factors connecting the lattice operators to the continuum counterparts defined in some renormalization scheme. In this article we make a perturbative calculation of the renormalization factors for the three- and four-quark operators consisting of physical quark fields in the domain-wall QCD(DWQCD). This work is an extension of the previous paper[@AIKT98], in which we developed a perturbative renormalization procedure for DWQCD demonstrating the calculation of the renormalization factors for quark wave function, mass and bilinear operators. We focus on whether or not the renormalization of the three- and four-quark operators in DWQCD is free from the notorious operator mixing problem. In the Wilson case it is well known that the mixing problem is not adequately manipulated by the perturbation theory, leading to an “incorrect” value for the $B_K$ matrix element.
This paper is organized as follows. In Sec. \[sec:model\] we briefly introduce the DWQCD action and the Feynman rules relevant for the present calculation to make this paper self-contained. In Sec. \[sec:4fermi\] our calculational procedure of the renormalization factors for the four-quark operators is described in detail. We also evaluate the renormalization factors for three-quark operators in Sec. \[sec:3fermi\]. In Secs. \[sec:4fermi\] and \[sec:3fermi\] numerical results for one-loop coefficients of the renormalization factors are given with and without the mean field improvement. In Sec. \[sec:bk\], using our results, we analyze a renormalization factor for $B_K$. Our conclusions are summarized in Sec. \[sec:concl\].
The physical quantities are expressed in lattice units and the lattice spacing $a$ is suppressed unless necessary. We take SU($N$) gauge group with the gauge coupling $g$ and the second Casimir $C_F = \displaystyle \frac{N^2-1}{2N}$, while $N=3$ is specified in the numerical calculations.
Action and Feynman rules {#sec:model}
========================
We take the Shamir’s domain-wall fermion action[@Shamir93], $$\begin{aligned}
S_{\rm DW} &=&
\sum_{n} \sum_{s=1}^{N_s} \Biggl[ \frac{1}{2} \sum_\mu
\left( {\overline{\psi}}(n)_s (-r+\gamma_\mu) U_\mu(n) \psi(n+\mu)_s
+ {\overline{\psi}}(n)_s (-r-\gamma_\mu) U_\mu^\dagger(n-\mu) \psi(n-\mu)_s \right)
{\nonumber}\\&&
+ \frac{1}{2}
\left( {\overline{\psi}}(n)_s (1+\gamma_5) \psi(n)_{s+1}
+ {\overline{\psi}}(n)_s (1-\gamma_5) \psi(n)_{s-1} \right)
+ (M-1+4r) {\overline{\psi}}(n)_s \psi(n)_s \Biggr]
{\nonumber}\\&+&
m \sum_n \left( {\overline{\psi}}(n)_{N_s} P_{+} \psi(n)_{1}
+ {\overline{\psi}}(n)_{1} P_{-} \psi(n)_{N_s} \right),
\label{eqn:action}\end{aligned}$$ where $n$ is a four dimensional space-time coordinate and $s$ is an extra fifth dimensional or “flavor” index, the Dirac “mass” $M$ is a parameter of the theory which we set $0 < M < 2$ to realize the massless fermion at tree level, $m$ is a physical quark mass, and the Wilson parameter is set to $r=-1$. It is important to notice that we have boundaries for the flavor space; $1 \le s \le N_s$. In our one-loop calculation we will take $N_s\to\infty$ limit to avoid complications arising from the finite $N_s$. $P_{R/L}$ is a projection matrix $P_{R/L}=(1\pm\gamma_5)/2$. For the gauge part we employ a standard four dimensional Wilson plaquette action and assume no gauge interaction along the fifth dimension.
In the DWQCD the zero mode of domain-wall fermion is extracted by the “physical” quark field defined by the boundary fermions $$\begin{aligned}
q(n) = P_R \psi(n)_1 + P_L \psi(n)_{N_s},
{\nonumber}\\
{\overline{q}}(n) = {\overline{\psi}}(n)_{N_s} P_R + {\overline{\psi}}(n)_1 P_L.
\label{eq:quark}\end{aligned}$$ We will consider the QCD operators constructed from this quark fields, since this field has been actually used in the previous simulations. Moreover our renormalization procedure is based on the Green functions consisting of only the “physical” quark fields, in which we have found that the renormalization becomes simple[@AIKT98].
Weak coupling perturbation theory is developed by expanding the action in terms of gauge coupling. The gluon propagator can be written as $$\begin{aligned}
G_{\mu\nu}^{AB}(k) =\delta_{\mu\nu}\delta_{AB} \frac{1}{4\sin^2(k/2)+\lambda^2}\end{aligned}$$ in the Feynman gauge with the infrared cut-off $\lambda^2$, where $\sin^2(k/2) =\sum_\mu \sin^2(k_\mu/2)$. Quark-gluon vertices are also identical to those in the $N_s$ flavor Wilson fermion. We need only one gluon vertex for our present calculation: $$\begin{aligned}
V_{1\mu}^A (k,p)_{st}
&=& -i g T^A \{ \gamma_\mu \cos(-k_\mu/2 + p_\mu/2)
-i r \sin(-k_\mu/2 + p_\mu/2) \} \delta_{st},\end{aligned}$$ where $k$ and $p$ represent incoming momentum into the vertex (see Fig. 1 of Ref.[@AIKT98]). $T^A$ $(A=1,\dots,N^2-1)$ is a generator of color SU($N$). The fermion propagator originally takes $N_s\times N_s$ matrix form in $s$-flavor space. In the present one-loop calculation, however, we do not need the whole matrix elements because we consider Green functions consisting of the physical quark fields. The relevant fermion propagators are restricted to following three types: $$\begin{aligned}
&&
{\langle q(-p) {\overline{q}}(p) \rangle} =
\frac{-i\gamma_\mu \sin p_\mu + \left(1-W e^{-\alpha}\right) m}
{-\left(1-e^{\alpha}W\right) + m^2 (1-W e^{-\alpha})}
\equiv S_q(p),
\label{eqn:phys-prop}
\\&&
{\langle q(-p) {\overline{\psi}}(p,s) \rangle}
=
\frac{1}{F}
\left( i\gamma_\mu \sin p_\mu - m \left(1 -W e^{-\alpha} \right)
\right)
\left( e^{-\alpha (N_s-s)} P_R + e^{-\alpha (s-1)} P_L \right)
{\nonumber}\\&&\qquad
+\frac{1}{F} \Bigl[
m \left(i\gamma_\mu \sin p_\mu -m \left(1-W e^{-\alpha}\right)\right)
- F \Bigr] e^{-\alpha}
\left( e^{-\alpha (s-1)} P_R + e^{-\alpha (N_s-s)} P_L \right),
\\&&
{\langle \psi(-p,s) {\overline{q}}(p) \rangle}
=
\frac{1}{F}
\left( e^{-\alpha (N_s-s)} P_L + e^{-\alpha (s-1)} P_R \right)
\left( i\gamma_\mu \sin p_\mu - m \left(1 - W e^{-\alpha} \right)
\right)
{\nonumber}\\&&\qquad
+\frac{1}{F}
\left( e^{-\alpha (s-1)} P_L + e^{-\alpha (N_s-s)} P_R \right) e^{-\alpha}
\Bigl[
m \left(i\gamma_\mu \sin p_\mu -m\left(1- We^{-\alpha}\right) \right)
- F \Bigr]\end{aligned}$$ with $$\begin{aligned}
W &=& 1-M -r \sum_\mu (1-\cos p_\mu),
\\
\cosh (\alpha) &=& \frac{1+W^2+\sum_\mu \sin^2 p_\mu}{2|W|},
\label{eq:alpha}
\\
F &=& 1-e^{\alpha} W-m^2 \left(1-W e^{-\alpha}\right),
\label{eq:F}\end{aligned}$$ where the argument $p$ in the factors $\alpha$ and $W$ is suppressed.
In the perturbative calculation of Green functions the external quark momenta and masses are assumed to be much smaller than the lattice cut-off, so that we can expand the external quark propagators in terms of them. We have the following expressions as leading term of the expansion: $$\begin{aligned}
\langle q\bar q \rangle (p) & = & \frac{1-w_0^2}{i{p\kern-1ex /}+ (1-w_0^2)m},
\\
\langle q \bar \psi_s \rangle (p) &=&
\langle q\bar q \rangle (p)
\left( w_0^{s-1}P_L + w_0^{N_s-s} P_R\right),
\label{eqn:qpsi}
\\
\langle \psi_s \bar q\rangle (p) &=&
\left( w_0^{s-1}P_R + w_0^{N_s-s} P_L\right)
\langle q\bar q \rangle (p),
\label{eqn:psiq}\end{aligned}$$ where $w_0 = 1-M$.
Renormalization factors for four-quark operators {#sec:4fermi}
================================================
We consider the following four-quark operators: $$\begin{aligned}
{\cal O}_\pm & = & \frac{1}{2} \left[
(\bar q_1 \gamma_\mu^L q_2)(\bar q_3 \gamma_\mu^L q_4)
\pm
(\bar q_1 \gamma_\mu^L q_4)(\bar q_3 \gamma_\mu^L q_2) \right], \\
{\cal O}_1 & = &
-C_F (\bar q_1 \gamma_\mu^L q_2)(\bar q_3 \gamma_\mu^R q_4)
+ (\bar q_1 T^A \gamma_\mu^L q_2)(\bar q_3 T^A \gamma_\mu^R q_4), \\
{\cal O}_2 & = &
\frac{1}{2N} (\bar q_1 \gamma_\mu^L q_2)(\bar q_3 \gamma_\mu^R q_4)
+ (\bar q_1 T^A \gamma_\mu^L q_2)(\bar q_3 T^A \gamma_\mu^R q_4) ,\end{aligned}$$ where $\gamma_\mu^{L,R} = \gamma_\mu P_{L,R}$. Summation over repeated indices such as $\mu$ and $A$ is assumed. We note that $q_i$ $(i=1,2,3,4)$ are boundary quark fields in DWQCD. For the convenience of calculation we rewrite the above operators as $$\begin{aligned}
{\cal O}_\pm & = & \frac{1}{2}
\left[1{\widetilde{\otimes}}1 \pm 1{\widetilde{\odot}}1\right]^{ab;cd}
\left[(\bar q_1^a \gamma_\mu^L q_2^b)
(\bar q_3^c \gamma_\mu^L q_4^d)\right], \\
{\cal O}_1 & = & \frac{1}{2}
\left[-N 1{\widetilde{\otimes}}1 + 1{\widetilde{\odot}}1\right]^{ab;cd}
\left[(\bar q_1^a \gamma_\mu^L q_2^b)
(\bar q_3^c \gamma_\mu^R q_4^d)\right], \\
{\cal O}_2 & = & \frac{1}{2}
\left[1{\widetilde{\odot}}1\right]^{ab;cd}
\left[(\bar q_1^a \gamma_\mu^L q_2^b)
(\bar q_3^c \gamma_\mu^R q_4^d)\right], \end{aligned}$$ where $a,b,c,d$ are color indices, and ${\widetilde{\otimes}}$, ${\widetilde{\odot}}$ represent the tensor structures in the color space: $$\begin{aligned}
\left[1 {\widetilde{\otimes}} 1\right]^{ab;cd} & \equiv & \delta_{ab}\delta_{cd}, \\
\left[1 {\widetilde{\odot}} 1\right]^{ab;cd} & \equiv & \delta_{ad}\delta_{cb}. \end{aligned}$$ To derive these formula, we have used the Fierz transformation for ${\cal O}_\pm$ and the formula $$\sum_A T^A{\widetilde{\otimes}} T^A = \frac{1}{2}\left[
-\frac{1}{N} 1{\widetilde{\otimes}} 1+1{\widetilde{\odot}} 1\right]$$ for ${\cal O}_{1,2}$ .
We calculate the following Green function: $$\langle {\cal O}_\Gamma \rangle_{\alpha\beta;\gamma\delta}^{ij;kl}
\equiv {\langle {\cal O}_\Gamma (q_1)_{\alpha}^i (\bar q_2)_{\beta}^j
(q_3)_{\gamma}^k (\bar q_4)_{\delta}^l \rangle},$$ where $\Gamma = \pm, 1,2$. Spinor indices are labeled by $\alpha,\beta,\gamma,\delta$ and color ones by $i,j,k,l$. Truncating the external quark propagators from $\langle {\cal O}_\Gamma \rangle$, where we multiply $\langle {\cal O}_\Gamma \rangle$ by ${i{p\kern-1ex /}_i + (1-w_0^2)m}$, we obtain the vertex functions, which is written in the following form up to the one-loop level $$(1-w_0^2)^4\left(\Lambda_\Gamma\right)_{\alpha\beta;\gamma\delta}^{ij;kl}
=(1-w_0^2)^4\left(\Lambda_\Gamma^{(0)}
+\Lambda_\Gamma^{(1)}\right)_{\alpha\beta;\gamma\delta}^{ij;kl},$$ where the superscript $(i)$ refers to the $i$-th loop level and the trivial factor $(1-w_0^2)^4$ is factored out for the convenience. We suppress the external momenta $p_i$ since the renormalization factor does not depend on them.
The tree level vertex functions $\Lambda_\Gamma^{(0)}$ are given by $$\begin{aligned}
&\Gamma = \pm, &\qquad
\frac{1}{2}\left[ \gamma_\mu^L
\otimes \gamma_\mu^L\right]_{\alpha\beta;\gamma\delta}
\left[1{\widetilde{\otimes}} 1 \pm 1{\widetilde{\odot}} 1\right]^{ij;kl},
\label{eq:lambda_pm_0}\\
&\Gamma = 1, &\qquad
\frac{1}{2}\left[ \gamma_\mu^L
\otimes \gamma_\mu^R \right]_{\alpha\beta;\gamma\delta}
\left[-N 1{\widetilde{\otimes}} 1+1{\widetilde{\odot}} 1\right]^{ij;kl},
\label{eq:lambda_1_0}\\
&\Gamma = 2, &\qquad
\frac{1}{2}\left[ \gamma_\mu^L
\otimes \gamma_\mu^R \right]_{\alpha\beta;\gamma\delta}
\left[1{\widetilde{\odot}} 1 \right]^{ij;kl},
\label{eq:lambda_2_0}\end{aligned}$$ where $\otimes$ acts on the Dirac spinor space representing $[\gamma_X \otimes \gamma_Y ]_{\alpha\beta;\gamma\delta} \equiv
(\gamma_X)_{\alpha\beta}(\gamma_Y)_{\gamma\delta}$.
The one-loop vertex corrections are illustrated by six diagrams in Fig. 1, the sum of which yields the one-loop level vertex function $$\Lambda_\Gamma^{(1)}=
\int_{-\pi}^{\pi}\frac{d^4 k}{(2\pi)^4}
\left(I_\Gamma^a+,\dots,+I_\Gamma^{c^\prime}\right).$$ In order to obtain the expressions for the integrands $I_\Gamma^a,\dots,I_\Gamma^{c^\prime}$ we should note that the internal quark propagators appearing in the diagrams are multiplied by the damping factor which comes from eqs.[(\[eqn:qpsi\])]{} and [(\[eqn:psiq\])]{}. The following formula are useful. $$\begin{aligned}
\langle q \bar \psi_s\rangle \left( w_0^{s-1}P_L + w_0^{N_s-s} P_R\right)
=\left( w_0^{s-1}P_R + w_0^{N_s-s} P_L\right) \langle \psi_s\bar q \rangle
&=& \frac{i\gamma_\mu \sin p_\mu}{{\widetilde{F}}\cdot {\widetilde{F}}_0 }\equiv {\overline{G}},
\label{eq:G_b} \\
\langle q \bar \psi_s\rangle \left( w_0^{s-1}P_R + w_0^{N_s-s} P_L\right)
=\left( w_0^{s-1}P_L + w_0^{N_s-s} P_R\right) \langle \psi_s\bar q \rangle
&=& -\frac{1}{{\widetilde{F}}_0 } \equiv {\widetilde{G}},
\label{eq:G_t}\end{aligned}$$ where ${\widetilde{F}}=e^{-\alpha}-W$ and ${\widetilde{F}}_0 = e^\alpha-w_0$. Here we set $m=p_i=0$ for the internal propagator.
The contribution from Fig. 1a takes the form $$\begin{aligned}
I_\Gamma^a &=& \frac{1}{2} J_a^{AB}
\left\{{\overline{V}}_\mu(k) {\overline{G}}(k) + {\widetilde{V}}_\mu (k){\widetilde{G}}(k)
\right\}
\Gamma_X
\left\{{\overline{G}}(k) {\overline{V}}_\nu(k) + {\widetilde{G}}(k) {\widetilde{V}}_\nu(k)\right\}
\otimes \Gamma_Y
\, G_{\mu\nu}^{AB}(k),\end{aligned}$$ where $\Gamma_X = \gamma_\mu^L$, $\Gamma_Y = \gamma_\mu^L$ or $\gamma_\mu^R$, and the interaction vertices are $${\overline{V}}_\mu = -i g \gamma_\mu \cos(k_\mu/2),\quad
{\widetilde{V}}_\mu = -rg \sin(k_\mu/2).
\label{eq:V_bt}$$ The color factors are represented by $J_a^{AB}$, which are listed in Table \[tab:color\]. In a similar way the contributions from Fig. 1b and Fig. 1c are given by $$\begin{aligned}
I_\Gamma^b &=& \frac{1}{2} J_b^{AB}
\left\{{\overline{V}}_\mu(k) {\overline{G}}(k) + {\widetilde{V}}_\mu(k) {\widetilde{G}}(k)\right\}
\Gamma_X\otimes
\left\{{\overline{V}}_\nu (-k) {\overline{G}}(-k) + {\widetilde{V}}_\nu (-k){\widetilde{G}}(-k)\right\}
\Gamma_Y
\, G_{\mu\nu}^{AB}(k), \\
I_\Gamma^c &=& \frac{1}{2} J_c^{AB}
\left\{{\overline{V}}_\mu(k) {\overline{G}}(k) + {\widetilde{V}}_\mu (k) {\widetilde{G}}(k)\right\}
\Gamma_X\otimes
\Gamma_Y
\left\{{\overline{G}}(k){\overline{V}}_\nu(k) + {\widetilde{G}}(k) {\widetilde{V}}_\nu(k) \right\}
\, G_{\mu\nu}^{AB}(k).\end{aligned}$$ After a little algebra the expressions of $I_\Gamma^{a,b,c}$ are reduced to $$\begin{aligned}
I_\Gamma^a &=& \frac{1}{2} g^2 J_a^{AA}
K \left[ T + A_{VA}\right]\left[\Gamma_X\otimes\Gamma_Y\right],
\label{eq:vtx_a}\\
I_\Gamma^b &=& - \frac{1}{2} g^2 J_b^{AA}
K \left[T \Gamma_X\otimes \Gamma_Y +
\cos^2(k_\mu/2) \sin^2 k_\alpha (\gamma_\mu\gamma_\alpha\Gamma_X)
\otimes (\gamma_\mu\gamma_\alpha\Gamma_Y) \right],
\label{eq:vtx_b}\\
I_\Gamma^c &=& \frac{1}{2} g^2 J_c^{AA}
K \left[T \Gamma_X \otimes \Gamma_Y +
\cos^2(k_\mu/2) \sin^2 k_\alpha (\gamma_\mu\gamma_\alpha\Gamma_X)
\otimes (\Gamma_Y\gamma_\alpha\gamma_\mu) \right],
\label{eq:vtx_c}\end{aligned}$$ where $$\begin{aligned}
&&
K =\displaystyle \frac{1}{{\widetilde{F}}^2{\widetilde{F}}_0^2 (4\sin^2(k/2)+\lambda^2)},
\label{eq:K}\\&&
T=r^2 \sin^2(k/2) {\widetilde{F}}^2+r \sin^2 k {\widetilde{F}},
\label{eq:T}\\&&
A_{VA}= \sum_\mu \cos^2(k_\mu/2) \sin^2 k_\mu.
\label{eq:A_VA}\end{aligned}$$
In order to rewrite the second term of $I_\Gamma^{b,c}$ we apply the Fierz transformation: $$\begin{aligned}
(\gamma_\mu\gamma_\alpha\gamma_\nu^L)\otimes
(\gamma_\mu\gamma_\alpha\gamma_\nu^L) &=&
-\gamma_\nu^L\odot\gamma_\nu^L
= \gamma_\nu^L\otimes\gamma_\nu^L,
\\
(\gamma_\mu\gamma_\alpha\gamma_\nu^L)\otimes
(\gamma_\nu^L\gamma_\alpha\gamma_\mu) &=&
-\gamma_\nu^L\odot\gamma_\nu^L (1-2\delta_{\alpha\nu})(1-2\delta_{\mu\nu})
\\&=&
-\gamma_\nu^L\otimes\gamma_\nu^L (1-2\delta_{\alpha\nu})
(1-2\delta_{\mu\nu})
+2\gamma_\nu^L\otimes\gamma_\nu^L \delta_{\alpha\mu},
\\
(\gamma_\mu\gamma_\alpha\gamma_\nu^L)\otimes
(\gamma_\mu\gamma_\alpha\gamma_\nu^R) &=&
2P_R\odot P_L \delta_{\mu\alpha}
= \gamma_\nu^L \otimes \gamma_\nu^R \delta_{\mu\alpha},
\label{eq:lr1_fierz}\\
(\gamma_\mu\gamma_\alpha\gamma_\nu^L)\otimes
(\gamma_\nu^R\gamma_\alpha\gamma_\mu) &=&
2P_R\odot P_L
= \gamma_\nu^L \otimes \gamma_\nu^R .
\label{eq:lr2_fierz}\end{aligned}$$ where $[\gamma_X \odot \gamma_Y ]_{\alpha\beta;\gamma\delta} \equiv
(\gamma_X)_{\alpha\delta}(\gamma_Y)_{\gamma\beta}$, and summation over $\nu$ is taken. The Fierz transformation is again used for the second equality. We omit the tensor term in eq.(\[eq:lr1\_fierz\]) since it vanishes in the integral.
Choosing $\Gamma_{X,Y} = \gamma_\mu^L$ in eqs.(\[eq:vtx\_a\]), (\[eq:vtx\_b\]) and (\[eq:vtx\_c\]) we first consider the case of ${\cal O}_{\pm}$. After simplifying the expressions of the color factors we obtain $$\begin{aligned}
I_\pm^a &=& \frac{1}{2} g^2
K (T+A_{VA})\left[\gamma_\nu^L\otimes\gamma_\nu^L\right]
\left[ (C_F\pm \frac{1}{2}) 1{\widetilde{\otimes}}1
\mp \frac{1}{2N} 1{\widetilde{\odot}}1 \right], \\
I_\pm^b &=& - \frac{1}{2} g^2
K (T+A_{SP})\left[\gamma_\nu^L\otimes\gamma_\nu^L\right]
\left[ (-\frac{1}{2N}\pm \frac{1}{2}) 1{\widetilde{\otimes}}1
+(\frac{1}{2}\mp \frac{1}{2N}) 1{\widetilde{\odot}}1 \right], \\
I_\pm^c &=& \frac{1}{2} g^2
K (T+A_{VA})\left[\gamma_\nu^L\otimes\gamma_\nu^L\right]
\left[ -\frac{1}{2N} 1{\widetilde{\otimes}}1
+(\frac{1}{2}\pm C_F) 1{\widetilde{\odot}}1 \right]. \end{aligned}$$ The total contribution becomes $$\begin{aligned}
I_+^a+I_+^b+I_+^c &=& \frac{1}{2} g^2 K
\left[\gamma_\nu^L\otimes\gamma_\nu^L\right]_{\alpha\beta;\gamma\delta}
\left[1{\widetilde{\otimes}}1+1{\widetilde{\odot}}1\right]^{ij;kl} \nonumber \\
&\times &
\{ T C_F + A_{VA}(C_F +\frac{1}{2} - \frac{1}{2N})
-A_{SP}(\frac{1}{2} - \frac{1}{2N}) \},
\label{eq:sumP}\end{aligned}$$ and $$\begin{aligned}
I_-^a+I_-^b+I_-^c &=& \frac{1}{2} g^2 K
\left[\gamma_\nu^L\otimes\gamma_\nu^L\right]_{\alpha\beta;\gamma\delta}
\left[1{\widetilde{\otimes}}1-1{\widetilde{\odot}}1\right]^{ij;kl} \nonumber \\
&\times &
\{ T C_F + A_{VA}(C_F -\frac{1}{2} - \frac{1}{2N})
+A_{SP}(\frac{1}{2} + \frac{1}{2N}) \},
\label{eq:sumM}\end{aligned}$$ where $$A_{SP}=\cos^2(k/2)\cdot \sin^2 k.
\label{eq:A_SP}$$ We should note that the other three contributions $I_\pm^{a^\prime,b^\prime,c^\prime}$ from Fig. 1$a^\prime$, 1$b^\prime$ and 1$c^\prime$ are equal to $I_\Gamma^{a,b,c}$ respectively, therefore the factors $1/2$ in eqs. (\[eq:sumP\]) and (\[eq:sumM\]) disappear in the total contributions of all.
Comparing the one-loop results to the tree level ones we obtain $$\begin{aligned}
\Lambda_+ &=&
\left[ 1 + g^2 \frac{N-1}{N}\left\{
<T> (N+1)+<A_{VA}>(N+2)-<A_{SP}> \right\}\right]
\Lambda_+^{(0)},
\label{eq:1loop_+}\\
\Lambda_- &=&
\left[ 1 + g^2 \frac{N+1}{N}\left\{
<T> (N-1)+<A_{VA}>(N-2)+<A_{SP}>\right\}\right]
\Lambda_-^{(0)}
\label{eq:1loop_-} \end{aligned}$$ with $$< X > = \int_{-\pi}^{\pi}\frac{d^4 k}{(2\pi)^4} K(k) X(k)
\label{eq:integ_X}$$ for $X = T, A_{VA}, A_{SP}$. We remark that $C_F< T+A_{VA} >$ and $C_F< T+A_{SP} >$ correspond to the one-loop vertex corrections to the (axial) vector current and the (pseudo) scalar density which are expressed as $(T_{VA}-1)$ and $(T_{SP}-1)$ in Ref. [@AIKT98]. The expressions of eqs.(\[eq:1loop\_+\]) and (\[eq:1loop\_-\]) show an important property of ${\cal O}_{\pm}$ in the DWQCD formalism: the one-loop vertex corrections are multiplicative. This is contrary to the Wilson case, in which the mixing operators with different chiralities appears at the one-loop level[@pt_w4].
We next turn to the case of ${\cal O}_{1,2}$. For ${\cal O}_1$ the vertex corrections of eqs.(\[eq:vtx\_a\]), (\[eq:vtx\_b\]) and (\[eq:vtx\_c\]) with $\Gamma_X = \gamma_\mu^L$ and $\Gamma_Y = \gamma_\mu^R$ are written as $$\begin{aligned}
I_1^a &=& \frac{1}{2} g^2 K (T+A_{VA})
\left[\gamma_\nu^L\otimes\gamma_\nu^R\right]
\left[ (-N C_F+ \frac{1}{2}) 1{\widetilde{\otimes}}1
- \frac{1}{2N} 1{\widetilde{\odot}}1 \right], \\
I_1^b &=& - \frac{1}{2} g^2 K (T+A_{VA})
\left[\gamma_\nu^L\otimes\gamma_\nu^R\right]
\left[ (\frac{1}{2}+ \frac{1}{2}) 1{\widetilde{\otimes}}1
+(-\frac{N}{2}- \frac{1}{2N}) 1{\widetilde{\odot}}1 \right], \\
I_1^c &=& \frac{1}{2} g^2 K (T+A_{SP})
\left[\gamma_\nu^L\otimes\gamma_\nu^R\right]
\left[ \frac{1}{2} 1{\widetilde{\otimes}}1
+(-\frac{N}{2}+ C_F) 1{\widetilde{\odot}}1 \right]. \end{aligned}$$ The total contribution including those from Fig. $1a^\prime$, $1b^\prime$ and $1c^\prime$ is given by $$\begin{aligned}
2(I_1^a+I_1^b+I_1^c) &=& 2\frac{1}{2} g^2
K \left[\gamma_\nu^L\otimes\gamma_\nu^R\right]_{\alpha\beta;\gamma\delta}
\left[- N 1{\widetilde{\otimes}} 1 +1{\widetilde{\odot}} 1\right]^{ij;kl} \nonumber
\\&\times &
\left[ T C_F + A_{VA}\frac{N}{2}-A_{SP}\frac{1}{2N} \right].\end{aligned}$$ Using the tree level result in eq.(\[eq:lambda\_1\_0\]) the vertex function up to the one-loop level is expressed as $$\begin{aligned}
\Lambda_1 &=&
\left[ 1 + g^2 \frac{1}{N}\left\{
<T> (N^2-1)+<A_{VA}>N^2-<A_{SP}>\right\}\right]
\Lambda_1^{(0)}.\end{aligned}$$ This result shows that the operator ${\cal O}_1$ is multiplicatively renormalizable in DWQCD, which is in contrast with the Wilson case[@pt_w4].
In a similar way we write the vertex corrections for ${\cal O}_2$. $$\begin{aligned}
I_2^a &=& \frac{1}{2} g^2 K (T+A_{VA})
\left[\gamma_\nu^L\otimes\gamma_\nu^R\right]
\left[ \frac{1}{2} 1{\widetilde{\otimes}}1
- \frac{1}{2N} 1{\widetilde{\odot}}1 \right], \\
I_2^b &=& - \frac{1}{2} g^2 K (T+A_{VA})
\left[\gamma_\nu^L\otimes\gamma_\nu^R\right]
\left[ \frac{1}{2} 1{\widetilde{\otimes}}1
- \frac{1}{2N} 1{\widetilde{\odot}}1 \right], \\
I_2^c &=& \frac{1}{2} g^2 K (T+A_{SP})
\left[\gamma_\nu^L\otimes\gamma_\nu^R\right]
\left[ C_F 1{\widetilde{\odot}}1 \right]. \end{aligned}$$ The total contribution including those from Fig. $1a^\prime$, $1b^\prime$ and $1c^\prime$ becomes $$\begin{aligned}
2(I_2^a+I_2^b+I_2^c) &=& 2\frac{1}{2}g^2
K \left[\gamma_\nu^L\otimes\gamma_\nu^R\right]_{\alpha\beta;\gamma\delta}
\left[1{\widetilde{\odot}} 1 \right]^{ij;kl} \nonumber \\
&\times &
C_F\left[ T + A_{SP}\right],\end{aligned}$$ which leads to $$\begin{aligned}
\Lambda_2 &=&
\left[ 1 + g^2 \frac{N^2-1}{N}\left\{ <T> +<A_{SP}>\right\}\right]
\Lambda_2^{(0)}.\end{aligned}$$ We again find that the vertex correction is multiplicative up to the one-loop level as opposed to the Wilson case[@pt_w4].
The contribution from the fermion self-energy has already been evaluated [@Aoki-Taniguchi; @AIKT98] and the total lattice renormalization factor is now obtained: $$Z_\Gamma^{lat} = (1-w_0^2)^2 Z_w^2 Z_2^2 V_\Gamma,$$ where $$\begin{aligned}
Z_2 &=& 1 + \frac{g^2}{16\pi^2} C_F \left[ \log (\lambda a)^2
+ \Sigma_1\right],
\\
V_\Gamma &=& 1 + \frac{g^2}{16\pi^2} \left[ -\delta_\Gamma \log (\lambda a)^2
+ v_\Gamma \right],\end{aligned}$$ $$\begin{aligned}
v_+ &=&\frac{16\pi^2(N-1)}{N}\left[ <T> (N+1)+<<A_{VA}>>(N+2)-<<A_{SP}>>
\right] +\delta_+ \log \pi^2,
\\
v_- &=&\frac{16\pi^2(N+1)}{N}\left[ <T> (N-1)+<<A_{VA}>>(N-2)+<<A_{SP}>>
\right] +\delta_- \log \pi^2,
\\
v_1 &=&\frac{16\pi^2}{N}\left[ <T> (N^2-1)+<<A_{VA}>>N^2-<<A_{SP}>>
\right] +\delta_1 \log \pi^2,
\\
v_2 &=&\frac{16\pi^2(N^2-1)}{N}\left[ <T> + <<A_{SP}>>
\right] +\delta_2 \log \pi^2 \end{aligned}$$ with $$\delta_\Gamma =
\left\{
\begin{array}{ll}
\displaystyle \frac{(N-1)(N-2)}{N} & \qquad \Gamma = + \\
& \\
\displaystyle \frac{(N+1)(N+2)}{N} & \qquad \Gamma = - \\
& \\
\displaystyle \frac{(N+2)(N-2)}{N} & \qquad \Gamma = 1 \\
& \\
\displaystyle \frac{4(N+1)(N-1)}{N} & \qquad \Gamma = 2
\end{array}
\right.$$ The infrared singularity of $< A_X>$ is subtracted as $$<< A_X >> = \int_{-\pi}^{\pi}\frac{d^4k}{(2\pi)^4}
\left[K(k) A_X(k) - c_X \frac{1}{(k^2)^2}\theta(\pi^2-k^2)\right]$$ with $c_{SP}=4$ and $c_{VA}=1$.
Numerical values of $v_\Gamma$ are evaluated by two independent methods. In one method the momentum integration is performed by a mode sum for a periodic box of a size $L^4$ after transforming the momentum variable through $k_\mu = q_\mu -\sin q_\mu$. We employ the size $L=64$ for integrals. In the other method the momentum integration is carried out by the the Monte Carlo integration routine VEGAS, using 20 samples of 1000000 points each. We find that both results agree very well. Numerical values of $v_\Gamma$ are presented in Table \[tab:4fermi\] as a function of $M$.
We have to also calculate the corresponding continuum wave-function renormalization factor and vertex corrections in the ${{\overline {\rm MS}}}$ scheme employing the same gauge and the same infrared regulator as the lattice case. For the present calculation it seems preferable to choose Dimensional Reduction(DRED) as the ultraviolet regularization, in which the loop momenta of the Feynman integrals are defined in $D<4$ dimensions while keeping the Dirac matrices in four dimensions. In the DRED scheme we can use the same calculational techniques for the vertex corrections as the lattice case thanks to applicability of the Fierz transformation for the Dirac matrices. For the wave-function renormalization factor a simple calculation gives $$Z_2^{{{\overline {\rm MS}}}} = 1+ \frac{g^2}{16\pi^2}C_F \left[
\log (\lambda/\mu)^2 -1/2\right],$$ where $\mu$ is a renormalization scale. This result leads to $\Sigma_1^{{{\overline {\rm MS}}}}=-1/2$. For the vertex corrections we obtain $$V_\Gamma^{{{\overline {\rm MS}}}} = 1+\frac{g^2}{16\pi^2} \delta_\Gamma \left[
-\log (\lambda/\mu)^2 +1\right],$$ giving $v_\Gamma^{{{\overline {\rm MS}}}}=\delta_\Gamma$. Here we should remark that the one-loop vertex corrections yield the evanescent operators which vanish in $D=4$ for the DRED scheme[@BW]. It is meaningless to give results without mentioning the definition of evanescent operators, because the constant terms at the one-loop level depend on the definition of the evanescent operators. Our choice is as follows: $$\begin{aligned}
E^{\rm DRED}_{\pm}=&
{\bar \delta}_{\mu\nu}\gamma_\mu(1-\gamma_5)
\otimes\gamma_\nu(1-\gamma_5)
-\frac{D}{4}\gamma_\mu(1-\gamma_5)\otimes\gamma_\mu(1-\gamma_5), \\
E^{\rm DRED}_{1,2}=&
{\bar \delta}_{\mu\nu}\gamma_\mu(1-\gamma_5)
\otimes\gamma_\nu(1+\gamma_5)
-\frac{D}{4}\gamma_\mu(1-\gamma_5)\otimes\gamma_\mu(1+\gamma_5),\end{aligned}$$ where ${\bar \delta}_{\mu\nu}$ is the $D$-dimensional metric tensor which emerges inevitably in the evaluation of the Feynman integrals.
Combining these results with the previous lattice ones we obtain $$\begin{aligned}
{\cal O}_\Gamma^{{{\overline {\rm MS}}}}(\mu) & = &\frac{1}{(1-w_0^2)^2 Z_w^2}
Z_\Gamma (\mu a ) {\cal O}_\Gamma^{lat} (1/a),\end{aligned}$$ where $$\begin{aligned}
Z_\Gamma (\mu a) &=& \frac{ (Z_2^{{{\overline {\rm MS}}}})^2 V_\Gamma^{{{\overline {\rm MS}}}}}
{(Z_2)^2 V_\Gamma} \nonumber \\
&=& 1 + \frac{g^2}{16\pi^2}\left[
(\delta_\Gamma - 2 C_F)\log (\mu a)^2 + z_\Gamma \right],
\label{eq:4fermi}
\\
z_\Gamma &=& v_\Gamma^{{{\overline {\rm MS}}}} - v_\Gamma + 2 C_F\{\Sigma_1^{{{\overline {\rm MS}}}}
-\Sigma_1\} .\end{aligned}$$ Numerical values of $z_\Gamma$ are given in Table \[tab:total\] and the results for the mean-field improved one, $z_\Gamma^{MF}$ , are also given in Table \[tab:totalMF\].
Although the results for the DRED scheme are presented here, it is an easy task to obtain those for the Naive Dimensional Regularization(NDR) scheme. In Appendix B we summarize the finite parts of the wave-function renormalization factor and vertex corrections in the NDR scheme.
Renormalization factors for three-quark operators {#sec:3fermi}
=================================================
The three-quark operators relevant to the proton decay amplitude are given by $$\left({\cal O}_{PD}\right)_\delta =\varepsilon^{abc}
\left((\bar q^C_1)^a \Gamma_X (q_2)^b\right) (\Gamma_Y (q_3)^c)_\delta,
\label{eq:O_PD}$$ where $\bar q^C = -q^T C^{-1}$ with $C=\gamma_0\gamma_2$ is a charge conjugated field of $q$ and $\Gamma_X\otimes\Gamma_Y = P_R\otimes P_R, P_R\otimes P_L,
P_L\otimes P_R, P_L\otimes P_L$. The summation over repeated color indices $a,b,c$ is assumed. We should note that the domain-wall fermion action [(\[eqn:action\])]{} is transformed identically into that with the conjugated field by using transpose and matrix $C$. The resultant action and Feynman rules for the conjugated field is obtained by the replacement that $$\begin{aligned}
ig T^A &\rightarrow & -ig (T^A)^T ,
$$ where the superscript $T$ means the transposed matrix.
In order to evaluate the vertex corrections we consider the following Green function: $$\langle \left({\cal O}_{PD}\right)_\delta
\rangle_{\alpha\beta\gamma}^{ijk}
\equiv {\langle \left({\cal O}_{PD}\right)_\delta
(q^C_1)_\alpha^i(\bar q_2)_\beta^j
(\bar q_3)_\gamma^k \rangle},$$ where $\alpha,\beta,\gamma$ and $i,j,k$ are spinor and color indices respectively. Truncating the external quark propagators of $\langle {\cal O}_{PD}\rangle$ we obtain the vertex function $$(1-w_0^2)^3\left(\Lambda_{PD}\right)_{\alpha\beta;\delta\gamma}^{ijk}
=(1-w_0^2)^3\left(\Lambda_\Gamma^{(0)}
+\Lambda_\Gamma^{(1)}\right)_{\alpha\beta;\delta\gamma}^{ijk},$$ where the trivial factor $(1-w_0^2)^3$ is factored out for the convenience. We suppress the external momenta $p_i$ since the renormalization factor does not depend on them.
At the tree level the vertex function takes the form $$\Lambda_\Gamma^{(0)}=\varepsilon^{ijk} \left[ \Gamma_X \otimes \Gamma_Y
\right]_{\alpha\beta;\delta\gamma},
\label{eq:PD_tree}$$ where $[\Gamma_X \otimes \Gamma_Y]_{\alpha\beta;\delta\gamma} \equiv
(\Gamma_X)_{\alpha\beta}(\Gamma_Y)_{\delta\gamma}$.
The one-loop vertex corrections are shown in Figs. 2a, 2b and 2c, the sum of which gives the one-loop level vertex function $$\Lambda_{PD}^{(1)}=
\int_{-\pi}^{\pi}\frac{d^4 k}{(2\pi)^4}
\left(I_{PD}^a+I_{PD}^b+I_{PD}^{c}\right).$$ Using the notations in eqs.(\[eq:G\_b\]), (\[eq:G\_t\]) and (\[eq:V\_bt\]) the integrands $I_{PD}^{a,b,c}$ are written as follows: $$\begin{aligned}
I_{PD}^a &=& \varepsilon^{abk}(-T^A)^T_{ia}T^B_{bj} {\nonumber}\\
&&\times \left\{{\overline{V}}_\mu(k) {\overline{G}}(k)
+ {\widetilde{V}}_\mu (k){\widetilde{G}}(k)\right\}
\Gamma_X
\left\{{\overline{G}}(k) {\overline{V}}_\nu(k)
+ {\widetilde{G}}(k) {\widetilde{V}}_\nu(k)\right\}
\otimes \Gamma_Y
G_{\mu\nu}^{AB}(k), \\
I_{PD}^b &=& \varepsilon^{ibc}T^A_{bj}T^B_{ck} {\nonumber}\\
&&\times \Gamma_X\left\{{\overline{G}}(k) {\overline{V}}_\nu(k)
+ {\widetilde{G}}(k) {\widetilde{V}}_\nu(k)\right\}
\otimes\Gamma_Y\left\{{\overline{G}}(-k) {\overline{V}}_\nu(-k)
+ {\widetilde{G}}(-k) {\widetilde{V}}_\nu(-k)\right\}
G_{\mu\nu}^{AB}(k), \\
I_{PD}^c &=& \varepsilon^{ajc}(-T^A)^T_{ia}T^B_{ck} {\nonumber}\\
&&\times \left\{{\overline{V}}_\mu(k) {\overline{G}}(k) + {\widetilde{V}}_\mu (k)
{\widetilde{G}}(k)\right\}
\Gamma_X\otimes
\Gamma_Y
\left\{{\overline{G}}(k){\overline{V}}_\nu(k)
+ {\widetilde{G}}(k){\widetilde{V}}_\nu(k) \right\}
G_{\mu\nu}^{AB}(k).\end{aligned}$$ A little algebra yields $$\begin{aligned}
I_{PD}^a &=& g^2\frac{N+1}{2N}\varepsilon^{ijk}
K \left[ T + A_{SP}\right]\left[\Gamma_X\otimes\Gamma_Y\right] , \\
I_{PD}^b &=& g^2\frac{N+1}{2N}\varepsilon^{ijk}
K \left[T (\Gamma_X\otimes\Gamma_Y) +
\cos^2 (k_\mu/2) \sin^2 k_\alpha (\Gamma_X\gamma_\alpha\gamma_\mu)
\otimes (\Gamma_Y\gamma_\alpha\gamma_\mu) \right], \\
I_{PD}^c &=& g^2\frac{N+1}{2N}\varepsilon^{ijk}
K \left[T (\Gamma_X \otimes\Gamma_Y) +
\cos^2 (k_\mu/2) \sin^2 k_\alpha (\gamma_\mu\gamma_\alpha\Gamma_X)
\otimes (\Gamma_Y\gamma_\alpha\gamma_\mu) \right],\end{aligned}$$ where $K$, $T$ and $A_{SP}$ are given in eqs.(\[eq:K\]), (\[eq:T\]) and (\[eq:A\_SP\]). It is noted that a sum of $I_{PD}^b$ and $I_{PD}^c$ becomes $$\begin{aligned}
& &g^2\frac{N+1}{2N}\varepsilon^{ijk}
K \left[2T (\Gamma_X \otimes\Gamma_Y) +
\cos^2 (k_\mu/2) \sin^2 k_\alpha (\gamma_\mu\gamma_\alpha\Gamma_X+
\Gamma_X\gamma_\alpha\gamma_\mu)
\otimes (\Gamma_Y\gamma_\alpha\gamma_\mu) \right] \nonumber \\
& = &
g^2\frac{N+1}{2N}\varepsilon^{ijk}
K \left[2T (\Gamma_X \otimes\Gamma_Y) +
\cos^2 (k_\mu/2) \sin^2 k_\alpha (\{\gamma_\mu\gamma_\alpha
+\gamma_\alpha\gamma_\mu \}\Gamma_X)
\otimes (\Gamma_Y\gamma_\alpha\gamma_\mu) \right ]\end{aligned}$$ for $\Gamma_X = P_R$ or $P_L$, therefore no Fierz transformation is necessary to simplify the spinor structure of the total contribution. Finally we obtain $$\begin{aligned}
I_{PD}^a+I_{PD}^b+I_{PD}^c &=& g^2\frac{N+1}{2N}\varepsilon^{ijk}
K \left[\Gamma_X\otimes\Gamma_Y\right]
\left[ 3T + A_{SP} + 2A_{VA} \right].\end{aligned}$$ Compared with the tree level result of eq.(\[eq:PD\_tree\]) we find that the vertex correction is multiplicative up to the one-loop level: $$\Lambda_{PD} =
\left[ 1 + g^2 \frac{N+1}{2N}\left\{
3<T> + <A_{SP}> + 2<A_{VA}> \right\}\right]
\Lambda_{PD}^{(0)},$$ where $<X>$ $(X = T, A_{VA}, A_{SP})$ are defined in eq.(\[eq:integ\_X\]). We remark that in the Wilson case ${\cal O}_{PD}$ mixes with other operators which have different chiral structures under renormalization[@pt_w3].
Taking account of the contribution of the wave function the lattice renormalization factor for ${\cal O}_{PD}$ is expressed as $$Z_{PD}^{lat} = (1-w_0^2)^{3/2}Z_w^{3/2} Z_2^{3/2}V_{PD},$$ where $$\begin{aligned}
V_{PD} &=& 1+ \frac{g^2}{16\pi^2}\left[-\delta_{PD}\log (\lambda a)^2 +
v_{PD}\right],\end{aligned}$$ $$\begin{aligned}
v_{PD} &=& \frac{16\pi^2 (N+1)}{2N}\left[ 3 <T> + <<A_{SP}>> +
2 <<A_{VA}>> \right]+\delta_{PD}\log\pi^2\end{aligned}$$ with $\delta_{PD}=\displaystyle\frac{6(N+1)}{2N}$. Numerical values for $v_{PD}$, evaluated as before, are given in Table \[tab:4fermi\] as a function of $M$.
The corresponding continuum renormalization factors in the ${{\overline {\rm MS}}}$ scheme are calculated employing the DRED scheme as the regularization in the Feynman gauge with the fictitious gluon mass $\lambda$. The vertex correction for ${\cal O}_{PD}$ is $$\begin{aligned}
V_{PD}^{{{\overline {\rm MS}}}} &=& 1+ \frac{g^2}{16\pi^2}\delta_{PD}
\left[-\log (\lambda/\mu)^2 +1 \right],\end{aligned}$$ giving $v_{PD}^{{{\overline {\rm MS}}}}=\delta_{PD}$. We remark that in this case the evanescent operator does not appear at the one-loop level.
Combining this result with the previous lattice one we finally obtain the relation between the operators ${\cal O}_{PD}^{{{\overline {\rm MS}}}}$ and ${\cal O}_{PD}^{lat}$: $$\begin{aligned}
{\cal O}_{PD}^{{{\overline {\rm MS}}}}(\mu) & = &\frac{1}{(1-w_0^2)^{3/2} Z_w^{3/2}}
Z_{PD} (\mu a ) {\cal O}_{PD}^{lat} (1/a),\end{aligned}$$ with $$\begin{aligned}
Z_{PD} (\mu a) &=& \frac{ (Z_2^{{{\overline {\rm MS}}}})^{3/2} V_{PD}^{{{\overline {\rm MS}}}}}
{(Z_2)^{3/2} V_{PD}} \nonumber \\
&=& 1 + \frac{g^2}{16\pi^2}\left[
(\delta_{PD} - 3 C_F/2)\log (\mu a)^2 + z_{PD} \right],
\\
z_{PD} &=& v_{PD}^{{{\overline {\rm MS}}}} - v_{PD} + \frac{3}{2} C_F\{\Sigma_1^{{{\overline {\rm MS}}}}
-\Sigma_1\}.\end{aligned}$$ We present numerical values for $z_{PD}$ in Table \[tab:total\] and those for the mean-field improved one, $z_{PD}^{MF}$, in Table \[tab:totalMF\].
Renormalization factor for $B_K$ {#sec:bk}
================================
As an application of results in the previous sections, we estimate a renormalization factor for the kaon $B$ parameter $B_K$, defined by $$B_K =\frac{\langle \overline{K}^0 \vert {\cal O}_+ \vert K^0 \rangle}
{\frac{8}{3} \langle \overline{K}^0 \vert A_4 \vert 0 \rangle
\langle 0 \vert A_4 \vert K^0 \rangle}$$ with $q_1=q_3 = s$ and $q_2=q_4= d$ in ${\cal O}_+$.
Denoting the renormalization factor between the continuum $B_K$ at scale $\mu$ and the lattice one at scale $1/a$ as $Z_{B_K}( \mu a )$, we obtain $$Z_{B_K}( \mu a ) =\frac{(1-w_0^2)^{-2} Z_w^{-2} Z_+ (\mu a )}
{(1-w_0)^{-2} Z_w^{-2} Z_A(\mu a)^2}
=\frac{Z_+ (\mu a )}{Z_A(\mu a)^2},$$ where $$Z_+(\mu a) = 1 + \frac{g^2}{16\pi^2}\left[ -4\log (\mu a) + z_+ \right]$$ from eq. (\[eq:4fermi\]) in this paper, and $$Z_A(\mu a) = 1 + \frac{g^2 C_F}{16\pi^2} z_A$$ from Ref. [@AIKT98], so that $$Z_{B_K}(\mu a) = 1 +\frac{g^2}{16\pi^2}\left[ -4\log (\mu a) + z_+
-2 C_F z_A \right] .$$ Note that $z_A$ in Ref. [@AIKT98] is evaluated in the NDR scheme while the DRED scheme is used for $z_+$ in this paper. From the result in Appendix B we have $$z_A({\rm DRED}) = z_A({\rm NDR}) +1/2, \qquad
z_+({\rm DRED}) = z_+({\rm NDR}) +3 .$$
In Ref.[@Blum-Soni] $B_K$ has been evaluated at $\beta = 5.85$, 6.0 with $ M= 1.7$ and $\beta = 6.3$ with $M=1.5$, using domain-wall QCD with the quenched approximation. Here we explicitly calculate $Z_{B_K}( \mu a )$ for these parameters. From Table \[tab:total\] and the previous result[@AIKT98], $z_+ = -41.854 (-42.399)$, $z_A = -17.039 (-16.827)$ and $z_+ -2C_F z_A = 3.583 (2.473)$ for $M=1.7 (1.5)$ in the DRED scheme, and $z_+ = -44.854 (-45.399)$, $z_A = -17.539 (-17.327)$ and $z_+ -2C_F z_A = 1.917 (0.8063)$ for $M=1.7 (1.5)$ in the NDR scheme. Taking $\mu =1/a$ and $g^2 = g^2_{\overline{MS}}(1/a)$, estimated by the formula $$\frac{1}{g^2_{\overline{MS}}}(1/a) = P\frac{\beta}{6} -0.13486$$ for the quenched QCD with $P$ being the average value of the plaquette, we have $Z_{B_K} = $ 1.053 (1.029), 1.049 (1.026) and 1.030 (1.010) at $\beta =$ 5.85, 6.0 and 6.3, respectively, in the DRED (NDR) scheme. Sizes of one-loop corrections for $B_K$ are not so large, $1-5\%$, at these $\beta$ values even without mean-field improvement, since the large contribution, which comes from a $(1-w_0) Z_w$ factor, cancels out in the ratio of ${\cal O}_+$ and $A_4^2$.
If we employ the mean-field improvement by replacing $M\rightarrow
{\widetilde{M}}=M+4(u-1)$ with $u = P^{1/4}$, we obtain $Z_{B_K} = $ 1.018 (0.994), 1.017 (0.994) and 1.009 (0.988) at $\beta =$ 5.85, 6.0 and 6.3, respectively, in the DRED (NDR) scheme. See Appendix A for some remarks.
Note that there is no mean-field improvement factor for $B_K$ in actual simulations since, as mentioned before, it is defined by the ratio. Therefore the difference between values of $Z_{B_K}$ with and without mean-field improvement comes from higher order ambiguity in perturbation theory.
Necessary informations for the analysis in this section are given in Table \[tab:bk\], together with values of $Z_{B_K}$.
Conclusion {#sec:concl}
==========
In this paper we have calculated the one-loop contributions for the renormalization factors of the three- and four-quark operators in DWQCD. We have demonstrated that the three- and four-quark operators in DWQCD can be renormalized without any operator mixing between different chiralities as opposed to the Wilson case. This desirable property in DWQCD would practically surpass the cost of the introduction of an unphysical fifth dimension. The numerical values for the finite parts $z_X$ with $X=\pm,1,2,PD$ settle in reasonable magnitude with the mean-field improvement, while unimproved values are rather large in general.
In this work we do not treat the operators which yield the so-called “penguin” diagram. It seems feasible to carry out the calculation of their renormalization factors, which we leave to future investigation.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported in part by the Grants-in-Aid for Scientific Research from the Ministry of Education, Science and Culture (Nos. 2373, 2375). T. I., Y. K. and Y. T. are supported by Japan Society for Promotion of Science.
Appendix A: Mean-field improvement {#appendix-a-mean-field-improvement .unnumbered}
==================================
The mean-field improvement[@MF] in our paper uses $$u= 1 - \frac{g^2 C_F}{2}T$$ with $T= 0.15493$, which is the value for the link in Feynman gauge. It may be better to use $u$ from $K_c$ or plaquette in DWQCD. In that case $$u = 1 - \frac{g^2 C_F}{2}(T + \delta T),$$ where $\delta T =0.00793$ for $K_c$, or $\delta T = -0.02993$ for plaquette. Accordingly we have to modify renormalization factors as follows: $$\begin{aligned}
z_w^{MF}(T+\delta T) &=& z_w^{MF}(T)+\frac{2w_0}{1-w_0^2}16\pi^2\times
2\delta T, \\
z_2^{MF}(T+\delta T) &=&z_2^{MF}(T)+16\pi^2\times \delta T /2, \\
z_{\Gamma,\rm bilinear}^{MF}(T+\delta T) &=& z_{\Gamma,\rm bilinear}^{MF}(T)
+16\pi^2 \times \delta T/2, \\
z_{\Gamma,\rm 4-quark}^{MF}(T+\delta T) &=& z_{\Gamma,\rm 4-quark}^{MF}(T)
+16\pi^2 \times\delta T/2 \times 2 C_F, \\
z_{\Gamma,\rm 3-quark}^{MF}(T+\delta T) &=& z_{\Gamma,\rm 3-quark}^{MF}(T)
+16\pi^2 \times\delta T/2 \times \frac{3}{2} C_F .\end{aligned}$$
Appendix B: Naive Dimensional Regularization(NDR) {#appendix-b-naive-dimensional-regularizationndr .unnumbered}
=================================================
In this appendix we compile the finite part of the renormalization constant in the ${{\overline {\rm MS}}}$ subtraction scheme with the Naive Dimensional Regularization: $$\begin{aligned}
\Sigma_1^{{{\overline {\rm MS}}}} & =& 1/2, \\
v_+^{{{\overline {\rm MS}}}} &=& \delta_+\times\{3/2 - \frac{2N+3}{N-2}\}, \\
v_-^{{{\overline {\rm MS}}}} &=& \delta_-\times\{3/2 - \frac{2N-3}{N+2}\}, \\
v_{ij}^{{{\overline {\rm MS}}}} & = &
\left(
\begin{array}{cc}
\delta_1 \times\left\{3/2-\displaystyle \frac{2(N^2-5)}{N^2-4}\right\}, &
\delta_2 \times 3/4 \\
\displaystyle\frac{1}{N}\times 3, & \delta_2\times 1/2
\end{array}
\right), \\
v_{PD}^{{{\overline {\rm MS}}}} & = & \delta_{PD}\times 2/3,\end{aligned}$$ where $v_{ij}$ with $i,j=1,2$ is a matrix, which represents the mixing of the finite part for ${\cal O}_{1,2}$. The one-loop vertex corrections for ${\cal O}_{\Gamma}$ $(\Gamma=\pm,1,2)$ require to specify their evanescent operators, which originates from the property that the Fierz transformation can not be defined in the NDR scheme. We employ $$\begin{aligned}
E^{\rm NDR}_{\pm}&=&
\gamma_\rho\gamma_\delta\gamma_\mu(1-\gamma_5)\otimes
\gamma_\mu(1-\gamma_5)\gamma_\delta\gamma_\rho
-{(2-D)^2}\gamma_\mu(1-\gamma_5)\otimes\gamma_\mu(1-\gamma_5),\\
E^{\rm NDR}_{1,2}&=&
\gamma_\rho\gamma_\delta\gamma_\mu(1-\gamma_5)\otimes
\gamma_\mu(1+\gamma_5)\gamma_\delta\gamma_\rho
-{D^2}\gamma_\mu(1-\gamma_5)\otimes\gamma_\mu(1+\gamma_5),\end{aligned}$$ where $D$ is the reduced space-time dimension. On the other hand, the evanescent operator does not appear in the one-loop vertex correction of ${\cal O}_{PD}$.
For later convenience, values of the finite part of quark bilinear operators are also given here. For NDR scheme $$z_{V,A}^{{{\overline {\rm MS}}}} = 0, \qquad z_{S,P}^{{{\overline {\rm MS}}}} = 5/2,
\qquad z_T^{{{\overline {\rm MS}}}} = 1/2,$$ while for DRED scheme $$z_{V,A}^{{{\overline {\rm MS}}}} = 1/2, \qquad z_{S,P}^{{{\overline {\rm MS}}}} = 7/2,
\qquad z_T^{{{\overline {\rm MS}}}} = -1/2,$$ where the evanescent operators are $$\begin{aligned}
E^{\rm DRED}_{\gamma_\mu}&=&
{\bar \delta}_{\mu\nu}\gamma_\nu-\frac{D}{4}\gamma_\mu, \\
E^{\rm DRED}_{\gamma_\mu\gamma_5}&=&
{\bar \delta}_{\mu\nu}\gamma_\nu\gamma_5-\frac{D}{4}\gamma_\mu\gamma_5.\end{aligned}$$
Y. Shamir, [[*[Nucl. Phys.]{}*]{} [**B406**]{} (1993) 90]{}.
V. Furman and Y. Shamir, [[*[Nucl. Phys.]{}*]{} [**B439**]{} (1995) 54]{}.
T. Blum and A. Soni, [[*[Phys. Rev.]{}*]{} [**D56**]{} (1997) 174]{} ; hep-lat/9706023 ; hep-lat/9712004 .
T. Blum, A. Soni and M. Wingate, hep-lat/9809065.
T. Blum, hep-lat/9810017 and references therein.
S. Aoki and Y. Taniguchi, hep-lat/9711004 (to appear in PRD).
S. Aoki, T. Izubuchi, Y. Kuramashi and Y. Taniguchi, hep-lat/9810020 (to appear in PRD).
G. Martinelli, [[*[Phys. Lett.]{}*]{} [**B141**]{} (1984) 395]{}; C. Bernard, T. Draper and A. Soni, [[*[Phys. Rev.]{}*]{} [**D36**]{} (1987) 3224]{}.
A. J. Buras and P. H. Weisz, [[*[Nucl. Phys.]{}*]{} [**B333**]{} (1990) 66]{}; S. Herrlich and U. Nierste, [[*[Nucl. Phys.]{}*]{} [**B455**]{} (1995) 39]{}.
D. G. Richards, C. T. Sachrajda and C. J. Scott, [[*[Nucl. Phys.]{}*]{} [**B286**]{} (1987) 683]{}.
G. P. Lepage and P. Mackenzie, [[*[Phys. Rev.]{}*]{} [**D48**]{} (1993) 2250]{}.
$\Gamma$ $J_a^{AB}$ $J_b^{AB}$ $J_c^{AB}$
---------- -------------------------------------------------------------------- ------------------------------------------------------------------ --------------------------------------------------------------------
$\pm$ $T^A T^B {\widetilde{\otimes}} 1 \pm T^A {\widetilde{\odot}} T^B$ $T^A {\widetilde{\otimes}} T^B \pm T^A {\widetilde{\odot}} T^B$ $T^A {\widetilde{\otimes}} T^B \pm T^A T^B {\widetilde{\odot}} 1$
$1$ $-N T^A T^B {\widetilde{\otimes}} 1 + T^A {\widetilde{\odot}} T^B$ $-N T^A {\widetilde{\otimes}} T^B + T^A {\widetilde{\odot}} T^B$ $-N T^A {\widetilde{\otimes}} T^B + T^A T^B {\widetilde{\odot}} 1$
$2$ $T^A {\widetilde{\odot}} T^B$ $T^A {\widetilde{\odot}} T^B$ $T^A T^B {\widetilde{\odot}} 1$
: Color factors for $I_\Gamma^{a,b,c}$ ($\Gamma=\pm,1,2$).[]{data-label="tab:color"}
$M$ $V_+$ $V_- $ $V_1$ $V_2$ $V_{PD}$
------ ------------ ----------- ------------- ----------- ----------
0.05 13.9096(8) 10.847(8) 13.3992(19) 8.805(12) 8.646(5)
0.10 13.5696 11.537 13.2309 10.182 8.992
0.15 13.2941 12.098 13.0948 11.301 9.273
0.20 13.0548 12.587 12.9768 12.275 9.518
0.25 12.8391 13.029 12.8708 13.155 9.740
0.30 12.6404 13.438 12.7734 13.970 9.946
0.35 12.4542 13.822 12.6822 14.734 10.139
0.40 12.2775 14.188 12.5960 15.462 10.323
0.45 12.1083 14.539 12.5135 16.160 10.499
0.50 11.9449 14.880 12.4341 16.837 10.671
0.55 11.7861 15.212 12.3571 17.496 10.838
0.60 11.6307 15.538 12.2819 18.143 11.002
0.65 11.4779 15.859 12.2081 18.780 11.164
0.70 11.3269 16.178 12.1354 19.412 11.325
0.75 11.1770 16.495 12.0634 20.041 11.485
0.80 11.0275 16.813 11.9917 20.670 11.645
0.85 10.8779 17.131 11.9201 21.300 11.806
0.90 10.7276 17.452 11.8484 21.935 11.968
0.95 10.5760 17.777 11.7762 22.578 12.133
1.00 10.4225 18.107 11.7033 23.230 12.300
1.05 10.2659 18.437 11.6278 23.885 12.466
1.10 10.1076 18.790 11.5547 24.579 12.646
1.15 9.9443 19.139 11.4768 25.269 12.822
1.20 9.7768 19.505 11.3981 25.990 13.007
1.25 9.6037 19.880 11.3165 26.732 13.198
1.30 9.4244 20.272 11.2323 27.504 13.396
1.35 9.2375 20.680 11.1446 28.309 13.603
1.40 9.0419 21.109 11.0530 29.153 13.820
1.45 8.8361 21.560 10.9567 30.042 14.049
1.50 8.6183 22.037 10.8547 30.983 14.291
1.55 8.3863 22.547 10.7464 31.987 14.550
1.60 8.1375 23.093 10.6301 33.063 14.827
1.65 7.8685 23.683 10.5043 34.227 15.127
1.70 7.5747 24.328 10.3668 35.496 15.454
1.75 7.2502 25.038 10.2149 36.897 15.814
1.80 6.8864 25.832 10.0440 38.463 16.216
1.85 6.4706 26.737 9.8482 40.247 16.675
1.90 5.9812 27.794 9.6168 42.337 17.210
1.95 5.3749 29.093 9.3281 44.907 17.867
: Numerical values for $V_\Gamma$ ($\Gamma=\pm,1,2,PD$) as a function of $M$.[]{data-label="tab:4fermi"}
$M$ $z_+$ $z_-$ $z_1$ $z_2$ $z_{PD}$
------ ------------- ------------- ------------ ------------- ------------
0.05 -49.908(10) -40.845(11) -48.397(8) -34.803(15) -32.144(7)
0.10 -49.332 -41.300 -47.994 -35.945 -32.314
0.15 -48.847 -41.651 -47.647 -36.853 -32.437
0.20 -48.416 -41.949 -47.338 -37.637 -32.539
0.25 -48.025 -42.215 -47.057 -38.341 -32.629
0.30 -47.663 -42.461 -46.796 -38.993 -32.713
0.35 -47.325 -42.693 -46.553 -39.605 -32.792
0.40 -47.007 -42.918 -46.326 -40.191 -32.870
0.45 -46.706 -43.137 -46.111 -40.758 -32.948
0.50 -46.419 -43.354 -45.908 -41.311 -33.027
0.55 -46.145 -43.571 -45.716 -41.855 -33.108
0.60 -45.883 -43.790 -45.534 -42.395 -33.191
0.65 -45.631 -44.012 -45.361 -42.933 -33.279
0.70 -45.388 -44.239 -45.196 -43.473 -33.370
0.75 -45.153 -44.472 -45.040 -44.017 -33.467
0.80 -44.927 -44.712 -44.891 -44.569 -33.570
0.85 -44.708 -44.961 -44.750 -45.130 -33.679
0.90 -44.496 -45.220 -44.617 -45.704 -33.795
0.95 -44.290 -45.491 -44.490 -46.292 -33.918
1.00 -44.091 -45.776 -44.372 -46.899 -34.051
1.05 -43.898 -46.070 -44.260 -47.517 -34.190
1.10 -43.710 -46.392 -44.157 -48.181 -34.347
1.15 -43.528 -46.723 -44.061 -48.853 -34.510
1.20 -43.352 -47.079 -43.973 -49.565 -34.688
1.25 -43.180 -47.457 -43.893 -50.308 -34.880
1.30 -43.014 -47.862 -43.822 -51.094 -35.088
1.35 -42.853 -48.296 -43.760 -51.924 -35.315
1.40 -42.697 -48.763 -43.708 -52.808 -35.561
1.45 -42.545 -49.269 -43.666 -53.751 -35.831
1.50 -42.399 -49.818 -43.635 -54.763 -36.127
1.55 -42.257 -50.417 -43.617 -55.857 -36.453
1.60 -42.119 -51.074 -43.611 -57.044 -36.813
1.65 -41.985 -51.800 -43.621 -58.343 -37.214
1.70 -41.854 -52.607 -43.646 -59.776 -37.663
1.75 -41.725 -53.513 -43.690 -61.372 -38.170
1.80 -41.595 -54.541 -43.753 -63.172 -38.748
1.85 -41.460 -55.727 -43.838 -65.237 -39.417
1.90 -41.311 -57.124 -43.946 -67.666 -40.207
1.95 -41.121 -58.840 -44.074 -70.652 -41.177
: Numerical values for $z_\Gamma$ ($\Gamma=\pm,1,2,PD$) as a function of $M$.[]{data-label="tab:total"}
$M$ $z_+^{MF}$ $z_-^{MF}$ $z_1^{MF}$ $z_2^{MF}$ $z_{PD}^{MF}$
------ ------------- ------------ ------------ ------------ ---------------
0.05 -17.287(10) -8.224(11) -15.776(8) -2.182(15) -7.679(7)
0.10 -16.712 -8.679 -15.373 -3.324 -7.848
0.15 -16.226 -9.030 -15.027 -4.232 -7.972
0.20 -15.796 -9.328 -14.718 -5.016 -8.074
0.25 -15.404 -9.594 -14.436 -5.720 -8.164
0.30 -15.042 -9.840 -14.175 -6.372 -8.247
0.35 -14.705 -10.073 -13.933 -6.98 -8.326
0.40 -14.386 -10.297 -13.705 -7.57 -8.404
0.45 -14.085 -10.516 -13.490 -8.13 -8.482
0.50 -13.799 -10.734 -13.288 -8.69 -8.561
0.55 -13.525 -10.951 -13.096 -9.23 -8.642
0.60 -13.262 -11.169 -12.913 -9.77 -8.726
0.65 -13.010 -11.391 -12.740 -10.3 -8.813
0.70 -12.767 -11.618 -12.575 -10.8 -8.905
0.75 -12.532 -11.851 -12.419 -11.3 -9.002
0.80 -12.306 -12.091 -12.270 -11.9 -9.104
0.85 -12.087 -12.340 -12.129 -12.5 -9.213
0.90 -11.875 -12.600 -11.996 -13.0 -9.329
0.95 -11.670 -12.871 -11.870 -13.6 -9.453
1.00 -11.470 -13.155 -11.751 -14.2 -9.585
1.05 -11.278 -13.449 -11.639 -14.8 -9.725
1.10 -11.089 -13.772 -11.536 -15.5 -9.882
1.15 -10.908 -14.103 -11.440 -16.2 -10.044
1.20 -10.731 -14.459 -11.352 -16.9 -10.223
1.25 -10.559 -14.836 -11.272 -17.6 -10.414
1.30 -10.393 -15.241 -11.201 -18.4 -10.623
1.35 -10.232 -15.675 -11.139 -19.3 -10.849
1.40 -10.076 -16.143 -11.087 -20.1 -11.096
1.45 -9.925 -16.648 -11.045 -21.13 -11.365
1.50 -9.778 -17.197 -11.014 -22.14 -11.661
1.55 -9.636 -17.796 -10.996 -23.23 -11.987
1.60 -9.498 -18.453 -10.991 -24.42 -12.347
1.65 -9.364 -19.179 -11.000 -25.72 -12.749
1.70 -9.233 -19.986 -11.025 -27.15 -13.198
1.75 -9.104 -20.892 -11.069 -28.75 -13.705
1.80 -8.975 -21.920 -11.132 -30.55 -14.283
1.85 -8.840 -23.106 -11.217 -32.61 -14.952
1.90 -8.690 -24.503 -11.326 -35.04 -15.742
1.95 -8.500 -26.219 -11.453 -38.03 -16.711
: Numerical value for $z_\Gamma^{MF}$ ($\Gamma=\pm,1,2,PD$) as a function of $M$.[]{data-label="tab:totalMF"}
---------------------------- --------- --------- --------- --------- --------- ---------
$\beta$
$M$
$P$
$g^2_{\overline{MS}}(1/a)$
$u$
${\widetilde{M}}$
DRED NDR DRED NDR DRED NDR
$z_+$ -41.854 -44.854 -41.854 -44.854 -42.399 -45.399
$z_A$ -17.039 -17.539 -17.039 -17.539 -16.827 -17.327
$z_+ - 2C_F z_A$ 3.583 1.917 3.583 1.917 2.473 0.806
$Z_{B_K}(\mu a=1)$ 1.053 1.029 1.049 1.026 1.030 1.010
$z_+^{MF}$ -17.033 -20.033 -17.033 -20.033 -17.580 -20.580
$z_A^{MF}$ -6.853 -7.353 -6.853 -7.353 -6.864 -7.364
$z_+^{MF} - 2C_F z_A^{MF}$ 1.242 -0.425 1.242 -0.425 0.724 -0.943
$Z_{B_K}^{MF}(\mu a=1)$ 1.018 0.994 1.017 0.994 1.009 0.988
---------------------------- --------- --------- --------- --------- --------- ---------
: Renormalization factor for $B_K$($1/a$) at some parameters.[]{data-label="tab:bk"}
(420,450)(0,20) (60,60)\[l\][${\rm a}^\prime$]{} (70,166)[4.5]{} (70,154)[4.5]{} (68,166)(8,226) (132,226)(72,166) (68,154)(8,94) (132,94)(72,154) (20,105)(120,105)[5]{}[8]{} (60,260)\[l\][a]{} (0,440)\[l\][$\alpha,i$]{} (115,440)\[l\][$\beta,j$]{} (0,285)\[l\][$\gamma,k$]{} (115,285)\[l\][$\delta,l$]{} (70,366)[4.5]{} (70,354)[4.5]{} (68,366)(8,426) (132,426)(72,366) (68,354)(8,294) (132,294)(72,354) (20,415)(120,415)[-5]{}[8]{} (200,60)\[l\][${\rm b}^\prime$]{} (210,166)[4.5]{} (210,154)[4.5]{} (208,166)(148,226) (272,226)(212,166) (208,154)(148,94) (272,94)(212,154) (260,213)(260,107)[-5]{}[8]{} (200,260)\[l\][b]{} (210,366)[4.5]{} (210,354)[4.5]{} (208,366)(148,426) (272,426)(212,366) (208,354)(148,294) (272,294)(212,354) (160,413)(160,307)[5]{}[8]{} (340,60)\[l\][${\rm c}^\prime$]{} (350,166)[4.5]{} (350,154)[4.5]{} (348,166)(288,226) (412,226)(352,166) (348,154)(288,94) (412,94)(352,154) (350,160)(30,51,230)[-5]{}[7]{} (340,260)\[l\][c]{} (350,366)[4.5]{} (350,354)[4.5]{} (348,366)(288,426) (412,426)(352,366) (348,354)(288,294) (412,294)(352,354) (350,360)(30,-50,130)[-5]{}[7]{}
(420,250)(0,20) (60,60)\[l\][a]{} (0,240)\[l\][$\alpha,i$]{} (115,240)\[l\][$\beta,j$]{} (115,85)\[l\][$\gamma,k$]{} (70,166)[4.5]{} (70,154)[4.5]{} (68,166)(8,226) (132,226)(72,166) (132,94)(72,154) (20,215)(120,215)[-5]{}[8]{} (200,60)\[l\][b]{} (210,166)[4.5]{} (210,154)[4.5]{} (208,166)(148,226) (272,226)(212,166) (272,94)(212,154) (260,213)(260,107)[-5]{}[8]{} (340,60)\[l\][c]{} (350,166)[4.5]{} (350,154)[4.5]{} (348,166)(288,226) (412,226)(352,166) (412,94)(352,154) (350,160)(30,-50,130)[-5]{}[7]{}
[^1]: On leave from Institute of Particle and Nuclear Studies, High Energy Accelerator Research Organization(KEK), Tsukuba, Ibaraki 305-0801, Japan
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
The Hopf algebra structure underlying Feynman diagrams which governs the process of renormalization in perturbative quantum field theory is reviewed. Recent progress is briefly summarized with an emphasis on further directions of research.\
---
\[section\] \[theorem\][Proposition]{} \[theorem\][Definition]{} \[theorem\][Lemma]{} \[theorem\][Corollary]{} \[section\] \[section\] \[theorem\][Remark]{} \[theorem\][Example]{} \[section\]
v
Ł ¶
ł
\
KURUSCH EBRAHIMI-FARD[^1]\
\
DIRK KREIMER[^2]\
\
December 6, 2005\
[`J. Phys. A: Math. Gen., 38, (2005), R385-R406.`]{}
\
Introduction and Overview
=========================
Quantum field theory (QFT) by now has a long and outstandingly successful history in all theories of physics. Merging the two major revolutionary achievements of early 20th century physics, quantum mechanics and special relativity, the founding fathers of QFT were setting out for an unified description of elementary particles phenomena. Its ideas and techniques found far reaching applications in different and very distinct areas of theoretical physics, and pure and applied mathematics.
Several approaches to QFT have been developed so far. Wightman’s early axiomatic [@StWh] setting leading to constructive QFT, together with Haag’s mathematically elegant and rigorous algebraic formulation of QFT in terms of von Neumann algebras [@Haag], best describes the nowadays common believe of what should be the general physical principles underlying any QFT. Still, despite the enormous and mathematically rigorous progress which has been made using these formulations, both approaches have several problems to make fruitful contact with experimental results, whilst they give a crucial insight into the structure of free quantum fields.
The perturbative approach to quantum field theory is the most successful. Theoretical predictions of physical quantities made by using their expansion in terms of –renormalized– Feynman graphs match experimental results with a vertiginous high precision. Nevertheless, in most, if not all, of the interesting and relevant 4-dimensional quantum field theories, performing even simple perturbative calculations one cannot avoid facing ill-defined integrals. The removal of these divergences in a sound way is the process of renormalization, better known by the illustrative description of ”sweeping them under the carpet". The basic idea of perturbative renormalization in QFT goes back to Kramers [@Brown], and was successfully applied for the first time in a 1947 seminal paper by Bethe [@Bethe], dealing with the concrete problem of the self energy contribution for the Lamb shift in perturbative quantum electrodynamics (QED). The latter can nowadays be regarded as one of the best tested physics theories. Its modern extension to the standard model of elementary particles represents one of the cornerstones of our present understanding of the physical world. Here again the perturbative treatment together with renormalization is the bread-and-butter of the practitioner in high energy physics.
Maintaining the physical principles of locality, unitarity, and Lorentz invariance, renormalization theory may be summed up by the statement that to all orders in perturbation theory the (ultraviolet) divergencies can be absorbed in a redefinition of the parameters defining the QFT. Here two distinct concepts enter, that of renormalizability, and the process of renormalization. The former distinguishes those theories with only a finite number of parameters, lending them considerably more predictive power. The process of renormalization instead works indifferently of the number of parameters.
Soon after Bethe’s paper on perturbative QED, there have been several approaches to establish that quantum field theories are renormalizable in general. Dyson [@Dyson1; @Dyson2] was the first to do so, using integral equations and skeleton expansions for Green’s functions. His work was then continued by Salam and Weinberg. Unfortunately, this attempt failed in the first instance, due to a problem related to a particular 14th order QED graph, but could be cured later. The second approach, based on earlier work by Stückelberg and Green, was taken by Bogoliubov and Parasiuk [@BP; @BS], using a recursive subtraction method, known as Bogoliubov’s $\mathrm{\bar{R}}$-map. Also their proof contained a loophole, but eventually found its final and satisfying form with the work of Hepp [@Hepp] and later Zimmermann [@Zimmermann]. This standard result is nowadays well-known under the name Bogoliubov–Parasiuk–Hepp–Zimmermann (BPHZ) renormalization prescription. Later, Epstein and Glaser [@EpGl] presented a rigorous proof of renormalizability situated in the realm of the axiomatic treatment of QFT. A fourth approach was taken by Blaer and Young [@BY], using the renormalization group equations, going back to a suggestion by Callan. At this point we refer the interested reader to consult the work by Caswell and Kennedy [@CaswellK], Brown [@Brown], Delamotte [@Delamotte], Collins [@Collins], and Wightman [@Wightman] for more references and details.
Notwithstanding its somewhat notorious reputation, renormalization theory, together with the gauge principle, forms the backbone of the perturbative approach to physically relevant quantum field theories. These days, the modern point of view, represented by the concept of Wilson’s renormalization group, elevates it even to a fundamental structure in the understanding of high energy physics.
Unfortunately, despite its accomplishments, renormalization theory was stigmatized, especially for its lack of a firm mathematical underpinning. Indeed, examining the current introductory and advanced literature on renormalization, as it is used in everyday applications in many branches of physics, one feels the need for a more conceptual picture unifying mathematical and computational aspects. A possible reason for this situation might have been the fact that its building blocks, the (one-particle irreducible) Feynman graphs in itself appeared to be unrelated to a sound mathematical structure that may underlie the renormalization prescription in perturbative QFT.
Almost five decades after Bethe’s work, this changed to a great extend with the original paper by Kreimer [@Kreimer1] introducing the notion of Hopf algebra. The ensuing work by Kreimer [@Kreimer2; @Kreimer6] and collaborators, especially those of Broadhurst and Kreimer [@BK1; @BK2; @BK3; @BK4], and Connes and Kreimer [@CK1; @CK2; @CK3] explored this new approach both in terms of its mathematical and physical content, as well as its computational aspects. The Hopf algebraic setting captures the combinatorial and algebraic aspects of the process for renormalization by organizing the Feynman graphs into a combinatorial Hopf algebra, $\mathcal{H}_F$, which is a connected graded commutative bialgebra, essentially characterized by its non-cocommutative coproduct structure map. The formulation of renormalization using Hopf algebras was completed in the work of Connes and Kreimer. It gives rise to an elegant and useful disentanglement of analytic and algebraic aspects of perturbative renormalization in general QFT, affirming the remark that ”Few physicists object nowadays to the idea that diagrams contain more truth than the underlying formalism\[...\]" made by Veltman and ’t Hooft in [@tHV].
In this review we will focus on an elementary introduction to the Hopf algebra structure on Feynman graphs combined with the description of a completely algebraic formulation of renormalization in terms of a factorization problem valid for any renormalization scheme, and based on a theorem by Atkinson and Spitzer’s identity [@EGK1; @EGK2].
Let us continue with some more details. The restricted dual of the Hopf algebra of Feynman graphs, denoted by $\mathcal{H}^*_{F}$, contains the group $\mathcal{G}:=char(\mathcal{H}_F,\mathbb{C})$ of characters, that is, algebra homomorphisms from $\mathcal{H}_F$ to the underlying base field $\mathbb{C}$. Feynman rules are understood as such linear and multiplicative maps, associating to each Feynman graph, seen as a collection of vertices and edges, its corresponding Feynman integral. This group of characters possesses a corresponding Lie algebra of derivations, or infinitesimal characters, $\mathcal{L}:=\partial
char(\mathcal{H}_F,\mathbb{C})$, which comes from a fundamental pre-Lie algebra structure on Feynman graphs.
The ill-defined Feynman integrals are plagued with ultraviolet divergences in general, and demand for a regularization prescription, where we replace the base field $\mathbb{C}$ by a (commutative and unital) algebra $A$ of Feynman amplitudes. Alternatively, we might consider Taylor expansions on the level of the integrands. Whichever way, this leads us to consider the space of $A$-valued, or regularized, linear maps ${\mathrm{Hom}}(\mathcal{H}_{F},A)$, which contains $\mathcal{G}_A:=char(\mathcal{H}_{F},A)$, the group of regularized characters, respectively its associated Lie algebra $\mathcal{L}_A:=\partial char(\mathcal{H}_{F},A)$. As a principal example serves dimensional regularization, where $A:=\mathbb{C}[\varepsilon^{-1},\varepsilon]]$, the field of Laurent series. In this context perturbative renormalization finds a compact formulation as a factorization problem in the group $\mathcal{G}_A$, to wit, the algebraic Birkhoff decomposition of Feynman rules [@CK1; @CK2]. The initial proof of the Connes–Kreimer factorization of regularized Feynman rules uses the property that Laurent series actually form a commutative Rota–Baxter algebra [@Baxter; @Rota1] with the pole part projection, $R:=R_{ms}$, as linear Rota–Baxter operator (minimal subtraction scheme map) fulfilling the Rota–Baxter relation (of weight 1) $$R(x)R(y)+R(xy)=R\big(R(x)y+xR(y)\big), \; \forall x,y \in A.$$ The linearity of $R$ permits to define a unital, but now non-commutative complete filtered Rota–Baxter algebra structure on the space ${\mathrm{Hom}}(\mathcal{H}_{F},A)$, with convolution as associative product. One of the fundamental results in the realm of commutative Rota–Baxter algebras is Spitzer’s classical identity [@EGK1; @EGK2; @RotaSmith; @Spitzer], and using its generalization to non-commutative Rota–Baxter algebras, together with Atkinson’s factorization theorem [@Atkinson] for Rota–Baxter algebras, one can show that the multiplicative factorization of Connes–Kreimer follows from an additive decomposition through the exponential map [@EGK1; @EGK2]. Hereby we realize Bogoliubov’s $\mathrm{\bar{R}}$-map as a special case of Spitzer’s identity.
We hope that this brief review will guide the reader to crucial aspects of the recent developments related to the Hopf algebraic description of renormalization theory. The long list of references is meant to indicate the rich spectrum of research directions triggered by this approach. The modest mathematical style, i.e. we do not strive for a rigorous theorem-proof presentation might help the interested and novice reader to get a glimpse of the new aspects which opened with the Hopf algebra point of view on perturbative renormalization. Some of the remarks made during the expository writing indicate points to be further developed. But, we should underline that this article is neither meant to be an introduction to (perturbative) quantum field theory nor to renormalization theory in general. Rather, we would like to focus on the by now well-understood and established combinatorial-algebraic picture that makes renormalization theory in perturbative QFT such a challenging and venerable subject from both, mathematical and physical perspectives. After reading this article and going back to the vast existing physics literature on renormalization theory in all its facets, the reader may get an idea of the interesting open questions related to its Hopf algebraic description.
In the following we will comment on assorted references with respect to their research directions in this field, in the hope to facilitate access to this developing subject. Due to the review character of this work and limited space none of these topics could be treated in full detail. We start by mentioning two recent papers [@FG3; @Manchon] devoted in great detail to the general Hopf algebra structure in renormalization theory. A more mathematical, but shorter summary was given in [@Boutet].
The initial discovery of the Hopf algebra structure grew out of a study of the number-theoretic properties of graphs with many subdivergences, see [@Kreimer9], which provides an overview of some of the results up to the year 1999, including the link between knot theory, Feynman graphs and number theory [@BK5].
The original work of Kreimer, and Connes and Kreimer [@CK1; @CK2; @CK3; @CK4; @Kreimer1; @Kreimer2] explores and settles the Hopf algebraic formulation of renormalization for general perturbative QFT, and links it to non-commutative geometry. It thereby establishes the Birkhoff decomposition for Feynman rules giving rise to an unexpected correspondence with the Riemann–Hilbert problem. In [@CORV; @FG1; @FG2; @GKM1; @Kastler1; @KW; @Kreimer4; @RV] further details were given. Malyshev’s work [@Malyshev1] shows the general character of Connes’ and Kreimer’s combinatorial Hopf algebra, applying it to Riemann surfaces in the context of ribbon Feynman graphs.
The link to Connes’ non-commutative geometry becomes evident in terms of a Hopf algebra of non-planar rooted trees [@CK1], solving a universal problem in Hochschild cohomology, and forming the role model for the Hopf algebraic structure of renormalization. This work renewed considerably the interest in Hopf algebraic aspects of combinatorics such as rooted tree Hopf algebras. In this context one must point out the work of G.-C. Rota [@Rota4] and his school, especially Rota’s and A. Joni’s seminal work [@JoniRota] from the late 1970ies, forming the starting point for the theory of incidence Hopf algebras, further developed in [@Ehrenborg; @Schmitt], see also [@FG3]. Holtkamp [@Holtkamp][^3] showed that the non-commutative version of Connes–Kreimer’s Hopf algebra of rooted trees is isomorphic to Loday and Ronco’s [@LodayRonco]. Aguiar and collaborators explored in more detail rooted tree Hopf algebras [@AguBS; @AguSot1; @AguSot2]. Hoffman in [@Hoffman] improved a result of Panaite [@Panaite], showing the isomorphism between the dual of Connes–Kreimer’s Hopf algebra and Grossmann–Larson’s [@G-L] rooted tree Hopf algebra (see also [@Foissy]). Brouder [@Brouder] explored the relation to Butcher’s seminal work on Runge–Kutta integration methods [@Butcher]. Turaev in [@Turaev1; @Turaev2] extended some combinatorial aspects of the Connes–Kreimer results, especially with respect to the notion of pre-Lie coalgebras, and thereby also gave a neat description of Connes–Kreimer’s rooted tree Hopf algebra.
Chapoton and Livernet [@ChapotonLivernet] described free pre-Lie algebras in terms of rooted tree operads. Mencattini and Kreimer [@KM1; @KM2] further analyzed the insertion and elimination Lie and pre-Lie algebraic structures of Feynman graphs [@CK4] in terms of infinite matrix representations.
In [@EGK1; @EGK2] the meaning of the Rota–Baxter relation in the context of Connes–Kreimer’s Birkhoff decomposition is investigated in detail, pointing out some parallels to the theory of classical integrable systems [@BBT; @STS1]. It thereby provides the algebraic underpinning for the factorization in terms of complete filtered Rota–Baxter algebras, Spitzer’s identity and Atkinson’s multiplicative decomposition theorem for Rota–Baxter algebras. This work was further extended in [@EG2; @EGGV] describing the combinatorics of renormalization in terms of unipotent triangular matrix representations, and their factorization capturing the process of renormalization analogously to the Birkhoff decomposition of Connes–Kreimer. An interesting application of renormalization techniques and Rota–Baxter algebras as described here to the iteration of symbols of pseudodifferential operators can be found in [@MP].
The work of Broadhurst and Kreimer [@BK1; @BK2; @BK3; @BK4] develops many computational and physics aspects. They show how to use the coproduct structure of the Hopf algebra to efficiently compute the forest formula and use the Hochschild cohomology of the algebra to resum the perturbative series. The latter two of the aforementioned references hence form the starting point for the latest work of Kreimer [@Kreimer7; @KreimerDS], and Bergbauer and Kreimer [@BergbauerKreimer2], putting emphasis on Hochschild cohomology of Hopf algebras as a source of locality, the Dyson–Schwinger equations, and even the Slavnov–Taylor identities for the couplings in generic gauge theories.
Finally, the authors in [@GKM2; @Malyshev2; @Sakakibara] started to analyze some aspects of renormalization group calculations in the Hopf algebra context.
The work [@BKK; @BW] (see also [@MUW]) is of more computational character, indicating the efficiency of the use of Hopf algebras in perturbative renormalization.
Several people [@BergbauerKreimer1; @Gudrun] investigated the link between the Hopf algebra of renormalization to the most rigorous approach to renormalization in perturbative QFT, provided by the Epstein–Glaser prescription.
Recently, progress has been made in the mathematical context of number theory, and motivic structures of Feynman integrals. The notion of equisingular connections was used to explore Tannakian categories and Galois symmetries in the spirit of differential Galois theory in [@CM1; @CM2; @CM3]. Underlying the notion of an equisingular connection is the locality of counterterms, which itself results from Hochschild cohomology. The resulting Dyson–Schwinger equation allows for gradings similar to the weight- and Hodge filtrations for the polylogarithm [@Kreimer3; @Kreimer8]. More concretely, the motivic nature of primitive graphs has been established very recently by Bloch, Esnault and Kreimer [@BlochEK].
Let us briefly outline the organization of the paper. Section \[section2\] introduces briefly the basic Lie and Hopf algebra structures in perturbative renormalization, including the pre-Lie composition of Feynman graphs and Bogoliubov’s $\bar{\mathrm{R}}$-operation. The next section uses perturbative QED as a simple example to manifest the aforementioned notions. In Section \[section4\] we formulate the process of renormalization in perturbative QFT as a factorization problem in purely algebraic terms. Emphasis is put on the freedom in choosing a particular regularization prescription, captured via the notion of commutative unital Rota–Baxter algebra. Section \[section5\] outlines the use of the Birkhoff decomposition introduced in the former section on the level of diffeomorphisms of coupling constants, in the realm of dimensional regularization together with the minimal subtraction scheme as a particular useful renormalization prescription. The review ends with a brief section on the role of Hochschild cohomology in perturbative renormalization. In an appendix we collect some general facts about Rota–Baxter algebras as they form the main ingredient for Section \[section4\].
From the Lie and Hopf algebras of graphs to Bogoliubov’s formula {#section2}
================================================================
In this section we describe the elementary Lie and Hopf algebra structures underlying perturbation theory. The reader looking for a mathematical rigorous and detailed presentation of Hopf algebras and related aspects is referred to the standard texts such as [@Abe; @FGV; @Kassel; @Sw] (see also [@FG3; @Manchon]).
Let $\mathbb{K}$ be a field of characteristic zero. All $\mathbb{K}$-algebras, denoted by a triple $(A,m,\eta)$, where $A$ is a $\mathbb{K}$-vector space with a product $m: A \otimes A \to
A$ and a unit map $\eta: \mathbb{K} \to A$, are supposed to be associative. Similarly for coalgebras, denoted by the triple $(C,\Delta, \bar{e})$, where the coproduct map $\Delta: C \to C
\otimes C$ fulfills coassociativity, i.e. $(\Delta \otimes {\mathrm{id}})
\circ \Delta=({\mathrm{id}}\otimes\Delta) \circ \Delta$, and $\bar{e}: C \to
\mathbb{K}$ is the counit map. The identity will be denoted by $\One$. All algebra homomorphisms are supposed to be unital. A bialgebra, denoted by a quintuple $\left( H =
\bigoplus_{i=0}^\infty H_{(i)}, m, \eta, \Delta, \bar{e} \right)$ consists of an algebra and coalgebra structure in a compatible way. Here, $\Delta: H \to H \otimes H$ is the coproduct, $m: H
\otimes H \to H$ the product. These maps together with the counit $\bar{e}: H\to \mathbb{K}$ fulfil the standard bialgebra axioms. See the above general references for details. It is called connected graded if $H_{(i)} H_{(j)} \subset H_{(i+j)}$ and $\Delta(H_{(i)}) \subset \bigoplus_{j+k=i} H_{(j)} \otimes
H_{(k)},$ and if $\Delta(\One)=\One\otimes\One$ and $H_{(0)}=\mathbb{K}\One$. The counit simply is $\bar{e}(\One) =
1\in\mathbb{K}$ and $\bar{e}=0$ on $\bigoplus_{i=1}^\infty
H_{(i)}.$
We call $\ker\bar{e}$ the augmentation ideal of $H$ and denote by $P$ the projection $H \rightarrow \ker\bar{e}$ onto the augmentation ideal, $P = {\mathrm{id}}- \eta\bar{e}.$ Furthermore, we use Sweedler’s notation $\Delta(h)=\sum h^\prime \otimes
h^{\prime\prime}$ for the coproduct. Let us define $${\rm Aug}^{(k)}= \big(\underbrace{P\otimes\cdots \otimes P}_{k\;{\rm times}}\big)\,
\Delta^{k-1},\;\; H \to \{\ker\bar{e}\}^{\otimes k},$$ as a map into the $k$-fold tensor product of the augmentation ideal. Here, $\Delta^{k-1}$ is defined inductively by $\Delta^0:={\mathrm{id}}$, and $\Delta^n:=(\Delta^{n-1}\otimes
{\mathrm{id}})\circ\Delta$ for $n>0$. We let $${\sl H}^{(k)}=\ker{\rm Aug}^{(k+1)}/\ker{\rm Aug}^{(k)},$$ $\forall k \geq 1$. All bialgebras considered here are bigraded in the sense that $$H=\bigoplus_{i=0}^\infty H_{(i)}=\bigoplus_{k=0}^\infty {\sl H}^{(k)},$$ where $H_{(k)} \subset \oplus_{j=1}^k {\sl H}^{(j)}$ for all $k
\geq 1$. $H_{(0)}\simeq {\sl H}^{(0)}\simeq \mathbb{K}$.
While these algebraic notions may seem rather abstract, they indeed govern the structure of quantum field theory. To understand how, we have first to study the pre-Lie algebra structure of one-particle irreducible Feynman graphs.
The Pre-Lie Structure of Feynman Graphs {#subsection1}
---------------------------------------
For each quantum field theory, we have an underlying free theory which provides propagators and hence Feynman rules for edges in Feynman graphs[^4]. The request for local interactions and a renormalizable theory then gives us Feynman rules for interactions. With those graphs come the sets of one-particle irreducible (1PI) graphs contributing to a chosen amplitude. The amplitudes are distinguished by the external fields, asymptotically free fields represented by external edges in the graphs. We call this an external leg structure, denoted by $\underline{r}$. For a renormalizable theory, there is a finite number of such external leg structures, one for each monomial in the Lagrangian.
For each such Feynman graph we hence have vertices as well as internal and external edges. External edges are edges which have an open end not connected to a vertex. They indicate the particles participating in the scattering amplitude under consideration and each such edge carries the quantum numbers of the corresponding free field. The internal edges and vertices form a graph in their own right. For an internal edge, both ends of the edge are connected to a vertex. For a graph $\Gamma$ we denote by $\Gamma^{[0]}$ its set of vertices and by $\Gamma^{[1]}:=\Gamma^{[1]}_{\rm int} \cup \Gamma^{[1]}_{\rm ext}$ its set of internal and external edges. Furthermore, $\omega_{\underline{r}}$ is the number of space-time derivatives appearing in the corresponding monomial in the Lagrangian.
We are considering 1PI Feynman graphs. By definition a graph $\Gamma$ is 1PI if and only if all graphs, obtained by removal of any one of its internal edges, are still connected. Such 1PI graphs are naturally graded by their number of independent loops, the rank of their first homology group $H_{[1]}(\Gamma,\mathbb{Z})$. We write $|\Gamma|$ for this degree of a graph $\Gamma$. Note that $|{\bf res}(\Gamma)|=0$, where we let ${\bf res}(\Gamma)$, called the residue of $\Gamma$, be the graph obtained when all edges in $\Gamma^{[1]}_{\rm int}$ shrink to a point. The graph we obtain in this manner consists of a single vertex, to which the edges $\Gamma^{[1]}_{\rm ext}$ are attached. In case the initial graph was a self-energy graph, we regard its residue as a single edge. We denote the set of all external leg structures $\underline{r}$ by $\mathfrak{R}$. For a renormalizable QFT it consists of the edges and vertices corresponding to the monomials in the Lagrangian.
Having specified free quantum fields and local interaction terms between them, one immediately obtains the set of 1PI graphs, and can consider for a given external leg structure $\underline{r}$ the set $M_{\underline{r}}$ of graphs with that external leg structure. The Green’s function for the corresponding amplitude is then obtained as the evaluation under the Feynman rules of the formal sum $$\Gamma^{\underline{r}}:= 1 + \sum_{{\bf res}(\Gamma)={\underline{r}}}
\alpha^{|\Gamma|}\frac{\Gamma}{{\rm sym}(\Gamma)},$$ where we divide by the symmetry factor ${{\rm sym}(\Gamma)}$ and $\alpha$ is a small parameter like (the square of) a coupling constant. These sums exhibit rich structure thanks to the algebraic structures of the single graphs [@KreimerDS] to be discussed now.
For a renormalizable theory, we can define a superficial degree of divergence $$\omega= \sum_{\underline{r}\in \Gamma^{[1]}_{\rm int}\cup
\Gamma^{[0]}}\omega_{\underline{r}}-4|H_{[1]}(\Gamma,\mathbb{Z})|,$$ for each such external leg structure: $\omega(\Gamma) =
\omega(\Gamma^\prime)$ if ${\bf res}(\Gamma) = {\bf
res}(\Gamma^\prime)$, all graphs with the same external leg structure have the same superficial degree of divergence. Only for a finite number of distinct external leg structures $\underline{r}
\in \mathfrak{R}$ will this degree indeed signify a divergence. Our first observation is that there is a natural pre-Lie algebra structure on 1PI graphs.
To see this, we define a bilinear operation on graphs $$\Gamma_1 * \Gamma_2 = \sum_\Gamma n(\Gamma_1,\Gamma_2;\Gamma)\Gamma,$$ where the sum is over all 1PI graphs $\Gamma$. Here, $n(\Gamma_1,\Gamma_2;\Gamma)$ is a section coefficient which counts the number of ways a subgraph $\Gamma_2$ in $\Gamma$ can be reduced to a point such that $\Gamma_1$ is obtained. The above sum is evidently finite as long as $\Gamma_1$ and $\Gamma_2$ are finite graphs, and the graphs which contribute necessarily fulfill $|\Gamma| = |\Gamma_1| + |\Gamma_2|$ and ${\bf res}(\Gamma) = {\bf
res}(\Gamma_1)$.
One then has:
The operation $\ast$ is pre-Lie: $$[\Gamma_1\ast\Gamma_2]\ast \Gamma_3 - \Gamma_1\ast[\Gamma_2\ast \Gamma_3] =
[\Gamma_1\ast \Gamma_3]\ast \Gamma_2 - \Gamma_1\ast[\Gamma_3\ast \Gamma_2].$$
This is evident when one rewrites the $*$ product in suitable gluing operations, using the dichotomy of inserting in nested or disjoint manner. See [@EMK; @Kreimer5; @Kreimer10; @Kreimer11] for more details.
Note that the equation claims that the lack of associativity in the bilinear operation $\ast$ is invariant under permutation of the elements indexed by $2,3$. This suffices to show that the anti-symmetrization of this map fulfils the Jacobi identity. Hence we get a Lie algebra ${\cal L}$ by anti-symmetrizing this operation: $$[\Gamma_1,\Gamma_2] = \Gamma_1\ast\Gamma_2-\Gamma_2\ast\Gamma_1.
\label{Lie}$$ This Lie algebra is graded and of finite dimension in each degree. Let us look at a couple of examples for pre-Lie products. We take graphs from quantum electrodynamics (QED) as a rather self-evident example. For the graphs $\gg$ and $\epemg$ with residues ${\bf
res}(\!\gg\!) = \!\!\begin{array}{c}\\[-.6cm] \scalebox{0.5}{\BOSONprop}\end{array}\!$ respectively ${\bf res}(\! \epemg
\!)=\!\!\begin{array}{c}\\[-.8cm] \!\scalebox{0.5}{\FERMprop}{}\end{array}\!$, we find $$\begin{aligned}
\gg\ast \epemg & = & 2\ggv.
\label{pre-LieExam}
\end{aligned}$$ Together with ${\cal L}$ one is led to consider the dual of its universal enveloping algebra ${\cal U}({\cal L})$ using the theorem of Milnor and Moore [@MilnorMoore]. For this we use the above grading by the loop number.
This universal enveloping algebra ${\cal U}({\cal L})$ is build from the tensor algebra $${\bf T}=\bigoplus_k T^k,\;T^k=\underbrace{{\cal L}\otimes\cdots\otimes {\cal L}}_{k\; {\rm times}},$$ by dividing out the ideal generated by the relations $$a\otimes b - b \otimes a = [a,b] \in {\cal L}.$$ Note that in ${\cal U}({\cal L})$ we have a natural concatenation product $m_*$. Even more, ${\cal U}({\cal L})$ carries a natural Hopf algebra structure with this product. For that, the Lie algebra ${\cal L}$ furnishes the primitive elements: $$\Delta_*(a)=a\otimes 1+1\otimes a,\;\forall a\in {\cal L}.$$ It is by construction a connected finitely graded Hopf algebra which is cocommutative but not commutative.
We can then consider its graded dual which will be a Hopf algebra ${\cal H}_{F}(m,\eta,\Delta,\bar{e})$ which is commutative but not cocommutative. One finds the coproduct $\Delta$ upon using a Kronecker pairing $$<Z_\Gamma,\delta_{\Gamma^\prime}>= \begin{cases} 1,\;\Gamma=\Gamma^\prime \\
{0,\; {\rm
else}}.\end{cases}$$ From there, one determines all other structure maps with ease, demanding that $$\langle Z_{[\Gamma_2,\Gamma_1]},\delta_\Gamma \rangle =
\langle Z_{\Gamma_1}\otimes Z_{\Gamma_2} - Z_{\Gamma_2}\otimes Z_{\Gamma_1},\Delta(\delta_\Gamma)\rangle.$$ In the above, we distinguished carefully between graphs $\Gamma$ as generators of the Lie algebra, denoted by $Z_\Gamma$, and graphs $\Gamma$ as generators of the Hopf algebra, denoted by $\delta_\Gamma$. The Lie algebra of graphs exponentiates to the character group of the Hopf algebra as explained below, eventually leading to Birkhoff factorization in that group.
The space of primitives of ${\cal U}({\cal L})$ is in one-to-one correspondence with the set ${\rm Indec}({\cal H}_{F})$ of indecomposables of ${\cal H}_{F}$, which is the linear span of its generators.
${\cal H}_{F}$ is a connected graded commutative Hopf algebra which describes renormalization theory. It operates on the superficially divergent 1PI Feynman graphs of the theory. The residues of these graphs are in one-to-one correspondence with the terms in the Lagrangian of a given theory. Often it is the case that several terms in a Lagrangian correspond to graphs with the same number and type of external legs, but match to different form-factor projections of the graph. In such cases, the above approach can be easily adopted. Below in Section \[section3\] we give an example for QED, incorporating its form-factor decomposition into our approach.
Bogoliubov’s recursive subtraction formula {#subsection2}
------------------------------------------
The above algebra structures are available once one has decided on the set of 1PI graphs of interest. Those one-particle irreducible graphs $\Gamma$ provide the generators $\delta_\Gamma$ of the Hopf algebra ${\cal H}_{F}=\oplus_{i=0}^\infty \H_{(i)}$, where ${\cal
H}_{ F, {\rm{lin}}}:={\rm span}(\delta_\Gamma)$, with their disjoint union providing the commutative product, which we denote by juxtaposition.
Let $\Gamma$ be a 1PI graph. The Hopf algebra ${\cal H}_{F}$ described above comes out to have a coproduct $\Delta:{\cal H}_{F}
\to {\cal H}_{F} \otimes {\cal H}_{F}$: $$\Delta(\Gamma) = \Gamma \otimes \One + \One \otimes\Gamma +
\sum_{\gamma{\subset}\Gamma}\gamma\otimes\Gamma/\gamma,$$ where the sum is over all unions of 1PI superficially divergent proper subgraphs, and we extend this definition to products of graphs, $\Delta(\Gamma_1\Gamma_2)=\Delta(\Gamma_1)\Delta(\Gamma_2)$, so that we get a bialgebra.
While the pre-Lie product respectively the Lie bracket inserted graphs into each other, dually the coproduct disentangles them. This is precisely what we make use of in renormalization theory: we have to render each subgraph finite before we can construct a local counterterm. Having a coproduct, two further structure maps of ${\cal H}_{F}$ are immediate, the counit and the antipode. The counit $\bar{e}$ vanishes on any non-trivial Hopf algebra element, $\bar{e}(X)=0$, $X \neq \One$, but $\bar{e}(\One )=1$. The antipode $S:{\cal H}_{F} \to {\cal H}_{F}$ is given by $$S(\Gamma) = -\Gamma- \sum_{\gamma{\subset}
\Gamma}S(\gamma)\Gamma/\gamma,$$ for $\Gamma \in \ker \bar{e}$, and $S(\One)=\One$. We can work out examples for the coproduct of a graph: $$\begin{aligned}
\Delta\big( \!\! \ggv \!\!\big) & = &\!\! \ggv\!\!\otimes \One + \One \otimes\!\! \ggv + 2\epemg\otimes \gg.
\end{aligned}$$ And an antipode: $$S\big(\!\!\!\! \begin{array}{l} \\[-0.4cm]
\epemfg \end{array} \!\!\!\!\big) = -\!\!\begin{array}{l} \\[-0.4cm]
\epemfg \end{array} \!\!+ \gg\epem.$$ We note in passing that the gluing operation underlying the pre-Lie insertion of graphs relies on gluing data which can be reconstructed from the subgraphs $\gamma$ and cographs $\Gamma/\gamma$ in the above coproduct. This is crucial in the proof of locality of counterterms upon studying the Hochschild cohomology of this Hopf algebra [@Kreimer10].
We have by now obtained a Hopf algebra generated by combinatorial elements, 1PI Feynman graphs. Its existence is indeed automatic once one has chosen interactions and free fields.
As disjoint scattering processes give rise to independent (divergent) amplitudes one is led to the study of regularized characters of the Hopf algebra, to say $A$-valued maps $\phi:
{\cal H}_{F} \to A$ such that $\phi \circ m=m_A \circ (\phi
\otimes \phi)$. Here $A$ denotes a unital commutative algebra.
Usually, Hopf algebra characters, i.e., linear multiplicative maps, assign to any element in the Hopf algebra an element in the base field, and form a group under convolution, denoted by $\mathcal{G}$. Motivated by the need for regularizing our theory, due to ultraviolet (UV) divergencies showing up in higher loop calculations, we take here a slightly more general point of view, replacing the base field as target space, say $\mathbb{C}$, by a suitable commutative and unital algebra, $A$, of -regularized- Feynman amplitudes. The group of regularized, or $A$-valued, Hopf algebra characters is denoted by $\mathcal{G}_A$, and the group law is given by the convolution product $$\phi_1 \star \phi_2 := m_A \circ (\phi_1 \otimes \phi_2) \circ \Delta,
\label{def:convol}$$ so that the coproduct, counit and coinverse (the antipode) give the product, unit and inverse of this group, as befits a Hopf algebra.
The study of tree-level amplitudes in lowest order perturbation theory justifies to assign to each edge a propagator and to each elementary scattering process a vertex which define the Feynman rules $\phi({\rm \bf res}(\Gamma))$ and the underlying Lagrangian, on the level of residues of these very graphs. With the Feynman rules providing a canonical character $\phi$, we will have to make one further choice: a renormalization scheme. The need for such a choice is no surprise: after all we are eliminating short-distance singularities in the graphs which renders their remaining finite part ambiguous, albeit in a most interesting manner.
We choose a $\mathbb{K}$-linear map $R:A \to A$, from which we obviously demand that it does not modify the UV-singular structure, i.e., $R^2=R$, and furthermore that it obeys $$R(x)R(y) + R(xy) = R\big(R(x)y) + R(xR(y)\big),
\label{RB}$$ an equation which guarantees the multiplicativity of renormalization and lies at the heart of the Birkhoff decomposition which emerges below: it tells us that elements in $A$ split into two parallel subalgebras given by the image and kernel of $R$. Algebras for which such a map exists fall into the class of Rota–Baxter algebras, well-known in mathematics. See below in Section \[section4\] (and Appendix \[appendix:RotaBaxter\]) for more details .
Let us take a shortcut for the moment and see how all the above structure comes together in renormalization theory. Starting with a regularized Feynman rules character $\phi$, we define a further character $S_ R^\phi$ which deforms $\phi \circ S$, the inverse of the $\phi$, slightly and delivers the counterterm for $\Gamma$ in the renormalization scheme $R$: $$\label{counterterm}
S_R^\phi(\Gamma)=-R\big[m_A(S_R^\phi\otimes\phi\circ P)\Delta(\Gamma)\big]=
-R[\phi(\Gamma)]-R\left[\sum_{\gamma{\subset}\Gamma}
S_R^\phi(\gamma)\phi(\Gamma/\gamma)\right],$$ for $\Gamma$ in $\ker\bar{e}$. Comparing with the undeformed inverse of $\phi$ $$\phi \circ S(\Gamma) = m_A(\phi \circ S \otimes\phi \circ P)\Delta(\Gamma)=
-\phi(\Gamma)-\sum_{\gamma{\subset} \Gamma}\phi\circ S (\gamma)\phi(\Gamma/\gamma)$$ allows to easily understand finiteness of renormalized quantities, thanks to the independence of counterterms on kinematical variables. Later, in Section \[section4\] we will fully derive Equation (\[counterterm\]) and the results below from a more mathematical point of view from the fact that $R$ is a Rota–Baxter operator.
We conclude that $S_R^\phi$ is an element of the group of regularized characters, ${\cal G}_{A}$, of the Hopf algebra, $S_R^\phi\in {\rm Spec}({\cal G}_{A})$. We now have determined the renormalized Lagrangian: $$Z^{\underline{r}}=S_R^\phi(\Gamma^{\underline{r}}).$$ The standard results of renormalization theory follow immediately using the group of regularized characters: the renormalization of a graph $\Gamma$ is obtained by the application of a renormalized character, $S_R^\phi \star \phi$ $$S_R^\phi\star\phi(\Gamma)=m_A(S_R^\phi\otimes\phi)\Delta(\Gamma)$$ for $\Gamma \in \ker \bar{e}$, and Bogoliubov’s $\bar{\mathrm{R}}$-operation is obtained as $$\label{BogosBARmap}
\bar{\mathrm{R}}(\Gamma)=m_A(S_R^\phi\otimes\phi)({\rm id}\otimes P)\Delta(\Gamma)=
\phi(\Gamma)+ \sum_{\gamma{\subset}\Gamma}
S_R^\phi(\gamma)\phi(\Gamma/\gamma).$$ In the following we write $\bar{R}(\Gamma)=:\bar{\phi}(\Gamma)$, so that we have $$S_R^\phi\star\phi(\Gamma)=\bar{\phi}(\Gamma) + S_R^\phi(\Gamma).$$ $S_R^\phi\star\phi$ is an element in the group of regularized characters, $\mathcal{G}_A$, of the Hopf algebra. This Lie group has indeed the previous Lie algebra ${\cal L}$ of graph insertions as its Lie algebra: ${\cal L}$ exponentiates to ${\cal G}_A$.
What we have achieved at this moment is a local renormalization of quantum field theory. Let $m_{\underline{r}}$ be a monomial in the Lagrangian $L$ of degree $\omega_{\underline{r}}$, $$m_{\underline{r}}=D_{\underline{r}}\{\varphi\},$$ for some suitable derivation $D_{\underline{r}}$ on the fields $\varphi$. Then one can prove using the Hochschild cohomology and induction over the augmentation degree in ${\cal H}_F$:
(locality) $$Z^{\underline{r}}D_{\underline{r}}\{\varphi\}=D_{\underline{r}}Z^{\underline{r}}\{\varphi\},$$ renormalization commutes with infinitesimal space-time variations of the fields.
Let us finally give the renormalization of a Feynman graph, say $\Gamma=\!\!\ggv$.
Example: QED {#section3}
============
The QED Lagrangian (density) for an electron coupled to the electromagnetic field in coordinate space reads (we skip the $1/2(\partial \cdot A)^2$ term) $$\label{def:QED}
L_{QED}(\alpha,m) = i\bar{\psi} \partial\!\!\!/ \psi
+ \bar{\psi} eA\!\!\!/ \psi
+ m \bar{\psi} \psi
+ \frac{1}{4} F^2,$$ with the electromagnetic field tensor $F_{\mu\nu}:=\partial_\mu
A_\nu - \partial_\nu A_\mu$ and $F^2:=F_{\mu\nu}F^{\mu\nu}$. The Euler-Lagrange equations for this term give the Maxwell equations. For the first and third term they give the Dirac equation. We use units defined by $c=\hbar=1$, for which the elementary charge squared, $e^2=4\pi \alpha_{QED}$. It is a dimensionless quantity depending on the fine-structure constant $\alpha:=\alpha_{QED}
\simeq \frac{1}{137}$. The first term in (\[def:QED\]) describes the kinetic energy of the electron, and together with the mass term, $m \bar{\psi} \psi$, it constitutes the free Lagrangian density for an electron. The second term in (\[def:QED\]) describes the minimal coupling of the electron with the electromagnetic field. The first two terms are dictated by local gauge invariance of the QED Lagrangian, i.e., invariance with respect to multiplying the electron field $\psi$ by a position dependent phase factor. Let us introduce for every term in $L_{QED}$ a 1PI Green’s function $G^{r}(p^2,\alpha,m,\mu)$, with $$\label{def:QEDmonomials}
r \in \left\{ i\bar{\psi} \partial\!\!\!/\psi ,\
\bar{\psi}e A\!\!\!/ \psi,\
m \bar{\psi} \psi ,\
\frac{1}{4}F^2
\right\},$$ all of which transform as scalars under the Lorentz group.
Usually, 1PI Green’s functions of QED are given in standard notation by the expressions
1. $\Gamma_{\nu}(p_1,p_2,m,\alpha,\mu)$ the vertex function;
2. $S_{\rm F}^{-1}(p,m,\alpha,\mu)$ the inverse fermion propagator;
3. $P^{-1}_{\nu\tau}(p,m,\alpha,\mu)$ the inverse photon propagator,
all depending on the bare parameters mass $m$ and coupling constant $\alpha$, and ’t Hooft’s parameter $\mu$ which sets the scale for the one-parameter groups of automorphisms of the Lie algebra which run the renormalization group [@CK2]. An extra parameter, such as ’t Hooft’s unit mass $\mu$, enters naturally in the context of regularization, and is needed for dimensional reasons [@Collins]. From symmetry considerations we find the following form factor decompositions, where we now assume the vertex at zero momentum transfer for simplicity:
1. $
\Gamma_{\nu}(p,p,m,\alpha,\mu)
= e G^{\bar{\psi} A\!\!\!/ \psi}(p^2,m,\alpha,\mu)\gamma_{\nu}
+ eH^{\bar{\psi} A\!\!\!/ \psi}(p^2,m,\alpha,\mu)\frac{p\!\!/ p_{\nu}}{p^2}
$;
2. $
S^{-1}(p,m,\alpha,\mu)=G^{\bar{\psi} p\!\!/ \psi}(p^2,m,\alpha,\mu)p\!\!/
+ G^{ m\bar{\psi} \psi}(p^2,m,\alpha,\mu)m\mathbb{I}
$;
3. $
P^{-1}_{\nu\tau}(p,m,\alpha,\mu)
= \Pi_{\rm tr}^{\frac{1}{4} F^2}(p^2,m,\alpha,\mu)\left\{g_{\nu\tau}p^2 - p_\nu p_\tau \right\}
+ \Pi_{\rm long}^{\frac{1}{2}(\partial\cdot A)^2}(p^2,m,\alpha,\mu)p_\nu p_\tau
$ ,
reintroducing a longitudinal term for completeness. Let us introduce a graphical notation for the monomials of the QED Lagrangian (\[def:QED\]), which will form the building blocks of our graphical Hopf algebra of Feynman graphs. $$\bar{\psi} \partial\!\!\!/ \psi \
\longleftrightarrow
\begin{array}{ccc}
\FERMprop\\[-0.1cm]
{\tiny{\partial}}
\end{array}
\qquad\quad
m \bar{\psi} \psi \
\longleftrightarrow
\begin{array}{c}
\FERMprop\\[-0.3cm]
{\tiny{m}}
\end{array}
\qquad\quad
\begin{array}{cc}
&\\
\bar{\psi} eA\!\!\!/ \psi \ \longleftrightarrow & \\ [-0.9cm]
& \QEDvertex
\end{array}$$ The first two graphs on the left represent the electron propagator parts, corresponding to the derivation and mass contribution in (\[def:QED\]), respectively. The third graph is the QED vertex, representing the interaction of the electromagnetic field with fermions in (\[def:QED\]). To ease the notation we suppressed spinorial indices.
Next we have the transversal respectively longitudinal parts of the bosonic photon propagator, $$ \frac{1}{4} F^2\
\longleftrightarrow
\begin{array}{c}
\BOSONprop\\[-0.3cm]
{\tiny{ t}}
\end{array}
\qquad \quad
1/2(\partial \cdot A)^2\
\longleftrightarrow
\begin{array}{c}
\BOSONprop\\[-0.3cm]
{\tiny{\rm long}}
\end{array}$$ In the following, we will work with a transversal photon propagator for concreteness. Once we have the graphical notation we define the -coordinate space- QED Feynman rules $\widehat{\phi}$ such that for any $\underline{r} \in
\mathfrak{R}_{QED}$ $$\label{def:QEDgraphs}
\mathfrak{R}_{QED}:=\left\{
\begin{array}{cccc}
\begin{array}{ccc}
\FERMprop\\[-0.1cm]
{\tiny{\partial}}
\end{array},
&
\begin{array}{c}
\FERMprop\\[-0.3cm]
{\tiny{m}}
\end{array},
&
&,\
\begin{array}{c}
\\[-0.2cm]
\BOSONprop\\[-0.3cm]
{\tiny{t}}
\end{array} \\[-1.3cm]
& & \QEDvertex\!\! &
\end{array}\right\}$$ we get back the corresponding coordinate space QED Lagrange monomial $r$ in (\[def:QEDmonomials\]) $$\widehat{\phi}(\underline{r}) = r.$$ We write the QED Lagrangian (\[def:QED\]) pictorially $$\label{def:graphQED}
L_{QED}(\alpha,m) = \sum_{\underline{r}\in \mathfrak{R}_{QED} }
\widehat{\phi}(\underline{r}).$$ With the QED free propagators and vertex in $\mathfrak{R}_{QED}$ at hand we have available the one-particle irreducible Feynman diagrams which provide amplitudes corresponding to these propagations and interaction, as before. We can now introduce partitions of unity for the form-factor decomposition of any Green’s function we are interested in. For example, if we do want to decompose the self-energy of the fermion into its kinetic energy and mass part We can easily incorporate this by promoting our Hopf algebra to pairs $(\Gamma,\sigma)$ [@CK1], where $\sigma$ indicates the desired form-factor obtained by composing the Feynman rules with a suitable projector as above.
The sum over all projectors defines a partition of unity $${\rm id}=\sum_\sigma P_\sigma.$$ This structure can be easily incorporated on the level of Hopf algebras, generalizing the study of external structures by setting for the pairs $(\Gamma,\sigma)$ $$\Delta(\Gamma,\sigma)=\sum (\Gamma^\prime,1)\otimes (\Gamma^{\prime\prime},\sigma).$$ Note that should we wish we can partition the unity on the left hand side above, $$(\Gamma^\prime,1)\to
(\Gamma^\prime,\sigma_R)$$ if we want to use information that only particular form-factors $\sigma_R$ need renormalization. Under the Feynman rules these pair of graphs then evaluate to the amplitudes corresponding to the structure functions defined by the projectors signified by the indicated external leg structures. The resulting Hopf algebras for such pairs $(\Gamma,\sigma)$ are decorated versions of the ones for graphs only, and define graph-like structures very similar to the tree-like structures of Turaev for the Hopf algebra of rooted trees [@Turaev1]. Examples can be found in [@Delbourgo-Kreimer; @BK1].
As an example, we might wish to renormalize the mass part of $\epemfg$, using the knowledge that the photon self-energy is transversal. We hence work out the coproduct $$\Delta\big((\!\!\!\!\begin{array}{l} \\[-0.4cm]
\epemfg \end{array}\!\!\!\!,\sigma_m)\big)= (\!\!\!\!\begin{array}{l} \\[-0.4cm]
\epemfg \end{array}\!\!\!\!,\sigma_m)\otimes \One+
\One\otimes(\!\!\!\!\begin{array}{l} \\[-0.4cm]
\epemfg \end{array}\!\!\!\!,\sigma_m)+(\gg,\sigma_{\rm trans})\otimes (\epem,\sigma_m).$$ Under the Feynman rules, we evaluate using the corresponding projectors and obtain the expected Lorentz scalar structure functions and counterterms following the routine as outlined in Eqs. (\[renorm\]).
Renormalization as a factorization problem {#section4}
==========================================
As we have seen the notion of connected graded commutative Hopf algebra appears naturally in the context of perturbative renormalization of 1PI Feynman graphs. Both, composing Feynman graphs in terms of the pre-Lie insertion product, where we replace vertices by Feynman graphs with compatible external leg structure, as well as their decomposition by eliminating subgraphs, i.e., replacing non-trivial 1PI subgraphs by their residues, look very familiar when inspecting the subtraction procedure encoded in the original BPHZ prescription [@BP; @BS; @CaswellK; @Collins]. The later was invented to extract the finite part of the Feynman integral corresponding to a Feynman graph via a regularized Feynman rules character, while maintaining fundamental physical principles, such as locality, unitarity, and Lorentz invariance.
The commutative Hopf algebra of Feynman graphs, $\mathcal{H}_{F}$, and its graded dual, $\mathcal{H}^*_{F}=\mathrm{Hom}(\mathcal{H}_{F},\mathbb{C})$ are intimately related by the Milnor–Moore theorem. The space $\mathcal{H}^*_{F}$ together with the convolution product and the counit map $\bar{e}: \mathcal{H}_{F} \to \mathbb{C}$ as unit forms a unital, associative and non-commutative $\mathbb{C}$-algebra, which contains the group of characters, $\mathcal{G}:=char(\mathcal{H}_{F},\mathbb{C})$, i.e., linear functionals $\phi \in \mathcal{H}^*_{F}$ from $\mathcal{H}_{F}$ to $\mathbb{C}$ respecting multiplication, $\phi(\Gamma_1
\Gamma_2)=\phi(\Gamma_1)\phi(\Gamma_2)$, $\Gamma_1$, $\Gamma_2 \in
\mathcal{H}_{F}$. This group of multiplicative maps possesses a corresponding Lie algebra, $\mathcal{L}=\partial
char(\mathcal{H}_{F},\mathbb{C}) \subset \mathcal{H}^*_{F}$, of derivations, or infinitesimal characters, i.e., linear maps $Z \in
\mathcal{H}^*_{F}$, satisfying Leibniz’ rule $$Z(\Gamma_1 \Gamma_2)= Z(\Gamma_1)\bar{e}(\Gamma_2)
+\bar{e}(\Gamma_1) Z(\Gamma_2)$$ for all $\Gamma_1$, $\Gamma_2 \in \mathcal{H}_{F}$. The grading of $\mathcal{H}_{F}$ implies a decreasing filtration on $\mathcal{H}^*_{F}$, which allows us to introduce a metric, and therefore a distance map. $\mathcal{H}^*_{F}$ is complete with respect to the induced topology. The exponential map $\mathrm{exp}^{\star}$ gives a bijection between the Lie algebra $\mathcal{L}$ and its corresponding group $\mathcal{G}$.
Using QED as an example we have seen that in general Feynman rules for any perturbative QFT form a subclass of characters. Also, we had to face the severe problem that the associated Feynman integrals for graphs beyond the tree level suffer from ultraviolet divergencies in the limit of large momenta, or equivalently small distances. Therefore, one is forced to invoke a regularization of such integrals, or more generally the Feynman rules themselves. Actually, there is no specific selection rule for such a regularization, indeed one must assure that the final physical result is independent of such an unphysical intermediate step. At the same time it is of vital importance that the regularization prescription used in calculations respects as many physical properties of the underlying theory as possible, such as gauge symmetries. We will ignore such subtleties and take the following stance. In the above Hopf algebraic setting, the regularization of our theory is achieved by replacing the base field $\mathbb{C}$ as target space of maps in $\mathcal{H}^*_{F}$ by an unital algebra $A$, of which we demand commutativity, and the existence of a linear map $R$ satisfying the Rota–Baxter relation $$\label{eq:RBrel}
R(x)R(y)+R(xy)=R\big(R(x)y\big)+\big(xR(y)\big)$$ for all $x,y \in A$. For $R$ being such a Rota–Baxter map, $\tilde{R}:={\mathrm{id}}_A - R$ also satisfies relation (\[eq:RBrel\]). Such algebras are well-known in mathematics under the name Rota–Baxter algebra (see Appendix \[appendix:RotaBaxter\] for more details). As a principal example we mention here dimensional regularization, where the image of Feynman rules lives in the field of Laurent series, $A=\mathbb{C}[\varepsilon^{-1},\varepsilon]]$, with the pole part projection $R:=R_{ms}$ as Rota–Baxter map. In the examples one encounters in QED renormalization calculations, where the regularization could have been for instance a simple cut-off of the momentum integrals, the map $R$ is given in terms of an evaluation at a specific fixed momentum (on-shell scheme), $R_{q}(f)(p):=f(q)$, and trivially satisfies relation (\[eq:RBrel\]), as it is an idempotent algebra homomorphism, i.e., the zeroth order in the Taylor expansion of the map $f$ at point $q$.
The space $\A:=\mathrm{Hom}(\mathcal{H}_{F},(A,R))$ of $A$-valued linear functionals, together with the convolution product (\[def:convol\]) and unit $e:=u_A \circ \bar{e}$, $e(\One)=1_A$, forms an associative non-commutative algebra, containing the group of regularized characters, $\mathcal{G}_A :=
char(\mathcal{H}_{F},A)$, and its corresponding Lie algebra, $\mathcal{L}_A := \partial char(\mathcal{H}_{F},A)$, of regularized derivations. The linearity of the Rota–Baxter map $R$ on the regularization target space $A$, gives rise to a Rota–Baxter algebra structure on $\mathrm{Hom}(\mathcal{H}_{F},(A,R))$, induced in terms of the linear map $\R$, which is defined for any $f \in \A$, by $\R(f):=R
\circ f \in \A$. As before, we can equip $\A$ with a decreasing filtration of Rota–Baxter ideals $$\A=\A_0 \supset \A_1 \supset \dots \supset \A_n \supset \dots$$ making it a complete filtered non-commutative Rota–Baxter algebra with convolution product as composition, $(\A,\R,\{\A_n\}_{n\geq
1})$, since $\R(\A_n) \subset \A_n$ for all $n$. Here we have $\mathcal{L}_A$ as a Lie subalgebra of $\A_1$, and $\mathcal{G}_A$ is a subgroup of $\widehat{\mathcal{G}}:=e +\A_1$, such that $$\begin{aligned}
\exp^{\star}: & \A_1 \to e+\A_1,\quad \exp^{\star}(Z):=
\sum_{n=0}^\infty\frac{Z^{\star n}}{n!}, \label{eq:exp} \\
\log^{\star}: & e+\A_1 \to \A_1,\quad \log^{\star}(e+Z):=
-\sum_{n=1}^\infty \frac{(-Z)^{\star n}}{n} \label{eq:log}\end{aligned}$$ are well-defined with respect to convolution and inverse to each other. Furthermore $\exp^{\star}$ restricts to a bijection between $\mathcal{L}_A$ and $\mathcal{G}_A$.
Atkinson’s [@Atkinson] (see Appendix \[appendix:RotaBaxter\]) factorization theorem for associative Rota–Baxter algebras implies in the above setting, that for a fixed $\phi = e + Z \in \mathcal{G}_A$ the solutions $X
\in e+\mathcal{R}(\A_1)$, $Y \in e+\tilde{\mathcal{R}}(\A_1)$ of the equations $$\label{eq:At1}
X = e - \R(X \star Z) \ \;\mathrm{ resp.}\ \; Y = e - \tilde{\R}(Z \star Y)$$ solve the factorization problem $$\label{eq:At2}
e + Z = \phi = X^{-1} \star Y^{-1},$$ which can be easily checked. If the Rota–Baxter map $R$ is idempotent, the decomposition in (\[eq:At2\]) is unique. In the following we denote $\phi_{-}:=X$ and $\phi_{+}=:Y^{-1}$. Spitzer’s classical identity [@Spitzer] for commutative Rota–Baxter algebras can be generalized to non-commutative Rota–Baxter algebras, thereby implying one of the main results of the Hopf algebraic approach to renormalization in QFT [@CK1], to wit the algebraic Birkhoff decomposition of Connes and Kreimer, which we formulate as a theorem.
[[[@EGK2]]{}]{} Let $\H_{F}$ be a connected graded Hopf algebra of Feynman graphs associated with a perturbatively treated renormalizable QFT. Let $A$ be a commutative unital Rota–Baxter algebra with an idempotent Rota–Baxter operator $R$. Let $\A$ be the complete filtered algebra ${\mathrm{Hom}}(\H_{F},A)$. \[thm:algBi\]
Equation (\[special2\]) follows by general arguments for Rota–Baxter algebras, as we can write for $\phi_{+}$ the following equation $$\phi_{+}=e + \tilde{\R}(\phi_{+} \star (e-\phi^{-1}))
=e + \tilde{\R}(\phi_{-} \star (\phi-e)).$$ The unique map $\chi: \A_1 \to \A_1$ in (\[eq:Spi\]) is the key-result for the generalization of Spitzer’s identity to non-commutative Rota–Baxter algebras, and satisfies the equation $$\chi(Z)= Z - BCH\Big(\R\big(\chi(Z)\big),\tilde{\R}\big(\chi(Z)\big)\Big),
\label{BCH-recur}$$ where $BCH(x,y)$ denotes the Baker–Campbell–Hausdorff relation $$\exp(x)\exp(y)=\exp\big(x+y+BCH(x,y)\big).$$ The non-linear map $\chi$ was introduced in [@EGK1; @EGK2], and was called the $BCH$-recursion. The reader may find it helpful to consult [@EG2] for more details, and [@Manchon] for a more conceptual proof in the context of Lie algebras.
As a proposition to this theorem, we mention without giving further details the fact that Bogoliubov’s $\bar{\mathrm{R}}$-operation (\[BogosBARmap\]) can be written as an exponential using the double Rota–Baxter convolution product, $\star_{\R}$, on $(\A,\R,\{\A_n\}_{n\geq 1})$ (see Eq. (\[def:doubleProd\]) in Appendix \[appendix:RotaBaxter\]): $$\bar{\mathrm{R}}(\Gamma)=\bar{\phi}(\Gamma)= \phi_{-}\star
(\phi-e)(\Gamma)=
-\exp^{\star_{\R}}\big(-\chi(Z) \big)(\Gamma)$$ for $\Gamma \in \ker \bar{e}$. Finally let us mention that the above notion of complete Rota–Baxter algebra and Theorem \[thm:algBi\] becomes very transparent for uni- and nilpotent upper (or lower) triangular matrices with entries in a commutative Rota–Baxter algebra [@EG2; @EGGV].
The above theorem presents a purely algebraic setting for the formulation of renormalization as a factorization problem in the group of regularized Hopf algebra characters, situated in the theory of non-commutative Rota–Baxter algebras with idempotent Rota–Baxter map. The formulae for the counterterm (\[special1\]) and renormalized character (\[special2\]) are completely dictated by a general decomposition structure, which characterizes Rota–Baxter algebras [@Atkinson]. The additional property of $R$ being a projector implies a direct decomposition of the algebra, hence the uniqueness of the factorization in (\[Birkhoff\]). We would like to emphasis the necessary freedom in the choice of the regularization prescription, encoded in the particular structure of the commutative Rota–Baxter algebra $A$ as target space of linear Hopf algebra functionals in ${\mathrm{Hom}}(\H_{F},A)$.
Specializing the target space Rota–Baxter algebra $A$ in the above theorem to the field of Laurent series, i.e., using dimensional regularization, we recover the original setting in [@CK1], opening a hitherto hidden geometric viewpoint on perturbative renormalization in terms of a correspondence to the Riemann–Hilbert problem. This approach was further extended in [@CM1; @CM2; @CM3].
Diffeomorphisms of physical parameters {#section5}
======================================
In the above, we obtained a unique Birkhoff decomposition of Feynman rules $\phi \in Spec({\cal G}_A)$ into two characters $\phi_- =: S_R^\phi \in Spec({\cal G}_A)$ and $\phi_+ = S_R^\phi
\star \phi \in Spec({\cal G}_A)$, for any idempotent Rota–Baxter map $R$. Thanks to Atkinson’s theorem this is possible for any renormalization scheme $R$. For the minimal subtraction scheme it amounts to the decomposition of the Laurent series $\phi(\Gamma)(\varepsilon)$, which has poles of finite order in the regulator $\varepsilon$, into a part holomorphic at the origin and a part holomorphic at complex infinity. This has a geometric interpretation upon considering the Birkhoff decomposition of a loop around the origin, providing the clutching data for the two half-spheres defined by that very loop, which is central in the work of Connes and Kreimer [@CK2; @CK3]. The geometric interpretation leads to motivic Galois theory upon studying the equisingularity of the corresponding connection in the Riemann–Hilbert correspondence [@CM1; @CM2], itself a result of the Hochschild cohomology of these Hopf algebras [@BergbauerKreimer1; @BergbauerKreimer2].
Our understanding of each term in the perturbative expansion and its renormalization have found hence satisfying mathematical interpretations. The character group ${\cal G}_A$ is a poorly understood object though, it is far too big. Fortunately renormalization can be captured by the study of diffeomorphisms of physical parameters, as by the very definition the range of allowed modifications in renormalization theory is the variation of the coefficients of monomials $\hat{\phi}(\underline{r})$ of the underlying Lagrangian $$L=\sum_{\underline{r}\in \mathfrak{R}}Z^{\underline{r}}\, \hat{\phi}(\underline{r}).$$ We can now eliminate the use of ${\cal G}_A$ as one can regain the Birkhoff decomposition at the level of diffeomorphisms of the coupling constants.
One proceeds by using that renormalized couplings provide a formal diffeomorphism $$g_{\rm new}=g_{\rm old}\;Z^g,$$ where $$Z^g= \frac{Z^v}{\prod_{e \in {\bf res}(v)^{[1]}_{\rm ext}}\sqrt{Z^e}},$$ for some vertex $v$, which obtains the new coupling in terms of a diffeomorphism of the old. This formula provides indeed a Hopf algebra homomorphism from the Hopf algebra of diffeomorphisms to the Hopf algebra of Feynman graphs, regarding $Z^g$, a series over counterterms for all 1PI graphs with the external leg structure corresponding to the coupling $g$, in two different ways: it is at the same time a formal diffeomorphism in the coupling constant $g_{\rm old}$ and a formal series in Feynman graphs. As a consequence, there are two competing coproducts acting on $Z^g$. Their consistency defines the required homomorphism, which transposes to a homomorphism from the largely unknown group of regularized characters of ${\cal H}_F$ to the one-dimensional diffeomorphisms of this coupling. Hence one concludes [@CK3]:
Let the unrenormalized effective coupling constant $g_{\rm eff}
(\varepsilon)$ viewed as a formal power series in $g$ be considered as a loop of formal diffeomorphisms and let $g_{\rm
eff} (\varepsilon)= (g_{{\rm eff}_-})^{-1}(\varepsilon) \, g_{{\rm
eff}_+}(\varepsilon)$ be its Birkhoff decomposition in the group of formal diffeomorphisms. Then the loop $g_{{\rm eff}_-}
(\varepsilon)$ is the bare coupling constant and $g_{{\rm eff}_+}
(0)$ is the renormalized effective coupling.
The role of Hochschild cohomology {#section6}
=================================
The Hochschild cohomology of the Hopf algebras of 1PI graphs illuminates the structure of 1PI Green’s functions in various ways:
- it gives a coherent proof of locality of counterterms –the very fact that $$[Z^{\underline{r}},D_{\underline{r}}]=0,$$ the coefficients in the Lagrangian remain independent of momenta, and hence the Lagrangian a polynomial expression in fields and their derivatives; [@BergbauerKreimer2; @Kreimer10; @KreimerDS]
- the quantum equation of motions take a very succinct form identifying the Dyson kernels with the primitives of the Hopf algebra [@BergbauerKreimer2; @Kreimer10; @KreimerDS], and hence replacing a sum over all graphs by a sum over all primitive graphs;
- sub-Hopf algebras emerge from the study of the Hochschild cohomology which connect the representation theory of these Hopf algebras to the structure of theories with internal symmetries leading to the Slavnov–Taylor identities for the couplings [@KreimerDS];
- these Hopf algebras are intimately connected to the structure of transcendental functions like the generalized polylogarithms which play a prominent role these days ranging from applied particle physics to recent developments in mathematics, in particular the primitive graphs which provide the Dyson kernels allow for a motivic interpretation [@BlochEK].
For more information, we refer the reader to the literature indicated.
Basic facts about general Rota–Baxter algebras {#appendix:RotaBaxter}
==============================================
For the reader’s convenience we collect some basic notions of more mathematical nature concerning mainly Rota–Baxter operators, in the hope that from the above presentation they become redundant. For more details we refer the reader to the standard literature, e.g. [@Atkinson; @Guo1; @RotaSmith]. Rota–Baxter operators (also known as Baxter operators in older mathematical references) were an active field of mathematical research in the late 1960s and early 1970s. After an almost three decades long period of dormancy they reappeared, as if on cue, in the mathematical literature in the context of dendriform algebras [@Aguiar; @K1; @JLL], number theory [@Guo2], generalizations of shuffle products [@EG1; @GuoKeigher], and Hopf algebras [@AndrewsGuo], as well as in theoretical physics in the seminal work of Kreimer and collaborators on the Hopf algebra of renormalization [@CK1; @EGK1; @EGK2; @Kreimer2].
In the following $\mathbb{K}$ denotes the base field of characteristics zero, over which all algebraic structures are defined. In general an algebra always means an associative unital $\KK$-algebra, not necessarily commutative. The algebra unit is simply denoted by $1$.
In Section \[section4\] we encountered Rota–Baxter operators respectively the Rota–Baxter relation in the context of renormalization schemes, i.e. subtraction operators for the BPHZ method. Let $A$ be an algebra together with a linear endomorphism $R: A \to A$. We call the tuple $(A,R)$ a Rota–Baxter algebra of weight $\theta \in \KK$, if the map $R$ fulfills the Rota–Baxter relation (of weight $\theta$) $$\label{eq:RBrelation}
R(x)R(y)+\theta R(xy)=R\big(R(x)y + xR(y)\big)$$ for all $x,y \in A$. Without proof we state the fact that the operator $\tilde{R}:=\theta {\mathrm{id}}_A - R$ is a Rota–Baxter operator of weight $\theta$, too, such that the mixed relation $$\label{eq:mixedRBrel}
R(x) \tilde{R}(y)=\tilde{R}\big(R(x)y\big) + R\big(x\tilde{R}(y)\big)$$ is satisfied for all $x,y \in A$. The map $B:=\theta{\mathrm{id}}_A-2R$ satisfies the modified Rota–Baxter relation $$B(x)B(y)+\theta^2 xy=B\big(B(x)y + xB(y)\big).$$ For $\theta \neq 0$, the normalized map $\theta^{-1} R$ is a Rota–Baxter operator of weight one. Therefore without lost of generality we may suppose in the following the canonical weight one case. A Rota–Baxter (left-) right-ideal $I$ is a (left-) right-ideal $I$ of $A$ such that $R(I) \subseteq I$. A Rota–Baxter ideal is a Rota–Baxter left- and right-ideal.
The American mathematician Glen Baxter introduced this relation 1960 in his probability studies in fluctuation theory [@Baxter]. Later, the Italian born American mathematician Gian-Carlo Rota [@RotaSmith; @Rota1; @Rota2; @Rota3], and others [@Cartier; @Kingman], notably F.V. Atkinson [@Atkinson], explored in detail Baxter’s work from different perspectives in analysis, algebra and combinatorics. The case $\theta = 0$ corresponds to the integration by parts property of the usual Riemann integral $I: \F\to\F$, $I[f](x) := \int_0^x
f(t)\,dt$ in the algebra $\F$ of continues functions on $\RR$, to wit, $$I[f_1]\,I[f_2] = I\bigl[ I[f_1] f_2 + f_1 I[f_2] \bigr],
\label{eq:ach-so}$$ for $f_1,f_2 \in \F$. We already encountered the pole part projection $R_{ms}$ in dimensional regularization as an example of an idempotent Rota–Baxter map of weight one. The images of $R$ as well as $\tilde{R}$ form subalgebras in $A$. Let $R$ be a projector on $A$. For $R$ to satisfy the Rota–Baxter relation is equivalent to a direct decomposition of $A=R(A)\oplus
\tilde{R}(A)$. This is just the special case of Atkinson’s additive decomposition theorem [@Atkinson], characterizing a general Rota–Baxter algebra $(A,R)$ as a subdirect difference of the images of $R$ and $\tilde{R}$.
The Lie algebra associated to $(A,R)$, with standard commutator bracket forms a Rota–Baxter Lie algebra, $(\L_A,R)$, with $R$ fulfilling $$\label{eq:RBLieRelation}
[R(x),R(y)]+\theta R([x,y])=R\big([R(x),y] + [x,R(y)]\big),$$ better known as (operator) classical Yang–Baxter equation. Let us mention here that it was rediscovered in this form in the early 1980ies by some Russian physicists in the context of classical integrable systems (see e.g. [@BBT; @STS1] for references and more details). This curious coincidence of Baxter and Baxter just happens to reveal the connections of Rota–Baxter operators with many areas of mathematics and physics.
The vector space underlying $A$, equipped with the new product $$\label{def:doubleProd}
x \star^{(1)}_R y:=xR(y) + R(x)y - xy$$ is again a Rota–Baxter algebra with Rota–Baxter map $R$, which we denote $(A_1,R)$. Hence, all Rota–Baxter algebras $(A,R)$, associative or non-associative, come with a whole hierarchy of so-called double Rota–Baxter algebras, $(A_n,R)$, $n\in\NN$. Relation (\[eq:RBrelation\]) naturally implies that $R(x
\star^{(1)}_R y)=R(x)R(y)$, i.e. $R$ is an algebra homomorphism from $A_1$ to $A$, or more generally from $A_n$ to $A_{n-1}$.
Baxter’s original motivation for his work was to prove that for a commutative algebra $A$ together with a linear map $R$ satisfying relation (\[eq:RBrelation\]), that later bore his name, the following identity for fixed $a \in A$ holds in the formal power series ring $\A:=A[[t]]$, with $t$ being a commuting parameter. This famous relation is called Spitzer’s classical identity, and appeared in Frank Spitzer’s 1956 paper [@Spitzer]. In [@EGK1] this was generalized to non-commutative Rota–Baxter algebras, based on a BCH-type recursion formula, the key result we used earlier, see Eq. (\[BCH-recur\]). A more general setting in which the above factorization may be regarded, and of which the algebra $A[[t]]$ is just a special case, is that of complete filtered Rota–Baxter algebras [@EGK2; @EG2; @EGGV].
The right hand side of (\[eq:SpitzerId\]) is the unique solution of the recursive equation $$\label{eq:Atkinson1}
X= 1 + tR(Xa).$$ This is a natural generalization of the recursion $f=1+I[gf]$ corresponding to the differential equation $f'=gf$, $f(0)=1$, solved by $\exp(I[g])$, in $(\F,I)$ where $I$ is the Riemann integral. It was again Atkinson [@Atkinson] in 1963, who observed that for any Rota–Baxter algebra, not necessarily commutative, a solution to Eq. (\[eq:Atkinson1\]), and its companion equation for $\tilde{R}$ $$\label{eq:Atkinson2}
Y= 1 + t\tilde{R}(aY)$$ in $A[[t]]$, solve the multiplicative decomposition problem $$\label{eq:AtkinsonFact}
(1-\theta a) = X^{-1}Y^{-1},$$ for any element $a \in A$. $R$ being a projector implies a unique decomposition of the element $(1-\theta a) \in A$. This genuine factorization property of Rota–Baxter algebras is further analyzed in [@EG2].
: The first author is very thankful for a PhD grant provided by the Ev. Studienwerk Villigst, as well as for an extra travel grant which allowed him to visit BIRS (Canada). Also, he would like to thank warmly the Theory Department at the Physics Institute of Bonn University for unlimited support and encouragement. Both authors would like to thank the Fields Institute, where this work was finished, for warm hospitality. The collaboration with Prof. L. Guo is acknowledged.
[0]{}
[^1]: [email protected]; currently visiting the Fields Institute, Toronto, Canada.
[^2]: [email protected] and [email protected], Center for Math.Phys., Boston University.
[^3]: See also L. Foissy’s PhD-thesis.
[^4]: We assume that the reader has already seen a Feynman graph, otherwise he might take a brief look on Eq. (\[pre-LieExam\]).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
This letter proposes a simple but accurate approximation to analytically model both the inter-Wireless Local Area Networks (WLANs) interactions and the negative effect of collisions in the performance of high density WLAN scenarios. Inter-WLANs interactions are characterized using a continuous time Markov chain (CTMC) model where states represent the set of active WLANs at a given time. Then, the effect of collisions is considered by analyzing the local dynamics between contending WLANs at every state of the CTMC. Simulation results confirm the accuracy of the presented approach.
[***Keywords*:**]{} IEEE 802.11, WLANs, CSMA/CA, dense networks
author:
- |
Boris Bellalta\
Dept. of Information and Communication Technologies\
Universitat Pompeu Fabra
bibliography:
- 'Bib.bib'
title: Throughput Analysis in High Density WLANs
---
Introduction
============
Analyzing scenarios with multiple IEEE 802.11 Wireless Local Area Networks (WLANs) operating in the same channel when not all of them are within each other’s carrier sense range is challenging. That is because in this case, the operation of the WLANs, which is governed by the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) protocol, is not fully coupled since not all of them observe the same events.
To characterize the interactions between WLANs, continuous time Markov chain (CTMC) models have recently[^1] been proposed [@baid2015understanding; @bellalta2016interactions; @michaloliakos2016performance][^2]. A CTMC model describes all possible combinations of active WLANs (states) as well as the transition rates between states. Thus, CTMC models allow, building upon the assumption of exponentially distributed backoff and packet transmission durations, to compute the long-term fraction of time the system remains at each state.
CTMC models are able to capture the interactions between neighboring WLANs but their accuracy decreases as collisions become increasingly significant, which has raised many concerns about their use in performance evaluation of next-generation, high density WLANs. The reason for this is the need in CTMC models to assume continuous backoff timers, which, as a consequence, results in negligible collision probability (i.e., the probability that the backoff timer of two or more nodes expire simultaneously). Hence, the use of CTMC models to characterize the performance of WLANs using a slotted backoff counter such as that defined in the IEEE 802.11 standard is only accurate in the case of small number of contenders or large backoff windows.
This letter provides a simple but accurate approximation to analytically model both the inter-WLANs interactions and the negative effect of collisions in scenarios with multiple adjacent WLANs. First, inter-WLANs interactions are characterized using a WLAN-centric CTMC model that allows us to derive the fraction of time each WLAN is active. Then, to compute the corresponding fraction of the WLAN’s active time that corresponds to successful transmissions, we analyze the local interactions at the node level between contending WLANs at every state of the CTMC using the well-known Bianchi IEEE 802.11 model[^3] [@bianchi2000performance]. Using Bianchi’s model we estimate the collision probability, which is then normalized to cope with the different temporal network dynamics of using a slotted backoff timer instead of a continuous one. Simulation results confirm the accuracy of the presented approach.
Modelling Inter-WLANs interactions {#Sec:interWLAN}
==================================
The analysis presented in this section is based in [@bellalta2016interactions], where the considerations required to derive a WLAN-centric CTMC model are discussed in detail. The work in [@bellalta2016interactions] also evaluates the computational advantage of the WLAN-centric CTMC approach compared to the node-centric CTMC one.
Multiple WLANs scenario
-----------------------
A scenario consisting of multiple WLANs deployed in a certain area is considered. WLAN $i$ consists of $N_i$ nodes, the access point (AP) and the $N_{i}-1$ associated stations (STAs). Following the CSMA/CA operation, whenever a transmission from another WLAN is detected, the backoff of WLAN $i$ is paused until the channel is detected free again, at which point the countdown is resumed. The expected active backoff duration for WLAN $i$, i.e., excluding the fraction of time the backoff is paused, is $1/\lambda_i$. At every transmission, WLAN $i$ sends $L$ bits and occupies the channel for an expected duration equal to $E[T]=1/\mu$. A full-buffer traffic model is considered for all nodes.
For simplicity, we assume that when coverage ranges from two WLANs overlap, the coverage ranges of all nodes in the WLANs involved also overlap. Such an assumption covers scenarios where there are multiple side-by-side WLANs, such as in an apartment building, where all STAs belonging to a WLAN are close to the AP.
CTMC model
----------
Let us define a feasible system state as the subset of WLANs that can transmit simultaneously without interference, i.e., WLANs with non-overlapping coverage ranges. Let $\Omega$ be the collection of all feasible system states. Then, the transition rates between two system states $s', s \in \Omega$ are $$q(s',s)=
\begin{cases}
\lambda_{i} & \text{ if } s=s' \cup \{i\} \in \Omega,\\
\mu & \text{ if } s=s' \setminus \{i\},\\
0 & \text{ otherwise}. \nonumber
\end{cases}$$
Let $S_t \in \Omega$ be the system state at time $t$. Assuming exponentially distributed backoff and transmission durations, $(S_t)_{t\geq 0}$ is a continuous-time Markov process on the state space $\Omega$. This Markov process is aperiodic, irreducible and thus positive recurrent, since the state space $\Omega$ is finite. Hence, it has a stationary distribution, which we denote by $\{\pi_s \}_{s \in \Omega}$. It follows from classical Markov chains results that $\pi_{s}$ is equal to the long-run fraction of time the system is in state $s \in \Omega$.
The process $(S_t)_{t\geq 0}$ has been proven to be a time-reversible Markov process in [@kelly1979reversibility]. In particular, detailed balance applies and the stationary distribution $\{\pi_s \}_{s \in \Omega}$ of the process $(S_t)_{t \geq 0}$ can be expressed in product form. The detailed balance relationship for two adjacent system states, $s$ and $s \cup \{i \}$, can be written as follows: $$\begin{aligned}
\frac{\pi_{s\cup \{i\}}}{\pi_{s}} = \frac{\lambda_{i}}{\mu}=\theta_{i}. \nonumber\end{aligned}$$
This relationship implies that for any $s \in \Omega$ $$\begin{aligned}
\pi_{s} = \pi_{\emptyset} \cdot \prod_{i \in s} \theta_{i}, \nonumber\end{aligned}$$ where $\emptyset$ denotes the system state where none of the WLANs is transmitting. The last equality, together with the normalizing condition $\sum_{s \in \Omega} \pi_s = 1$, yields $$\begin{aligned}
\pi_{\emptyset} = \frac{1}{\sum_{z \in \Omega} \prod_{j \in z} \theta_{j}}, \nonumber\end{aligned}$$ and $$\begin{aligned}
\pi_{s} = \frac{\prod_{i \in s} \theta_{i}}{\sum_{z \in \Omega} \prod_{j \in z} \theta_{j}}, \quad s \in \Omega. \nonumber\end{aligned}$$
Throughput computation {#Sec:Throughput}
======================
The throughput of WLAN $i$ is given by $$\begin{aligned}
\label{Eq:Throughput}
x_{i} &= \left(\sum_{s \in \Omega\,:\, i \in s}{\pi_s (1-\gamma_{i|s'\rightarrow s})}\right)\mu L,\end{aligned}$$ where $1-\gamma_{i|s'\rightarrow s}$ is the probability that the data transmitted by WLAN $i$ in state $s$ (moving from a predecessor state $s'$ in which WLAN $i$ is not active) is successfully received, and $L$ is the packet size. Since we do not consider transmission errors, the value of $\gamma_{i|s'\rightarrow s}$ is caused only by packet collisions. Following the IEEE 802.11 slotted backoff, a collision occurs when the backoff counter of two or more contending nodes reaches zero at the same time. We consider that all packets involved in a collision are lost.
Conditional throughput from state $s'$
--------------------------------------
Let us denote by $\mathcal{K}_{i|s'}$ the set of contending WLANs (i.e., with overlapping coverage ranges) with WLAN $i$ to access the channel at state $s'$, and $k_{i|s'}=\sum_{j \in \mathcal{K}_{i|s'}}{N_{j}}$ the total number of nodes in $\mathcal{K}_{i|s'}$. The achievable throughput by WLAN $i$ when it is contending with the WLANs in $\mathcal{K}_{i|s'}$ is $$\begin{aligned}
\label{Eq:y}
y_{i|s'}=\frac{d_{i|s'}L}{a_{i|s'}T_e+b_{i|s'}E[T]+c_{i|s'}E[T_c]},\end{aligned}$$ where $a_{i|s'}=(1-\tau_{i|s'})^{k_{i|s'}+N_{i}}$ is the probability that a backoff slot remains empty, $b_{i|s'}=(k_{i|s'}+N_{i})\tau_{i|s'}(1-\tau_{i|s'})^{k_{i|s'}+N_{i}-1}$ is the probability that a slot results in a successful transmission, $c_{i|s'}=1-a_{i|s'}-b_{i|s'}$ is the probability that a slot results in a collision, and $d_{i|s'}=N_{i}\tau_{i|s'}(1-\tau_{i|s'})^{k_{i|s'}+N_{i}-1}$ is the probability that a slot results in a successful transmission from WLAN $i$. $T_e$ and $E[T_c]$ are the duration of an empty slot and a collision, respectively.
To compute $\tau_{i|s'}$, the probability to transmit in a given slot, we solve the following system of equations using a fixed-point method: $$\begin{aligned}
\left \{\begin{array}{l}
E[B_{i|s'}]=\frac{1 - p_{i|s'} - p_{i|s'}(2p_{i|s'})^{m}}{(1-2p_{i|s'})}\frac{CW_{\min}}{2}-\frac{1}{2} \nonumber \\
p_{i|s'}=1-(1-\tau_{i|s'})^{k_{i|s'}+N_{i}-1} \nonumber\\
\tau_{i|s'} = \frac{1}{E[B_{i|s'}]+1} \nonumber
\end{array}\right.\end{aligned}$$ where $E[B_{i|s'}]$ is the expected backoff duration in slots, $p_{i|s'}$ is the collision probability, and $m=\log_2\left(\frac{\text{CW}_{\max}}{\text{CW}_{\min}}\right)$ is the relationship between the maximum and minimum backoff contention window as defined in the IEEE 802.11 standard.
Then, $\gamma_{i|s'\rightarrow s}$ is computed so as to result in the same throughput as (\[Eq:y\]) when (\[Eq:Throughput\]) is used to compute the throughput of $\mathcal{K}_{i|s'}+1$ WLANs contending for the channel, i.e., $$\begin{aligned}
\mu L\left(\frac{\theta_{i}}{1+\theta_i+\sum_{j \in \mathcal{K}_{i|s'}}{\theta_{j}}}\right) (1-\gamma_{i|s'\rightarrow s}) = y_{i|s'}, \nonumber\end{aligned}$$ from where the value of $\gamma_{i|s'\rightarrow s}$ is obtained: $$\begin{aligned}
\label{Eq:gamma}
\gamma_{i|s'\rightarrow s}=1-\frac{y_{i,s'}}{\mu L\left(\frac{\theta_{i}}{1+\theta_i+\sum_{{j} \in \mathcal{K}_{i|s'}}{\theta_{j}}}\right)}, \end{aligned}$$ with $\theta_i=\frac{\lambda_i}{\mu}$ and $\lambda_i=N_i\frac{2}{CW_{\min}-1}\frac{1}{T_e}$.
Relationship between $\gamma_{i|s'\rightarrow s}$ and $p_{i|s'}$
----------------------------------------------------------------
Intuitively, we could have considered that $\gamma_{i|s'\rightarrow s} = p_{i|s'}$ since $p_{i|s'}$ is the fraction of transmissions colliding in WLAN $i$. However, such an approximation for $\gamma_{i|s'\rightarrow s}$ results in very conservative predictions. The reason for this is that the rate at which packets are transmitted to the channel in the continuous-time backoff case is lower than when the IEEE 802.11 backoff is used, because collisions represent multiple concurrent transmissions. Therefore, removing the same fraction of transmissions in the former case as in the later is not accurate. Then, we need to normalize the value of $\gamma_{i|s'\rightarrow s}$ to the actual rate of packet transmissions in the continuous time backoff case, hence obtaining that $\gamma_{i|s'\rightarrow s} \leq p_{i|s'}$ (see Fig. \[Fig:Fig.0\]).
A note on “dominant states”-only consideration
----------------------------------------------
Considering typical IEEE 802.11 $\text{CW}_{\min}$ and $E[T]$ values, WLANs remain a high fraction of time (e.g., $\geq$ 95 %) in a subset of states of the CTMC to which we refer as *dominant states*. Therefore, the computational cost to solve (\[Eq:Throughput\]) can be reduced by considering only these dominant states and its predecessors without a significant effect on the accuracy of the obtained results. *Dominant states* correspond to the *maximal independent sets* defined in [@liew2010back] when the CSMA/CA operational parameters (e.g., channel width, Clear Channel Assessment (CCA), transmit power, etc.) are static. Otherwise, when these parameters are set depending on the system state, as it is shown in [@faridi2015analysis] for the case of Dynamic Channel Bonding, we obtain dominant states that are not maximal, thus it requires exploration of the full state space $\Omega$ to determine them.
Examples {#Sec:Examples}
========
To show the accuracy of the proposed method when computing the throughput of each WLAN in the system, we compare the throughput obtained by the proposed approach in Sec. \[Sec:Throughput\], with simulation results in the three scenarios depicted in Fig. \[Fig:Networks\]. For each scenario we indicate the position and coverage area of each WLAN, and include the CTMC model that represents the feasible system states.
The simulator was developed in C using the Component Oriented Simulation Toolkit libraries [@chen2002reusing]. It accurately reproduces the described scenarios and the operation of each WLAN. The parameters used in this section correspond to the IEEE 802.11ac amendment, with 40 MHz channels, 64-QAM modulation and 3/4 coding rate. Packet aggregation is considered, and each transmitted Aggregated MAC protocol data unit (A-MPDU) contains 64 packets of length $12000$ bits, resulting in a packet size of $L=768$ Kbits. Under these conditions, the expected duration of a successful transmission is given by $E[T]=6.63$ ms (for further details on how $E[T]$ is calculated, please refer to [@bellalta2016interactions]). Since the RTS/CTS mechanism is not used, we consider that $E[T_c]=E[T]$. Finally, unless otherwise specified, $\text{CW}_{\min}=32$ and $m=5$.
![The three scenarios (Scenario I, II and III) used in Sec. \[Sec:Examples\]. We consider all WLANs have the same number of nodes ($N$), i.e., 1 AP and $N-1$ STAs, and share the same values of $\lambda$ and $\mu$. In each CTMC, the states in orange represent the maximal states. Forward ($\lambda$) and backward ($\mu$) transition rates are represented over single double-headed arrows for Scenario I and II. For Scenario III they are omitted for space reasons.[]{data-label="Fig:Networks"}](Networks){width="0.7\columnwidth"}
The first scenario considers three WLANs with fully overlapping coverage ranges. From (\[Eq:Throughput\]), the throughput of WLANs A, B and C is given by $$\begin{aligned}
x_A=x_B=x_C = \mu L\frac{\theta(1-\gamma_{C|\emptyset\rightarrow C})}{1+3\theta}, \nonumber\end{aligned}$$ where $\theta_A=\theta_B=\theta_A=\theta$.
Fig. \[Fig:Fig.1\] shows a very good agreement between the throughput values provided by the analysis and simulation results. Notice the difference between the throughput estimated by the CTMC model (i.e., using (\[Eq:Throughput\]) with $\gamma_{i|s'\rightarrow s}=0$), which is independent of $m$ as it assumes negligible collision probability, with the throughput values obtained from the simulator and the proposed approximation (i.e., using (\[Eq:Throughput\]) with $\gamma_{i|s'\rightarrow s}$ from (\[Eq:gamma\])).
The second scenario consists of three WLANs placed in a line, with WLAN B in the middle of WLANs A and C. In this scenario, the throughput achieved by WLANs A, B and C is given by $$\begin{aligned}
x_A=x_C&=\mu L\frac{\theta(1-\gamma_{\text{C}|\emptyset\rightarrow \text{C}})+\theta^2(1-\gamma_{\text{C}|\text{A}\rightarrow \text{AC}})}{1+3\theta+\theta^2}, \text{and} \nonumber \\
x_B &= \mu L\frac{\theta(1-\gamma_{\text{B}|\emptyset\rightarrow \text{B}})}{1+3\theta+\theta^2}. \nonumber\end{aligned}$$
Fig. \[Fig:Fig.2\] shows the throughput obtained by WLANs A, B and C when the value of $\text{CW}_{\min}$ increases from $4$ to $8192$. It can be observed that the proposed approximation is very accurate for the two values of $N$. For $N=1$, the CTMC model shows good accuracy as well since collisions can be neglected. However, for $N=16$ and low $\text{CW}_{\min}$ values, the negative effect of collisions in the throughput of WLANs A and C, which the CTMC model fails to capture, is evident.
Scenario III extends Scenario II by adding WLANs D and E as shown in Fig. \[Fig:Networks\]. Note that WLANs D and E coverage ranges do not overlap, and therefore severely affect the transmission opportunities of WLAN C. In this case, WLAN C chances to access the channel are severely reduced due to the activity of WLANs D and E. Consequently, WLAN B is positively affected as it only (or almost only) contends with WLAN A. In the absence of WLANs D and E, on the contrary, WLANs C and A reduce WLAN B transmission opportunities as shown in Scenario II.
The throughput of each WLAN in Scenario III is given by (writing only the most significant states for each WLAN) $$\begin{aligned}
x_A & =\mu L \left(\frac{\ldots+\theta^3(1-\gamma_{\text{A}|\text{DE}\rightarrow \text{ADE}})}{1+5\theta+6\theta^2+2\theta^3}\right), \nonumber \\
x_B & =\mu L \left(\frac{\ldots+\theta^3(1-\gamma_{\text{B}|\text{DE}\rightarrow \text{BDE}})}{1+5\theta+6\theta^2+2\theta^3}\right), \nonumber \\
x_C & =\mu L\left(\frac{\ldots+\theta^2(1-\gamma_{\text{C}|\text{A}\rightarrow \text{AC})}}{1+5\theta+6\theta^2+2\theta^3}\right), \text{~and}\nonumber \\
x_D & =x_E= \nonumber \\ &= \mu L\left(\frac{\ldots+\theta^3(1-\gamma_{\text{E}|\text{BD}\rightarrow \text{BDE})}+\theta^3(1-\gamma_{\text{E}|\text{AD}\rightarrow \text{ADE}})}{1+5\theta+6\theta^2+2\theta^3}\right). \nonumber \end{aligned}$$
Fig. \[Fig:Fig.4\] confirms the accuracy of the presented approach in Scenario III. The accuracy of the approximation is reduced for $\text{CW}_{\min}=512$ compared to $\text{CW}_{\min}=16$. Indeed, large $\text{CW}_{\min}$ values affect the backoff dynamics of each contender when the system reaches state $s'$, not abiding to the stationary behavior assumed in (\[Eq:y\]).
Conclusions {#Sec:Conclusions}
===========
This letter presents a simple approach to approximate the achievable throughput of a group of WLANs operating in a given area, capturing both their mutual interactions and the negative effect of collisions in the system performance. It is based on decoupling the ’macro’ interactions caused by backoff pauses when the channel is detected busy, from the ’micro’ interactions caused by packet collisions resulting from two or more nodes’ backoff timers expiring at the same time. The accuracy of the results confirms that such ’macro’ and ’micro’ interactions have an important effect in high density WLAN deployments, and they can be considered as independent in most cases.
While in this letter we have evaluated the presented approach in several basic scenarios, there are many other aspects to consider in the future to determine under which conditions the presented approach is accurate, providing also further insight into the operation of high density WLANs. For example, future research extensions include: *i)* considering the presence of partially overlapping coverage ranges, encompassing the effect of hidden and exposed terminals, *ii)* taking account of the capture effect, and *iii)* incorporating miscellaneous traffic conditions. The proposed approach must also be validated for WLANs employing advanced features, such as static and dynamic channel bonding, CCA adaptation and multiuser communications [@bellalta2016ieee].
[^1]: The use of CTMC models to analyze CSMA/CA networks is not new. They have been extensively used in the last years to capture the interactions between nodes in multi-hop networks [@boorstyn1987throughput; @laufer2013capacity; @liew2010back].
[^2]: [@baid2015understanding; @michaloliakos2016performance] focus on the *maximal independent sets* approach presented in [@liew2010back] to compute the throughput of each WLAN, which only considers the maximal states of the CTMC model. Maximal states are those in which the maximum possible number of nodes are transmitting simultaneously.
[^3]: Note that other equivalent IEEE 802.11 analytical models could be used instead of Bianchi’s model. We have selected Bianchi’s model due to its accuracy and well-known conditions under which it can be applied.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Contributions of quantum interference effects occuring in quasicrystals are reported. First conversely to metallic systems, quasiperiodic ones are shown to enclose original alterations of their conductive properties while downgrading long range order. Besides, origin of localization mechanism are outlined within the context of the metal-insulator transition (MIT) found in these materials.'
address: '${\ }^{*}$ Department of Applied Physics, University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8654, Japan.'
author:
- Stephan Roche
title: Anomalous electronic conductance in quasicrystals
---
\#1[[$\backslash$\#1]{}]{}
\#1\#2\#3
Introduction
============
Despite sustained effort and concern, today’s understanding of exotic electronic properties of quasicrystals [@AlMn] remains unsatisfactory although quasicrytalline materials have already been implemented to miscellanous concrete applications[@QC-app; @QC-app2]. In particular, the role of quasiperiodic order on electronic localization and transport is believed to genuinely entail the most unexpected experimental features whereas so far, no coherent theoretical framework has been really successfully ascertained[@Berger; @RocheJMP; @Takeo; @RocheFS]. As a matter of fact, one of the unprecedented tendency of quasicrystals is the enhancement of their conductive ability upon increasing contribution of static (structural disorder) or dynamical excitations (phonons). This has been strongly supported by many experimental evidences[@Berger] and is often refered as an original property in the litterature. Notwithstanding theoretically, given heuristical arguments[@RocheJMP; @Sire] and numerical investigations (e.g. the Landauer conductance for quasiperiodic Penrose lattices [@Fujiwara-Land] or Kubo formula for 3D-quasiperiodic models [@RochePRL1] ) yield to uncomplete understanding of the observed properties which range from anomalously metallic behaviors to insulating ones [@Julien].
It is generally assumed that a specific “geometrical localization process” takes place in quasicrystals (sustained by critical states[@Kho; @Macia]) and that local disruptions of corresponding mesoscopic order reduce quantum interferences, resulting in an increase of conductivity. This issue has been takled for 1D quasiperiodic potential, tight-binding (TBM) as well as continuous Kronig-Penney models, and phason-type disorder has been shown to disclose manifestations of quantum interferences (QIE) in quasiperiodic order [@RocheFas].
Pionneer works of M. Kohmoto[@Kho] on multifractal properties of critical states in 1D-quasiperiodic chains have been recently followed by renewed focus on the relation between localization features of such states and their ability to convey current[@Macia]. Still, variety of critical states let the question about localisation properties and transport ability relation unresolved. The effect of disorder on top of these states is thus a complicated issue and very scarce rigorous results are available in the litterature. Attempts to rigorously establish analytical results in high-dimensional quasiperiodic systems have been facing some limitations despite encouraging early attempts [@Moulopoulos; @Moulopoulos-2].
In what follows, we review exact results carried out on 1D umperfected quasiperiodic systems. We also give additional and complementary results to enlarge the understanding of an early study[@RocheFas] Besides, the role of quantum interferences on both sides of the quasicrystalline MIT for higher dimensional materials is outlined in a second part. A general scenario to follow the metal-insulator is drawn in the ligth of recent results.
Intereference effects in 1D quasiperiodic systems with disorder
===============================================================
Introducing disorder could be done through typical randomizing of site or hopping energies, with the subsequent occurence of Anderson localization in the infinite chain limit. For finite systems, localization lengthes may be much larger than the characteristic size so that conductance fluctuations as a function of energy of tunneling electrons (from the leads to the system) keep its self-similar nature and still follow power law behavior for system size studied in [@Dasarma] even when introducing disorder of $10\%$ of the total bandwith. However, the particular order sustaining long range quasiperiodicity suggest the possible presence of unique defects, known as phason-type defects. Their geometrical definition and properties have been subjected to many studies[@Phasons], although some aspects remain controversial. From certain viewpoint it seems natural to consider how disruptions of quasiperiodic order inherent to such systems will degrade or improve transport properties. It is the aim of this work to contribute to such more general and fundamental understanding of electronic propagation in quasicrystals.
For 1D-quasiperiodic systems, we can define phasons that keep the essential characteristic of real systems, in the sense that they are a generic form of disorder which has no equivalent in usual metallic and periodic systems. In a preliminary study, such phason defects were introduced by K. Moulopoulos, as a main probe for investigating Landauer conductance[@RocheFas].
Tight-binding models (TBM) of perfect quasiperiodic chains have been intensively worked out both analytically and numerically only for some given energies, but the results are supposed to have provided typical features of localization in quasiperiodic structures, such as power-law decrease of wavefunctions or power-law bounded resistances [@Kho]. However, if leading to interesting analytical results, TBM do not allow to simply investigate energy-dependent properties of quantum dynamics and electronic transport. Following the works and discussions of Kollar and Süto[@Suto], and M. Baake et al.[@Baake], the limitations of TBM when may be shown when considering effects of phason-type disorder on electronic localization and propagation. Let us recall the main results of [@RocheFas]. Tight-binding expression of the hamiltonian usually reads ${\cal H}=\sum_{n}t_{n}(|n\>\<n+1|+|n+1\>\<n|)$ ($\gamma =t_{A}/t_{B}$ will stand for intensity of quasiperiodic potential, following a Fibonacci sequence ABAABABAABAAB…, whereas site energies are kept constant) and the Schrödinger equation on a localized basis gives
[( \_[n+1]{}\
\_[n]{}\
)=M\_[n]{} ( \_[n]{}\
\_[n-1]{}\
)=M\_[n]{}M\_[n-1]{}....M\_[1]{} ( \_[1]{}\
\_[0]{}\
)=P\_[n]{} ( \_[1]{}\
\_[0]{}\
)]{} with $\psi_{n}$ the component of wavefunction for energy E at site n M\_[n]{}= ( 0 & -\
[1]{} & [0]{}\
) P(n)=[\_[i=1]{}\^[n]{}]{} M\_[i]{}
\[figLF\_s\]
Overlap integrals are given by $t_{n}=f(\theta_{n})$ with $\theta_{n}$ [@HKubo] is a kind of phase and $n$ a link index, so that $\theta_{n}=\frac{n}{\tau}+\theta_{0} (mod 1)$ $$\begin{aligned}
f(\theta)&=& t_{A} \ \ \ \hbox{for}\ \ \ 1/\tau^{2}\leq \theta < 1
\nonumber \\
&=& t_{B} \ \ \ \hbox{for}\ \ \ 0 \leq \theta <
1/\tau^{2} \nonumber\end{aligned}$$ where $\theta_{0}=0$ for the usual Fibonacci chain. The quasiperiodic chain corresponding to an initial phase $\theta_{0}=m/\tau$ is associated with the Fibonacci sequence taken at $m^{\hbox{th}}$ site of the usual one. We define the phason as a abrupt geometrical transition between the two chains with $\theta_{0}=0,2/\tau$. The function $f(\theta_{n})$ will take two values $t_{A}$ or $t_{B}$, according to the link. This gives a deterministic way to construct phason defects, which break long range quasiperiodic order, but which is a kind of disorder which has no equivalent in periodic systems. Introducing a so-called local isomorphism class by changing initial phase $\theta_{0}$ randomly chosen between 0 and 1, statistical studies of the localization and transport properties of Fibonacci chains have been performed[@HKubo; @Wij].
Effect of phason-defects at E=0 with TBM
----------------------------------------
In a transmission study, we define the boundary conditions as $t_{N}=t_{0}=t_{ext}=t_{B}$ (Fig. \[figLF\_s\]), which lead to special values of the number of sites. From the two Fibonacci chains with $\theta=0,2/\tau$, the corresponding sequence is $N(i=1,2,3,4....)=4,12,17,25,33,38,46,51,59,...$ noticing that the difference between two consecutive numbers follow a Fibonacci sequence with numbers 8 and 5 (8,5,8,8,5,8,5,8...). $$\begin{aligned}
\hbox{I} &:&(B)\hbox{{\bf
ABAABABAABAABAB}}\hbox{AABABAABAABABAABAABABAA}(B) \ \theta_{0}=0
\nonumber\\
\hbox{II} &:&(B)\hbox{AABABAABAABABAA}\hbox{{\bf
BABAABAABABAABAABABAABA}}(B) \ \theta_{0}=2/\tau
\nonumber\\
\hbox{III} &:&(B)\hbox{ABAABABAABAABA}{\bf
BB}\hbox{ABAABAABABAABAABABAABA}(B)\nonumber\end{aligned}$$ In what follows, the N(i+1)-term of the sequence ${\cal S}$ will not correspond to N+1 sites but to the number of allowed sites as deduced from the above-mentionned Fibonacci sequence. Fig. \[figLF\_s\] shows a small chain with 5 sites. The construction of a phason is illustrated below for a N=29 site chain. Taking $\gamma =t_{A}/t_{B}$, transfer matrix can be evaluated analytically as well as the Landauer resistances. Defining the matrices ${\cal A}, {\cal B}, {\cal C}$ by = ( 0 & -\
& 0\
) [B]{}=( & 0\
0 &\
) [C]{}=[B]{}[A]{}= ( 0 & -1\
1 & 0\
) One shows that the sequence $P_{N}$ follows a Fibonacci sequence. Besides, noticing that ${\cal CA}={\cal C}$, ${\cal AC}={\cal
CA}^{-1}$ and ${\cal C}^{2}=-\one$, one shows that whatever $P_{N}$ there exist a recurrent structure given by ${\cal
C}^{t(N)}{\cal A}^{s(N)}$ P\_[N]{}= ( 0 & -1\
1 & 0\
)\^[t(N)]{} ( & 0\
0&\
)\^[s(N)]{} where $t(N)$ are integers and $s(N)$ are described by a recursive relation. For the Fibonacci sequence $N=\{F_{1},F_{2},\ldots,F_{N-1},F_{N}\}$, H. Kubo and M. Goda [@HKubo] have investigated the statistical properties of $s(N)$ which are directly related to the characteristical exponents of self-similar wavefunctions. Taking $(\psi_{0},\psi_{-1})=(-i,1)$, it is possible to show that $|\psi_{n}|^{2}=
{\displaystyle \gamma^{-2s(N)\times(-1)^{t(N)}}}$. The Landauer resistance of a finite chain N is given by $\rho_{N}=\frac{h}{e^{2}}\ \frac{R}{T}$ where T is the fraction of tunneling electrons transmitted from the system to the leads, and R is the reflected one. For a tight-binding model one relates the resistance to the total transfer matrix elements: $\rho_{N}=\frac{1}{4}(P^{2}_{N}(1,1)+P^{2}_{N}(1,2)
+P^{2}_{N}(2,1)+P^{2}_{N}(2,2)-2)$.
\[sN\]
and one write the Fibonacci chain’s resistance in a closed form $$\bigl(\rho_{N}\bigr{)}_{\hbox{I}}
=\frac{1}{4}(\gamma^{2s(N)}+\gamma^{-2s(N)})-\frac{1}{2}
=\biggl(\frac{\gamma^{s(N)}-\gamma^{-s(N)}}{ 2}\biggr{)}^{2}$$ The function $s(N)$ is illustrated on Fig.\[sN\] which manifest self-similar pattern. When $s({N}_{0})=0$, transmission is perfect $T=1$. If one considers the Lyapunov exponents for finite length systems we get some estimates of the associated localization lengthes at a given energy [@Suto2]. By definition effective Lyapunov exponents (since localization length can be much larger than system size in the finite limit) for a chain of N sites are given by $\rho_{N}(E)=\frac{h}{2e^{2}} exp(\gamma_{N}(E)\times N)$ which is equivalent to $\gamma_{N}(E)=\frac{1}{N}|P_{N}(1,2)|^{2}$ (given that $\rho_{N}(E)=|P_{N}(1,2)|^{2}$).
\[Lyapu\]
For energies lying outside the spectrum are easily identified with stronger Lyapunov exponent as exemplified on Fig.\[Lyapu\]. The cantor spectrum for eigenvalues is revealed by the corresponding self-similar pattern of localization lengthes distribution.
We now remind the effect of phason defect on the Landauer resistance obtained in [@RocheFas]. Taking the position of the defect, represented by a -BB- unit as an internal degree of freedom, at E=0, the properties of the matrices ${\cal A,B, C}$ are shown to be independent of the position henceforth called $x_{P}$[@RocheFas]. By simple manipulations of the transfer matrix, we further demonstrate that the total transfer matrix associated to the chain with one phason and with $N(i)$ links is the same as the one without phason but with N(i+2) links. Accordingly, a generic form can be written down : $P_{N}$ (with either $a=d=0$, or $b=c=0$)
P\_[N(i)]{}|\_[I]{}= ( a & b\
c & d\
) then P\_[N(i)]{}|\_[III]{}= P\_[N(i+2)]{}|\_[I]{}=- ( 0 & -\
1 & 0\
) ( 0 & -\
1 & 0\
) ( a & b\
c & d\
)
Landauer resistance was shown to be given by [@RocheFas]:
$$\bigl(\rho_{N}\bigr{)}_{\hbox{III}}
=\Biggl(\frac{\gamma^{s(N)-1}-\gamma^{-s(N)+1}}{ 2}\Biggr{)}^{2}$$
By comparing the two expressions for the Landauer resistance one concludes that at $E=0$, the sign of $\rho_{N}\bigr{)}_{\hbox{I}} -(\rho_{N}\bigr{)}_{\hbox{III}}$ is fluctuating as a function of chain-length, which means the phason defect does not alter the transport properties in the infinite limit.
We now consider the sequence constructed following the same algorithm but maximizing the number of phason defects as identified by ${\bf BB}$. We then calculated analytically the Landauer resistance $\rho_{N}|_{IV}$. For instance for chains with N=4 and N=17 links, one has :
$$\begin{aligned}
(B)&-&\hbox{ABBA}-(B)\nonumber\\
(B)&-&\hbox{ABBABBAABABBABBAA}-(B)\nonumber\end{aligned}$$
With the same transfer matrices previously introduced ${\cal A, B}$ and noticing that BAABAB and BAB are equal respectively to ${\cal A}$ and ${\cal -B}=\tilde{\cal B}$, we show that the sequence of $P_{N}|_{IV}$ is generated by the following Fibonacci sequence $P_{N}={\cal A}, \ {\cal A}\tilde{\cal B}, \ {\cal A}\tilde{B}{\cal
A}, {\cal A}\tilde{\cal B}{\cal AA} \tilde{\cal B}, {\cal A}\tilde{\cal B}{\cal AA}\tilde{\cal B}{\cal A}\tilde{\cal
B}{\cal A}$, ${\cal A}\tilde{\cal B}{\cal AA}\tilde{\cal B}
{\cal A}\tilde{\cal
B}{\cal AA}\tilde{\cal B}{\cal AA}\tilde{\cal B},...$. The analytical form of $\rho_{N}|_{IV}$ is shown to be given by
$$\bigl(\rho_{N}\bigr{)}_{\hbox{IV}}
=\Biggl(\frac{\gamma^{\tilde{s}(N)}-\gamma^{-\tilde{s}(N)}}{
2}\Biggr{)}^{2}$$
with $\tilde{s}(N)$ a new complex function of N calculated iteratively. The interesting point to observe is that the function $\rho_{N}|_{I}-\rho_{N}|_{III}$ for one phason defect changes its sign at each step $N\to N+1$, whereas $\rho_{N}|_{I}-\rho_{N}|_{IV}$ manifestes fluctuations on much larger range. As an illustration, chains with number of sites respectively equal to $N(i=1,2,3,4,5...17)=5-203$ sites follow $\rho_{N}|_{I}-\rho_{N}|_{IV}\geq 0$, whereas the behavior is opposite for chain from 330 to 456 sites, and so forth.
In conclusion, even for the highest density, such phason disruptions of quasiperiodic order are not able to break down the localization mechanism which do remain basically the same in the limit $N\to\infty$.
Landauer Resistance of a Krönig Penney model with phasons
=========================================================
Herafter we first rewrite the main steps of calculations as first described in [@RocheFas]. We then perfom power-spectra calculations of Landauer resistance interference patterns to clarify the differences between localization properties and transmission abilities of critical states. In the Krönig Penney model[@KP], the potential describing the interaction of the electron with the lattice is represented by a sum of Dirac distributions with intensity $V_{n}$ localized at $x_{n}$ $V(x)=\sum_{n} V_{n}\delta(x-x_{n})$, the $x_{n},V_{n}$ can be chosen as correlated variables, or uncorrelated. In between two successive scattering centers, the solution of the Schrödinger equation is a linear combination of two plane waves: $
\Psi(x)=A_{n}e^{ik(x-x_{n})}+B_{n}e^{-ik(x-x_{n})}$ $(x_{n}\leq
x\leq x_{n+1})$, the 1D wavevector $k>0$ is related to the energy $E$ through [ $E={\hbar}^{2}k^{2}/2m$]{}. In the Krönig-Penney (KP) [@KP] model, a solution of the problem is constructed by imposing continuity conditions for the wavefunction and its derivative. For sake of simplicity, we choose the case where the intensity of scattering potential is constant =$\lambda$ , whereas scattering centers are quasiperiodically spaced $\{({x}_{n}-{x}_{n-1}) \}= \{a,b \}= \{\tau,1\}$. The problem can then be written as followed :
( A\_[n+1]{}\
B\_[n+1]{}\
)=(n). ( A\_[n]{}\
B\_[n]{}\
)with
(n)= ( (1-[[i ]{}]{} ) e\^[i k (x\_[n+1]{}-x\_n)]{} & -[[i ]{}]{} e\^[i k (x\_[n+1]{}-x\_n)]{}\
& \
[[i ]{}]{} e\^[-i k (x\_[n+1]{}-x\_n)]{} & (1+[[i ]{}]{} ) e\^[-i k (x\_[n+1]{}-x\_n)]{}\
)
within this framework, Landauer resistance is given by $\rho_{N}= |P_{n}(1,2)|^2 \ \ \ \hbox{with} \ \ \ P_{n}=
\Lambda(F_n) ...\Lambda(1)$. Furthermore, a renormalization group associated with the Fibonacci chain enables to map the electronic spectrum to a dynamical system associated with the traces of the transfer matrices. The following energy-dependent invariant of the Krönig-Penney model is convenient to consider [@Invar]. In our case, we find $\hbox{I(k)}=\frac{\lambda^2 \sin^2 k(a-b)}{4 {k}^2}$ whose zeros are given by ${k_{s}}=n\pi/(a-b)$, with n integer. These points refered as conducting points are interesting since they correspond to the commutation of the transfer matrices, $[ P_{F_{n}},P_{F_{n+1}} ]=0 \ \ \hbox{given that} \ \ 4
\hbox{I}+2=\hbox{Tr}(P_{F_{n}}.P_{F_{n+1}}.P_{F_{n}}^{-1}.P_{F_{n+1}}^{-1})$. Besides elementary matrices also commute : = [[e\^[i k (a-b)]{}]{}]{} (1-e\^[2 i k (b-a)]{}) ( & -2 i k\
--2 i k& -\
) The resistance can be written down analytically at ${k_{s}}$ \_[N]{}|\_[\_[\_[k=k\_s]{}]{}]{}=( )\^2 with $\varphi$, depending on $k^{2}_{s}$ and $\lambda$, and is defined equivalently by $\cos \varphi=\cos k_s a+ \lambda/{2 k_s} \sin k_s a$ or $\cos \varphi=\cos k_s b+ \lambda/{2 k_s} \sin k_s b$. The anaytical form of [ $\rho_{N}|_{_{_{k=k_s}}}$]{} indicates that when $N\to\infty$, the Landauer resistante oscillates but remains bounded, so that the energies $k^{2}_{s}$ correspond to states that lead to best transmission, reminding that localized states will display exponantial increase of resistance. Following the discussion of [@Suto; @Baake], if ones restricts the study for energies $k^{2}=(k_{s}+\varepsilon)^{2}$ in the vicinity of $k_{s}$ where highest density of eigenvalues are expected in the infinite limit (hereafter $k_{s}=\pi/(\tau-1), (a=\tau, b=1)$ is taken without loss of generality). Two transfer matrices are defined \_[A]{}= ( (1-)[e]{}\^[ik]{} & - \^[ik]{}\
& \
\^[-ik]{} & (1+)[e]{}\^[-ik]{}\
) \_[B]{}= ( (1-)e\^[ik]{} & - \^[ik]{}\
& \
\^[-ik]{} & (1+) [e]{}\^[-ik]{}\
) Choosing $\lambda$ so that [ $\rho_{N}|_{_{_{k=k_s}}}=0$]{} exactly at $k={k_s}$, which lead to $N-1$ values for $\lambda$, given by ${\lambda}_{s}= {{2 k_s (\cos \varphi_s-\cos k_s )}\over {\sin k_s}}$ with the phase $\varphi_s$ defined through [${\varphi}_{s}=(m
\pi)/ N$]{} with m=1,...,N-1. The N-1 values $\varphi_s$ are symetrically distributed around $\pi/2$ and cover densely the half-upper trigonometric plane when increasing N [@RocheFas]. Reminding that $x_{P}$ is taken as the position of phason defect ([**BB**]{}) in the chain, with $x_{P}\in [1,P]$ an integer and $P$ the maximum number of position for a given chain, we know recall main important patterns [@RocheFas] and perform power spectra of $\rho_{N}(x_{P},\varphi_s,\varepsilon)$.
\[figdf\]
Numerical accuracy is checked through $P_{N}(1,1)^{2}-1=P_{N}(1,2)^{2}$ which gives the resolution. The Landauer resistance calculated in $k=k_{s}$ is always found to be $\sim 10^{-12}-10^{-13}$ and gives thus the numerical uncertainty on $\rho_{N}$. We now present the main results as sketched on the phase diagram in Fig.\[figdf\]. According to the variable $\varphi_{s}$ ones identifies several features. First, note that amazing pseudo-symmetry is found around $\pi/2$ for $\rho_{N}(x_{P},m,\varepsilon)$ which do not correspond to any symmetry in the scattering potential (see below Fig.\[figR34\]-\[fig1-5\] for illustration).
\[figR34\]
In the zones refered as zone-[**I**]{} the phason reduces the resistance for energies sufficiently close to the conducting points (see Fig.\[figR34\]-\[fig1-5\]). Two symmetrical zones are $ m < N-4P, \ \hbox{and} \ m > 4P$ with ${\rho_{N}}\mid_{\hbox{I}}(x_{P},m,\varepsilon)>
{\rho_{N}}\mid_{III}(x_{P},m,\varepsilon)$. In other words, for that values of parameters the Fibonacci chain becomes more conductive upon introduction of local phason. This effect is a pure result of quantum intereference at low temperature and has been shown for tunneling energies close to the ones of conducting points, so for the electrons that can better contribute to conduction mechanism. We illustrate these behaviors first on Fig.\[figR34\] which allow to write down all the 8 different inequivalent defected-chains as described below for clarity for a system with 35 sites:
FIBO&:& –B–ABAABABAABAABABAABABAABAABABAABAAB–B–\
Def1&:& –B–A[**BB**]{}ABAABAABABAABABAABAABABAABAABAB–B–\
Def2&:& –B–ABAABA[**BB**]{}AABABAABABAABAABABAABAABAB–B–\
Def3&:& –B–ABAABABAA[**BB**]{}ABAABABAABAABABAABAABAB–B–\
Def4&:& –B–ABAABABAABAABA[**BB**]{}ABAABAABABAABAABAB–B–\
Def5&:& –B–ABAABABAABAABABAABA[**BB**]{}AABABAABAABAB–B–\
Def6&:& –B–ABAABABAABAABABAABABAA[**BB**]{}ABAABAABAB–B–\
Def7&: &–B–ABAABABAABAABABAABABAABAABA[**BB**]{}AABAB–B–\
Def8&:& –B–ABAABABAABAABABAABABAABAABABAA[**BB**]{}AB–B–
Besides, the function ${\rho_{N}}\mid_{\hbox{III}}(k,m,x_{P})$ given a kind of interference pattern described by the position of phason defect ${\rho_{N}}\mid_{\hbox{III}}(k,m,x_{P})\sim \alpha(m/P)\
\sin\bigl( \frac{2m\pi}{P} x_{P}\bigr)$. This is further represented on Fig.\[fig1-5\] for chains with N=2000 links, and m=1,5. The curves for m=1 (Fig.\[figR34\] and Fig.\[fig1-5\](top)) and different chain lengths shows that the interference pattern encloses a memory of self-similarity.
\[fig1-5\]
The power spectrum of m=5 curve is given on Fig.\[fig6\], where the only eigenfrequency agrees with $m/P=5/472=0.0106$ in this case. Superimposed small oscillations are an unphysical effect due to a Fourier transform of a finite signal.
\[fig6\]
\[figr3\]
Let’s move forward to zone-[**II**]{}. for $N-4P<m<4P$ situations becomes more complex but recurrent simple ($m-P$)-periodic oscillatory patterns are found around the values $\varphi_{s}=
\{ (N-4P)/N , \ (N-3P)/N, \ (N-2P)/N , \ (N-P)/N \} \ \pi$ and symetrically for $\tilde{\varphi}_{s}= \{ 4P/N , \ 3P/N , \ 2P/N , \ P/N \} \pi$. In these regions of parameter space, there is an genuine transition from systematic increasing to decreasing (and vice versa) of the electronic resistance upon introduction of phason, as exemplified on Fig.\[figr3\].
\[figrc1\]
As far as zone-[**III**]{} is concerned, and related to some interval around $\varphi\sim\frac{\pi}{2}$, self similar patterns are observed which suggest that ${\rho_{N}}\mid_{\hbox{III}}(x_{P})$ reveal critical states which are robust against phason disorder as found in the TBM for E=0 case. The typical patterns represented on Fig.\[figc1\] actually encloses oscillations of resistance which smaller oscillations are described by some coefficients $s(n)$ as described previously.
\[figc3\]
On the Fig.\[figc3\], the case with 3000 sites is considered and differents values of the scattering potential in the vicinity of the symmetry point $\varphi=\frac{\pi}{2}$. For m=1497, the potential is $\lambda \sim 3.918$, for m=1498 $\lambda \sim 3.929$ and 3.941 for m=1499. Patterns exhibit small differences, and their Fourier spectrum is actually very similar (see below).
\[figc1\]
The power spectra of different similar patterns are given on Fig.\[figc1\] and Fig.\[figc2\] which show that superimposed frequencies are identical. The highest frequency is given by $\nu=0.5$ which is related to the change of $\rho_{N}(x_{P}\to
x_{P}+1)$. On respective figures five unambiguous frequencies have been located and named $\nu_{n=1,5}$. This prove that all these self-similar patterns are actually described by exactly the same function which is a superposition of several independent frequencies convoluate with a function defined by a serie of $s(N)$-type coefficients. Concerning all the discussed zones-[**I, II, III**]{}, simple forms for energies sufficiently close to conducting points. Then $\rho_{N}(x_{P},\varepsilon)$ result from a superposition of one or several frequencies. For larger energies one finds fluctuations as those analyzed in the inset of Fig.\[figc3\]. As energies get farther to conducting points ${k}_{s}=\pi/(\tau-1)$ the resistance is sharply increasing on several order of magnitudes. The power spectrum in this case reveal much more eigenfrequencies, with the interesting emergence of frequencies with similar amplitudes refered as $\nu$ and $\omega$. In all these cases the respective behaviors of the Landauer resistance of the perfect Fibonacci chain versus the imperfect one follows complicated random fluctuations.
\[figc2\]
The different behaviors found in these studies suggest that in some case local disruption of long range quasiperiodic order has improved the conductive ability of the chain in a systematic manner. Analyzing the interference pattern of the Landauer resistance as a function of phason defect suggest that extendedness (as a localization properties of available states at such energies) has also been jointly improved. This is shown by a bounded and simple oscillatory pattern for the resistance, common to what is found for extended eigenstates in a periodic systems.
\[fig3\]
To conclude this study, one stresses that first conducting points indeed seem to be the location of high density of eigenenergies as discussed in [@Suto; @Baake] and Krönig Penney model has been able to unravel specific quantum interference effects of phason disorder on localization and propagation properties, hidden in the treatment of TBM, and appealing in the context of quasicrystalline materials. Here following earlier results [@RocheFas], we have found that multi phason defects in TBM at E=0, do not alter the transmission abilities of corresponding states, and that interference patterns as revealed by $\rho_{N}(x_{P},\varepsilon)$ and their power-spectra analysis is an original way discovering how phason defects affect jointly localization and transport modes.
Quantum interference mechanisms in high-dimensionnal quasicrystals
==================================================================
Some effort to investigate quantum interferences effects in small quasiperiodic penrose approximants have been made [@Moulopoulos]. Here we propose how quantum interferences on both sides of a metal-insulator transition in real materials might be analyzed. Indeed, weak localization regime has been found in experiments for some quasicrystalline materials (AlCuFe,…) whereas other systems such as AlPdRe-quasicrystals behave differently being very close to a metal-insulator transition[@Berger]. Two different focus may be considered for a general understanding of quantum interferences in quasicrystals. First, as there exist approximant phases (periodic) sharing the same behavior, weak localization correction beyond Drude approximation should apply for that systems as well as for corresponding quasicrystals. The only recurrent approximation when solving the cooperon diffusion equation is to assume that scatters are uncorrelated i.e $\<{\cal U}(r){\cal U}(r')\>_{disorder}=cu^{2}\delta(r-r')$ (${\cal U}(r)= \sum_{i=1}^{N}\ {\cal U}(r-R_{i})$, c the impurity concentration, u typical strength). The calculation of the quantum correction of the conductivity in this regime is enclosed in phase factor interferences of the two-particle Green’s function $\<GG^{*}\>$ . By performing configuration averaging beyond the mean free path legnth scale, then $\< GG^{*} \>_{disorder}$ reduces phase interference to the ($k'=-k$)-Cooperon pole, as a consequence of time reversal symmetry, the possibility to have a coherent distribution of scatters is usually neglected.
However, assuming that the distribution of scatters is constrained to, let’s say for sake of illustration, a mirror-plane symmetry, i.e $\forall \alpha \in \{ R_{\alpha}, \alpha=1,N\}$ there is a site $R_{\beta}$ such that $R_{\beta}=-R_{\alpha}$, then without performing any diagrammatic expansion, we just notice that weak localization related with average of the potential scattering $\< {\cal U}(r){\cal U}(r'){\>}_{disorder}$ corresponding to phase factors
$$\sum_{\alpha\beta}\< {e}^{-i(k\ R_{\alpha}+p\
R_{\beta})}{\>}_{disorder} =\frac{1}{\Omega^{2}}\ \int d^{3}R_{\alpha}d^{3}R_{\beta} {e}^{-ikR_{\alpha}}
\ {e}^{-ipR_{\beta}}$$
will display new terms associated to above-mentionned symmetry, $\< {e}^{-i(k\ R_{\alpha}+p\ R_{\beta})} {\>}_{disorder}=
\< {e}^{i(p-k) R_{\alpha}} {\>}_{disorder}=\delta_{p-k,0}$ that will increase the contribution of phase interferences. Say in another way, if a double symmetrical loop crosses in the mirror plane and in a region with extension less or equal to $\lambda_{F}$, then four equivalent pathes will interfere at the returning point instead of the usual two of the weak localization scheme, resulting in a total interference amplitude will be 4 times stronger ($\mid {\cal A}_{I}+{\cal A}_{II}+{\cal A}_{III}+{\cal A}_{IV}\mid
^{2}=16\mid {\cal A}\mid ^{2} $). Similar ideas have been already introduced in context of mesoscopic physics [@Baranger] but we propose here that they may serve as a path to follow a metal-insulator in quasicrystals in which even disorder (such as phasons) may keep some strong correlated properties[@Roche-un].
The second focus is suggested by the close proximity of a metal-insulator transition and gives a very dissimilar weight to quantum interferences. Any critical point of an electronic localisation-delocalisation transition can essentially be described by its anomalous diffusion which means that two-particle Green function reads $\< |G^{+}(r,r';E)|^{2}\>\sim|r-r'|^{-\eta +2-D}$ with $\eta$ a critical exponent (and the propagator representes the transition probability in real space of an electron of E energy from site $|r\>$ to $|r'\>$), and which directly affect the conductivity of the system since
\_[DC]{}=\_[0]{} 4d\^[D]{}r r\^[2]{} |G\^[+]{}(r,r’;E)|\^[2]{}\_[disorder]{} \[eqs\]
In the case where all states are localized (insulating side), the average is carried out over random configurations of frozen disorder and given that $\<|G^{+}(r,r';E)|^{2}\>_{imp}\sim \exp^{-|r-r'|/\xi_{E}}$ ones easily recovers that $\sigma_{DC}\to 0$ at zero temperature. At the critical point, the power law describing electronic propagation has been illustrated in quantum Hall regime (D=2) for length scales in the regime of multifractality, i.e $l\ll r\ll\xi$, $\eta\simeq 0.38\pm 0.04$[@QHE] or analytically in hierarchical potential [@Jona] and quasiperiodic systems[@Bell] and numerically in 3D quasiperiodic systems[@RocheJMP; @RochePRL1]. No characteristic length scales, such as the mean free path, can be defined and interference effects are intrinsically defining the dominant transport mechanism. A simple analysis of the equation 12 shows that according to the strength of criticality, as manifested by the value of the exponent, conductivity can be zero or infinite at zero temperature for a perfect system. Elastic scattering allow to work out the exponents of the anomalous Drude formula[@Bell; @RochePRL1] at very low temperature.
At finite temperature, a dissipation mechanism results in a cut-off in the integral, and one always gets a finite value for the resistance. Since the interesting point occurs when multifractality dominate the transport regime, it is then important to estimate temperature dependent transport. In quasicrystals, critical states are associated with the so-called anomalous Drude law ($\sigma\sim \tau^{2\beta-1}$) [@RocheJMP; @Bell], where $\tau$ is supposed to be the relevant inelastic scattering rates at a given temperature. From a scaling analysis of the Kubo formula on periodic quasicrystalline approximants, and from the knowledge of exact eigenstates a scaling behavior was revealed ${\displaystyle
\Delta\sigma(T)\sim T^{\eta},\ \ \hbox{\small with}\ \ \eta\simeq 1.25}$ in good agreement with the values obtained experimentally for several quasicrystalline and approximant phases at temperature $\geq 10K$ [@Gignoux-I; @Kimura-I; @Lalla-I].
In both side of the metal-insulator transition, the presence of phasons as demonstrated in 1D systems may weeken the interference effects while destroying quasiperiodic long range order, resulting thus in a unconventional mechanism. Working this out analytically in realistic models remains a great challenge.
Conclusion
==========
Several features occurring in the electronic transport of 1D quasicrystals have been reviewed. Quantum interferences have been shown to be interestingly altered by phason defects which may be a natural disruption of quasicrystalline disorder. In quasicrystals QIE may be sustained by different mechanisms as discussed in the second part. Revealed pattern in 1D-systems may keep some generality in higher dimension.
S.R. is indebted to the European Commission for financial support (Contract ERBIC17CT980059), and to Prof. T. Fujiwara from Department of Applied Physics of Tokyo University for hospitality. Kostas Moulopoulos is deeply acknowledged for many enlightening discussions and encouragements.
[12345678]{}
D. Shechtman, I. Blech, D. Gratias et J.W. Cahn, [*Phys. Rev. Lett.*]{} [**51**]{}, 1951 (1984).
J.M. Dubois et al. Annales de Chimie [**19**]{}, 3 (1994). S. Roche, Annales de Chimie 23, 953 (1998). E. Maciá, in Revista Espa$\tilde{n}$ola de Física (may 1998).
New Horizons in Quasicrystals: Research and Applications, ed. A.I. Goldman, D.J. Sordelet, P.A. Thiel and J.M. Dubois, (World Scientific Singapore, 1997).
C. Berger, [*Seminar 3*]{}, E. Akkermans, G. Montambaux, J.L. Pichard and J. Zinn-Justin eds. Les Houches, Session LXI, [*Mesoscopic Quantum Physics*]{} (1994).
S. Roche, G. Trambly de Laissardière and D. Mayou, J. of Math. Phys. [**38**]{}, 1794 (1997).
T. Fujiwara, in Physical Properties of Quasicrystals, Springer Series in Solid-State Sciences [**126**]{}, Editor Z.M. Stadnik, 169 (1999)
S. Roche and T. Fujiwara, Phys. Rev. B 58, 11338 (1998).
C. Sire in “Lectures on Quasicrystals", ed. F. Hippert and D. Gratias (Les Editions de Physique Les Ulis, 1994).
S. Yamamoto and T. Fujiwara, Phys. Rev. B [**51**]{}, 8841 (1995).
S. Roche and D. Mayou, Phys. Rev. Lett. [**79**]{}, 2518 (1997).
M. Kohmoto, L. P. Kadanoff, and Ch. Tang, Phys. Rev. Lett. [**50**]{} 1870 (1983); ibid. [**50**]{} 1873 (1983). M. Kohmoto, Intern. J. of Mod. Phys. B [**1**]{} 31 (1986). M. Kohmoto and J. R. Banavar, Phys. Rev. B [**34**]{} 563 (1986).
E. Maciá and F. Domínguez-Adame, Phys. Rev. Lett. [**76**]{}, 2957 (1996). and references therein.
S. Das Sarma and X.C. Xie, Phys. Rev. B [**37**]{}, 1097 (1988).
T.C. Lubensky, J.E.S. Socolar, P.J. Steinhardt, P.A. Bancel and P.A. Heiney, Phys. Rev. Lett. [**57**]{}, 1440 (1986). G. Coddens, R. Bellissent, Y. Calvayrac, J.P. Ambroise, Europhys. Lett. [**16**]{} 271 (1991). A. Trub and H.R. Trebin, [*J. Phys. France*]{} [**4**]{} 1855-1866 (1994). H. Klein, M. Audier, M. Boudarda and M. de Boissieu, L. Behara and D. Duneau, Phil. Magazine A [**73**]{}, 309-331 (1996).
K. Moulopoulos and S. Roche, Phys. Rev. B [**53**]{}, 212 (1996).
K. Moulopoulos and S. Roche, Journal of Condensed Matter C: [**7**]{} 8883 (1995).
K. Moulopoulos, unpublished.
J. Kollar and A. Süto, Physics Letters A [**117**]{}, 203 (1986).
M. Baake, D. Joseph and P. Kramer, Phys. Lett. A [**168**]{}, 203 (1986).
M. Goda and H. Kubo, J. Phys. Soc. Jpn. [**58**]{} 3624 (1989) ; H. Kubo and M. Goda, J. Phys. Soc. Jpn. [**60**]{} 2729 (1991) ; J. Phys. Soc. Jpn. [**62**]{} 2601 (1993).
F. Wijnands, J. Phys. A [**22**]{}, 3267 (1989).
A. Süto, in [*Beyond quasicrystals*]{}, edited by F.Axel and D. Gratias (Les Editions de Physique, France, 1994) 483.
R. de L. Krönig and W.G. Penney, in Mathematical Physics in One Dimension, eds. E.H. Lieb and D.C. Mattis, p. 243.
M. Baake, U. Grimm and D. Joseph, Int. Journ. of Moder. Phys. B, [**7**]{} Nos. 6&7 (1993) 1527.
S. Roche, unpublished.
G. Jona-Lasinio, F. Martinelli and E. Scoppola, J. Phys. A: Math. Gen., [**17**]{} (1984) L635; ibid Ann. Inst. Henri Poincaré [**42**]{} (1985) 73.
H. Schulz-Baldes and J. Bellissard, Rev. Math. Phys., [**10**]{} 1 (1998). H. Schulz-Baldes, Phys. Rev. Lett. [**78**]{} , 2176 (1997).
M. Janssen, Intern. Journ. of Mod. Phys. B [**8**]{}, 943 (1994).
M. B. Hastings, A. D. Stone and H. U. Baranger, Phys. Rev. B [**50**]{} 8230 (1994). H.U. Baranger, R.A. Jalabert and A.D. Stone, Phys. Rev. Lett. [**70**]{} 3876 (1993).
C. Gignoux, C. Berger, G. Fourcaudot, J.L. Grieco and H. Rakoto, Europhys. Lett. [**39**]{}, 171-176 (1997).
R. Tamura, H. Sawada, K. Kimura and H. Ino, [*Proceedings of the 6th International Conference on Quasicrystals*]{}, Editors S. Takeuchi and T. Fujiwara, (World Scientific, Singapore 1998).
N.P. Lalla, R.S. Tiwari and O.N. Srivastava, J. Phys.: Condens. Matter [**7**]{}, 2409 (1995).
Figure captions
Fig1 Chain $N=4$: ABAA connected with perfect leads.
Fig2 Multifractal distribution of $|s(N)|$ for the Fibonacci chain of 800 sites.
Fig3 Lyapunov exponents as a function of energy of a Fibonacci chain.
Fig4 Energy dependent phase diagram.
Fig5 Landauer resistance as a function of phason position $x_{P}$, for $\varepsilon=10^{-4}$, $N=35$ sites and m=1.
Fig6 Landauer resistances as a function of phason position ${x}_{P}$ (N=2001,P=472) for $\varepsilon=[10^{-9},10^{-6}]$ and m=5. For higher values of $\varepsilon$, then $\rho_{N} \to \infty$ which indicates that energy lies within a gap, or may be associated with a localized state.
Fig7 Power spectrum ofthe pattern given in the inset for m=5 and same parameters as previous figure (bottom).
Fig8 Regular evolution of interference pattern for $m=P, P+1,P+2$ and $N=2000$, $P=472$. For $\varphi_{s}=P\pi/N$, Fibonacci chain is always less resistive that the imperfected one whereas transition occurs for $m=P+1,P+2,...$. dashed curves are the values for Fibonacci chains with same parameter m as imperfected one.
Fig9 Landauer resistance for values of scattering potential which unravel self-similar patterns as described in the text.
Fig10 Interferences pattern for several values of scattering potentials close to $\varphi\sim\frac{\pi}{2}$.
Fig11 Self similar interference pattern for N=3207, P=757 and m=1605. Stability is ensured from $\varepsilon=10^{-3}$ down to the numerical resolution limit.
Fig12 Same as previous figure with different m.
Fig13. Landauer resistance interference pattern for $\varepsilon=0.5$, N=2000 and P=472.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the setting of 2-step nilpotent Lie groups in the particular case that its type $(p,q)$ is not exceptional. We demonstrate that, generically, the orbits of $\mathbb R^{>0}\times {\operatorname{Aut}}_0$ in $GL(n)/O(n)$ are congruent even when a Ricci soliton metric does exists. In doing so, we provide a counterexample to a conjecture of Taketomi-Tamaru.'
author:
- Michael Jablonski
date: 'October 18, 2018'
title: 'Counterexample to a conjecture of Taketomi-Tamaru'
---
[^1]
dotsep[5]{}
Among solvable and nilpotent groups, perhaps the most natural distinguished Riemannian metrics are those left-invariant metrics which are either Einstein or Ricci soliton. These metrics are known to minimize natural functionals [@Heber; @LauretNilsoliton] and have maximal symmetry when compared to other left-invariant metrics [@Jablo:ConceringExistenceOfEinstein; @Jablo:MaximalSymmetryAndUnimodularSolvmanifolds; @GordonJablonski:EinsteinSolvmanifoldsHaveMaximalSymmetry]. The pursuit of algebraic and geometric criteria which guarantee or preclude the existence of these metrics has been a long standing avenue of investigation, with the lion’s share of the attention given to the algebraic side, see e.g. [@Nikolayevsky:EinsteinSolvmanifoldsandPreEinsteinDerivation].
Recently there have been several works approaching this question from the geometric side. Given a Lie group $G$ with Lie algebra $\mathfrak g$, one can study the left-invariant metrics on $G$ by studying inner products on $\mathfrak g$; thus, one considers the set of inner products which is naturally presented as the symmetric space $GL(n)/O(n)$, where $n=\dim \mathfrak g$.
The subgroup $\mathbb R^*\times {\operatorname{Aut}}(\mathfrak g) \subset GL(n)$ acts on the space of inner products $GL(n)/O(n)$. Any two inner products in the same ${\operatorname{Aut}}(\mathfrak g)$-orbit are isometric and so inner products in the same $\mathbb R^*\times {\operatorname{Aut}}(\mathfrak g)$-orbit are isometric up to scaling. Note, in the nilpotent setting the ${\operatorname{Aut}}(\mathfrak g)$-orbits are precisely the isometry classes, but this is not necessarily true for solvable, non-nilpotent groups [@GordonWilson:IsomGrpsOfRiemSolv].
Given an inner product ${\langle \cdot, \cdot \rangle} \in GL(n)/O(n)$, the orbit $\mathbb R^*\times {\operatorname{Aut}}(\mathfrak g) \cdot {\langle \cdot, \cdot \rangle}$ has recently been dubbed the *corresponding submanifold*. We collect some recent results on the geometry of this corresponding submanifold.
- Let $G$ be a 3-dimensional solvable Lie group with left-invariant metric ${\langle \cdot, \cdot \rangle} \in GL(3)/O(3)$. Then $(G,{\langle \cdot, \cdot \rangle})$ is a Ricci soliton if and only if the corresponding submanifold $\mathbb R^*\times {\operatorname{Aut}}(\mathfrak g)\cdot {\langle \cdot, \cdot \rangle}$ in $GL(3)/O(3)$ is a minimal submanifold [@Hashinaga-Tamaru:Three-dimensionalSolvsolitonsAndTheMinimalityOfTheCorrespondingSubmanifolds].
- Let $G$ be a 4-dimensional nilpotent Lie group with left-invariant metric ${\langle \cdot, \cdot \rangle}\in GL(4)/O(4)$. Then $(G,{\langle \cdot, \cdot \rangle})$ is a Ricci soliton if and only if the corresponding submanifold $\mathbb R^*\times {\operatorname{Aut}}(\mathfrak g)\cdot {\langle \cdot, \cdot \rangle}$ in $GL(4)/O(4)$ is a minimal submanifold. Furthermore, it is shown that this result is false for some solvable groups in dimension 4 [@Hashinaga-OnTheMinimalityOfTheCorrespondingSubmanifoldsToFour-DimensionalSolvsolitons] .
- Let $G$ be an n-dimensional nilpotent Lie group. It is conjectured that if $\mathbb R^*\times {\operatorname{Aut}}(\mathfrak g)$ does not act transitively on $GL(n)/O(n)$ and all the orbits are congruent, then $G$ does not admit a Ricci soliton [@Taketomi-Tamaru:OnTheExistenceOfLeftInvariantRicciSolitons-AConjectureAndExamples Conj. 1.2]. The authors go on to verify the conjecture for some classes of nilpotent groups in every dimension - the algebras there have large automorphism groups.
Recall, two orbits being congruent means that there is an isometry $\phi \in GL(n)$ of the symmetric space $GL(n)/O(n)$ which sends one orbit to the other.
In the present work, we consider the connected components of the corresponding submanifolds, i.e. we will look at the orbits of the connected group $\mathbb R^{>0}\times {\operatorname{Aut}}(\mathfrak g)_0$ in $GL(n)/O(n)$.
\[thm: main theorem\] There exists a 9-dimensional nilpotent Lie group $G$ such that
1. $G$ admits a Ricci soliton metric and
2. $\mathbb R^{>0}\times {\operatorname{Aut}}(\mathfrak g)_0$ does not act transitively on $GL(9)/O(9)$ and the orbits are all congruent.
This essentially provides a counterexample to the conjecture of Taketomi-Tamaru; note that our result only concerns the connected group $\mathbb R^{>0}\times {\operatorname{Aut}}(\mathfrak g)_0$ and not the full group $\mathbb R^{*}\times {\operatorname{Aut}}(\mathfrak g)$. However, our result does demonstrate that really there is no criterion on the geometry of the corresponding submanifolds which can determine the soliton condition, or likely any other distinguished condition. Furthermore, the phenomenon occurring in the theorem above seems to be more common than not, see Theorem \[thm: generic points being soliton and small derivations\].
We note that we do not take up the remaining interesting question of whether or not it is possible that conditions such as the corresponding submanifold being minimal or not either guarantee or preclude the existence of a soliton metric. It would be interesting to know if the corresponding submanifolds appearing in Theorem A are indeed minimal submanifolds.
outline of proof {#sec: outline of proof}
================
Our strategy is to work with Lie algebras whose algebra of derivations is very small. Recall, the Lie algebra of the automorphism group is the algebra of derivations.
Consider a 2-step nilpotent Lie algebra $\mathfrak n = \mathfrak v + \mathfrak z$, where $\mathfrak z=[\mathfrak n,\mathfrak n]$ is the commutator subalgebra and $\mathfrak v$ is a complement of $\mathfrak z$. As $\mathfrak n$ is 2-step nilpotent, $\mathfrak z=[\mathfrak n,\mathfrak n]$ is central, but might not be all of the center of $\mathfrak n$. The algebra $\mathfrak n$ comes equipped naturally with two kinds of derivations. First we have the (1,2)-derivation $$\label{eqn: 1,2 derivation}
D = \begin{bmatrix}
Id_\mathfrak v & 0\\
0 & 2 \ Id_\mathfrak z \end{bmatrix} ;$$ then we have derivations which vanish on $\mathfrak z$ and map $\mathfrak v $ to $\mathfrak z$, i.e. ones of the form $$\label{eqn: der v to z}
\begin{bmatrix}
0 & 0\\
* & 0 \end{bmatrix} .$$ These are precisely the derivations valued in $\mathfrak z$ and they form an ideal in ${\operatorname{Der}}(\mathfrak n)$. We denote this set of derivations by ${\operatorname{Der}}_{\mathfrak v \to \mathfrak z}$. Combining the above, we have a subalgebra $\mathbb R (D) \ltimes {\operatorname{Der}}_{\mathfrak v \to\mathfrak z}$ of ${\operatorname{Der}}(\mathfrak n)$. In fact, one can argue that this is an ideal of ${\operatorname{Der}}(\mathfrak n)$, though we won’t need this fact. In general, the set of derivations is $$\label{eqn: derivations split}
{\operatorname{Der}}(\mathfrak n) = {\operatorname{Der}}(\mathfrak n)\cap (\mathfrak{gl(v)}\oplus \mathfrak{gl(z)}) \oplus {\operatorname{Der}}_{\mathfrak v \to \mathfrak z};$$ and at the automorphism level one has $${\operatorname{Aut}}(\mathfrak n) = ({\operatorname{Aut}}(\mathfrak n) \cap GL(\mathfrak v)\times GL(\mathfrak z)) \ \exp ({\operatorname{Der}}_{\mathfrak v\to\mathfrak z}),$$ where the subgroup $\exp({\operatorname{Der}}_{\mathfrak v\to\mathfrak z})$ is a normal subgroup.
We are interested in the case when ${\operatorname{Der}}(\mathfrak n) = \mathbb R (D) \ltimes {\operatorname{Der}}_{\mathfrak v \to\mathfrak z}$ and for the rest of this section we make this assumption going forward. By Lie’s Theorem, we know there is some upper triangular matrix $\mathfrak s \supset {\operatorname{Der}}(\mathfrak n)$; this is easily seen using a basis of $\mathfrak v$ concatenated with a basis of $\mathfrak z$. Observe that ${\operatorname{Der}}(\mathfrak n)$ is then an ideal of $\mathfrak s$. If $S$ denotes the subgroup of $GL(n,\mathbb R)$ with Lie algebra $\mathfrak s$ then we have
- $\mathbb R^{>0}\times {\operatorname{Aut}}(\mathfrak n)_0$ is a normal, proper subgroup of $S$ and
- $S$ acts transitively on $GL(n,\mathbb R)/O(n)$.
Using normality of $\mathbb R^{>0}\times{\operatorname{Aut}}(\mathfrak n)_0$ and transitivity of $S$ above, we immediately see that all the $\mathbb R^{>0}\times{\operatorname{Aut}}(\mathfrak n)_0$-orbits in $GL(n,\mathbb R)/O(n)$ are congruent; furthermore, as $\mathbb R^{>0}\times {\operatorname{Aut}}(\mathfrak n)_0$ is a proper subgroup of $S$, it does not act transitively on $GL(n)/O(n)$.
Below we detail for which types of 2-step nilpotent Lie groups the above arguments come together. In a generic sense, many algebras simultaneously
- admit a soliton metric and
- have smallest possible derivation algebra.
These two facts combined allow us to realize our counterexample.
The $j$-map and algebras of type $(p,q)$
========================================
Consider a 2-step nilpotent algebra $\mathfrak n = \mathfrak v + \mathfrak z$ where $\mathfrak z = [\mathfrak n , \mathfrak n]$. We say $\mathfrak n$ is of type $(p,q)$ if $\dim \mathfrak z = p$ and $\dim \mathfrak v = q$. Given an algebra of type $(p,q)$, with an inner product ${\langle \cdot, \cdot \rangle}$, we may consider the so-called $j$-map considered by Eberlein and others [@Eberlein:2step]: $$j:\mathfrak z \to \mathfrak{so(v)},$$ defined by $${\langle j(z)v,w \rangle} = {\langle [v,w],z \rangle}.$$ Taking an orthonormal basis $\{z_1,\dots, z_p\}$ of $\mathfrak z$, one may associate to $\mathfrak n$ a $p$-tuple of skew-symmetric matrices $$C = (C_1, \dots, C_p) \in \mathfrak{so}(q) ^p = \mathfrak{so}(q) \otimes \mathbb R^p$$ via $C_i = j(z_i) \in \mathfrak{so(v)}$. The set of 2-step nilpotent algebras of type $(p,q)$ forms a Zariski open set $V_{pq}^0$ in $\mathfrak{so}(q) \otimes \mathbb R^p$, being those $p$-tuples whose entries are linearly independent. We note that the constraint of linear independence forces us to have $1\leq p \leq \frac 1 2 q (q-1) = \dim \mathfrak{so}(q)$.
Interestingly, both the automorphism group and the ismorphism classes of 2-step nilpotent algebras of type $(p,q)$ can be read off from a natural $GL(q)\times GL(p)$ action on $\mathfrak{so}(q) \otimes \mathbb R^p$. This action is given as follows. For $(g,h)\in GL(q)\times GL(p)$ and $M\otimes v\in\mathfrak{so}(q)\otimes \mathbb R^p$, $$(g,h)\cdot M\otimes v = gMg^t\otimes hv,$$ where the $GL(p)$ action on $\mathbb R^p$ is the standard one; of course, one extends linearly. We note that there is an induced Lie algebra action of $\mathfrak{gl}(q)\oplus \mathfrak{gl}(p)$ given by $$\label{eqn: gl x gl Lie algebra action}
(X,Y) \cdot M\otimes v = (XM + MX^t)\otimes v + M\otimes Yv,$$ for $X\in\mathfrak{gl}(q)$ and $Y\in\mathfrak{gl}(p)$.
This action and its relationship to nilpotent geometry have been explored in depth by Eberlein [@EberleinModuli]. We record some useful facts here, cf. Equation \[eqn: derivations split\].
- Two algebras of type $(p,q)$ are isomorphic if and only if their corresponding $p$-tuples lie in the same $GL(q)\times GL(p)$ orbit.
- If $\mathfrak n$ corresponds to $C\in\mathfrak{so}(q)\otimes \mathbb R^p$, then $${\operatorname{Der}}(\mathfrak n)\cap (\mathfrak{gl}(q)\oplus \mathfrak{gl}(p)) \simeq (\mathfrak{gl}(q)\oplus \mathfrak{gl}(p))_C ,$$ where the right-hand side is the stabilizer of the Lie algebra action at $C$.
The isomorphism above is given by $$(X,Y)\in (\mathfrak{gl}(q)\oplus \mathfrak{gl}(p))_C \quad \mbox{ if and only if } \quad (-X^t,Y)\in{\operatorname{Der}}(\mathfrak n)\cap (\mathfrak{gl}(q)\oplus \mathfrak{gl}(p)).$$
\[definition of exception pairs (p,q)\] We say that $(p,q)$ is exceptional if either $(p,q)$ or $(\frac 1 2 q (q-1) - p, q)$ appears in the list below. Note, $\frac 1 2 q(q-1) = \dim \mathfrak{so}(q)$.
> $(1,q)$ for $q \geq 2$\
> $(\frac 1 2 q(q-1), q)$ for $q \geq 2$\
> $(2,k)$ for $k \geq 3$\
> $(3,k)$ for $4\leq k\leq 6$
As we will see below, the exceptional types $(p,q)$ are when the generic derivation algebras are larger than the minimal possible one appearing in the next theorem. Interestingly, most of the go-to examples that people work with fall into the exceptional cases above and so do not reflect the nature of generic 2-step nilpotent geometry.
\[thm: generic derivations\] For non-exceptional types, a generic algebra has as its derivation algebra the minimal possible derivation algebra, i.e. $${\operatorname{Der}}= \mathbb R(D) \oplus {\operatorname{Der}}_{\mathfrak v \to \mathfrak z},$$ where $D$ is the $(1,2)$-derivation given in Eqn. \[eqn: 1,2 derivation\]. More precisely, there exists a Zariski open set in $\mathfrak{so}(q)^p$ with the property above.
Take $C\in \mathfrak{so}(q)^p$. If $SL(q)\times SL(p) \cdot C$ is closed, then the stabilizer of the $\mathfrak{gl}(q)\oplus \mathfrak{gl}(p)$ action at $C$ is $\mathbb R (D) \oplus (\mathfrak{sl}( q)\oplus \mathfrak{sl}(p))_C$.
Conditions are needed on the orbit to ensure the result in the lemma. For example, for type $(2,2k+1)$ it is known that there is one generic orbit of the $GL(2k+1)\times GL(2)$ action and this orbit is open. As such, one can see that the stabilizer of $\mathfrak{gl}(2k+1)\oplus \mathfrak{gl}(2)$ is bigger than the stabilizer of $\mathfrak{sl} (2k+1)\oplus \mathfrak{sl}(2)$ extended by the (1,2) derivation given in Eqn. \[eqn: 1,2 derivation\].
We write $\mathfrak{gl}(q)\oplus \mathfrak{gl}(p)$ as $\mathbb R \oplus \mathbb R \oplus \mathfrak{sl}(q)\oplus \mathfrak{sl}(p)$. Each $\mathbb R$ factor acts by scaling on $C$ (cf. Eqn. \[eqn: gl x gl Lie algebra action\]). Given $X\in (\mathfrak{gl}(q)\oplus \mathfrak{gl}(p))_C$, we see that it may be written as $X=X_1 + X_2$ with $X_1$ acting by scaling and $X_2\in \mathfrak{sl}(q)\oplus \mathfrak{sl}(p)$. Thus $$X_2 \cdot C = r C,$$ for some $r\in\mathbb R$. This implies $$\exp(tX_2)\cdot C = e^{rt} C,$$ for $t\in\mathbb R$. However, since $SL(q)\times SL(p)\cdot C$ is closed it cannot contain the origin in its boundary, thence we see that $r=0$ and $X_2$ stabilizes $C$. In turn, $X_1$ must stabilize $C$ and so is a multiple of the (1,2)-derivation $D$.
The proof of the theorem above now follows immediately from the lemma combined with some general theory on certain representations of $SL(q)\times SL(p)$ which we present in the following section.
The action of $SL_p\times SL_q$
===============================
For a detailed discussion on the following facts and their relationship to nilpotent geometry, we refer the interested reader to [@Jablo:Thesis Chapter 7].
From the work of Knop-Littelman [@KnopLittelman:DerGradErzeugenderFunktionenVonInvariantenringen], we know that in the case of non-exceptional types $(p,q)$, the generic stabilizers of the the $SL(q)\times SL(p)$ representation above are finite (i.e. they have trivial connected component). From [@PopovBook], we then have that generic orbits of $SL(q)\times SL(p)$ are closed. We summarize this information below.
For non-exceptional types $(p,q)$, there is a Zariski open set in $\mathfrak{so}(q)^p$ whose elements $C$ satisfy
1. $SL(q)\times SL(p) \cdot C$ is closed, and
2. $(\mathfrak{sl}(q)\times \mathfrak{sl}(p) )_C = \{0\}$
Combining part (ii) of this theorem with the lemma above, the proof of Theorem \[thm: generic derivations\] is complete.
Application to nilsolitons
==========================
It was recognized in [@Eberlein:prescribedRicciTensor] that if the orbit $SL(q)\times SL(p) \cdot C$ is closed, then the associated 2-step nilpotent Lie group of type $(p,q)$ admits a soliton metric. Applying the above facts to nilpotent geometry, we now have the following.
\[thm: generic points being soliton and small derivations\] Consider a non-exceptional type $(p,q)$. There exists a Zariski open set of $\mathfrak{so}(q)^p$ such that the corresponding 2-step nilpotent Lie groups $N$ satisfy
1. $N$ admits a nilsoliton metric and
2. the derivation algebra of $\mathfrak n = Lie~N$ is the minimal one described in Theorem \[thm: generic derivations\].
The smallest dimension where these facts hold is dimension 9 and this is for algebras of type $(4,5)$. We now have all the necessary conditions to carry out the construction in Section \[sec: outline of proof\] for building the counterexample to the conjecture by Taketomi-Tamaru.
[GW88]{}
Patrick Eberlein, *Geometry of $2$-step nilpotent groups with a left invariant metric*, Ann. Sci. École Norm. Sup.(4) **27** (1994), no. 5, 611–660.
[to3em]{}, *The moduli space of 2-step nilpotent [L]{}ie algebras of type $(p,q)$*, Explorations in complex and Riemannian geometry, Contemporary Mathematics, American Mathematical Society **332** (2003), 37–72.
[to3em]{}, *Riemannian 2-step nilmanifolds with prescribed [R]{}icci tensor*, Geometric and probabilistic structures in dynamics, Contemp. Math., vol. 469, Amer. Math. Soc., Providence, RI, 2008, pp. 167–195.
Carolyn Gordon and Michael Jablonski, *Einstein solvmanifolds have maximal symmetry*, Journal of Differential Geometry - in press (2018).
Carolyn S. Gordon and Edward N. Wilson, *Isometry groups of [R]{}iemannian solvmanifolds*, Trans. Amer. Math. Soc. **307** (1988), no. 1, 245–269.
Takahiro Hashinaga, *On the minimality of the corresponding submanifolds to four-dimensional solvsolitons*, Hiroshima Math. J. **44** (2014), no. 2, 173–191. [MR ]{}[3251821]{}
Jens Heber, *Noncompact homogeneous [E]{}instein spaces*, Invent. Math. **133** (1998), no. 2, 279–352.
Takahiro Hashinaga and Hiroshi Tamaru, *Three-dimensional solvsolitons and the minimality of the corresponding submanifolds*, Internat. J. Math. **28** (2017), no. 6, 1750048, 31. [MR ]{}[3663797]{}
Michael Jablonski, *Real geometric invariant theory and [R]{}icci soliton metrics on two-step nilmanifolds*, Thesis (May 2008).
[to3em]{}, *Concerning the existence of [E]{}instein and [R]{}icci soliton metrics on solvable lie groups*, Geometry & Topology **15** (2011), no. 2, 735–764.
[to3em]{}, *Maximal symmetry and unimodular solvmanifolds*, to appear in Pacific Journal of Mathematics (2018).
Friedrich Knop and Peter Littelman, *Der grad erzeugender funktionen von invariantenringen. ([G]{}erman) \[the degree of generating functions of rings of invariants\]*, Math. Z. **196** (1987), no. 2, 211–229.
Jorge Lauret, *Ricci soliton homogeneous nilmanifolds*, Math. Ann. **319** (2001), no. 4, 715–733.
Y. Nikolayevsky, *Einstein solvmanifolds and the pre-[E]{}instein derivation*, Trans. Amer. Math. Soc. **363** (2011), 3935–3958.
V.L. Popov and E.B. Vinberg, *Algebraic geometry [IV]{}: [II]{}. invariant theory*, Springer-Verlag, Berlin Heidelberg, 1994.
Y. Taketomi and H. Tamaru, *On the nonexistence of left-invariant [R]{}icci solitons—a conjecture and examples*, Transform. Groups **23** (2018), no. 1, 257–270. [MR ]{}[3763948]{}
[^1]: MSC2010: 53C25, 53C30, 22E25\
This work was supported in part by NSF grant DMS-1612357.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recently, Li and Liu have studied global monoole of tachyon in a four dimensional static space-time. We analyze the motion of massless and massive particles around tachyon monopole. Interestingly, for the bending of light rays due to tachyon monopole instead of getting angle of deficit we find angle of surplus. Also we find that the tachyon monopole exerts an attractive gravitational force towards matter.'
author:
- 'M.Kalam$^{\ddag}$, F.Rahaman$^*$ and S.Mondal$^*$'
title:
- Particle Motion Around Tachyon Monopole
- '1. Introduction: '
- '2. Tachyon Monopole Revisited: '
- '3. The Geodesics: '
- '**3.1. Motion of Massless Particle ( L=0 ):** '
- '**3.2. Motion of Massive Particles ( L=1 ):** '
- '4. Bending of Light rays: '
- '5. Motion of test particle: '
- '6. Concluding remarks: '
---
At the early stages of its evolution, the Universe has underwent a number of phase transitions. During the phase transitions, the symmetry has been broken. According to the Quantum field theory, these types of symmetry-breaking phase transitions produces topological defects \[1\]. These are namely domain walls, cosmic strings, monopoles and textures. Monopoles are point like defects that may arise during phase transitions in the early universe. In particular , $ \pi_2 ( M ) \neq I $ ( M is the vacuum manifold ) i.e. M contains surfaces which can not be continuously shrunk to a point, then monopoles are formed \[2\].\
A typical symmetry - breaking model is described by the Lagrangian, $$L = \frac{1}{2}\partial_\mu \Phi^a \partial^\mu
\Phi^a - V ( f )
\label{Eq1}$$ Where $\Phi^a$ is a set of scalar fields, $ a = 1, 2, ….., N,
f=\sqrt{\Phi^a \Phi^a}$ and $V( f )$ has a minimum at a non zero value of $f$. The model has $ 0(N)$ symmetry and admits domain wall, string and monopole solutions for $ N = 1, 2 $ and $
3 $ respectively. It has been recently suggested by Cho and Vilenkin(CV) \[3,4\] that topological defects can also be formed in the models where $ V(f)$ is maximum at $f = 0$ and it decreases monotonically to zero for $ f \rightarrow \infty $ without having any minima. For example, $$\begin{aligned}
V ( f ) = \lambda M^{4+n} ( M^n + f^n )^{-1}
\end{aligned}$$ where $ M, \lambda $ and $ n $ are positive constants. This type of potential can arise in non-perturbative superstring models. Defects arising in these models are termed as “ vacuumless defects ”. Recently, several authors have studied vacuumless topological defects in alternative theory of gravity \[5\].
Barriola and Vilenkin \[6\] were the pioneer who studied the gravitational effects of global monopole. It was shown by considering only gravity that the linearly divergent mass of global monopole has an effect analogous to that of a deficit solid angle plus that of a tiny mass at the origin \[6\]. Later it was studied by Harari and Loustò \[7\], and Shi and Li \[8\] that this small gravitational potential is actually repulsive. Recently, Sen \[9\] showed in string theories that classical decay of unstable D-brane produces pressureless gas which has non-zero energy density. The basic idea is that though the usual open string vacuum is unstable, there exists a stable vacuum with zero energy density.This state is associated with the condensation of electric flux tubes of closed string \[10\]. By using an effective Born-Infeld action, these flux tubes could be explained \[11\]. Sen also proposed the tachyon rolling towards its minimum at infinity as a dark matter candidate \[10\]. Sen have also analyzed the Dirac-Born-Infeld Action on the Tachyon Kink and Vortex\[12\]. Gibbons actually initiated the study of ”tachyon cosmology”. He took the coupling into gravitational field by adding an Einstein-Hilbert term to the effective action of the tachyon on a brane \[13\]. In the cosmological background, several scientists have studied the process of rolling of the tachyon \[14, 15\].
Different kinds of cold stars such as Q-stars have been proposed to be a candidate for the cold dark matter \[16-25\]. A new class of cold stars named as D-stars(defect stars) have been proposed by Li et.al.\[26\]. Compared to Q-stars, the D-stars have a peculiar phenomena, that is, in the absence of the matter field the theory has monopole solutions, which makes the D-stars behave very differently from the Q-stars. Moreover, if the universe does not inflate and the tachyon field T rolls down from the maximum of its potential, the quantum fluctuations produced various topological defects during spontaneous symmetry breaking. That is why it is so crucial to investigate the property and the gravity of the topological defects of tachyon, such as Vortex \[27\], Kink \[28\] and monopole, in the static space time. Recently, Li and Liu \[29\] have studied gravitational field of global monopole of tachyon.
In this paper, we will discuss the behavior of the motion of massless and massive particles around Tachyon Monopole. We will calculate the amount of deficit angle for the bending of light rays. Also we will investigate the nature of gravitational field of tachyon monopole towards matters by using Hamilton-Jacobi method.\
Let us consider, a general static, spherically-symmetric metric as $$ds^2= A(r) dt^2 - B(r)dr^2 - r^2 (d\theta^2 + sin^2\theta d\phi^2 )$$
The Lagrangian density of rolling tachyon can be written in Born-Infeld form as $$\begin{aligned}
L = L_R + L_T
= \sqrt{-g}\left[\frac{R}{2\kappa}-V(|T|)\sqrt{1-g^{\mu
\nu}\partial_\mu T^a \partial_\nu T^a }\right]
\end{aligned}$$ where $ T^a $ is a triplet of tachyon fields, $ a = 1, 2, 3 $ and $ g_{\mu\nu}$ is the metric coefficients. One can consider the monopole as associated with a triplet of scalar field as $$\begin{aligned}
T^a = f ( r )\frac{x^a}{r}
\end{aligned}$$ where $ x^a x^a = r^2 $. Now using the Lagrangian density, L, the metric and the scalar field, the Einstein equations take the following forms as $$\begin{aligned}
\frac{1}{r^2}-\frac{1}{B}\left(\frac{1}{r^2}+\frac{B^\prime}{r
B}\right) = \kappa T^0_0
\end{aligned}$$ $$\begin{aligned}
\frac{1}{r^2}-\frac{1}{B}\left(\frac{1}{r^2}+\frac{A^\prime}{r
A}\right) = \kappa T^1_1
\end{aligned}$$ where the prime denotes the derivative with respect to r and energy momentum tensor $ T^\mu_\nu $ are given by $$\begin{aligned}
T^0_0 = V(f)\sqrt{1+\frac{f^{\prime 2}}{B}+\frac{2 f^2}{r^2}}
\end{aligned}$$ $$\begin{aligned}
T^1_1 = \frac{ V(f)(1+\frac{2 f^2}{r^2})}{\sqrt{1+\frac{f^{\prime 2}}{B}+\frac{2 f^2}{r^2}}}
\end{aligned}$$ $$\begin{aligned}
T^2_2 = T^3_3 = \frac{ V(f)(1+\frac{f^{\prime 2}}{B}+\frac{f^2}{r^2})}{\sqrt{1+\frac{f^{\prime 2}}{B}+\frac{2 f^2}{r^2}}}
\end{aligned}$$ and the rest are zero. So, the system depends on the tachyon potential $ V(T)$. According to Sen \[9\], the potential should have an unstable maximum at $T = 0$ and decay exponentially to zero when $T \rightarrow \infty $.
One can choose the tachyon potential which satifies the above two conditions as follows: $$\begin{aligned}
V(f) = M^4 \left(1+3\lambda f^4 \right)^{\frac{1}{6}} exp
( - \lambda f^4 )
\end{aligned}$$ where $M$ and $\lambda$ are positive constants.
In flat space-time, the Euler-Lagrange equation will take the following form: $$\begin{aligned}
\frac{1}{V}\left(\frac{dV}{df}\right) + \frac{2f}{r^2} =
f^{\prime\prime}+\frac{2f^\prime}{r}-f^\prime \left[\frac{f^\prime f^{\prime\prime}+\frac{2f}{r^2}\left(f^\prime - \frac{f}{r}\right)}{1+f^{\prime
2}+\frac{2f^2}{r^2}}\right]
\end{aligned}$$ and the energy density of the system can be written as $$\begin{aligned}
T^0_0 = V(f)\sqrt{1+f^{\prime 2}+\frac{2 f^2}{r^2}}
\end{aligned}$$ For the above mentioned tachyon potential, $ V(f)$ the Euler-Lagrange equation has a simple exact solution $$\begin{aligned}
f (r) = \lambda ^ {- \frac{1}{4} } \left(\frac {\delta}{r}\right )
\end{aligned}$$ where $\delta = \lambda ^ {- \frac{1}{4} }$ is the size of the monopole core and corresponding energy density becomes $$\begin{aligned}
T^0_0 = M^4 \left [ 1 + 3 \left ( \frac{\delta}{r}\right)^4 \right ]^{\frac{2}{3}} exp \left [-\left (\frac{\delta}{r}\right)^4 \right ]
\end{aligned}$$ Considering the Newtonian approximation, the Newtonian potential can be written as $$\begin{aligned}
\nabla^2 \Phi = \frac{\kappa}{2}(T^0_0 - T^i_i)
\end{aligned}$$ At $ r \gg \delta $, $$\begin{aligned}
T^0_0 - T^i_i \simeq -2M^4 .
\end{aligned}$$ Therefore, the solution of the above equation is $$\begin{aligned}
\Phi (r) \simeq - \frac{4 \pi M^4}{3 \lambda M_p^2 f^2 }
\end{aligned}$$ where $M_p$ is the Planck mass and the parameter M should satisfies the condition $ M \leq 10^{-3} $ eV in order to avoid conflicting present cosmological observations. The linearized approximation applies for $ |\Phi(r)| \ll 1 $, which is equivalent to $ f \gg \sqrt{\frac{4\pi}{3\lambda}}
\frac{M^2}{M_p} $.
Now, one can express the metric coefficients A(r) and B(r) as $$\begin{aligned}
A(r) = 1 + \alpha(r), B(r) = 1 + \beta(r).
\end{aligned}$$ Linearizing in $\alpha(r)$ and $\beta(r)$, and using the flat space expression for $f(r)$, the Einstein equations becomes $$\begin{aligned}
\frac{\alpha^\prime}{r} + \frac{\beta^\prime}{r} = \kappa M^4 \left(\frac{\delta}{r}\right)^4\left[1+3\left(\frac{\delta}{r}\right)^4\right]
^{-\frac{1}{3}} exp \left [-\left (\frac{\delta}{r}\right)^4 \right ]
\end{aligned}$$ and $$\begin{aligned}
\alpha^{\prime\prime}+\frac{2\alpha^\prime}{r} = -\kappa M^4 \left[2+3\left(\frac{\delta}{r}\right)^4\right]\left[1+3\left(\frac{\delta}{r}
\right)^4\right]^{-\frac{1}{3}} exp \left [-\left (\frac{\delta}{r}\right)^4 \right ]
\end{aligned}$$ After solving one can write the solution of the external metric as $$A(r) = \left(1-\frac{\kappa M^4}{3}r^2\right); B(r) = \left(1+ \frac{\kappa M^4}{3}r^2 -\frac{\kappa M^4}{2\lambda
r^2 }\right)$$
Let us now write down the equation for the geodesics in the metric (2) . From $$\frac{d^2 x^\mu}{d\tau^2} + \Gamma^\mu_{\nu\lambda}
\frac{d x^\nu}{d\tau}\frac{d x^\lambda}{d\tau}=0$$ we have $$B(r)\left(\frac{d r}{d\tau}\right)^2 = \frac{E^2}{A(r)} - \frac{J^2}{r^2} - L$$ $$r^2\left(\frac{d \phi}{d\tau}\right) = J$$ $$\frac{d t}{d\tau} = \frac{E}{A(r)}$$
where the motion is considered in the $ \theta = \frac{\pi}{2}$ plane and constants E and J are identified as the energy per unit mass and angular momentum, respectively , about an axis perpendicular to the invariant plane $ \theta = \frac{\pi}{2}$. Here $\tau$ is the affine parameter and L is the Lagrangian having values 0 and 1, respectively, for massless and massive particles.\
The equation for radial geodesic ( $ J =0$):
$$\dot{r}^2 \equiv \left(\frac{dr}{d\tau}\right)^2 = \frac{E^2}{A(r)B(r)} - \frac{L}{B(r)}$$
Using equation(7) we get $$\left(\frac{dr}{dt}\right)^2 = \frac{A(r)}{B(r)} - \frac{A^2(r)L}{E^2B(r)}$$
From equation(3), we can write $$\left(\frac{dr}{dt}\right)^2 = \left(1- \frac{\kappa M^4}{3} r^2\right)\left(1+ \frac{\kappa M^4}{3} r^2 - \frac{\kappa M^4}{2\lambda
r^2}\right)^{-1} - \frac{L}{E^2}\left(1+ \frac{\kappa M^4}{3} r^2 - \frac{\kappa M^4}{2\lambda
r^2}\right)^{-1}\left(1- \frac{\kappa M^4}{3} r^2\right)^2$$ Expanding the expression binomially and neglecting the higher order of $\kappa M^4$ ( as $\kappa M^4$ is very small ) we get $$\left(\frac{dr}{dt}\right)^2 = \left (1- \frac{2\kappa M^4}{3} r^2 + \frac{\kappa M^4}{2\lambda
r^2}\right) - \frac{L}{E^2}\left(1- \kappa M^4 r^2 + \frac{\kappa M^4}{2\lambda
r^2}\right)$$
In this case, $$\left(\frac{dr}{dt}\right)^2 = \left(1- \frac{2\kappa M^4}{3} r^2 + \frac{\kappa M^4}{2\lambda r^2}\right)$$ After integrating, we get $$\pm t = \int \frac{r dr}{\sqrt{\left(r^2- \frac{2\kappa M^4}{3} r^4 + \frac{\kappa M^4}{2\lambda
}\right)}}$$ This gives the $t - r $ relationship as $$\pm t = - \frac{1}{\sqrt{ \frac{8\kappa M^4}{3}}} sin^{-1} \left( \frac{1-\frac{4\kappa M^4}{3} r^2}{ \sqrt{1+\frac{4\kappa^2M^8}{3\lambda }}}\right)$$ The $t - r$ relationship is depicted in Fig. 1.\
![$t - r$ relationship for massless particle( choosing $\kappa M^4=573.95\times10^{-12}$, $\lambda =1$ ) []{data-label="fig:monopole"}](tm1.eps)
Again, from equation (8) we get $$\dot{r}^2 \equiv \left(\frac{dr}{d\tau}\right)^2 = \frac{E^2}{A(r)B(r)}$$
After integrating, we get $$\pm E \tau = \int \sqrt{\left(1- \frac{\kappa M^4}{3} r^2\right)\left(1+ \frac{\kappa M^4}{3} r^2 - \frac{\kappa M^4}{2\lambda
r^2}\right)}dr$$ This gives the $\tau - r $ relationship as $$\pm E\tau = \left( r + \frac{\kappa M^4}{4\lambda r}\right)$$ ( neglecting the higher order of $ \kappa M^4 $ ).
We show graphically (see Fig. 2 ) the variation of proper-time ($\tau$) with respect to radial co-ordinates (r) .\
![$\tau - r$ relationship for massless particle ( choosing $ \kappa M^4 = 573.95\times10^{-12} $, $ \lambda = 1, E = 0.5 $ )[]{data-label="fig:monopole"}](tm2.eps)
In this case, $$\left(\frac{dr}{dt}\right)^2 = \left(1- \frac{2\kappa M^4}{3}r^2+\frac{\kappa M^4}{2\lambda r^2}\right)-\frac{1}{E^2}\left(1- \kappa M^4 r^2+\frac{\kappa M^4}{2\lambda r^2}\right)$$ After integrating, we get $$\pm t = \int \frac{E r dr}{\sqrt{\left(\kappa M^4-\frac{2\kappa M^4E^2}{3}\right) r^4 +\left(E^2-1\right)r^2 +
\frac{\kappa M^4}{2\lambda}\left(E^2-1\right) }}$$ This gives the $t - r $ relationship as (see graphical Fig. (3))
$
\pm t =\frac{E/2}{\sqrt{\kappa M^4\left(1-\frac{2}{3}E^2\right)}} \ln [
2\sqrt{\left(\kappa M^4\left(1-\frac{2}{3}E^2\right)\right)\left(\kappa M^4\left(1-\frac{2}{3}E^2\right)r^4\right)+\left(E^2-1\right)r^2
+ \frac{\kappa M^4}{2\lambda}\left(E^2-1\right) }
\linebreak
+ 2\kappa M^4\left(1-\frac{2}{3}E^2\right)r^2 + \left(E^2-1\right) ]
$
![$t - r$ relationship for massive particle( choosing $ \kappa M^4 = 573.95\times10^{-12} $, $ \lambda = 1, E = 0.5 $ )[]{data-label="fig:monopole"}](tm3.eps)
Again, from equation (8) we get $$\begin{aligned}
\dot{r}^2 \equiv \left(\frac{dr}{d\tau}\right)^2 =
\frac{E^2}{A(r)B(r)} - \frac{1}{B(r)}\end{aligned}$$
Neglecting the higher order of $ \kappa M^4 $ , we get $$\begin{aligned}
\pm \int d\tau = \int \frac{\left(1-\frac{\kappa M^4}{4\lambda r^2}\right) dr }{\sqrt{E^2-1+\frac{\kappa M^4}{3}r^2}}\end{aligned}$$
This gives the $\tau - r $ relationship as $$\begin{aligned}
\pm \tau = \sqrt{\frac{3}{\kappa M^4}} \cosh^{-1}\left[\frac{r}{\sqrt{\frac{3(1-E^2)}{\kappa M^4}}}\right]
- \frac{(\kappa M^4)^{3/2}}{4\sqrt{3}\lambda (1-E^2)} \frac{\sqrt{r^2- \frac{3(1-E^2)}{\kappa M^4}}}{r}\end{aligned}$$
We show graphically (see Fig. 4 ) the variation of proper-time ($\tau$) with respect to radial co-ordinates (r) .\
![$\tau - r$ relationship for massive particle ( choosing $ \kappa M^4 = 573.95\times10^{-12} $, $ \lambda = 1, E = 0.5 $ )[]{data-label="fig:monopole"}](tm7.eps)
For photons ( L=0 ), the trajectory equations (5) and (6) yield $$\left(\frac{dU}{d\phi}\right)^2 = \frac{a^2}{A(r)B(r)} - \frac{U^2}{B(r)}$$ where $ U = \frac{1}{r}$ and $ a^2 =\frac{E^2}{J^2}$.
Equation (20) and (3) gives $$\phi = \int \frac{\pm dU}{\sqrt{\left(a^2 + \frac{\kappa M^4}{3}\right)-\left(1- \frac{a^2 \kappa M^4}{2 \lambda}\right)U^2 }}$$ ( neglecting the higher order of $ \kappa M^4 $ and the product of $ \kappa M^4 $ $\times$$U^4$ terms ).\
This gives $$\phi = \frac{1}{\sqrt{\left(1- \frac{a^2 \kappa M^4}{2 \lambda}\right)}} cos^{-1}\frac{U}{A}$$ where $ A = \frac{a^2 +
\frac{\kappa M^4}{3}}{1-\frac{a^2\kappa M^4}{2\lambda}}$.\
![We Plot U vs. $\phi$ ( choosing $\kappa M^4 = 573.95\times10^{-12}$, $\lambda = 1, a^2 = 0.5$) []{data-label="fig:monopole"}](tm6.eps)
For $ U \rightarrow 0 $, one gets $$2\phi = \pi \left(1+ \frac{a^2 \kappa M^4}{4 \lambda}\right)$$ and bending comes out as $$\Delta\phi = \pi - 2\phi = \pi- \pi\left(1+ \frac{a^2 \kappa M^4}{4\lambda}\right) = - \frac{a^2 \kappa M^4}{4\lambda} \pi$$ which is nothing but angle of surplus\[30\].
![We plot Deflection vs. Mass ( choosing $ \kappa = 25.12 $, $ \lambda = 1, a^2 = 0.5 $ ) []{data-label="fig:monopole"}](tm4.eps)
![We plot Deflection vs. E/J ( choosing $\kappa M^4=573.95\times10^{-12}$,$\lambda=1$ )[]{data-label="fig:monopole"}](tm5.eps)
Let us consider a test particle having mass $m_0$ moving in the gravitational field of the tachyon monopole described by the metric ansatz(2). So the Hamilton-Jacobi \[ H-J \] equation for the test particle is \[31\] $$g^{ik}\frac{\partial S}{\partial x^i} \frac{\partial S}{\partial x^k}+ m_0^2 = 0$$
where $ g_{ik}$ are the classical background field (2) and S is the standard Hamilton’s characteristic function .
For the metric (2) the explicit form of H-J equation (25) is \[31\] $$\frac{1}{A(r)}\left(\frac{\partial S}{\partial t}\right)^2 - \frac{1}{B(r)}\left(\frac{\partial S}{\partial
r}\right)^2- \frac{1}{r^2} \left(\frac{\partial S}{\partial \theta}\right)^2-\frac{1}{r^2\sin^2\theta}\left(\frac{\partial S}{\partial \varphi}\right)^2
+ m_0^2 = 0$$
where $A(r)$ and $B(r)$ are given in equation (3) .
In order to solve this partial differential equation, let us choose the $H-J$ function $ S $ as \[32\] $$S = - E.t + S_1(r) + S_2(\theta) + J.\varphi$$ where $E$ is identified as the energy of the particle and $J$ is the momentum of the particle.
The radial velocity of the particle is ( for detailed calculations, see $ref.[32]$ ) $$\frac{dr}{dt} = \frac{A(r)}{E\sqrt{B(r)}} \sqrt{\frac{E^2}{A(r)} +m_0^2 - \frac{p^2}{r^2} }$$ where $p$ is the separation constant.
The turning points of the trajectory are given by $\left(\frac{dr}{dt}\right) = 0 $ and as a consequence the potential curve are $$\frac{E}{m_0} = \sqrt{A(r) \left(\frac{p^2}{m_0^2r^2} - 1\right)} \equiv V(r)$$ In a stationary system, $ E $ i.e. $ V(r)$ must have an extremal value. Hence the value of $r$ for which energy attains it extremal value is given by the equation $$\frac{dV}{dr} = 0$$
Hence we get $$\frac{2\kappa M^4}{3} r^4 = \frac {2 p^2}{m^2}
\Rightarrow r = \left(\frac{3 p^2}{\kappa M^4 m^2}\right)^{\frac{1}{4}}$$ So this equation has at least one positive real root. Therefore, it is possible to have bound orbit for the test particle i.e. the test particle can be trapped by the tachyon monopole. In other words, the tachyon monopole exerts an attractive gravitational force towards matter.
In this paper, we have investigated the behavior of a massless and massive particles in the gravitational field of a tachyon monopole. The tachyon monopole, in compare to the ordinary monopole, are very diffuse objects whose energy distributed at large distances from the monopole core, their space-time is vastly different from the ordinary monopole. The figures (1) and (2) indicate that the nature of ordinary time and proper time for the massless particle in the gravitational field of tachyonic monopole is opposite to each other. Here, one can see that ordinary time decreases with increase of radial distance where as the proper time increases with increase of radial distance. Figures (3) and (4) show that in case of massive particle, the ordinary time and proper time have the same nature. According to Li and Liu \[29\], tachyon monopole has a small gravitational potential of repulsive nature, corresponding to a negative mass at origin. In the analysis of the bending of light rays, we get angle of surplus instead of angle of deficit. So, we may conclude that it has a property of short range repulsive force. From eqn.(31), we see that $ r = \frac{1}{M}\left(\frac{3 p^2}{\kappa
m^2}\right)^{\frac{1}{4}} $ i.e. $r$ would be very large as $M$ is very small, in other words, particle can be trapped at a large distance from the monopole core. This implies tachyon monopole would have effect on particles far away from its core. That means tachyon monopole has a long range gravitational field which is sharply contrast to ordinary monopole.
[ Acknowledgments ]{}
F.R. is thankful to DST , Government of India for providing financial support. MK has been partially supported by UGC, Government of India under MRP scheme.\
[99]{} T.W.B. Kibble, J. Phys. A9, 1387(1976). A. Vilenkin and E.P.S. Shellard Cosmic String and other Topological Defects (Camb. Univ. Press) (1994). I. Cho and A. Vilenkin, Phys.Rev.D59, 021701(1999). I. Cho and A. Vilenkin, Phys.Rev.D59, 063510(1999). F. Rahaman et al, arXiv: gr-qc/0610086; F. Rahaman et al, Fizika B12, 291(2003); A.A. Sen, Int.J.Mod.Phys.D10, 515(2001); L.C. Garcia de Andrade, arXiv: gr-qc/9902078; F. Rahaman et al, arXiv: gr-qc/0702147. M. Barriola and A. Vilenkin, Phys. Rev. Lett.63, 341(1989). D. Harari and C. Loustò, Phys. Rev. D42, 2626(1990).
X. Shi and X. Z. Li, Class. Quantum Grav. 8, 761(1991).
A. Sen, JHEP 9806, 007(1998); A. Sen, JHEP 9808, 010(1998); A. Sen, JHEP 9808, 012(1998).
A. Sen, JHEP 0204, 048(2002); A. Sen, JHEP 0207, 065(2002).
A. Sen J. Math. Phys. 42, 2844(2001); G. W. Gibbons, K. Hori and P. Yi, Nucl. Phys. B596, 136(2001).
A. Sen, Phys.Rev. D68, 066008(2003). G. W. Gibbons, Phys. Lett. B537, 1(2002); M. Sami and T. Padmanabhan Phys. Rev. D67, 083509(2003). X. Z. Li, J. G. Hao and D. J. Liu, Chin. Phys. Lett. 19, 1584(2002); J. G. Hao and X. Z. Li, Phys. Rev. D66, 087301(2002); J. G. Hao and X. Z. Li, Phys. Rev. D68, 067501(2003); D. J. Liu and X. Z. Li, Phys. Rev. D70, 123504(2004); J. G. Hao and X. Z. Li, Phys. Rev. D70, 043529(2004). M. Sami, P. Chingangbam, T. Qureshi, Phys.Rev. D66, 043530(2002); D. Choudhury et al, Phys.Lett. B544, 231(2002); X. Z. Li and X. H. Zhai, Phys. Rev.D67, 067501(2003); S. Mukohyama, Phys. Rev. D66, 123512(2002); J. S. Bagla, H. K. Jassal and T. Padmanabhan,Phys. Rev. D67, 063504(2003); T. Padmanabhan, Phys. Rept. 380, 235(2003); P. Singh, M. Sami, N. Dadhich, Phys. Rev. D68, 023522(2003). R. Ruffini and S. Binazzola, Phys.Rev. 187, 1767(1969). M. Colpi, S.L. Shapiro and I. Wasserman, Phys. Rev. Lett. 57, 2485(1986). P. Jetzer and J. J. Van Der Bij, Phys. Lett. B227, 341(1989). T. D. Lee, Phys. Rev. D35, 3637(1987). R. Friedberg, T.D. Lee and Y. Pang, Phys. Rev. D35, 3640(1987). R. Friedberg, T.D. Lee and Y. Pang, Phys. Rev. D35, 3658(1987).
T.D. Lee and Y. Pang, Phys. Rev. D35, 3678(1987). B.W. Lynn, Nucl. Phys. B321, 465(1989). S. Bahcall, B.W. Lynn and S.B. Selipsky, Nucl. Phys. B331, 67(1990). S. Coleman, Nucl. Phys. B262, 263(1985). X. Z. Li and J. Z. Lu, Phys. Rev. D62, 107501(2000). D. J. Liu and X. Z. Li, Chin. Phys. Lett. 20, 1678(2003). C. Kim, Y. Kim and C. O. Lee, JHEP 0305, 020(2003). X. Z. Li and D. J. Liu, Int.J.Mod.Phys.A20, 5491(2005) arXiv: gr-qc/0510116. C. Dyer et al , Phys. Rev. D13, 5588(1995);\
F.Rahaman et al , Mod. Phys. Lett. A20, 1627(2005). L. Landau and E. Lifschitz , Classical theory of fields (Pergamon Press, Oxford) 1975. S. Chakraborty, Gen. Rel. Grav. 28, 1115(1996);\
S. Chakraborty, F. Rahaman, Pramana 51, 689(1998).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Assuming that the mechanism proposed by Gell-Mann and Hartle works as a mechanism for decoherence and classicalization of the metric field, we formally derive the form of an effective theory for the gravitational field in a semiclassical regime. This effective theory takes the form of the usual semiclassical theory of gravity, based on the semiclassical Einstein equation, plus a stochastic correction which accounts for the back reaction of the lowest order matter stress-energy fluctuations.'
---
7ex
5ex
Rosario Martín and Enric Verdaguer[^1]\
[*Departament de Física Fonamental, Universitat de Barcelona, Av. Diagonal 647*]{},\
[*, Spain*]{}
. Introduction {#sec:introduction}
==============
In the semiclassical theory of gravity, the gravitational field is treated classically, but the matter fields are quantum. The key equation of the theory is the semiclassical Einstein equation, a generalization of the Einstein equation where the expectation value of the stress-energy tensor of quantum matter fields is the source of curvature.
One expects that semiclassical gravity could be derived from a fundamental quantum theory of gravity as a certain approximation, but, in the absence of such a fundamental theory, the scope and limits of the semiclassical theory are not very well understood. It seems clear, nevertheless, that it should not be valid unless gravitational fluctuations are negligibly small. This condition may break down when the matter stress-energy has appreciable quantum fluctuations, since one would expect that fluctuations in the stress-energy of matter would induce gravitational fluctuations [@ford82]. A number of examples have been recently studied, both in cosmological and flat spacetimes, where, for some states of the matter fields, the stress-energy tensor have significant fluctuations [@stress-en_fluctu]. To account for such fluctuations, it is necessary to extend the semiclassical theory of gravity.
To address this problem, or analogous problems in quantum mechanics or quantum field theory, different approaches have been adopted in the literature. The present paper attempts to unify, at least conceptually, two of these approaches in a formal derivation of an effective theory for the gravitational field in the semiclassical regime. The common feature of these two approaches is the idea of viewing the metric field as the system of interest and the matter fields as being part of its environment. This idea was first proposed by Hu [@hu89] in the context of semiclassical cosmology. Both approaches make use of the influence functional formalism, introduced by Feynman and Vernon [@feynman-vernon] to deal with a system-environment interaction in a full quantum theory. In this formalism, the integration of the environmental variables in a path integral yields the influence functional, from which one can define an effective action for the dynamics of the system .
The first of these two approaches has been extensively used in the literature, not only in the framework of semiclassical cosmology [@calzettahu; @humatacz; @husinha; @cv96; @lomb-mazz; @ccv97; @campos-hu], but also in the context of analogous semiclassical regimes for systems of quantum mechanics [@caldeira; @hu-matacz94; @hu-paz-zhang2] and of quantum field theory [@greiner; @matacz; @morikawa; @shaisultanov; @gleiser]. It makes use of the closed time path (CTP) functional technique, due to Schwinger and Keldysh [@schwinger]. This is a path integral technique designed to obtain expectation values of field operators in a direct way [@ctp]. In the semiclassical regime, a tree level approximation is performed in the path integrals involving the system variables. In this approximation, the equation of motion for the expectation value of the system field operator is the semiclassical equation, which can be directly derived from the effective action of Feynman and Vernon [@calzettahu; @greiner; @cv96; @ccv97; @campos-hu; @shaisultanov]. When computing this effective action perturbatively up to quadratic order in its variables, one usually finds some imaginary terms which do not contribute to the semiclassical equation. The key point of this approach is the formal identification of the contribution of such terms to the influence functional with the characteristic functional of a Gaussian stochastic source. Assuming that in the semiclassical regime this stochastic source interacts with the system variables, and, thus, these become stochastic variables, equations of the Langevin type are derived for these variables. However, since this approach relies on a purely formal identification, doubts can be raised on the physical meaning of the derived equations.
The second approach is based on the description of the transition from quantum to classical behavior in the framework of the consistent histories formulation of a quantum theory. The consistent histories formulation, proposed by Griffiths [@griffiths], and developed by Omnès [@omnes] and by Gell-Mann and Hartle [@gell-mann-hartle; @hartle], was designed to deal with quantum closed ([*i.e.*]{}, isolated) systems. It is thus believed to be an appropriate approach to quantum cosmology, where the quantum system is the whole universe. The main goal of this formulation is the study of the conditions under which a set of quantum mechanical variables become decoherent, which means that these variables can be described in a probabilistic way [@gell-mann-hartle; @hartle; @halliwell93; @histories; @paz-zurek]. When the closed system consists on a distinguished subsystem (the “system”, which is also often called an “open system”) interacting with its environment, Gell-Mann and Hartle proposed a mechanism for decoherence and classicalization of suitably coarse-grained system variables [@gell-mann-hartle; @hartle]. This approach allows to evaluate the probability distribution functional associated to such decoherent variables and, under some approximations, to derive effective quasiclassical equations of motion of the Langevin type for such variables [@gell-mann-hartle; @hartle; @halliwell93; @dowker; @halliwell].
In Sec. \[sec:classicalization\] we show that that these two approaches can in fact be related. In this way, we see that, on the one hand, the second approach sheds light into the physical meaning of the first one. On the other hand, the first approach provides a tool for computing effective Langevin-type equations to the second one. A large portion of this section consists of reformulating the mechanism for decoherence and classicalization of Gell-Mann and Hartle in the language of the CTP functional formalism.
In Sec. \[sec:Einstein-Langevin\], we use the results of this analysis to formally derive effective equations of motion for the gravitational field in a semiclassical regime. This derivation relies heavily on the results of the previous section. We find that, in the semiclassical regime, gravity might be described by a background metric, solution of the semiclassical Einstein equation, plus some stochastic metric perturbations. The equation for these perturbations, the semiclassical Einstein-Langevin equation, is seen to incorporate the effect of the lowest order matter stress-energy fluctuations on the gravitational field.
In this paper we use the $(+++)$ sign conventions and the abstract index notation of Ref. [@wald84], and we work in units in which $c=\hbar =1$.
. Effective equations of motion from environment-induced classicalization {#sec:classicalization}
=========================================================================
. The CTP functional formalism for a system-environment interaction {#subsec:CTP}
-------------------------------------------------------------------
We start this section by sketching the CTP functional formalism [@schwinger] applied to a system-environment interaction and its relation with the influence functional formalism of Feynman and Vernon [@feynman-vernon]. For more detailed reviews of the CTP functional formalism, see Refs. [@ctp; @campos-hu], and for the influence functional formalism of Feynman and Vernon, see Refs. . For simplicity, we shall work in this section with a model of quantum mechanics, but all the formalism can also be formally applied to field theory. It is instructive to maintain in this section the explicit dependence on $\hbar$. Let us consider a model of quantum mechanics which describe the interaction of two subsystems: one, called the “system”, with coordinates $q$, and the other, called the “environment”, with coordinates $Q$.[^2] We write the action for this model as $S[q,Q]=S_s[q]+S_{se}[q,Q]$.[^3] Let $\hat{q}(t)$ and $\hat{Q}(t)$ be the Heisenberg picture coordinate operators, which are assumed to be self-adjoint, [*i.e.*]{}, $\hat{q}^{\dag}\!=\!\hat{q}$ and $\hat{Q}^{\dag}\!=\!\hat{Q}$, and let $\hat{q}^{\rm \scriptscriptstyle S}$ and $\hat{Q}^{\rm \scriptscriptstyle S}$ be the corresponding Schrödinger picture operators. Suppose that we are only interested in describing the physical properties of system observables from some initial time $t_i$ until some final time $t_f>t_i$. Working in the Schrödinger picture, the state of the full system ([*i.e.*]{}, system plus environment) at the initial time $t\!=\!t_i$ will be described by a density operator $\hat{\rho}^{\rm \scriptscriptstyle S}(t_i)$. Let $\left\{ |q,Q\rangle^{\rm \scriptscriptstyle S} \right\}$ be the basis of eigenstates of the operators $\hat{q}^{\rm \scriptscriptstyle S}$ and $\hat{Q}^{\rm \scriptscriptstyle S}$. The matrix elements of the initial density operator in this basis will be written as $\rho(q,Q;q^{\prime},Q^{\prime};t_i)\equiv
\mbox{}^{\rm \scriptscriptstyle S}
\langle q,Q|\:
\hat{\rho}^{\rm \scriptscriptstyle S}(t_i)
\:|q^{\prime},Q^{\prime}\rangle^{\rm \scriptscriptstyle S}$. For simplicity, we shall assume that the initial density operator can be factorized as $\hat{\rho}^{\rm \scriptscriptstyle S}(t_i) \!=\!
\hat{\rho}_s^{\rm \scriptscriptstyle S}(t_i)\otimes
\hat{\rho}_e^{\rm \scriptscriptstyle S}(t_i)$, in such a way that its matrix elements in coordinate representation can be written as $\rho(q,Q;q^{\prime},Q^{\prime};t_i)\!=\!\rho_s(q,q^{\prime};t_i)\,
\rho_e(Q,Q^{\prime};t_i)$. However, the formalism can be generalized to the most general case of a non-factorizable initial density operator [@hakim; @grabert; @gell-mann-hartle]. We are interested in computing expectation values of operators related to the system variables only, for times $t$ between $t_i$ and $t_f$. The dynamics of the system in this sense can be completely characterized by the knowledge of the whole family of Green functions of the system. Working in the Heisenberg picture, these Green functions can be defined as expectation values of products of $\hat{q}(t)$ operators. These Green functions can be derived from a CTP generating functional in which only the system variables are coupled to external sources $j_+(t)$ and $j_-(t)$ [@calzettahu; @greiner; @cv96; @campos-hu; @morikawa; @shaisultanov]. This CTP generating functional can be written as the following path integral[^4] Z\[j\_+,j\_-\] = \[q\_+\]\[q\_-\] \_s (q\_[+\_[i]{}]{},q\_[-\_[i]{}]{};t\_i ) (q\_[+\_[f]{}]{}-q\_[-\_[f]{}]{}) e\^[[i ]{} (S\_[eff]{}\[q\_+,q\_-\]+ dt j\_+ q\_+ - dt j\_- q\_- )]{}, \[generating functional\] with S\_[eff]{}\[q\_+,q\_-\]S\_s\[q\_+\]-S\_s\[q\_-\]+S\_[IF]{}\[q\_+,q\_-\], \[effective action\] where $S_{\rm IF}$ is the influence action of Feynman and Vernon, which is defined in terms of the influence functional ${\cal F}_{\rm IF}$ as \_[IF]{}\[q\_+,q\_-\]&& e\^[[i ]{} S\_[IF]{}\[q\_+,q\_-\]]{}\
&& \[Q\_+\]\[Q\_-\] \_e (Q\_[+\_[i]{}]{},Q\_[-\_[i]{}]{};t\_i ) (Q\_[+\_[f]{}]{}-Q\_[-\_[f]{}]{}) e\^[[i ]{} (S\_[se]{}\[q\_+,Q\_+\]-S\_[se]{}\[q\_-,Q\_-\] )]{}. \[influence functional\] We shall call $S_{\rm eff}[q_+,q_-]$ the effective action of Feynman and Vernon. In these expressions we use the notation $q_{+_{\scriptstyle i}} \!\!\equiv\! q_+(t_i)$, $q_{+_{\scriptstyle f}} \!\!\equiv\! q_+(t_f)$, $Q_{+_{\scriptstyle i}} \!\!\equiv\! Q_+(t_i)$, $Q_{+_{\scriptstyle f}} \!\!\equiv\! Q_+(t_f)$, and similarly for $q_-$ and $Q_-$. All the integrals in $t$, including those that would define the actions $S_s[q]$ and $S_{se}[q,Q]$ in terms of the corresponding Lagrangians, have to be understood as integrals between $t_i$ and $t_f$. The CTP generating functional has the properties Z\[j,j\]=1, Z\[j\_-,j\_+\]=Z\^\[j\_+,j\_-\], | Z\[j\_+,j\_-\]|1. \[generating funct properties\] From this generating functional, we can derive the following Green functions for the system: \[(t\_1\^) (t\_s\^)\] [T]{}\[(t\_1) (t\_r)\] = . [Z\[j\_+,j\_-\] ij\_+(t\_1) ij\_+(t\_r) (-i)j\_-(t\_1\^) (-i)j\_-(t\_s\^) ]{}|\_[j\_=0]{} , \[green functions\] where $t_1,\dots ,t_r, t_1^{\prime},\dots ,t_s^{\prime}$ are all between $t_i$ and $t_f$, ${\rm T}$ and $\tilde{\rm T}$ mean, respectively, time and anti-time ordering. The expectation value is taken in the Heisenberg picture state corresponding to the Schrödinger picture state described by $\hat{\rho}^{\rm \scriptscriptstyle S}(t_i)$ at the initial time $t\!=\!t_i$. The influence functional (\[influence functional\]) can actually be interpreted as a CTP generating functional for quantum variables $Q$ coupled to classical time-dependent sources $q(t)$ through the action $S_{se}[q,Q]$ [@su]. Let us consider the quantum theory for the variables $Q$ in presence of classical sources $q(t)$ corresponding to this action, and assume that the initial Schrödinger picture state for the quantum variables $Q$ is described by the density operator $\hat{\rho}_e^{\rm \scriptscriptstyle S}(t_i)$. For this theory, let $\hat{\cal U}[q](t,t^{\prime })$ be the unitary time-evolution operator, which can be formally written as $\hat{\cal U}[q](t,t^{\prime })\!=\! {\rm T} \exp \!
\left[-{i\over \hbar} \int_{t^{\prime }}^{t} dt^{\prime\prime }
\hat{H}^{\rm \scriptscriptstyle S}[q](t^{\prime\prime})\right]$, for $t\!>\!t^{\prime }$, where $\hat{H}^{\rm \scriptscriptstyle S}[q](t)$ is the Hamiltonian operator in the Schrödinger picture. This Hamiltonian operator depends on $t$ as a function of $q(t)$ and their derivatives $\dot{q}(t)$, and this gives a functional dependence on $q$ in the operator $\hat{\cal U}$. It is easy to see that [@gell-mann-hartle; @hartle; @humatacz; @hu-matacz94; @greiner; @hakim] \_[IF]{}\[q\_+,q\_-\]= [Tr]{} = \^\[q\_-\](t\_[f]{},t\_[i]{}) \[q\_+\](t\_[f]{},t\_[i]{}) \_[\_e\^[S]{}(t\_i)]{}, \[influence funct representation\] where we use $\langle \hspace{1.5ex} \rangle_{\!\hat{\rho}_e^{\rm S}(t_i)}$ to denote an expectation value in the state described by $\hat{\rho}_e^{\rm \scriptscriptstyle S}(t_i)$. From this expression, it follows that the influence functional satisfies \_[IF]{}\[q,q\]=1, \_[IF]{}\[q\_-,q\_+\]= [F]{}\_[IF]{}\^\[q\_+,q\_-\], |\_[IF]{}\[q\_+,q\_-\]|1, \[influence funct properties\] or, equivalently, in terms of the influence action, S\_[IF]{}\[q,q\]=0, S\_[IF]{}\[q\_-,q\_+\]= -S\_[IF]{}\^\[q\_+,q\_-\], S\_[IF]{}\[q\_+,q\_-\] 0, \[influence action properties\] and similar properties follow for $S_{\rm eff}[q_+,q_-]$. A decoherence functional for the system, where the environment variables have been completely integrated out, can now be introduced as the functional Fourier transform of the CTP generating functional in the external sources: Z\[j\_+,j\_-\] \[q\_+\]\[q\_-\]\[q\_+,q\_-\] e\^[i dt (j\_+ q\_+ - j\_- q\_- )]{}, \[decoherence functional\] that is, from (\[generating functional\]) we have that \[q\_+,q\_-\]= \_s (q\_[+\_[i]{}]{},q\_[-\_[i]{}]{};t\_i ) (q\_[+\_[f]{}]{}-q\_[-\_[f]{}]{}) e\^[[i ]{} S\_[eff]{}\[q\_+,q\_-\]]{}. \[decoherence functional 2\] In the consistent histories approach to quantum mechanics, ${\cal D}[q_+,q_-]$ is known as the decoherence functional for fine-grained histories of the system [@gell-mann-hartle; @hartle; @halliwell93; @histories; @dowker; @halliwell].
The environment of a system has to be understood as characterized by all the quantum degrees of freedom which can affect the dynamics of the system, but which are “not accessible” in the observations of that system. This environment includes in general an “external” environment (variables representing other particles, or, in the context of field theory, other fields) and an “internal” environment (some degrees of freedom which, from the fundamental quantum theory point of view, would be associated to the same physical object as the “system” variables, but which are not directly probed in our observations of the system) [@zurek; @omnes]. For instance, a problem which has been studied using the influence functional method is that of quantum Brownian motion . In this problem, one is interested in the dynamics of a macroscopic particle interacting with a medium composed by a large number of other particles. In this example, one considers that the only “observable” system degree of freedom is the center of mass position of the macroscopic particle, whereas the remaining microscopic degrees of freedom of the macroscopic particle are considered as environmental variables. Such “internal” environment degrees of freedom, and also those of the particles of the medium (the “external” environment), are usually modelized as an infinite set of harmonic oscillators. In the context of field theory, one would typically consider as “inaccessible” to the observations the modes of the field of interest with characteristic momenta higher than some cut-off momentum [@lombardo; @greiner; @matacz]. In the case of the gravitational field, this has been considered by Whelan [@whelan] in a toy model designed to investigate the decoherence mechanism for gravity.
It is convenient at this stage to distinguish between these two kinds of environmental variables, so let $Q$ represent the coordinates of the “external” environment (the coordinates of “other particles”) and $q_{\mbox{}_{\rm U}}$ the “unobservable system” coordinates (the coordinates of the “internal” environment). As before, $q$ will represent the “true” system coordinates. One could now simply replace $Q$ by $(Q,q_{\mbox{}_{\rm U}})$ in the previous expressions. However, for convenience, we shall do the integrations in the environmental variables in two steps. The action of the full system will be now written as $S[q,q_{\mbox{}_{\rm U}},Q]$, and, as before, we shall assume a totally factorizable initial density operator $\hat{\rho}^{\rm \scriptscriptstyle S}(t_i)=
\hat{\rho}_s^{\rm \scriptscriptstyle S}(t_i)\otimes
\hat{\rho}_{\mbox{}_{\rm U}}^{\rm \scriptscriptstyle S}(t_i)
\otimes \hat{\rho}_e^{\rm \scriptscriptstyle S}(t_i)$, which leads to an initial density matrix in coordinate representation of the form $\rho(q,q_{\mbox{}_{\rm U}},Q;
q^{\prime},q_{\mbox{}_{\rm U}}^{\prime},Q^{\prime};t_i)=
\rho_s(q,q^{\prime};t_i)\,
\rho_{\mbox{}_{\rm U}}(q_{\mbox{}_{\rm U}},
q_{\mbox{}_{\rm U}}^{\prime};t_i)\,
\rho_e(Q,Q^{\prime};t_i)$ (notice that we are now using the subindex $e$ for the “external” environment). Such a factorization is based on the assumption that the interactions between the three subsystems can be neglected for times $t \leq t_{i}$. Unfortunately, in most situations, this assumption does not seem to be very physically reasonable, especially for the “true” system-“internal” environment interactions. One would need to consider the generalization of the formalism to a non-factorizable initial density operator mentioned above and the analysis would be more complicated. We start defining && e\^[[i ]{}( S\_[s]{}\^[eff]{}\[q\_+\]-S\_[s]{}\^[eff]{}\[q\_-\] +S\_[se]{}\^[eff]{}\[q\_+,Q\_+;q\_-,Q\_-\] )]{}\
&& \[q\_[\_[U]{}+]{}\]\[q\_[\_[U]{}-]{}\] \_[\_[U]{}]{} (q\_[[\_[U]{}+]{}\_[i]{}]{}, q\_[[\_[U]{}-]{}\_[i]{}]{};t\_i ) (q\_[[\_[U]{}+]{}\_[f]{}]{}- q\_[[\_[U]{}-]{}\_[f]{}]{}) e\^[[i ]{} (S\[q\_+,q\_[\_[U]{}+]{},Q\_+\] -S\[q\_-,q\_[\_[U]{}-]{},Q\_-\] )]{}, \[s-e effective actions\] where the effective action for the system $S_{s}^{\rm eff}[q]$ is chosen to be real and local. Notice that the effective action $S_{se}^{\rm eff}[q_+,Q_+;q_-,Q_-]$ has analogous properties to those of $S_{\rm IF}$ in (\[influence action properties\]). We introduce now an effective influence functional and an effective influence action as \^[eff]{}\_[IF]{}\[q\_+,q\_-\] e\^[[i ]{} S\^[eff]{}\_[IF]{}\[q\_+,q\_-\]]{} \[Q\_+\]\[Q\_-\] \_e (Q\_[+\_[i]{}]{},Q\_[-\_[i]{}]{};t\_i ) (Q\_[+\_[f]{}]{}-Q\_[-\_[f]{}]{}) e\^[[i ]{} S\_[se]{}\^[eff]{}\[q\_+,Q\_+;q\_-,Q\_-\]]{}. \[effective influence functional\] With these definitions, the effective action of Feynman and Vernon, $S_{\rm eff}[q_+,q_-]$, which appears in expression (\[generating functional\]) can be written as S\_[eff]{}\[q\_+,q\_-\]S\_s\^[eff]{}\[q\_+\]-S\_s\^[eff]{}\[q\_-\] +S\^[eff]{}\_[IF]{}\[q\_+,q\_-\]. \[effective action 2\] Note that, since $S_{\rm eff}[q_+,q_-]$ satisfies the same properties as $S_{\rm IF}$ in (\[influence action properties\]), it follows from the last expression that $S^{\rm eff}_{\rm IF}$ has also these properties.
. The “naive” semiclassical approximation {#subsec:naive semiclassical}
-----------------------------------------
The usual “naive” semiclassical approximation for the system variables consists in performing a “tree level” approximation in the path integrals involving the $q$ variables in expression (\[generating functional\]) [@calzettahu; @greiner; @cv96; @ccv97; @campos-hu; @shaisultanov]. Therefore, the CTP generating functional is approximated by Z\[j\_+,j\_-\] e\^[[i ]{} (S\_[eff]{}+ dt j\_[\_+]{} |[q]{}\_+\^[(0)]{} \[j\]- dt j\_[\_-]{} |[q]{}\_-\^[(0)]{} \[j\] )]{}, \[semiclass approx\] where $\bar{q}_{\pm}^{\scriptscriptstyle (0)}\hspace{-0.2ex}[j]
\!\equiv\!
\bar{q}_{\pm}^{\scriptscriptstyle (0)}\hspace{-0.2ex}[j_+,j_-]$ are solutions of the classical equations of motion for the action $S_{\rm eff}[q_+,q_-]+
\hbar \int\! dt\, j_+ q_+ -\hbar \int\! dt\, j_- q_-$, that is, = j\_(t), \[semiclass eqs with j’s\] which satisfy the boundary condition $\bar{q}_+^{\scriptscriptstyle (0)}(t_f)
\!=\!\bar{q}_-^{\scriptscriptstyle (0)}(t_f)$. Whenever this approximation is valid, we can see from (\[semiclass approx\]), (\[semiclass eqs with j’s\]) and (\[green functions\]) that $\langle \hat{q}(t) \rangle \simeq q^{\scriptscriptstyle (0)}(t)$, with $q^{\scriptscriptstyle (0)} \equiv
\bar{q}_{+}^{\scriptscriptstyle (0)}\hspace{-0.2ex}
[j_+\!=\!j_-\!=\!0]=
\bar{q}_{-}^{\scriptscriptstyle (0)}\hspace{-0.2ex}
[j_+\!=\!j_-\!=\!0]$, that is, $q^{\scriptscriptstyle (0)}(t)$ is a solution of the two equivalent equations: . [S\_[eff]{}\[q\_+,q\_-\] q\_+(t)]{} |\_[q\_+=q\_-=q\^[(0)]{}]{} =0, . [S\_[eff]{}\[q\_+,q\_-\] q\_-(t)]{} |\_[q\_+=q\_-=q\^[(0)]{}]{} =0. \[semiclassical eq\] One can see that these two equations are actually the same equation, and that this equation is real. This is the semiclassical equation for the system variables. In a naive way, one would think that, when the above semiclassical approximation is valid, the system would behave as a classical system described by the coordinate functions $q^{\scriptscriptstyle (0)}(t)$, [*i.e.*]{}, that one could substitute the description of the system in terms of the operators $\hat{q}(t)$ by a classical description in terms of the functions $q^{\scriptscriptstyle (0)}(t)$. However, one can see from (\[semiclass approx\]), (\[semiclass eqs with j’s\]) and (\[green functions\]) that, in general, \[(t\_1\^) (t\_s\^)\] [T]{}\[(t\_1) (t\_r)\] / q\^[(0)]{}(t\_1) q\^[(0)]{}(t\_r) q\^[(0)]{}(t\_1\^) q\^[(0)]{}(t\_s\^). Thus, in general, whenever the above approximations are valid, we can only interpret the solutions of the semiclassical equation as representing the expectation value of the operators $\hat{q}(t)$.
. Further coarse-graining and decoherence {#subsec:decoherence}
-----------------------------------------
Decoherence takes place in a set of quantum-mechanical variables when the quantum interference effects are (in general, approximately) suppressed in the description of the properties of a physical system which are associated to that variables. When this happens, such decoherent variables can be described in an effective probabilistic way. In the Heisenberg picture, we will say that a set of variables decohere when the description in terms of the operators corresponding to these variables can be replaced by an effective description in terms of a set of classical random variables, in the sense that the quantum Green functions for such operators become approximately equal to the moments of the classical random variables. For the Green functions (\[green functions\]), it is easy to see that this would hold in an exact way if the CTP generating functional (\[generating functional\]) depended on the sources $j_{\pm}$ only as a functional $\Phi_q[j_+\!-\!j_-]$ of the difference $j_+ - j_-$, or, equivalently, if the decoherence functional (\[decoherence functional\]) could be written as ${\cal D}[q_+,q_-]= {\cal P}_q[q_+] \, \delta[q_+-q_-]$. However, in practice, one finds that this condition is usually too strong to be satisfied, even in an approximate way . One needs to introduce further coarse-graining in the system degrees of freedom in order to achieve decoherence. Let us then introduce coarse-grained system operators, which correspond to imprecisely specified values of the system coordinates. In the Heisenberg picture, such operators can be defined as \_c(t) \_[|[q]{}]{} |[q]{} \_[|[q]{}]{}(t), where $\hat{P}_{\bar{q}}(t)$ is a set of projection operators, labeled by some variables $\bar{q}$ (these are often discrete variables), of the form \_[|[q]{}]{}(t) = dq dq\_[\_[U]{}]{} dQ (q-|[q]{}) |q,q\_[\_[U]{}]{} ,Q,t q,q\_[\_[U]{}]{} ,Q,t|. Here $\left\{ |q,q_{\mbox{}_{\rm U}} ,Q,t \rangle \right\}$ is the basis of eigenstates of the operators $\hat{q}(t)$, $\hat{ q}_{\mbox{}_{\rm U}} (t)$ and $\hat{Q}(t)$, and $\gamma$ is a real function. We shall assume coarse-grainings of characteristic sizes $\sigma$, that is, such that the function $\gamma(q-\bar{q})$ vanishes or has negligible values for $q$ outside a cell $I_{\bar{q}}$ of sizes $\sigma$ centered around $\bar{q}$. This means that dq (q-|[q]{}) f(q) \_[I\_[|[q]{}]{}]{} dq (q-|[q]{}) f(q), \[c-g characteristic sizes\] for any function $f(q)$. In addition, the function $\gamma$ must be chosen in such a way that the set of projection operators is (at least, approximately) exhaustive and mutually exclusive, which means that \_[|[q]{}]{} \_[|[q]{}]{}(t)= , \_[|[q]{}]{}(t) \_[|[q]{}\^]{}(t)= \_[|[q]{} |[q]{}\^]{} \_[|[q]{}]{}(t), \[proj properties\] where $\hat{I}$ is the identity operator. For specific examples of operators satisfying the above properties in an exact or in an approximate way, see Refs. [@dowker; @halliwell].
Next, we can introduce a family of decoherence functions for coarse-grained histories of the system . In order to do so, let us consider a set $\{t_1, \dots , t_N \}$ of $N$ instants of time, such that $t_k < t_{k+1}$, $k = 0, \dots , N$, with $t_0 \equiv t_i$ and $t_{N+1} \equiv t_f$. Introducing two sets of values of $\bar{q}$ associated to such set of instants, $\{ \bar{q}_+ \} \equiv \{ \bar{q}_{+_1}, \dots , \bar{q}_{+_N}\}$ and $\{ \bar{q}_- \} \equiv \{ \bar{q}_{-_1}, \dots , \bar{q}_{-_N}\}$, the decoherence function for this pair of “coarse-grained histories” of the system is defined as \_c( { |[q]{}\_+ },{ |[q]{}\_-} )\_[(t\_1, …, t\_N)]{} , \[c-g decoh funct\] where $\hat{\rho}$ is the density operator describing the state of the entire system (system plus environment) in the Heisenberg picture (${\cal D}_c$ is often called decoherence “functional” in the literature, but, for each set $\{t_1, \dots , t_N \}$, this is actually a function of $2N$ variables). These decoherence functions can be written in a path integral form as \_c( { |[q]{}\_+ },{ |[q]{}\_-} )\_[(t\_1, …, t\_N)]{} = \[q\_+\][D]{}\[q\_-\] \_[k=1]{}\^N (q\_+(t\_k) - |[q]{}\_[+\_k]{}) (q\_-(t\_k) - |[q]{}\_[-\_k]{})\[q\_+,q\_-\], \[c-g decoh funct 2\] where ${\cal D}[q_+,q_-]$ is the decoherence functional for fine-grained histories of the system (\[decoherence functional\]). From the definition (\[c-g decoh funct\]) and the properties (\[proj properties\]), one can show that these decoherence functions have the properties \_[{ |[q]{}\_+ }]{} \_[{ |[q]{}\_- }]{} [D]{}\_c( { |[q]{}\_+ },{ |[q]{}\_-} ) = 1, \_c( { |[q]{}\_- },{ |[q]{}\_+} ) = [D]{}\_c\^( { |[q]{}\_+ },{ |[q]{}\_-} ), \[decoh funct properties\] and that the diagonal elements of the decoherence functions (the values of those functions in the limit $\bar{q}_{-_k}\!\rightarrow \! \bar{q}_{+_k}$) are positive. For $N \!>\! 1$, we can also see that, if we divide the set $\{t_1, \dots , t_N \}$ into a subset of $M \!<\! N$ instants, $\{t_1^{\prime}, \dots , t_M^{\prime} \}
\!\subset\! \{t_1, \dots , t_N \}$, with $t_1^{\prime} < \cdots <t_M^{\prime}$, and the subset of the remaining $L \!\equiv\! M\!-\!N$ instants, denoted as $\{t_1^{\prime \prime}, \dots , t_L^{\prime \prime} \}$ \[[*i.e.*]{}, $\{t_1, \dots , t_N \}\!=\! \{t_1^{\prime}, \dots , t_M^{\prime} \}
\cup \{t_1^{\prime \prime}, \dots , t_L^{\prime \prime} \}$\], then \_c( { |[q]{}\_+ }\_[\_[M]{}]{}, { |[q]{}\_-}\_[\_[M]{}]{} )\_[(t\_1\^, …, t\_M\^)]{} = \_[ { |[q]{}\_+ }\_[\_[L]{}]{} ]{} \_[{ |[q]{}\_- }\_[\_[L]{}]{} ]{} [D]{}\_c( { |[q]{}\_+ }\_[\_[N]{}]{}, { |[q]{}\_-}\_[\_[N]{}]{} )\_[(t\_1, …, t\_N)]{}, \[decoh funct prop 4\] with $\{ \bar{q}_{\pm} \}_{\!\mbox{}_{M}} \!\equiv\!
\{ \bar{q}_{\pm}(t_1^{\prime}), \dots,
\bar{q}_{\pm}(t_M^{\prime}) \}$, $\{ \bar{q}_{\pm} \}_{\!\mbox{}_{L}} \!\equiv\!
\{ \bar{q}_{\pm}(t_1^{\prime\prime}), \dots,
\bar{q}_{\pm}(t_L^{\prime\prime}) \}$, where we use the notation $\bar{q}_{\pm}(t_k) \!\equiv\! \bar{q}_{\pm_k}$, for $k \!=\! 1, \dots, N$, and $\{ \bar{q}_{\pm} \}_{\!\mbox{}_{N}} \!\equiv\!
\{ \bar{q}_{{\pm}_1}, \dots, \bar{q}_{{\pm}_N} \}$.
To make contact with the CTP formalism, let us introduce now, in analogy with (\[decoherence functional\]), a family of generating functions for the coarse-grained system degrees of freedom as the following Fourier series: Z\_c( { j\_+ },{ j\_- } )\_[(t\_1, …, t\_N)]{} \_[ { |[q]{}\_+ }]{} \_[ { |[q]{}\_- }]{} [D]{}\_c( { |[q]{}\_+ },{ |[q]{}\_-} )\_[(t\_1, …, t\_N)]{} e\^[i \_[k=1]{}\^[N]{} (j\_[+\_k]{} |[q]{}\_[+\_k]{} - j\_[-\_k]{} |[q]{}\_ [-\_k]{})]{} , \[c-g generating funct\] where $\{ j_{\pm} \} \equiv \{ j_{\pm_1}, \dots , j_{\pm_N}\}$. Note that the properties (\[decoh funct properties\]) for the decoherence functions are equivalent to Z\_c( {0 },{ 0 } )=1, Z\_c( { j\_- },{ j\_+ } ) =Z\_c\^( { j\_+ },{ j\_- } ). \[c-g generating funct properties\] From the generating function (\[c-g generating funct\]), we can compute the Green functions $$G_{c \; m_1 \cdots \, m_s}^{n_1 \cdots \, n_r}
(t_1^{\prime}, \dots, t_r^{\prime};
t_1^{\prime \prime}, \dots, t_s^{\prime \prime})
\equiv
\left\langle\, \tilde{\rm T}[\hat{q}_c^{m_1}(t_1^{\prime \prime})
\cdots
\hat{q}_c^{m_s}(t_s^{\prime \prime})] \,
{\rm T}[\hat{q}_c^{n_1}(t_1^{\prime}) \cdots
\hat{q}_c^{n_r}(t_r^{\prime})]\, \right\rangle,$$ with $n_1, \dots, n_r, m_1, \dots, m_s \!\in \!
{\rm I\hspace{-0.4 ex}N}$, $\{t_1^{\prime}, \dots , t_r^{\prime} \} \!\subseteq \!
\{t_1, \dots , t_N \}$ and $\{t_1^{\prime \prime}, \dots , t_s^{\prime \prime} \} \!\subseteq \!
\{t_1, \dots , t_N \}$ (thus, $r,s \leq N$): G\_[c m\_1 m\_s]{}\^[n\_1 n\_r]{} (t\_1\^, …, t\_r\^; t\_1\^, …, t\_s\^)= . [ (-i ) \^[n\_1+ +n\_r+ m\_1+ +m\_s]{} Z\_c( { j\_+ },{ j\_- } )\_[(t\_1, …, t\_N)]{} \^[ n\_1]{} \^[ n\_r]{} \^[ m\_1]{} \^[ m\_s]{} ]{} |\_[{j\_}={0} ]{} , \[c-g green funct\] where $j_\pm (t_k) \!\equiv\! j_{\pm_k}$, for $k \!=\! 1, \dots, N$. The property (\[decoh funct prop 4\]) can also be written in terms of the corresponding generating functions as Z\_c({ j\_+ }\_[\_[M]{}]{}, { j\_-}\_[\_[M]{}]{} )\_[(t\_1\^, …, t\_M\^)]{} =. Z\_c( { j\_+ }\_[\_[N]{}]{}, { j\_- }\_[\_[N]{}]{} )\_[(t\_1, …, t\_N)]{} |\_[{j\_}\_[\_[L]{}]{}={0} ]{}, \[c-g generating funct prop 4\] with the notation $\{ j_{\pm} \}_{\!\mbox{}_{M}} \!\equiv\!
\{ j_{\pm}(t_1^{\prime}), \dots,
j_{\pm}(t_M^{\prime}) \}$, and similarly for $\{ j_{\pm} \}_{\!\mbox{}_{L}}$ and $\{ j_{\pm} \}_{\!\mbox{}_{N}}$. Notice that this last property is consistent with (\[c-g green funct\]), in the sense that, for instance, $G_c^{n_1 n_2}(t_1^{\prime},t_2^{\prime})$ can be equally computed either from $Z_c(\{ j_+ \}_{\mbox{}_{2}},
\{ j_-\}_{\mbox{}_{2}} )_{(t_1^{\prime},t_2^{\prime})}$, or from $Z_c( \{ j_+ \}_{\!\mbox{}_{N}},
\{ j_- \}_{\!\mbox{}_{N}} )_{(t_1, \dots , t_N)}$, with $N>2$.
Having introduced the coarse-grained description of the system in terms of the operators $\hat{q}_c(t)$, we can now sketch the decoherence mechanism for them. For the Green functions (\[c-g green funct\]), one can show that the decoherence condition described above holds in an exact way if the generating function (\[c-g generating funct\]) depends on the sources $j_{\pm_k}$ only as a function of the differences $j_{+_k}\!-\!j_{-_k}$, [*i.e.*]{}, as $\Phi_{\bar{q}}( \{ j_+\!-\!j_- \} )_{(t_1, \dots , t_N)}$. Then, introducing the Fourier series corresponding to $\Phi_{\bar{q}}$, we can write Z\_c( { j\_+ },{ j\_- })\_[(t\_1, …, t\_N)]{}= \_[|[q]{}]{}( { j\_+ - j\_- } ) \_[(t\_1, …, t\_N)]{} \_[{ |[q]{}} ]{} [P]{}\_[|[q]{}]{}( { |[q]{} } )\_[(t\_1, …, t\_N)]{} e\^[i \_[k=1]{}\^[N]{} |[q]{}\_k ( j\_[+\_k]{} - j\_[-\_k]{}) ]{}. \[decoherence condition\] Note from the last expression that, if we interpret the function ${\cal P}_{\bar{q}}$ as the probability distribution for a set of random variables $\bar{q}_k$, $k \!=\! 1, \dots, N$, associated to the instants $t_k$, then $\Phi_{\bar{q}}$ is the corresponding characteristic function. Therefore, from (\[c-g green funct\]), we get && G\_[c m\_1 m\_s]{}\^[n\_1 n\_r]{} (t\_1\^, …, t\_r\^; t\_1\^, …, t\_s\^)= . [ (-i ) \^[n\_1+ +n\_r+ m\_1+ +m\_s]{} \_[|[q]{}]{}( { j } )\_[(t\_1, …, t\_N)]{} \^[n\_1]{} \^[n\_r]{} \^[m\_1]{} \^[m\_s]{} ]{} |\_[{j }={0} ]{}\
&& = \_[{ |[q]{}} ]{} \_[|[q]{}]{}( { |[q]{} } ) \_[(t\_1, …, t\_N)]{} |[q]{}\^[n\_1 ]{}(t\_1\^) |[q]{}\^[n\_r ]{}(t\_r\^) |[q]{}\^[m\_1 ]{}(t\_1\^) |[q]{}\^[m\_s ]{}(t\_s\^) |[q]{}\^[n\_1 ]{}(t\_1\^) |[q]{}\^[n\_r ]{}(t\_r\^) |[q]{}\^[m\_1 ]{}(t\_1\^) |[q]{}\^[m\_s ]{}(t\_s\^) \_[ c]{} ,\
\[correlation functions\] where $\langle \hspace{1.5ex} \rangle_c$ means statistical average of the random variables, and we use the notation $\bar{q}(t_k) \!\equiv\! \bar{q}_k$, $j(t_k) \!\equiv\! j_k$, for $k \!=\! 1, \dots, N$. Note that, if (\[decoherence condition\]) is satisfied, then the property (\[c-g generating funct prop 4\]) reduces to \_[|[q]{}]{}( { j }\_[\_[M]{}]{} ) \_[(t\_1\^, …, t\_M\^)]{} = . \_[|[q]{}]{}( { j }\_[\_[N]{}]{} )\_[(t\_1, …, t\_N)]{} |\_[{ j }\_[\_[L]{}]{}={0} ]{}, \[prop 4\] or, equivalently, \_[|[q]{}]{}( { |[q]{} }\_[\_[M]{}]{} ) \_[(t\_1\^, …, t\_M\^)]{} = \_[ { |[q]{} }\_[\_[L]{}]{}]{} [P]{}\_[|[q]{}]{}( { |[q]{} }\_[\_[N]{}]{} ) \_[(t\_1, …, t\_N)]{}. \[prop 4 bis\] This last property is a necessary condition for the probabilistic interpretation (\[correlation functions\]) to be consistent.
The conditions for decoherence (\[decoherence condition\]) can be written in terms of the corresponding decoherence functions as \_c( { |[q]{}\_+ },{ |[q]{}\_-} )\_[(t\_1, …, t\_N)]{} =[P]{}\_[|[q]{}]{}( { |[q]{}\_+ } )\_[(t\_1, …, t\_N)]{} \_[k=1]{}\^N \_[|[q]{}\_[+\_k]{} |[q]{}\_[-\_k]{}]{}. \[decoherence condition 2\] These are actually the conditions for decoherence of coarse-grained system variables as stated in the consistent histories formulation of quantum mechanics . Notice, from (\[proj properties\]), that (\[decoherence condition 2\]) is always satisfied for a single instant of time ([*i.e.*]{}, when $N \!=\! 1$) [@dowker].
We can now check that the interpretation of ${\cal P}_{\bar{q}}$ as a probability function is actually correct. From the second of the properties (\[decoh funct properties\]), we have that ${\cal P}_{\bar{q}}^{\displaystyle \ast}( \{ \bar{q} \} )
={\cal P}_{\bar{q}}( \{ \bar{q} \} )$, [*i.e.*]{}, ${\cal P}_{\bar{q}}$ is real. Since the diagonal elements of the decoherence functions are positive, ${\cal P}_{\bar{q}}( \{ \bar{q} \} )$ is also positive. These two properties of ${\cal P}_{\bar{q}}( \{ \bar{q} \} )_{(t_1, \dots , t_N)}$, together with (\[prop 4 bis\]), are enough to guarantee that it can be properly interpreted as the probability distribution for a set of random variables associated to the instants $t_1, \dots , t_N$. From the first of the relations (\[decoh funct properties\]), which yields $\sum_{\{ \bar{q}\} } {\cal P}_{\bar{q}}( \{ \bar{q} \} )=1$, it follows that this probability distribution is normalized.
In practice, the conditions for decoherence described above will be usually only satisfied in an approximate way. Approximate decoherence is typically achieved through a mechanism which was proposed by Gell-Mann and Hartle [@gell-mann-hartle; @hartle]. To see how this works, note that, if we assume coarse-grainings of characteristic sizes $\sigma$ \[see (\[c-g characteristic sizes\])\], and using (\[decoherence functional 2\]), we can write the decoherence function (\[c-g decoh funct 2\]) as &&\_c( { |[q]{}\_+ },{ |[q]{}\_-} )\_[(t\_1, …, t\_N)]{} \_[ \_[ { I\_[|[q]{}\_[+]{}]{} }, { I\_[|[q]{}\_[-]{}]{} } ]{}]{} \[q\_+\^[(0)]{}\][D]{}\[q\_-\^[(0)]{}\] \_[k=1]{}\^N [D]{}\[q\_+\^[(k)]{}\][D]{}\[q\_-\^[(k)]{}\] \_s (q\^[(0)]{}\_[+\_[i]{}]{},q\^[(0)]{}\_[-\_[i]{}]{} ;t\_i ) (q\^[(N)]{}\_[+\_[f]{}]{}-q\^[(N)]{}\_[-\_[f]{}]{} )\
&& (q\^[(k-1)]{}\_+(t\_k) - q\^[(k)]{}\_+(t\_k)) (q\^[(k-1)]{}\_-(t\_k) - q\^[(k)]{}\_-(t\_k)) (q\^[(k)]{}\_+(t\_k) - |[q]{}\_[+\_k]{}) (q\^[(k)]{}\_-(t\_k) - |[q]{}\_[-\_k]{}) \_[k=0]{}\^N e\^[[i ]{} S\_[eff]{}\[q\_+\^[(k)]{} ,q\_-\^[(k)]{}\]]{},\
\[c-g decoh funct 3\] where each path integration $\int {\cal D}[q_{\pm}^{(k)}]$, for $k=0, \dots, N$, is over paths $q_{\pm}^{(k)}(t)$ with $t \in [t_k,t_{k+1}]$, being $t_0 \equiv t_i$ and $t_{N+1} \equiv t_f$, and we have used a notation to indicate that these paths are restricted to pass through the cells $I_{\bar{q}_{\pm _k}}$ at the instants $t_k$, for $k=1, \dots, N$. From (\[effective action 2\]), the modulus of each factor $\exp \bigr(
{{i \over \hbar}\hspace{0.2ex} S_{\rm eff}[q_+^{\hspace{-0.1ex}(k)}
,q_-^{\hspace{-0.1ex}(k)}]} \bigl)$ in the last expression is $\exp \bigr(
{-{1 \over \hbar} \hspace{0.2ex} {\rm Im}\, S_{\rm IF}^{\rm eff}
[q_+^{\hspace{-0.1ex}(k)} ,q_-^{\hspace{-0.1ex}(k)}]} \bigl)$. Then, if for every $k=0, \dots, N$, ${\rm Im}\, S_{\rm IF}^{\rm eff}
[q_+^{\hspace{-0.1ex}(k)} ,q_-^{\hspace{-0.1ex}(k)}]$, which is always positive or zero, is much larger than $\hbar$ whenever the differences $|q_+^{\hspace{-0.1ex}(k)}\!-q_-^{\hspace{-0.1ex}(k)}|$ are larger than some “cut-off” sizes $d^{(k)}$, the integrand in (\[c-g decoh funct 3\]) will be only non-negligible for $|q_+^{\hspace{-0.1ex}(k)}\!-q_-^{\hspace{-0.1ex}(k)}| \leq
d^{(k)}$. If the characteristic sizes $\sigma$ of the coarse-graining satisfy $\sigma \!\gg \!d^{(k)}$, then the “off-diagonal” elements of ${\cal D}_c( \{ \bar{q}_+ \},\{ \bar{q}_-\} )_{(t_1, \dots , t_N)}$ are negligible and one has approximate decoherence [@gell-mann-hartle; @hartle]. We should stress that $S_{\rm IF}^{\rm eff}[q_+,q_-]$ is the result of integrating out both the “external” environment degrees of freedom and also the system degrees of freedom which are “not accessible” to the observations (the “internal” environment). In general, these two integrations play an important role in the achievement of this sufficient condition for approximate decoherence. A characterization of the degree of approximate decoherence has been given in Ref. [@dowker] (see also Refs. [@histories; @halliwell93]).
Typically, $d^{(k)}$ can be estimated in terms of $\Delta t_k \!\equiv \! t_{k+1} \!-\! t_k$. When this is the case, one usually finds that the Gell-Mann and Hartle mechanism for approximate decoherence works provided all the time intervals satisfy $\Delta t_k \!\geq\! \Delta t_c$, $k=0, \dots, N$, where $\Delta t_c$ is sufficiently larger than some characteristic decoherence time scale $t_{\scriptscriptstyle \! D}$ ($t_{\scriptscriptstyle \! D}$ can be written in terms of $\sigma$ and some parameters characterizing the environment and the system-environment couplings) [@gell-mann-hartle; @hartle; @paz-zurek]. For $\Delta t_c$ one should take the smallest value compatible with a specified degree of approximate decoherence. In this sense, we can think of a coarse-graining as characterized both by the sizes $\sigma$ and by the time scale $\Delta t_c$.
. Effective equations of motion for the system {#subsec:effective eqs}
----------------------------------------------
Assuming that the mechanism for approximate decoherence described in the previous subsection works, an approximate effective description of the coarse-grained system variables in terms of a set of random variables \[in the sense of Eq. (\[correlation functions\])\] is available, at least for instants of time satisfying $\Delta t_k \!\geq\! \Delta t_c$, for $k=0, \dots, N$. The corresponding probability distribution ${\cal P}_{\bar{q}}( \{ \bar{q} \} )_{(t_1, \dots , t_N)}$ is given by the diagonal elements of the decoherence function (\[c-g decoh funct\]). We shall next make an estimation of this probability distribution. This follows essentially the derivation of Gell-Mann and Hartle in Refs. [@gell-mann-hartle; @hartle]. For alternative derivations for more specific models, see Refs. [@halliwell93; @dowker; @halliwell]. Introducing the new variables $q_{\scriptscriptstyle \Delta}\!\equiv\! q_+\!-q_-$ and $q_{\scriptscriptstyle \Sigma}\!\equiv\! {1 \over 2}\, (q_+\!+\!q_-)$, and similarly for $\bar{q}_{\pm_k}$, and assuming that $\sigma \gg d^{(k)}$, note first, from (\[c-g decoh funct 3\]), that the restrictions on the integration over $q_{\scriptscriptstyle \Delta}$ coming from the coarse-graining can be neglected in the diagonal elements of this decoherence function. Therefore, using (\[c-g decoh funct 2\]) and (\[decoherence functional 2\]), and writing $S_{\rm eff}[q_+,q_-]\equiv S_{\rm eff}
[q_{\scriptscriptstyle \Delta},q_{\scriptscriptstyle \Sigma}]$, we get \_[|[q]{}]{}( { |[q]{}\_ } ) \_[(t\_1, …, t\_N)]{} \[q\_\] \_[k=1]{}\^N \^2 (q\_(t\_k) - |[q]{}\_[\_k]{}) \_[f]{}\[q\_\], \[probability 2\] where \_[f]{}\[q\_\] \_[\_[\_[ q\_[\_]{}(t\_f)=0 ]{}]{}]{} \[q\_\] \_s (q\_[\_[i]{}]{} + q\_[\_[i]{}]{}, q\_[\_[i]{}]{} - q\_[\_[i]{}]{} ;t\_i ) e\^[[i ]{} S\_[eff]{} \[q\_[\_]{},q\_[\_]{}\]]{}. \[probability 3\]
At this stage, we introduce two simplifications in our analysis. First, we restrict our evaluation to coarse-grained system variables having significance only up to certain scales, larger enough than $\sigma$, so that the random variables $\bar{q}_k$ can be well approximated by continuous random variables. This approximation can be implemented with the use of a set of approximate projection operators $\hat{P}_{\bar{q}}(t)$, with $\bar{q}$ being continuous variables, which satisfy the properties (\[proj properties\]) in an approximate way (see Refs. [@dowker; @halliwell] for an example). Then, all the sums $\sum_{\{ \bar{q}\} }$ can be replaced by integrals $\int \hspace{-0.2ex} \prod_{k=1}^N d\bar{q}_k$ and the functions ${\cal P}_{\bar{q}}( \{ \bar{q} \} )_{(t_1, \dots , t_N)}$ become probability densities. Second, as long as we are only interested in the dynamics of the system on time scales much larger than $\Delta t_c$ ($\Delta t_c$ is proportional to the decoherence time scale $t_{\scriptscriptstyle \! D}$, which is typically extremely small, see Refs. [@omnes; @zurek; @joos-zeh] for some examples), we can take the continuous time limit in (\[probability 2\]). In order to do so, consider the instants $t_k \equiv t_i+k \, \Delta t$, $k = 0, \dots , N+1$, with $\Delta t \equiv (t_f-t_i)/(N+1)$. Introducing functions $\bar{q}(t)$, such that $\bar{q}(t_k) =
\bar{q}_k$ (assumed now to be continuous variables), and letting $N \!\rightarrow \! \infty$ in (\[probability 2\]) \[replace $\bar{q}_{{\scriptscriptstyle \Sigma}_k}$ by $\bar{q}_k$\], with $(t_f-t_i)$ maintained finite (thus, $\Delta t \!\rightarrow \! 0$), we get a probability distribution functional associated to some stochastic variables $\bar{q}(t)$ [@feynman-hibbs]: \_[|[q]{}]{}\[|[q]{}\] \[q\_\] \^2\[q\_ -|[q]{}\] \_[f]{}\[q\_\], \[prob functional\] where $\gamma[q]$ is the functional corresponding to $\prod_{k=1}^N \gamma (q(t_k))$ in the limit $N \!\rightarrow \!\infty$ (some redefinitions in the parameters entering in the function $\gamma (q)$ may be needed in order that such limit is well defined, see Refs. [@halliwell; @halliwell93] for an explicit example of how this limit is taken). Notice that, if we take the limit to the continuous in time and in the variables $\bar{q}_k$ in (\[decoherence condition\]), we get a functional $\Phi_{\bar{q}}[j]$ which is the functional Fourier transform of ${\cal P}_{\bar{q}}[\bar{q}]$. Hence, $\Phi_{\bar{q}}[j]$ can be interpreted as the characteristic functional for the stochastic variables $\bar{q}(t)$ [@feynman-hibbs]. From the probability functional (\[prob functional\]) or, equivalently, from the associated characteristic functional \[by functional derivation with respect to the sources $j(t)$\], we can compute the Green functions $G_{c \; m_1 \cdots \, m_s}^{n_1 \cdots \, n_r}
(t_1^{\prime}, \dots, t_r^{\prime};
t_1^{\prime \prime}, \dots, t_s^{\prime \prime})$ with each of the instants in $\{t_1^{\prime}, \dots , t_r^{\prime} \}$ being separated from $t_i$ and from the remaining instants in this set by intervals larger enough than $\Delta t_c$, and similarly for the instants in $\{t_1^{\prime \prime}, \dots , t_s^{\prime \prime} \}$.
We can get a good approximation to the path integral (\[probability 3\]) by expanding $S_{\rm eff}
[q_{\scriptscriptstyle \Delta},q_{\scriptscriptstyle \Sigma}]$ in powers of $q_{\scriptscriptstyle \Delta}$ and neglecting higher than quadratic terms, [*i.e.*]{}, we make a Gaussian approximation in this path integral. This expansion can be made using (\[effective action 2\]) and writing $S_{\rm IF}^{\rm eff}[q_+,q_-]\equiv
S_{\rm IF}^{\rm eff}
[q_{\scriptscriptstyle \Delta},q_{\scriptscriptstyle \Sigma}]$. In this expansion, the dependence of $S_{\rm eff}
[q_{\scriptscriptstyle \Delta},q_{\scriptscriptstyle \Sigma}]$ on the velocities $\dot{q}_{\scriptscriptstyle \Delta}(t)$[^5] (we assume that there is no dependence on time derivatives of higher order) gives rise, after integration by parts, to boundary terms proportional to $q_{{\scriptscriptstyle \Delta}_{\scriptstyle i}}$ (we use that $q_{{\scriptscriptstyle \Delta}_{\scriptstyle f}}=0$). For instance, assuming that $S_{s}^{\rm eff}[q]= \int\! dt \, L_s(q(t),\dot{q}(t),t)$, in the expansion of the terms $S_{s}^{\rm eff}$ we find a boundary term $- p_{s}(q_{\scriptscriptstyle \Sigma_{\scriptstyle i}},
\dot{q}_{\scriptscriptstyle \Sigma_{\scriptstyle i}},t_i) \,
q_{\scriptscriptstyle \Delta_{\scriptstyle i}}$, where $p_{s} \equiv \partial L_s/ \partial \dot{q}$ are the canonical momenta. Similarly, if $S_{\rm IF}^{\rm eff}$ depends on $\dot{q}_{\scriptscriptstyle \Delta}(t)$, its expansion will contain some boundary terms. However, since, in general, $S_{\rm IF}^{\rm eff}$ depends non-locally on $q_{\scriptscriptstyle \Delta}(t)$ and $q_{\scriptscriptstyle \Sigma}(t)$, these terms will be more complicated. Note that we are considering models slightly more general than the ones studied by Gell-Mann and Hartle in Refs. [@gell-mann-hartle; @hartle], since we allow for the possibility of an influence action depending on $\dot{q}_{\scriptscriptstyle \Delta}(t)$ and $\dot{q}_{\scriptscriptstyle \Sigma}(t)$. The motivation for considering such a generalization is that we are interested in field theory actions with interaction terms depending on the derivatives of the fields.
One can show that, when expanding up to quadratic order in $q_{\scriptscriptstyle \Delta}$, the general form for the boundary terms in $S_{\rm IF}^{\rm eff}$ is $- F_1[q_{\scriptscriptstyle \Sigma}](t_i) \,
q_{\scriptscriptstyle \Delta_{\scriptstyle i}}+
i F_2[q_{\scriptscriptstyle \Sigma}](t_i) \,
q_{\scriptscriptstyle \Delta_{\scriptstyle i}}^2
+i \!\int\! dt \, q_{\scriptscriptstyle \Delta}(t) \,
F_3[q_{\scriptscriptstyle \Sigma}](t,t_i) \,
q_{\scriptscriptstyle \Delta_{\scriptstyle i}}$, where $F_1$, $F_2$ and $F_3$ are real functionals of $q_{\scriptscriptstyle \Sigma}$, which vanish when $S_{\rm IF}^{\rm eff}$ does not depend on $\dot{q}_{\scriptscriptstyle \Delta}(t)$. Finally, we get the following expansion: && S\_[eff]{} \[q\_,q\_\]= S\_s\^[eff]{}\[q\_ + q\_\] -S\_s\^[eff]{}\[q\_ - q\_\] +S\_[IF]{}\^[eff]{} \[q\_,q\_\] = -p\_1\[q\_\](t\_i) q\_[\_[i]{}]{}+ i F\_2\[q\_\](t\_i) q\_[\_[i]{}]{}\^2\
&& + i dt q\_(t) F\_3\[q\_\](t,t\_i) q\_[\_[i]{}]{} + dt q\_(t) C\[q\_\](t) +[i 2]{} dt dt\^ q\_(t) q\_(t\^) C\_2\[q\_\](t,t\^) +O (q\_\^3 ) ,\
\[eff action expansion\] with p\_1\[q\](t\_i) p\_s (q\_i,\_i,t\_i)+F\_1\[q\](t\_i), C\[q\](t) +C\_1\[q\](t), \[C\] and C\_k\[q\_\](t\_1,...,t\_k)( [i]{} )\^[ k-1]{} . [\^k S\_[IF]{}\^[eff]{} \[q\_,q\_\] q\_(t\_1) q\_(t\_k)]{} |\_[q\_[\_]{}=0]{}, \[C’s\] where the functional derivatives with respect to $q(t)$ are defined for variations which keep the value of $q(t)$ fixed at $t=t_i$ and $t=t_f$.
Substituting the expansion (\[eff action expansion\]) into Eq. (\[probability 3\]), we get a Gaussian path integral, which can be calculated. Note that, since ${\rm Im}\, S_{\rm IF}^{\rm eff} \geq 0$, $C_2[q](t,t^{\prime})$ is positive semi-definite. In order that the Gaussian approximation that we have carried out is valid, we must assume in addition that $C_2[q](t,t^{\prime})$ is strictly positive definite and, thus, $\det C_2[q] \neq 0$. We get \_[f]{}\[q\] N W\_i\[q\] \^[-1/2]{} e\^[-[1 2]{} dt dt\^ C\[q\](t) C\_2\^[-1]{}\[q\](t,t\^) C\[q\](t\^) ]{}, \[approx probability\] where $N$ is a normalization constant, $C_2^{-1}$ is the inverse of $C_2$ defined by dt\^ C\_2(t,t\^) C\_2\^[-1]{}(t\^,t\^)= (t-t\^), $W_i[q] \equiv W\! \left(q(t_i),p[q](t_i),\Pi[q](t_i);t_i \right)$, with W(q,p,;t\_i) e\^[-[i ]{} q\_0 p]{} e\^[-[1 ]{} q\_0\^2 ]{} \_s (q\_[i]{}+ q\_0, q\_[i]{}- q\_0;t\_i ), \[Wigner funct\] and &&p\[q\](t\_i) p\_1\[q\](t\_i)+ dt dt\^ F\_3\[q\](t,t\_i) C\_2\^[-1]{}\[q\](t,t\^) C\[q\](t\^),\
&&(t\_i) F\_2\[q\](t\_i) - [2]{} dt dt\^ F\_3\[q\](t,t\_i) C\_2\^[-1]{}\[q\](t,t\^) F\_3\[q\](t\^,t\_i). Note that the function $W$ defined in (\[Wigner funct\]) is a generalization of the Wigner function associated to the initial state of the system, and it reduces to the ordinary Wigner function for $\Pi =0$ [@wigner]. Note that, in expression (\[approx probability\]), the momenta $p[q](t_i)$ in this generalized Wigner function are in general different from the canonical momenta $p_s (q_i,\dot{q}_i,t_i)$. In the case of $S^{\rm eff}_{\rm IF}$ non-depending on the velocities $\dot{q}_{\scriptscriptstyle \Delta}(t)$, one has $p[q](t_i)=p_s (q_i,\dot{q}_i,t_i)$ and $\Pi[q](t_i)=0$, thus, $W_i[q]$ is the standard Wigner function. From the definition (\[C’s\]), and using the properties of $S_{\rm IF}^{\rm eff}[q_+,q_-]$, we can see that C\_1\[q\](t)&=& .[ [Re]{} S\_[IF]{}\^[eff]{}\[q\_+,q\_-\] q\_+(t) ]{}|\_[q\_+=q\_-=q]{} =.[S\_[IF]{}\^[eff]{}\[q\_+,q\_-\] q\_+(t) ]{}|\_[q\_+=q\_-=q]{},\
C\_2\[q\](t,t\^)&=& . |\_[q\_+=q\_-=q]{}, \[C’s 2\] and then, from (\[C\]) and (\[effective action 2\]), we have C\[q\](t)=.[S\_[eff]{}\[q\_+,q\_-\] q\_+(t) ]{}|\_[q\_+=q\_-=q]{}.
Substituting (\[approx probability\]) into (\[prob functional\]), we see that the only non-negligible contribution to the path integral in (\[prob functional\]) come from those paths which are not very far deviated from the paths $q^{\scriptscriptstyle (0)}(t)$ which satisfy $C[q^{\scriptscriptstyle (0)}](t)=0$, that is, which satisfy the semiclassical equation (\[semiclassical eq\]). This implies that only those paths $\bar{q}(t)$ which remain always near from the semiclassical paths $q^{\scriptscriptstyle (0)}(t)$ will give a non-negligible value to ${\cal P}_{\bar{q}}[\bar{q}]$. In this sense, the mechanism proposed by Gell-Mann and Hartle is a mechanism for decoherence and classicalization of coarse-grained system variables. However, we see that, in general, ${\cal P}_{\bar{q}}[\bar{q}]$ has a complicated functional dependence on $\bar{q}(t)$.
Let us then study the deviations from a specific solution of the semiclassical equation, that is, we shall now restrict our considerations to those paths $\bar{q}(t)$ which are distributed around a given solution $q^{\scriptscriptstyle (0)}(t)$ of the semiclassical equation. We can now introduce stochastic variables $\Delta q(t) \equiv \bar{q}(t)- q^{\scriptscriptstyle (0)}(t)$ which describe the deviations from $q^{\scriptscriptstyle (0)}(t)$. The associated probability distribution functional ${\cal P}_{\!\Delta q}[\Delta q]$ is equal to ${\cal P}_{\bar{q}}[q^{\scriptscriptstyle (0)}\!+\!\Delta q]$ up to a normalization factor, which, from (\[prob functional\]), is given by \_[|[q]{}]{}\[q\^[(0)]{}+q\] \[q\] \^2\[q\] \_[f]{}\[q\^[(0)]{} +q+q\]. \[probability 4\]
In practice, it is difficult to work out the explicit dependence of the probability distribution functional on the characteristic parameters of the coarse-graining, $\sigma$ and $\Delta t_c$, even in simple models [@halliwell93; @halliwell]. Nevertheless, if such parameters are small enough so that the values of ${\cal P}_{\rm f}[q^{\scriptscriptstyle (0)} +\Delta q+q]$ do not change very much for the different paths $q(t)$ which give a non-negligible contribution in (\[probability 4\]), the functional (\[probability 4\]) can be approximated by ${\cal P}_{\rm f}[q^{\scriptscriptstyle (0)} +\Delta q]$. We can make a further approximation by expanding ${\cal P}_{\rm f}[q^{\scriptscriptstyle (0)}+\Delta q]$ around $q^{\scriptscriptstyle (0)}$. This can be done by setting $q_{\scriptscriptstyle \Sigma}= q^{\scriptscriptstyle (0)}
+\Delta q$ in (\[eff action expansion\]), expanding in $\Delta q$, and substituting the result for this expansion in (\[probability 3\]). The result to lowest non-trivial order is \_[q]{}\[q\] N\[q\^[(0)]{}\] W\_i\[q\^[(0)]{}+ q\] e\^[-[1 2]{} dt dt\^ C\_L\[q\^[(0)]{}+q\](t) C\_2\^[-1]{}\[q\^[(0)]{}\](t,t\^) C\_L\[q\^[(0)]{}+q\](t\^) ]{}, \[approx gaussian probability\] where $N[q^{\scriptscriptstyle (0)}]$ is a normalization factor and $C_L[q^{\scriptscriptstyle (0)}+\Delta q]$ is the expansion of $C[q^{\scriptscriptstyle (0)}+\Delta q]$ to linear order in $\Delta q$. Notice that, in this probability functional, the factor $W_i[q^{\scriptscriptstyle (0)}\!+\!
\Delta q]$ contains all the contribution arising from the initial state of the system. This generalized Wigner function, even if computed expanding around $q^{\scriptscriptstyle (0)}$, will have in general a complicated non-local dependence on $\Delta q$, except when $S^{\rm eff}_{\rm IF}$ is independent of $\dot{q}_{\scriptscriptstyle \Delta}$, in which case it reduces to the standard Wigner function for the initial state of the system and depends only on $\Delta q_i$ and $\Delta \dot{q}_i$. If the deviations from $q^{\scriptscriptstyle (0)}$ are small enough, we can approximate $W_i[q^{\scriptscriptstyle (0)}+
\Delta q] \simeq W_i[q^{\scriptscriptstyle (0)}]$. Then, with these approximations, the variables $\Delta q$ are distributed in such a way that $C_L[q^{\scriptscriptstyle (0)}\!+\!\Delta q](t)$ are Gaussian stochastic variables characterized by C\_L\[q\^[(0)]{}+q\](t) \_c=0, C\_L\[q\^[(0)]{}+q\](t) C\_L\[q\^[(0)]{}+q\](t\^) \_c= C\_2\[q\^[(0)]{}\](t,t\^) . \[gaussian correlators\] Thus, the equation of motion for $\Delta q$ is the Langevin equation C\_L\[q\^[(0)]{}+q\](t) +(t)=0, \[langevin eq\] where $\xi(t)$ is a Gaussian stochastic source with (t) \_c=0, (t) (t\^) \_c= C\_2\[q\^[(0)]{}\](t,t\^) . \[gaussian correlators 2\]
We should mention that there are very simple models for quantum Brownian motion in which all the actions involved are quadratic in their variables and the interaction terms are independent of the velocities . For such models, assuming that the environment is in an initial state of thermal equilibrium, the influence functional can be computed exactly and it is Gaussian. The effective action of Feynman and Vernon in these cases is exactly of the form (\[eff action expansion\]), with $C_1[q_{\scriptscriptstyle \Sigma}](t)$ linear in $q_{\scriptscriptstyle \Sigma}$, $C_2(t,t^{\prime})$ independent of $q_{\scriptscriptstyle \Sigma}$ and $F_1\!=\!F_2\!=\!F_3\!=\!0$. Thus, for these models, expression (\[approx probability\]) is actually exact. In these cases, with the approximation ${\cal P}_{\bar{q}}[\bar{q}] \simeq {\cal P}_{\rm f}[\bar{q}]$, one can derive a Langevin equation for the stochastic variables $\bar{q}(t)$, without need of introducing a specific solution $q^{\scriptscriptstyle (0)}$ of the semiclassical equation. This Langevin equation is simply $C[\bar{q}](t)\!+\!\xi(t)\!=\!0$, being $\xi(t)$ a Gaussian stochastic source with $\left\langle \xi(t)\right\rangle_c=0$ and $\left\langle \xi(t) \, \xi(t^{\prime})\right\rangle_c=
C_2(t,t^{\prime})$. However, for models with more complicated actions, we are only able to derive effective equations of motion for the deviations $\Delta q$ around a given solution $q^{\scriptscriptstyle (0)}$ of the semiclassical equation.
. A quick method to obtain the Langevin equation {#subsec:quick method}
------------------------------------------------
Starting with the effective action of Feynman and Vernon (\[effective action 2\]), there is a quick way to obtain the Langevin equation (\[langevin eq\]) for the deviations $\Delta q$ around a specific solution of the semiclassical equation. This method has actually been extensively used in the literature, in the context of quantum Brownian motion [@caldeira; @hu-matacz94; @hu-paz-zhang2], and also in the context of field theory [@greiner; @matacz; @morikawa; @shaisultanov; @gleiser], including some models for gravity interacting with a scalar field [@calzettahu; @humatacz; @husinha; @cv96; @lomb-mazz; @ccv97; @campos-hu]. One starts with an expansion of this effective action around a solution $q^{\scriptscriptstyle (0)}(t)$ of the semiclassical equation up to quadratic order in perturbations $\Delta q_{\pm}$ satisfying $\Delta q_+(t_i)=\Delta q_-(t_i)$ and $\Delta q_+(t_f)=\Delta q_-(t_f)$ (in the simplest models, in which this effective action is exactly quadratic in $q_+$ and $q_-$, one works directly with the exact expression). From (\[eff action expansion\]), it is easy to see that the expansion for the influence action reads && S\^[eff]{}\_[IF]{}\[q\^[(0)]{}+q\_+ , q\^[(0)]{}+q\_-\]= dt (q\_+(t)-q\_-(t) ) C\_1\[q\^[(0)]{}+ (q\_+ + q\_-)\](t)\
&& +[i 2]{} dt dt\^ (q\_+(t)-q\_-(t) ) C\_2\[q\^[(0)]{}\](t,t\^) (q\_+(t\^)-q\_-(t\^) ) +O (q\^3 ), \[influence action expansion\] where it is understood that $C_1$ has to be expanded up to linear order. Using the identity, which follows from a Gaussian path integration, e\^[-[1 2\^2]{} dt dt\^ (q\_+(t)-q\_-(t) ) C\_2\[q\^[(0)]{}\](t,t\^) (q\_+(t\^)-q\_-(t\^) )]{}= \[\]\_\[\] e\^[[i ]{} dt (t) (q\_+(t)-q\_-(t) ) ]{}, \[identity\] where ${\cal P}_{\xi}[\xi]$ is the Gaussian probability distribution functional for the Gaussian stochastic variables $\xi(t)$ characterized by (\[gaussian correlators 2\]), that is, \_\[\]= , \[xi probability\] we can write in this approximation |\^[eff]{}\_[IF]{} \[q\^[(0)]{}+q\_+, q\^[(0)]{}+q\_-\] |= e\^[-[1 ]{}[Im]{} S\^[eff]{}\_[IF]{} \[q\^[(0)]{}+q\_+, q\^[(0)]{}+q\_-\]]{}= e\^[[i ]{} dt (t) (q\_+(t)-q\_-(t) ) ]{} \_c, \[identity2\] where $\langle \hspace{1.5ex} \rangle_c$ means statistical average over the stochastic variables $\xi(t)$. Thus, the effect of the imaginary part of the influence action (\[influence action expansion\]) on the corresponding influence functional is equivalent to the averaged effect of the stochastic source $\xi(t)$ coupled linearly to the perturbations $\Delta q_{\pm}$ (note that, in the above expressions, the perturbations $\Delta q_{\pm}$ are deterministic functions). Notice that expression (\[identity\]) or, equivalently, (\[identity2\]) give the characteristic functional of the stochastic variables $\xi(t)$ [@feynman-hibbs]. The influence functional, in the approximation (\[influence action expansion\]), can then be written as an statistical average over $\xi$: \^[eff]{}\_[IF]{} \[q\^[(0)]{}+q\_+, q\^[(0)]{}+q\_-\]= e\^[[i ]{} [A]{}\^[eff]{}\_[IF]{}\[q\_+,q\_-;\] ]{} \_c, where \^[eff]{}\_[IF]{}\[q\_+,q\_-;\] S\^[eff]{}\_[IF]{} \[q\^[(0)]{}+q\_+, q\^[(0)]{}+q\_-\] + dt (t) (q\_+(t)-q\_-(t) ) +O (q\^3 ), where ${\rm Re}\, S^{\rm eff}_{\rm IF}$ can be read from expression (\[influence action expansion\]). The Langevin equation (\[langevin eq\]) can be easily derived from the action \_[eff]{}\[q\_+,q\_-;\] S\_s\^[eff]{}\[q\^[(0)]{}+q\_+\]- S\_s\^[eff]{}\[q\^[(0)]{}+q\_-\]+ [A]{}\^[eff]{}\_[IF]{}\[q\_+,q\_-;\], \[new effective action\] where $S_s^{\rm eff}[q^{\scriptscriptstyle (0)}\!+\!\Delta q_{\pm}]$ has to be expanded up to second order in the perturbations $\Delta q_{\pm}$. That is, . |\_[q\_+=q\_-=q]{}=0 \[quick Langevin eq\] leads to Eq. (\[langevin eq\]).
. Effective equations of motion for the gravitational field {#sec:Einstein-Langevin}
===========================================================
In this section, we shall apply the results of the previous section to derive effective equations of motion for the gravitational field in a semiclassical regime. In order to do so, we will consider the simplest case of a linear real scalar field $\Phi$ coupled to the gravitational field. We shall restrict ourselves to the case of fields defined on a globally hyperbolic manifold ${\cal M}$. In this case, we would consider the metric field $g_{ab}(x)$ as the system degrees of freedom, and the scalar field $\Phi(x)$ and also some “high-momentum” gravitational modes, considered as inaccessible to the observations, as the environment variables. Unfortunately, since the form of a complete quantum theory of gravity interacting with matter is unknown, we do not know what these “high-momentum” gravitational modes are. Such a fundamental quantum theory might not even be a field theory, in which case the metric and scalar fields would not be fundamental objects. Thus, in this case, we cannot attempt to evaluate the effective actions in Eq. (\[s-e effective actions\]) starting from the fundamental quantum theory and integrating out the “high-momentum” gravitational modes . What we can do instead is to adopt the usual procedure when dealing with an effective quantum field theory. That is, we shall take for the actions $S_{s}^{\rm eff}[g]$ and $S_{se}^{\rm eff}[g^+,\Phi_+;g^-,\Phi_-]$ the most general local form compatible with general covariance and with the properties of $S_{se}^{\rm eff}$ \[these properties are analogous to those of $S_{\rm IF}$ in Eq. (\[influence action properties\])\] [@weinberg; @donoghue]. The general form for $S_{s}^{\rm eff}[g]$ is S\_[s]{}\^[eff]{}\[g\]= d\^4 x , \[grav action\] where $R$ and $C_{abcd}$ are, respectively, the scalar curvature and the Weyl tensor associated to the metric $g_{ab}$, $1/G_B$, $\Lambda_B /G_B$, $\alpha_B$ and $\beta_B$ are bare coupling constants and the dots represent terms of higher order in the curvature \[because of the Gauss-Bonnet theorem in four spacetime dimensions, there is no need of considering terms of second order in the curvature different from those written in Eq. (\[grav action\])\]. Since ${\cal M}$ is a globally hyperbolic manifold, we can foliate it by a family of Cauchy hypersurfaces $\Sigma_{t}$, labeled by a time coordinate $t$. We use the notation ${\bf x}$ for spatial coordinates on each of these hypersurfaces, and $t_{i}$ and $t_{f}$ for some initial and final times, respectively. The integration domain for all the action terms must now be understood as a compact region ${\cal U}$ of the manifold ${\cal M}$, bounded by the hypersurfaces $\Sigma_{t_i}$ and $\Sigma_{t_f}$ ([*i.e.*]{}, as in the previous section, integrals in $t$ are integrals between $t_i$ and $t_f$).
For the matter part of the effective action, let us consider the following ansatz: S\_[se]{}\^[eff]{}\[g\^+,\_+;g\^-,\_-\]=S\_m\[g\^+,\_+\] -S\_m\[g\^-,\_-\], \[effective action ansatz\] with S\_m\[g,\] -[12]{} d\^4x , \[scalar field action\] where $\xi$ is a dimensionless coupling parameter of the field to the scalar curvature, and the dots stand for terms of higher order in the curvature and in the number of derivatives of the scalar field. Self-interaction terms for the scalar field could also be included but, for simplicity, we shall ignore them in this paper. One can see that general covariance and the properties of $S_{se}^{\rm eff}[g^+,\Phi_+;g^-,\Phi_-]$ imply that imaginary terms and terms mixing the “plus” and “minus” fields in this action must be necessarily non-local. Thus, within a local approximation, the ansatz (\[effective action ansatz\]) is the most general form for this action. We shall comment below some limitations of this local approximation.
In order to simplify the analysis, we neglect the contributions of the higher order terms not written in Eqs. (\[grav action\]) and (\[scalar field action\]). Assuming that the mass of the scalar field is much smaller than the Planck mass, this is a good approximation in a regime where all the characteristic curvature scales are far enough from the Planck scales. The terms in the gravitational Lagrangian density proportional to $R^2$ and $C_{abcd}C^{abcd}$ need to be considered in order to renormalize the matter one-loop ultraviolet divergencies.
Assuming the form (\[effective action ansatz\]) for the matter part of the effective action, we can now introduce the corresponding effective influence functional as in Eq. (\[effective influence functional\]). Let us assume that the state of the scalar field in the Schrödinger picture at the initial time $t\! =\! t_{i}$ is described by a density operator $\hat{\rho}^{\rm \scriptscriptstyle S}(t_{i})$ (in the notation of the previous section, this was $\hat{\rho}^{\rm \scriptscriptstyle S}_e(t_{i})$, but, here, we drop the subindex $e$ to simplify the notation). If we now consider the theory of a scalar field quantized in a classical background spacetime $({\cal M},g_{ab})$ through the action (\[scalar field action\]), to this state it would correspond a state in the Heisenberg picture described by a density operator $\hat{\rho}[g]$. Let $\left\{ |\varphi(\mbox{\bf x})\rangle^{\rm \scriptscriptstyle S}
\right\}$ be the basis of eigenstates of the Schrödinger picture scalar field operator $\hat{\Phi}^{\rm \scriptscriptstyle S}({\bf x})$: $\hat{\Phi}^{\rm \scriptscriptstyle S}({\bf x})
\, |\varphi\rangle ^{\rm \scriptscriptstyle S}=
\varphi(\mbox{\bf x})
\, |\varphi\rangle^{\rm \scriptscriptstyle S}$. The matrix elements of $\hat{\rho}^{\rm \scriptscriptstyle S}(t_{i})$ in this basis will be written as $\rho_{i} \!\left[\varphi,\tilde{\varphi}\right] \equiv
\mbox{}^{\rm \scriptscriptstyle S}
\langle \varphi|\,\hat{\rho}^{\rm \scriptscriptstyle S}(t_{i})
\, |\tilde{\varphi}\rangle^{\rm \scriptscriptstyle S}$. We can now introduce the effective influence functional as \^[eff]{}\_[IF]{}\[g\^+,g\^-\] \[\_+\]\[\_-\] \_i e\^[i(S\_[m]{}\[g\^+,\_+\]-S\_[m]{}\[g\^-,\_-\]) ]{}, \[path integral\] and the effective influence action will be given by ${\cal F}^{\rm eff}_{\rm IF}[g^+,g^-] \equiv
e^{i S^{\rm eff}_{\rm IF}[g^+,g^-]}$.
Of course, trying to show how the mechanism for decoherence and classicalization of the previous section can work in this case would involve some technical difficulties, such as introducing diffeomorphism invariant coarse-grainings and eliminating properly the gauge redundancy (with the use of some suitable Faddeev-Popov method) in the path integrals. We are not going to deal with such issues in this paper. We shall rather assume that they can be suitably implemented without changing the main results for the effective equations of motion.
Expression (\[path integral\]) is actually formal, it is ill-defined and must be regularized in order to get a meaningful quantity for the influence functional. We shall formally assume that we can regularize it using dimensional regularization, that is, that we can give sense to Eq. (\[path integral\]) by dimensional continuation of all the quantities that appear in this expression. We should mention that, however, when performing specific calculations, the dimensional regularization procedure may not be in all cases the most suitable one. In this sense, one should understand the following derivation as being formal. Using dimensional regularization, we must substitute the action $S_m$ in (\[path integral\]) by some generalization to $n$ spacetime dimensions. This can be taken as S\_m\[g,\_[n]{}\] = -[12]{} d\^n x , \[scalar action\] where we use a notation in which we write a subindex $n$ in all the quantities that have different physical dimensions than the corresponding physical quantities in the spacetime of four dimensions. The quantities that do not carry a subindex $n$ have the same physical dimensions than the corresponding ones in four spacetime dimensions, although they should not be confused with such physical quantities. A quantity with a subindex $n$ can always be associated to another one without a subindex $n$; these are related by some mass scale $\mu$, for instance, it is easy to see that $\Phi_{n}=\mu^{{n-4\over 2}}\,\Phi$.
In order to write the effective equations for the metric field in dimensional regularization, we need to substitute the action (\[grav action\]) by some suitable generalization to $n$ spacetime dimensions. We take S\_[s]{}\^[eff]{}\[g\]=\^[n-4]{} d\^n x , \[grav action in n\] where $R_{abcd}$ is the Riemann tensor and, again, the mass parameter $\mu$ has been introduced in order to get the correct physical dimensions. Using the Gauss-Bonnet theorem in four spacetime dimensions, one can see that the action obtained by setting $n\!=\!4$ in (\[grav action in n\]) is equivalent to (\[grav action\]). The form of the action (\[grav action in n\]) is suggested from the Schwinger-DeWitt analysis of the divergencies in the stress-energy tensor in dimensional regularization [@bunch]. The effective action of Feynman and Vernon (\[effective action 2\]) is in our case given by $S_{\rm eff}[g^+,g^-]= S_{s}^{\rm eff}[g^+]-S_{s}^{\rm eff}[g^-]
+S^{\rm eff}_{\rm IF}[g^+,g^-]$. Since the action terms (\[scalar action\]) and (\[grav action in n\]) contain second order derivatives of the metric, one should also add some boundary terms to them [@wald84; @humatacz]. The effect of these boundary terms is simply to cancel out the boundary terms that appear when taking variations of $S_{\rm eff}[g^+,g^-]$ that keep the value of $g^+_{ab}$ and $g^-_{ab}$ fixed on the boundary of ${\cal U}$. They guarantee that we can obtain an expansion for $S_{\rm eff}[g^+,g^-]$ analogous to (\[eff action expansion\]), with no extra boundary terms coming from the integration by parts of terms containing second order derivatives of $g_{ab}^{\scriptscriptstyle \Delta} \equiv g^+_{ab}-g^-_{ab}$. Alternatively, in order to obtain the effective equations for the metric \[equations analogous to (\[semiclassical eq\]) and (\[langevin eq\])\], we can work with the action terms (\[scalar action\]) and (\[grav action in n\]) (without boundary terms) and neglect all boundary terms when taking variations with respect to $g^{\pm}_{ab}$. From now on, all the functional derivatives with respect to the metric must be understood in this sense.
. The semiclassical Einstein equation {#subsec:semiclassical Einstein}
-------------------------------------
From the action (\[scalar action\]), we can define the stress-energy tensor functional in the usual way T\^[ab]{}\[g,\_[n]{}\](x) , \[s-t functional\] which yields T\^[ab]{}\[g,\_n\]=\^[a]{}\_n \^[b]{} \_n- [12]{} g\^[ab]{} \^[c]{} \_n \_[c]{} \_n -[12]{} g\^[ab]{} m\^2 \_n\^2 +( g\^[ab]{} -\^[a]{} \^[b]{} + G\^[ab]{} ) \_n\^2 \[class s-t\] where $\bigtriangledown_{\!a}$ is the covariant derivative associated to the metric $g_{ab}$, $\Box \!\equiv\! \bigtriangledown_{\!a}\bigtriangledown^{a}$, and $G_{ab}$ is the Einstein tensor. Working in the Heisenberg picture, we can now formally introduce the stress-energy tensor operator for a scalar field quantized in a classical spacetime background, regularized using dimensional regularization, as \_[n]{}\^[ab]{}\[g\] T\^[ab]{}\[ g,\_[n]{}\[g\]\], \^[ab]{}\[g\] \^[-(n-4)]{} \_[n]{}\^[ab]{}\[g\], \[regul s-t\] where $\hat{\Phi}_{n}[g](x)$ is the Heisenberg picture field operator in $n$ spacetime dimensions, which satisfies the Klein-Gordon equation ( -m\^2- R ) \_[n]{}=0, \[Klein-Gordon in n\] and where we use a symmetrical ordering (Weyl ordering) prescription for the operators. Using Eq. (\[Klein-Gordon in n\]), one can write the stress-energy operator in the following way: \_[n]{}\^[ab]{}\[g\] = [12]{} { \^[a]{}\_[n]{}\[g\] , \^[b]{}\_[n]{}\[g\] } + [D]{}\^[ab]{}\[g\] \_[n]{}\^2\[g\], \[regul s-t 2\] where ${\cal D}^{ab}[g]$ is the differential operator \^[ab]{}\_[x]{} (-[14]{}) g\^[ab]{}(x) \_[x]{}+ ( R\^[ab]{}(x)- \^[a]{}\_[x]{} \^[b]{}\_[x]{} ), \[diff operator\] being $R_{ab}$ the Ricci tensor. From the definitions (\[path integral\]), (\[s-t functional\]) and (\[regul s-t\]), one can see that . [2]{} |\_[g\^+=g\^-=g]{} =\_n\^[ab]{}(x) , \[s-t expect value\] where the expectation value is taken in the $n$-dimensional spacetime generalization of the state described by $\hat{\rho}[g]$.
As in Eq. (\[semiclassical eq\]), if we derive $S_{\rm eff}[g^+,g^-]$ with respect to $g^+_{ab}$ and then set $g^+_{ab}=g^-_{ab}=g_{ab}$, we get the semiclassical Einstein equation in dimensional regularization: ( G\^[ab]{}\[g\]+ \_[B]{} g\^[ab]{} )- ([43]{} \_[B]{} D\^[ab]{} +2 \_[B]{} B\^[ab]{}) \[g\] = \^[-(n-4)]{} \_[n]{}\^[ab]{} \[g\], \[semiclassical eq in n\] where the tensors $D^{ab}$ and $B^{ab}$ are defined as $$\begin{aligned}
&&\hspace{-6ex}
D^{ab} \equiv
{1\over\sqrt{- g}} \frac{\delta}{\delta g_{ab}}
\int \! d^n x \,\sqrt{- g} \left(R_{cdef}R^{cdef}-
R_{cd}R^{cd} \right)
= {1\over2}\, g^{ab} \! \left( R_{cdef} R^{cdef}-
R_{cd}R^{cd}+\Box \hspace{-0.2ex} R \right) \nonumber \\
&& \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;
\hspace{1ex}
-\,2R^{acde}{R^b}_{cde}-2 R^{acbd}R_{cd}+4R^{ac}{R_c}^b
-3 \hspace{0.2ex}\Box \hspace{-0.2ex} R^{ab}
+\bigtriangledown^{a}\!\bigtriangledown^{b}\! \hspace{-0.2ex} R,
\label{D}\end{aligned}$$ and B\^[ab]{} d\^n x R\^2 = [12]{} g\^[ab]{} R\^2-2 R R\^[ab]{}+2 \^[a]{} \^[b]{} R -2 g\^[ab]{} R. \[B in n\] From equation (\[semiclassical eq in n\]), after renormalizing the coupling constants in order to eliminate the divergencies in $\mu^{-(n-4)}\langle \hat{T}_{n}^{ab}\rangle [g]$ in the limit $n\!\rightarrow \! 4$ and then taking this limit, we will get the semiclassical Einstein equation in the physical spacetime of four dimensions: ( G\^[ab]{}\[g\]+ g\^[ab]{} )- 2 ( A\^[ab]{}+B\^[ab]{} )\[g\]= \_[R]{}\^[ab]{}. \[semiclassical Einstein eq\] In the last equation $1/G$, $\Lambda /G$, $\alpha$ and $\beta$ are renormalized coupling constants, $\langle\hat{T}_{R}^{ab}\rangle [g]$ is the renormalized expectation value of the stress-energy tensor operator, and we have used that, for $n\!=\!4$, $D^{ab}=(3/2) A^{ab}$, being $A^{ab}$ the local curvature tensor obtained by functional derivation with respect to the metric of the action term corresponding to the Lagrangian density $C_{abcd}C^{abcd}$.
. The semiclassical Einstein-Langevin equation {#subsec:Einstein-Langevin}
----------------------------------------------
According to the results of the previous section, assuming that some suitably coarse-grained metric field satisfies the conditions for approximate decoherence and that the approximations of subsection \[sec:classicalization\]\[subsec:effective eqs\] are valid in a certain regime, small deviations from a given solution $g_{ab}$ of the semiclassical Einstein equation (\[semiclassical Einstein eq\]) can be described by linear stochastic perturbations $h_{ab}$ to that semiclassical metric. These perturbations satisfy a Langevin equation of the form (\[langevin eq\]), which shall be called the semiclassical Einstein-Langevin equation. Our next step will be to write the semiclassical Einstein-Langevin equation in dimensional regularization. Let us assume that $g_{ab}$ is a solution of Eq. (\[semiclassical eq in n\]) in $n$ spacetime dimensions. The semiclassical Einstein-Langevin equation in dimensional regularization has then the form $$\begin{aligned}
{1\over 8 \pi G_{B}}\biggl( G^{ab}_L[g\!+\!h]+
\Lambda_{B} \left(g^{ab}\!-\!h^{ab}\right) \biggr)\!
&\!\!\!\!\!\!\!-\!\!\!\!\!\!\!&
\!\left({4\over 3}\, \alpha_{B} D^{ab}_L
+ 2 \beta_{B} B^{ab}_L \right)\![g \!+\! h] = \mu^{-(n-4)}
\left\langle \hat{T}_{n}^{ab}\right\rangle
\!\!_{\mbox{}_{\scriptstyle L}}[g\!+\!h]\nonumber \\
&&\hspace{30 ex}\!+\,2 \mu^{-(n-4)} \xi_n^{ab},
\label{Einstein-Langevin eq in n} \end{aligned}$$ where $h_{ab}$ is a linear stochastic perturbation to $g_{ab}$, $h^{ab}\!\equiv\! g^{ac}g^{bd}h_{cd}$, that is, $g^{ab}\!-h^{ab}\!+0(h^2)$ is the inverse of the metric $g_{ab}\!+\!h_{ab}$, and, as in the previous section, we use a subindex ${\scriptstyle L}$ to denote an expansion up to linear order in $h_{ab}$. In this equation, $\langle \hat{T}_{n}^{ab}\rangle [g+h]$ is the expectation value of $\hat{T}_{n}^{ab}[g\!+\!h]$ in the $n$-dimensional spacetime generalization of the state described by $\hat{\rho}[g+h]$, and $\xi_n^{ab}$ is a Gaussian stochastic tensor characterized by the correlators \_n\^[ab]{}(x) \_[c]{}= 0, \_n\^[ab]{}(x)\_n\^[cd]{}(y) \_[c]{}= N\_n\^[abcd]{}\[g\](x,y), \[correlators in n\] with \[see Eqs. (\[gaussian correlators 2\]) and (\[C’s 2\])\] 2 N\_n\^[abcd]{}\[g\](x,y) . [1 ]{} |\_[g\^+=g\^-=g]{}. \[noise in n\]
We can write Eq. (\[Einstein-Langevin eq in n\]) in a more explicit way by working out the expansion $\langle \hat{T}_{n}^{ab}\rangle
\!_{\mbox{}_{\scriptstyle L}}[g+h]$. Since, from Eq. (\[s-t expect value\]), we have that \_n\^[ab]{}(x) = . [2]{} |\_[h\^+=h\^-=h]{}, \[perturb s-t expect value\] this expansion can be obtained from an expansion of the influence action $S^{\rm eff}_{\rm IF}[g+h^+,g+h^-]$ up to second order in $h^{\pm}_{ab}$ (in this expansion, we can neglect boundary terms). At the same time, we can obtain a more explicit expression for the noise kernel (\[noise in n\]). To perform this expansion for the influence action, we have to compute the first and second order functional derivatives of $S^{\rm eff}_{\rm IF}[g^+,g^-]$ and then set $g^+_{ab}\!=\!g^-_{ab}\!=\!g_{ab}$. If we do so using the path integral representation (\[path integral\]), we can interpret these derivatives as expectation values of operators in the Heisenberg picture for a scalar field quantized in a classical spacetime background $({\cal M},g_{ab})$ as, for instance, in expression (\[s-t expect value\]). The relevant second order derivatives are . [1 ]{} |\_[g\^+=g\^-=g]{} &=& -H\_[\_[n]{}]{}\^[abcd]{}\[g\](x,y) -K\_n\^[abcd]{}\[g\](x,y)+ i N\_n\^[abcd]{}\[g\](x,y),\
. [1 ]{} |\_[g\^+=g\^-=g]{} &=& -H\_[\_[n]{}]{}\^[abcd]{} \[g\](x,y) -i N\_n\^[abcd]{}\[g\](x,y), \[derivatives\] with N\_n\^[abcd]{}\[g\](x,y) &= & { \_n\^[ab]{}(x)- \_n\^[ab]{}(x) , \_n\^[cd]{}(y)- \_n\^[cd]{}(y) } ,\
H\_[\_[n]{}]{}\^[abcd]{} \[g\](x,y) &= & \^ ( \_n\^[ab]{}(x) \_n\^[cd]{}(y) ) ,\
H\_[\_[n]{}]{}\^[abcd]{} \[g\](x,y) &= & -[i4]{} ,\
K\_n\^[abcd]{}\[g\](x,y) &= & . [-1 ]{} |\_[\_[n]{}=\_[n]{}]{}, \[kernels\] using again a symmetrical ordering (Weyl ordering) prescription for the operators in the last of these expressions. All the expectation values in these expressions are in the $n$-dimensional spacetime generalization of the state described by $\hat{\rho}[g]$. In the above equations, $\{ \; , \: \}$ and $[ \; , \: ]$ mean, respectively, the anticommutator and the commutator, and we use the symbol ${\rm T}^{\displaystyle \ast}$ to denote that, first, we have to time order the field operators $\hat{\Phi}_{n}$ and then apply the derivative operators that appear in each term of the product $T^{ab}(x) T^{cd}(y)$, where $T^{ab}$ is the functional (\[class s-t\]). For instance, \^ ( \^[a]{}\_[ \_[x]{}]{} \_[n]{}(x) \^[b]{}\_[ \_[x]{}]{}\_[n]{}(x) \^[c]{}\_[ \_[y]{}]{}\_[n]{}(y) \^[d]{}\_[ \_[y]{}]{}\_[n]{}(y) ) =\_[ x\_1,x\_2 x\_[ \_[\_[\_[\_ [\_[x\_3,x\_4 y]{}]{}]{}]{}]{}]{} ]{} \^[a]{}\_[ \_[x\_1]{}]{} \^[b]{}\_[ \_[x\_2]{}]{} \^[c]{}\_[ \_[x\_3]{}]{} \^[d]{}\_[ \_[x\_4]{}]{} (\_[n]{}(x\_1)\_[n]{}(x\_2) \_[n]{}(x\_3)\_[n]{}(x\_4) ), \[T star\] where ${\rm T}$ is the usual time ordering. Notice that all the kernels that appear in expressions (\[derivatives\]) are real.
In fact, from (\[kernels\]), we see that the noise kernel $N_n^{abcd}$, and also the kernel $H_{\scriptscriptstyle \!{\rm A}_{\scriptstyle n}}^{abcd}$, are free of ultraviolet divergencies in the limit $n \!\rightarrow \!4$. This is because, for a linear quantum field, the ultraviolet divergencies in $\left\langle\hat{T}_n^{ab}(x)
\hat{T}_n^{cd}(y)\right\rangle$ are the same ones as those of $\left\langle\hat{T}_n^{ab}(x)\right\rangle
\left\langle\hat{T}_n^{cd}(y)\right\rangle$. Therefore, in the semiclassical Einstein-Langevin equation (\[Einstein-Langevin eq in n\]), one can perform exactly the same renormalization procedure as the one for the semiclassical Einstein equation (\[semiclassical eq in n\]). After this renormalization procedure, Eq. (\[Einstein-Langevin eq in n\]) will yield the semiclassical Einstein-Langevin equation in the physical spacetime ($n\!=\!4$). It can be written as ( G\^[ab]{}\_L\[g+h\]+ (g\^[ab]{}-h\^[ab]{}) )- 2 ( A\^[ab]{}\_L+B\^[ab]{}\_L ) \[g+h\]=\_[R]{}\^[ab]{}\_[\_[L]{}]{} \[g+h\] +2 \^[ab]{} , \[Einstein-Langevin eq\] being $\xi^{ab}$ is a Gaussian stochastic tensor with \^[ab]{}(x) \_c = 0, \^[ab]{}(x)\^[cd]{}(y) \_c = N\^[abcd]{}\[g\](x,y), \[correlators\] where $N^{abcd} \equiv \lim_{n \rightarrow 4} \mu^{-2 (n-4)}
N_n^{abcd}$. Notice from (\[kernels\]) that the noise kernel $N^{abcd}[g](x,y)$ gives a measure of the lowest order fluctuations of the scalar field stress-energy tensor around its expectation value. Thus, the stochastic metric perturbations $h_{ab}$, solution of the semiclassical Einstein-Langevin equation (\[Einstein-Langevin eq\]), account for the back reaction of such matter stress-energy fluctuations on the spacetime geometry. For a more detailed analysis of the semiclassical Einstein-Langevin equation and some of its applications, see Ref. [@mv98].
Going back to the expressions in dimensional regularization, which may be useful for calculational purposes, we can now write the expansion of the influence action around a given metric $g_{ab}$. From (\[s-t expect value\]) and (\[derivatives\]), taking into account that $S^{\rm eff}_{\rm IF}[g,g]=0$ and that $S^{\rm eff}_{\rm IF}[g^-,g^+]=
-S^{\rm eff {\displaystyle \ast}}_{\rm IF}[g^+,g^-]$, we get && S\^[eff]{}\_[IF]{}\[g+h\^+,g+h\^-\] =[12]{} d\^nx \_[n]{}\^[ab]{}(x)(h\^+\_[ab]{}(x)-h\^-\_[ab]{}(x) )\
&& -[12]{} d\^nx d\^ny (H\_[\_[n]{}]{}\^[abcd]{} \[g\](x,y)+K\_n\^[abcd]{}\[g\](x,y) ) (h\^+\_[ab]{}(x)h\^+\_[cd]{}(y)-h\^-\_[ab]{}(x)h\^-\_[cd]{}(y) )\
&& -[12]{} d\^nx d\^ny H\_[\_[n]{}]{}\^[abcd]{} \[g\](x,y) (h\^+\_[ab]{}(x)h\^-\_[cd]{}(y)-h\^-\_[ab]{}(x)h\^+\_[cd]{}(y) )\
&& +[i2]{} d\^nx d\^ny N\_n\^[abcd]{}\[g\](x,y) (h\^+\_[ab]{}(x)-h\^-\_[ab]{}(x) ) (h\^+\_[cd]{}(y)-h\^-\_[cd]{}(y) )+0(h\^3).\
\[expansion 1\] From (\[kernels\]), it is easy to see that the kernels satisfy the symmetry relations H\_[\_[n]{}]{}\^[abcd]{}(x,y)= H\_[\_[n]{}]{}\^[cdab]{}(y,x), H\_[\_[n]{}]{}\^[abcd]{}(x,y)= -H\_[\_[n]{}]{}\^[cdab]{}(y,x), K\_n\^[abcd]{}(x,y) = K\_n\^[cdab]{}(y,x). \[symmetries\] Using these relations, and defining H\_n\^[abcd]{}(x,y) H\_[\_[n]{}]{}\^[abcd]{}(x,y) +H\_[\_[n]{}]{}\^[abcd]{}(x,y), \[H\] we can write the expansion (\[expansion 1\]) as S\^[eff]{}\_[IF]{}\[g+h\^+,&g&+h\^-\] =[12]{} d\^nx \_[n]{}\^[ab]{}(x)\
&& -[12]{} d\^nx d\^ny (H\_n\^[abcd]{}\[g\](x,y) +K\_n\^[abcd]{}\[g\](x,y) ) { h\_[cd]{}(y) }\
&& +[i2]{} d\^nx d\^ny N\_n\^[abcd]{}\[g\](x,y) +0(h\^3), \[expansion 2\] where we have used the notation h\^+\_[ab]{}-h\^-\_[ab]{}, { h\_[ab]{}} h\^+\_[ab]{}+h\^-\_[ab]{}. \[notation\] Using this expansion and noting, from (\[kernels\]), that K\_n\^[abcd]{}\[g\](x,y)= -[14]{} \_[n]{}\^[ab]{}(x) \^n(x-y)-[12]{}[1]{} . |\_[\_[n]{}=\_[n]{}]{}, \[K\] we get, from (\[perturb s-t expect value\]), \_n\^[ab]{}(x) \_[\_[L]{}]{} \[g+h\] = \_n\^[ab]{}(x) + \_n\^[[(1)]{} ab]{} \[g;h\](x) - 2 d\^ny H\_n\^[abcd]{}\[g\](x,y) h\_[cd]{}(y), \[s-t expect value expansion\] where the operator $\hat{T}_n^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}$ is defined from the term of first order in the expansion $T^{ab}_L[g+h,\Phi_{n}]$ as T\^[ab]{}\_L\[g+h,\_[n]{}\]=T\^[ab]{}\[g,\_[n]{}\]+ T\^[[(1)]{} ab]{}\[g,\_[n]{};h\], \_n\^[[(1)]{} ab]{} \[g;h\]T\^[[(1)]{} ab]{}\[g,\_[n]{}\[g\];h\], \[T(1)\] using, as always, a Weyl ordering prescription for the operators in the last definition. Note that the third term in the right hand side of Eq. (\[s-t expect value expansion\]) is due to the dependence on $h_{cd}$ of the field operator $\hat{\Phi}_{n}[g+h]$ and of the dimensional regularized version of the density operator $\hat{\rho}[g+h]$.
Substituting (\[s-t expect value expansion\]) into (\[Einstein-Langevin eq in n\]), and taking into account that $g_{ab}$ satisfies the semiclassical Einstein equation (\[semiclassical eq in n\]), we can write the Einstein-Langevin equation (\[Einstein-Langevin eq in n\]) as &&( G\^[[(1)]{} ab]{} \[g;h\](x)- \_[B]{} h\^[ab]{}(x) ) - [43]{} \_[B]{} D\^[[(1)]{} ab]{} \[g;h\](x) -2\_[B]{} B\^[[(1)]{} ab]{} \[g;h\](x)\
&&- \^[-(n-4)]{} \_n\^[[(1)]{} ab]{} \[g;h\](x) + 2 d\^ny \^[-(n-4)]{} H\_n\^[abcd]{}\[g\](x,y) h\_[cd]{}(y) = 2 \^[-(n-4)]{} \_n\^[ab]{}(x).\
\[Einstein-Langevin eq 2\] In the last equation we have used the superindex ${\scriptstyle (1)}$ to denote the terms of first order in the expansions $G^{ab}_L[g+h]$, $D^{ab}_L[g+h]$ and $B^{ab}_L[g+h]$. Thus, for instance, $G^{ab}_L[g+h]\!=\!G^{ab}[g]+
G^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}[g;h]$. The explicit expressions for the tensors $T^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}[g,\Phi_{n};h]$, $G^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}[g;h]$, $D^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}[g;h]$ and $B^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}[g;h]$ are given in the Appendix. From $T^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}[g,\Phi_{n};h]$, we can write an explicit expression for the operator $\hat{T}_n^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}$. Using the Klein-Gordon equation (\[Klein-Gordon in n\]), and expressions (\[regul s-t 2\]) and (\[diff operator\]) for the stress-energy operator, we can write this operator as \_n\^[[(1)]{} ab]{} \[g;h\]=([12]{} g\^[ab]{}h\_[cd]{}-\^a\_c h\^b\_d- \^b\_c h\^a\_d ) \_[n]{}\^[cd]{}\[g\] +[F]{}\^[ab]{}\[g;h\] \_[n]{}\^2\[g\], \[T(1) operator\] where ${\cal F}^{ab}[g;h]$ is the differential operator \^[ab]{} && (-) (h\^[ab]{}- g\^[ab]{} h\^c\_c ) + [2]{} ,\
\[diff operator F\] and it is understood that indices are raised with the background inverse metric $g^{ab}$ and that all the covariant derivatives are associated to the metric $g_{ab}$. Substituting expression (\[T(1) operator\]) into Eq. (\[Einstein-Langevin eq 2\]), and using the semiclassical equation (\[semiclassical eq in n\]) to get an expression for $\mu^{-(n-4)}
\langle \hat{T}_{n}^{ab}\rangle [g]$, we can finally write the semiclassical Einstein-Langevin equation in dimensional regularization as &&[18 G\_[B]{}]{}(x) - [43]{} \_[B]{}( D\^[[(1)]{} ab]{}\
&&. -[12]{} g\^[ab]{} D\^[cd]{} h\_[cd]{}+ D\^[ac]{} h\^b\_c+D\^[bc]{} h\^a\_c ) (x) -2\_[B]{}( B\^[[(1)]{} ab]{}- g\^[ab]{} B\^[cd]{} h\_[cd]{}+ B\^[ac]{} h\^b\_c+B\^[bc]{} h\^a\_c ) (x)\
&&- \^[-(n-4)]{} [F]{}\^[ab]{}\_x \_[n]{}\^2(x) +2 d\^ny \^[-(n-4)]{} H\_n\^[abcd]{}\[g\](x,y) h\_[cd]{}(y) =2 \^[-(n-4)]{} \^[ab]{}\_n(x),\
\[Einstein-Langevin eq 3\] where the tensors $G^{ab}$, $D^{ab}$ and $B^{ab}$ are computed from the semiclassical metric $g_{ab}$, and where we have omitted the functional dependence on $g_{ab}$ and $h_{ab}$ in $G^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}$, $D^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}$, $B^{{\scriptscriptstyle (1)}\hspace{0.1ex} ab}$ and ${\cal F}^{ab}$ to simplify the notation. Notice that, in Eq. (\[Einstein-Langevin eq 3\]), all the ultraviolet divergencies in the limit $n \!\rightarrow \!4$, which shall be removed by renormalization of the coupling constants, are in $\left\langle \hat{\Phi}_{n}^2(x) \right\rangle$ and the symmetric part $H_{\scriptscriptstyle \!{\rm S}_{\scriptstyle n}}^{abcd}(x,y)$ of the kernel $H_n^{abcd}(x,y)$, whereas, as we have pointed out above, the kernels $N_n^{abcd}(x,y)$ and $H_{\scriptscriptstyle \!{\rm A}_{\scriptstyle n}}^{abcd}(x,y)$ are free of ultraviolet divergencies. Once we have performed such a renormalization procedure, setting $n \!= \!4$ in this equation will yield the physical semiclassical Einstein-Langevin equation, Eq. (\[Einstein-Langevin eq\]). Note that, due to the presence of the kernel $H_n^{abcd}(x,y)$ in Eq. (\[Einstein-Langevin eq 3\]), such Einstein-Langevin equation will be non-local in the metric perturbation.
. Discussion
------------
We have seen that effective equations of motion for the metric field of the form (\[semiclassical Einstein eq\]) and (\[Einstein-Langevin eq\]) follow from the local approximation (\[effective action ansatz\]) for the effective action describing the “effective interaction” of the metric and the scalar field. A more realistic evaluation of this effective action starting from a fundamental theory of quantum gravity would certainly lead to some real and imaginary non-local terms in this action. In some situations, the contribution of these terms to the effective equations of motion for the metric (note that they would also give some extra terms in the semiclassical equation) might not be negligible and, in any case, one would expect that their role in the decoherence mechanism for the metric field would be important. This would represent non trivial effects coming from the “high-momentum” modes of quantum gravity, which are not part of the gravitational field described by the classical stochastic metric $g_{ab}+h_{ab}$, but which can be source of this gravitational field in the same way as the matter fields. The contribution of these neglected terms to the equations for the background metric $g_{ab}$ and for the stochastic metric perturbation $h_{ab}$ would be similar to the contribution of the scalar field through its stress-energy operator, but with this operator replaced with some “effective” stress-energy operator of such primordial “high-momentum” gravitational modes coupled to the scalar field. These equations would take the form (\[semiclassical Einstein eq\]) and (\[Einstein-Langevin eq\]) only when the effect of this “effective” stress-energy tensor on the classical spacetime geometry can be neglected. A way of partially modelizing this effect would consist on replacing the stress-energy operator $\hat{T}_{n}^{ab}[g]$ by $\hat{T}_{n}^{ab}[g]+\hat{t}_{n}^{ab}[g]$, where $\hat{t}_{n}^{ab}[g]$ is the stress-energy tensor of gravitons quantized in classical spacetime background $({\cal M},g_{ab})$ [@wald84].
We end this paper with some comments on the relation between the semiclassical Einstein-Langevin equation (\[Einstein-Langevin eq\]) and the Langevin-type equations for stochastic metric perturbations recently derived in the literature [@calzettahu; @humatacz; @husinha; @cv96; @lomb-mazz; @ccv97; @campos-hu]. In these previous derivations, one starts with the influence functional (\[path integral\]), with the state of the scalar field assumed to be an “in” vacuum or an “in” thermal state, and computes explicitly the expansion for the corresponding influence action around a specific metric background. One then applies the method of subsection \[sec:classicalization\]\[subsec:quick method\] to derive a Langevin equation for the perturbations to this background. As we have seen in subsection \[sec:classicalization\]\[subsec:quick method\], this method yields the same equations as the one used in this section. However, in most of the previous derivations, one starts with a “mini-superspace” model and, thus, the metric perturbations are assumed from the beginning to have a restrictive form. In those cases, the derived Langevin equations do not correspond exactly to our equation, Eq. (\[Einstein-Langevin eq\]), but to a “reduced” version of this equation, in which only some components of the noise kernel in Eq. (\[correlators\]) (or some particular combinations of them) influence the dynamics of the metric perturbations. Only those equations which have been derived starting from a completely general form for the metric perturbations are actually particular cases, computed explicitly, of the semiclassical Einstein-Langevin equation (\[Einstein-Langevin eq\]) [@cv96; @lomb-mazz; @campos-hu].
Acknowledgments {#acknowledgments .unnumbered}
===============
We are grateful to Esteban Calzetta, Antonio Campos, Bei-Lok Hu and Albert Roura for very helpful suggestions and discussions. This work has been partially supported by the CICYT Research Project number , and the European Project number .
[**Appendix: Expansions around a background metric**]{}
For a metric of the form $\tilde{g}_{ab}\equiv g_{ab}+h_{ab}$, where $h_{ab}$ is a small perturbation to a background metric $g_{ab}$, we list the expansions of metric functionals around the background metric up to linear order in the perturbation. In the following expressions, all the tilded quantities refer to functionals constructed with the metric $\tilde{g}_{ab}$, whereas that the analogous untilded ones are constructed with the background metric $g_{ab}$. In particular, $\taderiv_{\!a}$ and $\bigtriangledown_{\!a}$ are respectively the covariant derivatives associated to the metric $\tilde{g}_{ab}$ and to the metric $g_{ab}$, and $\taderiv^a \equiv \tilde{g}^{\,ab} \taderiv_{\!b}$, $\tBox \equiv \tderiv^a \! \taderiv_{\!a}$, $\bigtriangledown^a \equiv g^{ab}\bigtriangledown_{\!b}$, $\Box \equiv \bigtriangledown^a \bigtriangledown_{\!a}$, where $\tilde{g}^{\,ab}$ and $g^{ab}$ are respectively the inverses of $\tilde{g}_{ab}$ and $g_{ab}$. We shall also raise indices in the metric perturbation with the inverse background metric $g^{ab}$: $h^a_b \equiv g^{ac}h_{cb}$ and $h^{ab} \equiv g^{ac} g^{bd}h_{cd}$.
&&\^[ab]{}=g\^[ab]{}-h\^[ab]{}+O(h\^2),\
&& = ( 1+[1 2]{} h\^a\_a +O(h\^2) ),\
&&\^[c]{}\_[ab]{}=\^[c]{}\_[ab]{}+[1 2]{} (\_[a]{}h\^c\_b+\_[b]{}h\^c\_a- \^[c]{}h\_[ab]{} )+O(h\^2).\
&&,\
&&\_[a]{} \_[b]{} f= \_[a]{} \_[b]{} f -[1 2]{} \^[c]{} f (\_[a]{}h\_[bc]{}+\_[b]{}h\_[ac]{}- \_[c]{}h\_[ab]{})+O(h\^2),\
&&f=f - \^a \^b f h\_[ab]{}- \^a f (\^b h\_[ab]{} -[1 2]{}\_[a]{}h\^b\_b )+O(h\^2),\
&&\^a \^b f= \^a \^b f -\^a \^c f h\^b\_c -\^b \^c f h\^a\_c -[1 2]{} \^[c]{} f (\^a h\^b\_c +\^b h\^a\_c-\_[c]{} h\^[ab]{} ) +O(h\^2).\
&&,\
&&t\^[ab]{}=t\^[ab]{} -\^c\^d t\^[ab]{} h\_[cd]{} +(g\^[ae]{}\^c t\^[db]{} +g\^[be]{}\^c t\^[ad]{} -[1 2]{}g\^[cd]{}\^e t\^[ab]{} ) (\_[c]{} h\_[de]{}+\_[d]{} h\_[ce]{} -\_[e]{} h\_[cd]{} )\
&& +[1 2]{} ( g\^[ac]{}t\^[db]{} +g\^[bc]{}t\^[ad]{}) ( \^e\_[d]{} h\_[ce]{}+ h\_[cd]{}- \^e\_[c]{} h\_[de]{}) +O(h\^2),\
&&\
&&\_[ab]{}=R\_[ab]{}+[1 2]{} ( \^[c]{}\_[a]{} h\_[bc]{} +\^[c]{}\_[b]{} h\_[ac]{} -h\_[ab]{} -\_[a]{}\_[b]{} h\^c\_c ) +O(h\^2),\
&& \^a\_b=R\^a\_b-R\^c\_bh\^a\_c+[1 2]{} ( \^[c]{}\_[b]{} h\^a\_c +\^[c]{}\^[a]{} h\_[bc]{} - h\^a\_b -\_[b]{}\^[a]{} h\^c\_c ) +O(h\^2),\
&&=R-R\^[ab]{}h\_[ab]{} +\^a \^b h\_[ab]{}-h\^a\_a +O(h\^2),\
&& \^[ab]{}=R\^[ab]{}-R\^[ac]{} h\^b\_c-R\^[bc]{} h\^a\_c +[1 2]{} (\^c \^a h\^b\_c +\^c \^b h\^a\_c -h\^[ab]{} -\^a \^b h\^c\_c ) +O(h\^2),\
&&\^[ab]{}=G\^[ab]{} +G\^[[(1)]{} ab]{}+O(h\^2),\
&& G\^[[(1)]{} ab]{}= -R\^[ac]{} h\^b\_c-R\^[bc]{} h\^a\_c +[1 2]{} =-G\^[ac]{} h\^b\_c-G\^[bc]{} h\^a\_c +[1 2]{} ,\
&& \^a\_[bcd]{}=R\^a\_[bcd]{}+[1 2]{} ( \_[c]{}\_[b]{} h\^a\_d +\_[c]{}\_[d]{} h\^a\_b +\_[d]{}\^a h\_[bc]{} -\_[c]{}\^a h\_[bd]{} -\_[d]{}\_[b]{} h\^a\_c\
&& -\_[d]{}\_[c]{} h\^a\_b )+O(h\^2),\
&&\_[abcd]{}=R\_[abcd]{}+ [1 2]{} ( R\^e\_[bcd]{}h\_[ae]{} +R\^[e]{}\_[acd]{}h\_[be]{} ) +[1 2]{} ( \_[c]{}\_[b]{} h\_[ad]{} +\_[d]{}\_[a]{} h\_[bc]{} -\_[c]{}\_[a]{} h\_[bd]{}\
&& -\_[d]{}\_[b]{} h\_[ac]{} )+O(h\^2),\
&&\^[abcd]{}=R\^[abcd]{} -[1 2]{} (2 R\^[abce]{}h\^d\_e +2 R\^[abed]{}h\^c\_e +R\^[aecd]{}h\^b\_e +R\^[ebcd]{}h\^a\_e )\
&& +[1 2]{} ( \^c \^b h\^[ad]{} +\^d \^a h\^[bc]{} -\^c \^a h\^[bd]{} -\^d \^b h\^[ac]{} )+O(h\^2),\
&&\^a \^b = \^a \^b R -\^a \^b (R\^[cd]{}h\_[cd]{}) +\^a \^b \^c \^d h\_[cd]{} -\^a \^b h\^c\_c -\^a \^c R h\^b\_c\
&& -\^b \^c R h\^a\_c -[1 2]{} \^c R (\^a h\^b\_c+\^b h\^a\_c -\_[c]{}h\^[ab]{} )+O(h\^2),\
&& =R- (R\^[ab]{}h\_[ab]{}) +\^a \^b h\_[ab]{} -\^2 h\^a\_a - \^a \^b R h\_[ab]{}\
&& -\^a R (\^b h\_[ab]{} -[1 2]{}\_[a]{}h\^b\_b ) +O(h\^2),\
&&\^[ab]{}=R\^[ab]{} -(R\^[ac]{}h\^b\_c+R\^[bc]{}h\^a\_c) -\^c \^d R\^[ab]{} h\_[cd]{} -\^c R\^[ab]{} ( \^d h\_[cd]{} -[1 2]{} \_[c]{} h\^d\_d )\
&& +\^c R\^[ad]{} ( \_[c]{} h\^b\_d +\_[d]{} h\^b\_c -\^b h\_[cd]{} ) +\^c R\^[bd]{} ( \_[c]{} h\^a\_d +\_[d]{} h\^a\_c -\^a h\_[cd]{} )\
&& +[1 2]{} R\^[ac]{} ( \^d \_[c]{} h\^b\_d+ h\^b\_c-\^d \^b h\_[cd]{} ) +[1 2]{} R\^[bc]{} ( \^d \_[c]{} h\^a\_d+ h\^a\_c-\^d \^a h\_[cd]{} )\
&& +[1 2]{} ( \^c \^a h\^b\_c +\^c \^b h\^a\_c -\^2 h\^[ab]{} -\^a \^b h\^c\_c )+O(h\^2),\
&&\^2=R\^2-2R R\^[ab]{}h\_[ab]{} +2R\^a \^b h\_[ab]{} -2R h\^a\_a +O(h\^2),\
&& \^[ab]{}=R R\^[ab]{} -R R\^[ac]{} h\^b\_c -R R\^[bc]{} h\^a\_c -R\^[ab]{}R\^[cd]{}h\_[cd]{} +[1 2]{}R ( \^c \^a h\^b\_c +\^c \^b h\^a\_c\
&& -h\^[ab]{} -\^a \^b h\^c\_c ) +R\^[ab]{} ( \^c \^d h\_[cd]{} -h\^c\_c ) +O(h\^2),\
&&\^[ab]{}\_[ab]{}=R\^[ab]{} R\_[ab]{} -2R\^[ab]{}R\_a\^ch\_[bc]{}+ R\^[ab]{} (2 \^[c]{} \_[a]{} h\_[bc]{} -h\_[ab]{} -\_[a]{}\_[b]{} h\^c\_c )+ O(h\^2),\
&& \^[ac]{}\^b\_c=R\^[ac]{}R\^b\_c-R\^[ac]{}R\^[bd]{}h\_[cd]{} -R\^[cd]{} ( R\^a\_c h\^b\_d +R\^b\_c h\^a\_d ) +[1 2]{}R\^[ac]{} ( \^d \_[c]{} h\^b\_d +\^d \^b h\_[cd]{}\
&& -h\^b\_c -\_[c]{}\^b h\^d\_d ) +[1 2]{}R\^[bc]{} ( \^d \_[c]{} h\^a\_d +\^d \^a h\_[cd]{} -h\^a\_c -\_[c]{}\^a h\^d\_d ) + O(h\^2),\
&& \^[abcd]{} \_[abcd]{}=R\^[abcd]{}R\_[abcd]{} -2 R\^[abcd]{}R\_[abce]{} h\^e\_d +4 R\^[abcd]{} \_[c]{}\_[b]{} h\_[ad]{} +O(h\^2),\
&&\^[acbd]{}\_[cd]{}=R\^[acbd]{} R\_[cd]{} +[1 2]{}R\_[cd]{} ( R\^[acde]{} h\^b\_e +R\^[bcde]{} h\^a\_e -2 R\^[acbe]{} h\^d\_e -2 R\^[bcae]{} h\^d\_e )\
&& +[1 2]{} R\^[acbd]{} ( \^e \_[c]{} h\_[de]{} +\^e \_[d]{} h\_[ce]{} -h\_[cd]{} -\_[c]{}\_[d]{} h\^e\_e ) -[1 4]{} R\_[cd]{} ( 2 \^c \^d h\^[ab]{}\
&& +\^a \^b h\^[cd]{} +\^b \^a h\^[cd]{} -2 \^a \^c h\^[bd]{} -2 \^b \^c h\^[ad]{} )+O(h\^2),\
&& \^[acde]{} \^b\_[cde]{}=R\^[acde]{} R\^b\_[cde]{} -[1 2]{} ( R\^[acde]{}R\^f\_[cde]{}h\^b\_f +R\^[bcde]{}R\^f\_[cde]{}h\^a\_f ) -2 R\^[acde]{} R\^b\_[cdf]{} h\^f\_e\
&& +[1 2]{} R\^[acde]{} ( \_[d]{}\_[c]{} h\^b\_e +\_[e]{}\^b h\_[cd]{} -\_[e]{}\_[c]{} h\^b\_d -\_[d]{}\^b h\_[ce]{} )\
&& +[1 2]{} R\^[bcde]{} ( \_[d]{}\_[c]{} h\^a\_e +\_[e]{}\^a h\_[cd]{} -\_[e]{}\_[c]{} h\^a\_d -\_[d]{}\^a h\_[ce]{} )+O(h\^2),\
&& \^[ab]{}=B\^[ab]{}+B\^[[(1)]{} ab]{} +O(h\^2),\
&&B\^[[(1)]{} ab]{}= -[1 2]{}R\^2 h\^[ab]{} -g\^[ab]{}RR\^[cd]{}h\_[cd]{} +2 R (R\^[ac]{}h\^b\_c+R\^[bc]{}h\^a\_c ) + R\^a \^b h\^c\_c +2 R\^[ab]{} (R\^[cd]{}h\_[cd]{}+h\^c\_c\
&& -\^c \^d h\_[cd]{} ) +g\^[ab]{} \^c \^d R h\_[cd]{} +2 R h\^[ab]{} -2 \^c \^a R h\^b\_c -2 \^c \^b R h\^a\_c +g\^[ab]{} \^c \^d (R h\_[cd]{})\
&& -\^c +2 g\^[ab]{} (R\^[cd]{}h\_[cd]{}+h\^c\_c -\^c \^d h\_[cd]{} )\
&& -2 \^a \^b (R\^[cd]{}h\_[cd]{}+h\^c\_c -\^c \^d h\_[cd]{} ),\
&& \^[ab]{}=D\^[ab]{}+D\^[[(1)]{} ab]{} +O(h\^2),\
&&D\^[[(1)]{} ab]{}= [1 2]{} (R\^[cd]{}R\_[cd]{}-R\^[cdef]{}R\_[cdef]{}) h\^[ab]{} +2 R\_[cdef]{} (R\^[acde]{} h\^[bf]{}+R\^[bcde]{} h\^[af]{} ) -R\_[cd]{} (4 R\^[ac]{}h\^[bd]{}\
&& +4 R\^[bc]{}h\^[ad]{}+R\^[acde]{} h\^b\_e+R\^[bcde]{} h\^a\_e -2 R\^[acbe]{} h\^d\_e-2 R\^[bcae]{} h\^d\_e ) +g\^[ab]{}(R\^[cf]{}R\_[cg]{}-R\^[cdef]{}R\_[cdeg]{} ) h\^g\_f\
&& -4 R\^[ac]{}R\^[bd]{} h\_[cd]{} +4 R\^[acde]{} R\^b\_[cdf]{} h\^f\_e +[1 2]{} R\^[cd]{} ( 2 \_[c]{}\_[d]{} h\^[ab]{} +\^a \^b h\_[cd]{} +\^b \^a h\_[cd]{}\
&& -2 \^a \_[c]{} h\^b\_d -2 \^b \_[c]{} h\^a\_d ) +[1 2]{} R\^[ac]{} ( \^d \_[c]{} h\^b\_d -7 h\^b\_c +7 \^d \^b h\_[cd]{} -4 \_[c]{} \^b h\^d\_d )\
&& +[1 2]{} R\^[bc]{} ( \^d \_[c]{} h\^a\_d -7 h\^a\_c +7 \^d \^a h\_[cd]{} -4 \_[c]{} \^a h\^d\_d ) -R\^[acde]{} ( \_[d]{}\_[c]{} h\^b\_e +\_[e]{}\^b h\_[cd]{}\
&& -\_[d]{}\^b h\_[ce]{} -\_[e]{}\_[c]{} h\^b\_d ) -R\^[bcde]{} ( \_[d]{}\_[c]{} h\^a\_e +\_[e]{}\^a h\_[cd]{} -\_[d]{}\^a h\_[ce]{} -\_[e]{}\_[c]{} h\^a\_d )\
&& -[1 2]{} g\^[ab]{} R\^[cd]{} ( 2 \^e \_[c]{} h\_[de]{} -h\_[cd]{} -\_[c]{}\_[d]{} h\^e\_e ) +2 g\^[ab]{} R\^[cdef]{} \_[e]{}\_[d]{} h\_[cf]{} -R\^[acbd]{} ( \^e \_[c]{} h\_[de]{}\
&& +\^e \_[d]{} h\_[ce]{} -h\_[cd]{} -\_[c]{}\_[d]{} h\^e\_e ) +[1 2]{} \^c R ( \_[c]{} h\^[ab]{} -\^a h\^b\_c -\^b h\^a\_c ) -3 \^c R \^[ad]{} ( \_[c]{} h\^b\_d\
&& +\_[d]{} h\^b\_c -\^b h\_[cd]{} ) -3 \^c R \^[bd]{} ( \_[c]{} h\^a\_d +\_[d]{} h\^a\_c -\^a h\_[cd]{} ) -[1 4]{} ( g\^[ab]{} \^c R -6 \^c R\^[ab]{} ) ( 2 \^d h\_[cd]{}\
&& -\_[c]{} h\^d\_d ) -[1 2]{} R h\^[ab]{} -\^a \^c R h\^b\_c -\^b \^c R h\^a\_c -[1 2]{} g\^[ab]{} \^c \^d R h\_[cd]{} +3 \^c \^d R\^[ab]{} h\_[cd]{}\
&& + [3 2]{} ( 2 R\^[ac]{} h\^b\_c+ 2 R\^[bc]{} h\^a\_c +h\^[ab]{} + \^a \^b h\^c\_c - \^c \^a h\^b\_c - \^c \^b h\^a\_c ) -[1 2]{} g\^[ab]{} ( R\^[cd]{}h\_[cd]{} +h\^c\_c\
&& -\^c \^d h\_[cd]{} ) -\^a \^b ( R\^[cd]{}h\_[cd]{} +h\^c\_c -\^c \^d h\_[cd]{} ).\
&&\
&& ,\
&& T\^[ab]{}\[g,\_[n]{}\]\^a \_[n]{} \^b \_[n]{}- [12]{} g\^[ab]{} \^[c]{}\_[n]{} \_[c]{}\_[n]{} -[12]{} g\^[ab]{} m\^2 \_[n]{}\^2 +( g\^[ab]{} -\^a \^b +G\^[ab]{} ) \_[n]{}\^2,\
&&T\^[ab]{}\[,\_[n]{}\]=T\^[ab]{}\[g,\_[n]{}\]+ T\^[[(1)]{} ab]{}\[g,\_[n]{};h\] +0(h\^2),\
&& T\^[[(1)]{} ab]{}\[g,\_[n]{};h\]= -T\^[ac]{}\[g,\_[n]{}\] h\^b\_c-T\^[bc]{}\[g,\_[n]{}\] h\^a\_c -[1 2]{} ( \^[c]{}\_[n]{} \_[c]{}\_[n]{} +m\^2 \_[n]{}\^2 ) h\^[ab]{}\
&& +[1 2]{} g\^[ab]{} \^c \_[n]{} \^d \_[n]{} h\_[cd]{} +[2]{} \_[n]{}\^2.
[11]{}
L. H. Ford, [*Ann. Phys. *]{}[**144**]{}, 238 (1982).
C.-I. Kuo and L. H. Ford, [*Phys. Rev. *]{}[**D47**]{}, 4510 (1993); N. G. Phillips and B.-L. Hu, [*Phys. Rev. *]{}[**D55**]{}, 6123 (1997).
B.-L. Hu, [*Physica* ]{}[**A158**]{}, 399 (1989).
R. P. Feynman and F. L. Vernon, [*Ann. Phys. *]{}[**24**]{}, 118 (1963).
R. P. Feynman and A. R. Hibbs, [*Quantum mechanics and path integrals*]{} (McGraw-Hill, New York, 1965).
E. Calzetta and B.-L. Hu, [*Phys. Rev. *]{}[**D49**]{}, 6636 (1994).
B.-L. Hu and A. Matacz, [*Phys. Rev. *]{}[**D51**]{}, 1577 (1995).
B.-L. Hu and S. Sinha, [*Phys. Rev. *]{}[**D51**]{}, 1587 (1995).
A. O. Caldeira and A. J. Legget, [*Physica* ]{}[**A121**]{}, 587 (1983).
B.-L. Hu, J. P. Paz and Y. Zhang, [*Phys. Rev. *]{}[**D45**]{}, 2843 (1992).
B.-L. Hu and A. Matacz, [*Phys. Rev. *]{}[**D49**]{}, 6612 (1994).
C. Greiner and B. Müller, [*Phys. Rev. *]{}[**D55**]{}, 1026 (1997).
A. Campos and E. Verdaguer, [*Phys. Rev. *]{}[**D53**]{}, 1927 (1996).
F. C. Lombardo and F. D. Mazzitelli, [*Phys. Rev. *]{}[**D55**]{}, 3889 (1997).
A. Campos and E. Verdaguer, [*Int. J. Theor. Phys. *]{}[**36**]{}, 2525 (1997); E. Calzetta, A. Campos and E. Verdaguer, [*Phys. Rev. *]{}[**D56**]{}, 2163 (1997).
A. Campos and B.-L. Hu, [*Phys. Rev. *]{}[**D58**]{}, 125021 (1998).
B.-L. Hu, J. P. Paz and Y. Zhang, [*Phys. Rev. *]{}[**D47**]{}, 1576 (1993).
A. Matacz, [*Phys. Rev. *]{}[**D55**]{}, 1860 (1997).
M. Morikawa, [*Phys. Rev. *]{}[**D33**]{}, 3607 (1986); D.-S. Lee and D. Boyanovsky, [*Nucl. Phys. *]{}[**B406**]{}, 631 (1993).
R. Zh. Shaisultanov, hep-th/9509154; hep-th/9512144.
M. Gleiser and R. O. Ramos, [*Phys. Rev. *]{}[**D50**]{}, 2441 (1994); D. Boyanovsky, H. J. de Vega, R. Holman, D.S-Lee and A. Singh, [*Phys. Rev. *]{}[**D51**]{}, 4419 (1995); E. Calzetta and B.-L. Hu, [*Phys. Rev. *]{}[**D55**]{}, 3536 (1997); M. Yamaguchi and J. Yokoyama, [*Phys. Rev. *]{}[**D56**]{}, 4544 (1997); S. A. Ramsey, B.-L. Hu and A. M. Stylianopoulos, [*Phys. Rev. *]{}[**D57**]{}, 6003 (1998).
J. Schwinger, [*J. Math. Phys. *]{}[**2**]{}, 407 (1961); [*Phys. Rev. *]{}[**128**]{}, 2425 (1962); P. M. Bakshi and K. T. Mahanthappa, [*J. Math. Phys. *]{}[**4**]{}, 1 (1963); [*ibid. *]{}[**4**]{}, 12 (1963); L. V. Keldysh, [*Zh. Eksp. Teor. Fiz. *]{}[**47**]{}, 1515 (1964) \[[*Sov. Phys. JETP* ]{}[**20**]{}, 1018 (1965)\]; J. Schwinger, [*Particles, Sources, and Fields*]{} (Addison-Wesley, 1970).
K.-C. Chou, Z.-B. Su, B.-L. Hao and L. Yu, [*Phys. Rep. *]{}[**118**]{}, 1 (1985); N. P. Landsman and Ch. G. van Weert, [*Phys. Rep. *]{}[**145**]{}, 141 (1987).
R. B. Griffiths, [*J. Stat. Phys. *]{}[**36**]{}, 219 (1984).
R. Omn\` es, [*Rev. Mod. Phys. *]{}[**64**]{}, 339 (1992), and references therein; R. Omn\` es, [*The Interpretation of Quantum Mechanics*]{} (Princeton Series in Physics, Princeton University Press, Princeton, 1994).
M. Gell-Mann and J. B. Hartle, [*Phys. Rev. *]{}[**D47**]{}, 3345 (1993).
J. B. Hartle, in [*Gravitation and Quantizations*]{}, proceedings of the 1992 Les Houches Summer School, edited by B. Julia and J. Zinn-Justin (North-Holland, Amsterdam, 1995), gr-qc/9304006, and references therein.
J. J. Halliwell, in [*Stochastic Evolution of Quantum States in Open Systems and Measurement Processes*]{}, edited by L. Diósi (World Scientific, Singapore, 1994), gr-qc/9308005.
J. J. Halliwell, [*Ann. N. Y. Acad. Sci. *]{}[**755**]{}, 726 (1995), and references therein.
J. P. Paz and W. H. Zurek, [*Phys. Rev. *]{}[**D48**]{}, 2728 (1993).
H. F. Dowker and J. J. Halliwell, [*Phys. Rev. *]{}[**D46**]{}, 1580 (1992).
J. J. Halliwell, [*Phys. Rev. *]{}[**D48**]{}, 4785 (1993); [*ibid.*]{}, [**D57**]{}, 2337 (1998).
R. M. Wald, [*General Relativity*]{} (The University of Chicago Press, Chicago, 1984).
E. S. Abers and B. W. Lee, [*Phys. Rep. *]{}[**9C**]{}, 1 (1973).
S. Weinberg, [*The Quantum Theory of Fields*]{}, vols. I and II (Cambridge University Press, Cambridge, England, 1995 and 1996).
V. Hakim and V. Ambegaokar, [*Phys. Rev. *]{}[**A32**]{}, 423 (1985); C. Morais Smith and A. O. Caldeira, [*Phys. Rev. *]{}[**A36**]{}, 3509 (1987).
H. Grabert, P. Schramm and G.-L. Ingold, [*Phys. Rep. *]{}[**168**]{}, 115 (1988).
Z.-B. Su, L.-Y. Chen, X.-T. Yu and K.-C. Chou, [*Phys. Rev. *]{}[**B37**]{}, 9810 (1988).
W. H. Zurek, in [*Conceptual Problems of Quantum Gravity*]{}, based on the proceedings of the 1988 Osgood Hill Conference, edited by A. Ashtekar and J. Stachel (Birkh" auser, Boston, 1991); [*Phys. Today* ]{}[**44**]{} (10), 36 (1991); [*Vistas in Astronomy* ]{}[**37**]{}, 185 (1993).
T. Brun, [*Phys. Rev. *]{}[**D47**]{}, 3383 (1993); J. P. Paz, S. Habib and W. H. Zurek, [*Phys. Rev. *]{}[**D47**]{}, 488 (1993).
F. Lombardo and F. D. Mazzitelli, [*Phys. Rev. *]{}[**D53**]{}, 2001 (1996).
J. T. Whelan, [*Phys. Rev. *]{}[**D57**]{}, 768 (1998); gr-qc/9702003.
E. Joos and H. D. Zeh, [*Z. Phys. *]{}[**B59**]{}, 223 (1985).
E. Wigner, [*Phys. Rev. *]{}[**40**]{}, 749 (1932); M. Hillery, R. F. O’Connell, M. O. Scully, E. P. Wigner, [*Phys. Rep. *]{}[**106**]{}, 121 (1984); J. J. Halliwell, [*Phys. Rev. *]{}[**D36**]{}, 3626 (1987).
J. F. Donoghue, [*Phys. Rev. Lett. *]{}[**72**]{}, 2996 (1994); [*Phys. Rev. *]{}[**D50**]{}, 3874 (1994); [*Helv. Phys. Acta *]{}[**69**]{}, 269 (1996); in [*Advanced School on Effective Theories*]{}, edited by F. Cornet and M. J. Herrero (World Scientific, Singapore, 1996), gr-qc/9512024; gr-qc/9712070.
T. S. Bunch, [*J. Phys. *]{}[**A12**]{}, 517 (1979).
R. Martín and E. Verdaguer, gr-qc/9811070.
[^1]: Institut de Física d’Altes Energies (IFAE)
[^2]: Even if, in order to simplify the notation, we do not write indices in these coordinates, $q$ and $Q$ have to be understood as representing an arbitrary number of degrees of freedom (which, in particular, can be an infinite number of degrees of freedom).
[^3]: We shall assume that the action $S[q,Q]$ is the one that appears in the path integral formulas for the model, which, in general, needs not to coincide with the classical action for the model [@abers; @weinberg].
[^4]: A way of generalizing the formalism to a non-factorizable initial density operator consists in the following [@hakim; @gell-mann-hartle]. One writes the initial density matrix in coordinate representation as $\rho(q,Q;q^{\prime},Q^{\prime};t_i)=\rho_s(q,q^{\prime};t_i)\,
\rho_{se}(q,Q;q^{\prime},Q^{\prime};t_i)$, where $\rho_s$ is chosen in such a way that $\int\! dq\, \rho_s(q,q;t_i)=1$. Then, the CTP generating functional can be written as (\[generating functional\]), with $$e^{{i \over \hbar}\,S_{\rm eff}[q_+,q_-]}\equiv
\int\! {\cal D}[Q_+]\;{\cal D}[Q_-]\;
\rho_{se} (q_{+_{\scriptstyle i}},Q_{+_{\scriptstyle i}};
q_{-_{\scriptstyle i}},Q_{-_{\scriptstyle i}};t_i ) \:
\delta(Q_{+_{\scriptstyle f}}\!-Q_{-_{\scriptstyle f}}) \;
e^{{i \over \hbar}\, \left(\,S[q_+,Q_+]-S[q_-,Q_-]\, \right)}.$$
[^5]: We understand that a term depends on $\dot{q}_{\scriptscriptstyle \Delta}(t)$ if it does so before any integration by parts.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In these lectures we provide a basic introduction into the topic of dispersion relation and analyticity. The properties of 2-point functions are discussed in some detail from the viewpoint of the Källén-Lehmann and general dispersion relations. The Weinberg sum rules figure as an application. The analytic structure of higher point functions in perturbation theory are analysed through the Landau equations and the Cutkosky rules.'
bibliography:
- 'input3.bib'
---
-----------------------------
CP3-Origins-2016-042 DNRF90
-----------------------------
1.5cm
[**A brief Introduction to Dispersion Relations and Analyticity**]{}[^1]\
[ Roman Zwicky\
Higgs Centre for Theoretical Physics, School of Physics and Astronomy,\
University of Edinburgh, Edinburgh EH9 3JZ, Scotland ]{}\
E-Mail: `[email protected]`.
Prologue
========
Dispersion relations are a powerful non-perturbative tool which have originated in classical electrodynamics in the theory of Kramers-Kronig dispersion relations. Analytic properties follow from causality and the use of Cauchy’s theorem allows to obtain the real part of an amplitude from the knowledge of the imaginary part which is often better accessible. This is the idea of the S-matrix program from the fifties and sixties. Dispersion relations are sparsely discussed in modern textbooks as the focus is on other aspects of Quantum Field Theory (QFT). There are some excellent older textbooks on analyticity e.g. [@S-matrix; @Todorov; @Barton], some modern textbooks devote some chapters to the topic e.g.[@IZ; @Weinberg1], as well as some lecture notes [@deRafael]. I would hope that a student who has followed an introductory course on QFT or has read some chapters of a QFT textbook would be able to largely follow the presentation below.
Introduction {#sec:prologue}
------------
In the fifties and sixties QFT has found a big success in describing quantum electrodynamics (QED) thanks to the successful renormalisation program carried out by Dyson, Feynman, Schwinger, Tomonaga and others [@Schweber]. The description of the strong force with QFT proved to be difficult and there was some prejudice that a solution outside field theory had to be found. Two such approaches are dispersion theory using analytic properties [@S-matrix] (Heisenberg, Chew, …) and Wilson’s operator product expansion [@Wilson69]. As Weinberg remarks in his book [@Weinberg1] both of these approaches later became a part of QFT! By analytic properties we mean analyticity in the external momenta. In QFT analytic continuation is inherent in the field description (second quantisation). Let us remind ourselves how this is related to scattering matrix elements of particles.
![Schematic diagram for $2\to 2$ scattering corresponding to the matrix element in Eq. . Time goes from left to right.[]{data-label="fig:2to2"}](figs-pdf/fig1.pdf){width="2.5in"}
A primary goal of particle physics is to describe scattering of $n$-particles via the so-called S-matrix. For the scattering of $ 2 \to 2$ particles this reads (cf. Fig.\[fig:2to2\]) $$\label{eq:Smatrix}
\phantom{a}_{\rm out}{\langle p_1,p_2|} q_1 q_2 \rangle_{\rm in} = \phantom{a}_{\rm out}{\langle p_1,p_2|}
S |q_1 q_2 \rangle_{\rm out} \;,$$ where we have assumed the particles to be of spin $0$. In the case where they are all of equal mass this implies the following on-shell conditions: $p_1^2 = p_2^2 = q_1^2 = q_2^2 = m^2$. Hence one might wonder how analytic properties come into play. The answer is through the celebrated *Lehman-Symmanzik-Zimmermann (LSZ) formula* whose derivation can be found in most textbooks e.g. [@IZ]. For our case it reads $$\begin{aligned}
\label{eq:LSZ}
\phantom{a}_{\rm out}{\langle p_1,p_2|} q_1 q_2 \rangle_{\rm in} &=& - (i Z^{-1/2})^4 \int_{x_1,x_2,y_1,y_2}
e^{- i ( q_1 \cdot x_1 + q_2 \cdot x_2- p_1\cdot y_1 - p_2\cdot y_2 )} K_{x_1} K_{x_2} K_{y_1} K_{y_2} \times\nonumber \\
& & {\langle T \phi(x_1) \phi(x_2)\phi^\dagger(y_1)\phi^\dagger(y_1) \rangle} + \text{disconnected terms} \;,\end{aligned}$$ where $\int_x = \int d^4 x$ hereafter, $T$ is the time ordering, ${\langle \dots \rangle}$ is the vacuum expectation value (VEV), the quanta are assumed to carry a charge (complex conjugation for outgoing particle), $K$ is the Klein-Gordon operator $K_{x_1} = \Box + m^2 \to -(q_1^2-m^2)$ and the $Z$ factor results from the *asymptotic condition*, $$\label{eq:asym}
{\langle 0 |\phi(x)|q_1\rangle} \stackrel{x_0 \to \mp \infty}{\to} Z^{1/2} {\langle 0 |\phi_{\text{in(out)}}(x)|q_1\rangle} \;,$$
The asymptotic condition is the key idea of the LSZ-approach. Namely that when the particles are well separated from each other all that remains is the self-interaction which is parameterised by the renormalisation factor $Z$. The field $\phi$ is what is known as an interacting field whereas $ \phi_{\text{in(out)}}$ are free fields in which case the right-hand side of the equation above equals $\sqrt{Z /(2\pi)^3}e^{-i q_1 \cdot x}$.[^2][^3] The disconnected part corresponds, for example, to the case where particle $q_1 \to p_1$ and $q_2 \to p_2$ without any interaction which is of no interest to us. From we conclude that
- The scattering of $n$-particles ($n = n_{\rm in } + n_{\rm out}$) is described by $n$-point functions (or $n$-point correlators). The study of the latter is therefore of primary importance.
- The $n$-point correlators are functions of the external momenta e.g. $p_{1,2}^2, q_{1,2}^2,
p_1 \cdot p_2 ,\dots $. First and foremost they are defined for real values or more precisely for real values with a small imaginary part e.g. $p_1^2 = {\rm Re}[p_1^2] + i0$.[^4] From there they can be analytically continued into the complex plane. Hence it is the second quantisation, describing particles with fields, that allows us to go off-shell for correlation functions.[^5]
The course consists of three parts. Analytic properties of 2-point functions (section \[sec:2pt\]), which comes with definite answer in terms of the non-perturbative Källén-Lehmann spectral representation. Applications of 2-point function in section \[sec:app\]. Last a short discussion of the analytic properties of higher point function in perturbation theory (PT) e.g. Landau equations and Cutkosky rules in section \[sec:higher\].
2-point Function {#sec:2pt}
================
Dispersion Relation from $1^{\text{st}}$-principles: Källén-Lehmann Representation {#sec:KL}
----------------------------------------------------------------------------------
Let us define the Fourier transform of the 2-point correlator as follows $$\Gamma(p^2) = i \int_x e^{i p \cdot x} {\langle T \phi(x) \phi^\dagger(0) \rangle} \;.$$ What determines the analytic structure of $\Gamma(p^2)$? By analytic structure we mean the singularities e.g. poles, branch points and the associated branch cuts. The Källén-Lehmann representation [@K; @L] gives a very definite answer to this question. The presentation is straightforward and can be found in most textbooks e.g. [@Weinberg1].
The 2-point function in the free and interacting case can be written as $$\Gamma(p^2) = \left\{
\begin{array}{ll}
\label{eq:Ga}
\frac{1}{m^2 - p^2 - i 0} = - \Delta_F(p^2,m^2) & \text{free} \\[0.3cm]
\frac{ Z({\lambda})}{m^2 - p^2 - i 0} + f({\lambda},p^2) & \text{interacting}
\end{array}
\right. \;.$$ The function $Z({\lambda})$ and $f({\lambda},p^2)$, where ${\lambda}$ is the coupling constant e.g. ${\cal L}_{\rm int} = {\lambda}\phi^3 + {\rm h.c.}$, obey $$Z({\lambda}) \stackrel{{\lambda}\to 0}{\to} 1 \;, \quad f({\lambda},p^2)\stackrel{{\lambda}\to 0}{\to} 0 \;,$$ in order to reproduce the free field theory limit. In what follows it is our goal to determine the properties of $f({\lambda},p^2)$ in more detail. At the end of this section we are going to make remarks about the possible ranges of the $Z({\lambda})$-function. The first lesson to be learnt from the free field theory case is that it is the mass (i.e. the spectrum) which determines the analytic properties cf. Fig. \[fig:analytic-structure\](left). As we shall see this generalises to the interacting case.
![(left) analytic structure for free field theory propagator with spectral function underneath (right) idem for an interacting theory with a stable 1-particle state and a multiparticle-threshold[]{data-label="fig:analytic-structure"}](figs-pdf/fig2.pdf){width="4.6in"}
For technical reason it is advantageous to first study the positive frequency distribution $$\label{eq:T}
{\langle \phi(x) \phi^\dagger(0) \rangle} = \left\{
\begin{array}{ll}
\Delta_+(x^2,m^2) = \int \frac{d^4 p}{(2 \pi)^3} e^{- i p \cdot x}\delta^+(p^2-m^2) & \text{free} \\[0.3cm]
(*) & \text{interacting}
\end{array}
\right.$$ where $\delta^+(p^2-m^2) \equiv \delta(p^2-m^2)\theta(p_0)$ assures that energies are positive and that the momenta are on the mass-shell. It is the quantity $(*)$ that we intend to study. First we use the formal decomposition of the identity into a complete set of states $\mathbb{1} =
\sum_n {|n\rangle}{\langle n|}$ which follow from unitarity. Inserting this relation and using translation invariance one gets $$(*) = \sum_n e^{-i p_n \cdot x} |\underbrace{ {\langle 0| \phi(0)|n(p_n)\rangle}}_{ \equiv f_n}|^2 \;.$$ Further using $1 = \frac{1}{(2 \pi)^4}\int_p e^{- i p \cdot x} \int_x e^{ i p \cdot x}$ and interchanging the $\sum_n$ and the $\int_x$[^6] leads to $$\label{eq:spec}
(*) = \int d^4 p e^{- i p \cdot x} \underbrace{ \equiv \sum_n {\delta}(p-p_n) |f_n|^2}_{(2 \pi)^{-3} \rho(p^2) \theta(p_0) } \;,$$ where $\rho(p^2)$ is known as the *spectral function*, $(2\pi)^{-3}$ a convenient normalisation factor and $\theta(p_0)$ assures positive energies which come from the positive energy condition on the external momentum. Upon using $\int_p F(p) = \int_p \int ds {\delta}(s -p^2) F(s)$ and exchanging the $ds$ and $d^4 p$ integration one finally gets $$(*) = \int_0^\infty ds \rho(s) \Delta_+(x^2,s) \;,$$ a spectral representation.
From and it seems plausible that this spectral representation generalises to the time ordered 2-point function as follows
$$\label{eq:KL}
\Gamma(p^2) = \int_0^\infty ds \rho(s) (-\Delta_F(s,p^2)) = \int_0^\infty ds \frac{\rho(s)}{s-p^2 - i0} \;.$$
Eq. is referred to as the *Källén-Lehmann (spectral) representation*.
At this stage we can make many relevant comments.
1. The Källén-Lehmann representation is a special case of dispersion relation. It shows that dispersion representation follow from first principles in QFT.
2. The analytic properties of $\Gamma(p^2)$ are in one-to-one correspondence with the spectrum of the theory which is the answer to the question what determines the analytic properties of the 2-point function. Hence for the 2-point function there are no other singularities on the first sheet (known as the physical sheet)[^7] other than on the positive real axis determined by the spectrum. The analytic structure is depicted in Fig. \[fig:analytic-structure\](right). An example of an unphysical singularity (not on the physical sheet) is given in section \[sec:example\].
3. The spectral function $\rho(s) \geq 0 $ is positive definite as a direct consequence of unitarity. \[As a homework question you could try to show that for a non-unitary theory with negative normed states (i.e. $\langle gh|gh \rangle = -1$ where “gh" stands for ghost) $\rho(s)$ loses positive definiteness.\]
4. Often the spectral function is decomposed into a pole part[^8] $$\label{eq:rho}
\rho(s) = Z \delta(s-m^2) + \theta(s-s_0) \sigma(s)$$ and continuum part $\sigma(s)$. The latter is the concrete realisation of the function $f({\lambda},s)$ in . In many applications $f_0$, the residue of the lowest state, $$\label{eq:NP}
\Gamma(p^2) = \frac{|f_0|^2}{m^2 -p^2 -i0} + \int_{s_0}^\infty \frac{\sigma(s)}{m^2 -p^2 -i0} \;,$$ is the non-perturbative quantity that is to be extracted. The left-hand side is computed and the $\sigma$-part is then either estimated or suppressed by applying an operation to the equation. This technique is the basis of QCD sum rules [@SVZ] and lattice QCD [@DeGrand:2006zz] extraction of low-lying hadronic parameters. In the former case the $\sigma$-part is suppressed by a Borel-transformation and in lattice QCD $\sigma$-part is exponentially suppressed in euclidian time.
5. The Källén-Lehmann representation straightforwardly applies to the case of a non-diagonal correlation function e.g. ${\langle \phi_A(x) \phi_B^\dagger(0) \rangle}$ but clearly positive definiteness is, in general, lost since $|f_n|^2 \to f_n^A (f_n^B)^*$.
6. As promised we return to the issue of interchanging various sums and integrals. This is of no consequence as long as there are no UV-divergences. As is well-known most field theories show UV-divergences so care has to be taken. UV-divergences demand regularisations and a prescription to renormalise the ambiguities which arise from removing the infinities. There are two ways to formally handle this problem. First, assuming a logarithmic divergence, we may amend as $$\label{eq:subtract}
\Gamma(p^2) = \int_0^{\Lambda_{\rm UV}} ds \frac{\rho(s)}{s-p^2 - i0} + A \;,$$ where the so-called subtraction constant is adjusted to cancel the logarithmic divergence coming form the integral: $A = A_0 \ln (\Lambda_{\rm UV}^2/\mu_0^2) + A_1$ with $\mu_0$ being some arbitrary reference scale. The constant $A_1$ has either to be taken from experiment in the case where $\Gamma(p^2) $ is physical (which implies scheme-independence) or is dependent on the scheme. The dependence in the latter case has to disappear when physical information is extracted from $\Gamma(p^2)$. A more elegant way, in my opinion, is to handle the problem with a once subtracted dispersion $$\label{eq:subtractDR}
\Gamma(p^2) = \Gamma(p_0^2) + (p^2-p_0^2)
\int_0^{\infty} ds \frac{\rho(s)}{(s-p^2 - i0)(s-p_0^2)} \;.$$ It is observed that the integral is now convergent due to the extra $1/(s-p_0^2)$ factor. The same remarks apply to $\Gamma(p_0^2)$ as for the previously discussed $A_1$. To derive the above expression one writes an unsubtracted dispersion relation for $\Gamma(p^2)$ and $\Gamma(p_0^2)$ separately takes the difference and combines the fraction. They key point is that the divergent parts are the same and cancel each other.
7. Following the presentation in Weinberg’s book [@Weinberg1]: imposing the canonical commutation relation $[ \partial_t \phi^\dagger(x) , \phi(0)]_{x_0=0} =
-i {\delta}( \vec{x})$ (in $\hbar = 1$ units) leads to the sum rule $$\label{eq:SR}
\int_0^\infty ds \rho(s) = 1 \;,$$ from where one deduces that:
- $Z= 1$ for a free theory
- $0 \leq Z \leq 1$ for an interacting theory
- $Z = 0$ if $\phi$ is a confined field
The last case does not follow directly from but is an important result due to Weinberg. An example is given by the quark propagator for which we do not expect a residue since it is a confined (coloured) particle. The fact that $Z_{\rm quark} \neq 0$ in each order in PT is a sign that the latter is not suited to describe the phenomenon of confinement.
8. By using causality, i.e. ${\langle [\phi(x), \phi^\dagger(0) ] \rangle}=0$ for $x^2 < 0$ spacelike, it follows that $\bar{\rho}(s) = \rho(s)$ where $\bar{\rho}(s)$ is the antiparticle spectral function associated with $\Delta_-(x^2,m^2) = {\langle \phi^\dagger(x) \phi(0) \rangle}$. This is a special case of the CPT theorem. Related to this matter it was Gell-Mann, Goldberger and Thirring [@GGT] in 1954 who derived analyticity properties from causality, for ${\gamma}+ N \to {\gamma}+N$, justifying dispersion relations from a non-perturbative viewpoint.
Dispersion Relations and Cauchy’s Theorem {#sec:cauchy}
-----------------------------------------
![Standard sketch of optical theorem. The right-hand side is the sum over all intermediate states. It is a particular case of the cutting rules to be discussed in section \[sec:CR\].[]{data-label="fig:optical"}](figs-pdf/fig3.pdf){width="4.0in"}
It is our goal to characterise the spectral function $\rho(s)$ in other ways than through the spectrum. In preparation to the general case we are going to recite the optical theorem for the S-matrix. The S-matrix is conveniently parameterised as $$S = \mathbb{1} + i T \;,$$ where $T$ is the non trivial part of the scattering operator. From the unitarity of the S-matrix it follows that $$\mathbb{1} = S S^\dagger = \mathbb{1} + \underbrace{i (T - T^\dagger)}_{- 2 {\rm Im}[T]} + |T|^2 \;,$$ if and only if
$$\label{eq:optical}
2 {\rm Im}[T] = |T|^2 = T^\dagger \sum_n {|n\rangle}{\langle n|} T \;.$$
Eq. is the celebrated *the optical theorem* depicted in Fig. \[fig:optical\].
The right-hand side of the equation above is reminiscent of the spectral function in the case where $T$ is associated with $\phi$. Hence the expectation that $\rho(s)$ is related to an imaginary part is not unexpected from the viewpoint of the optical theorem. Below we are going to see that this is the case on very general grounds.
To do so we first take a little detour to discuss integral representations of arbitrary analytic functions by the use of Cauchy’s theorem. Let $f(p^2)$ be an analytic function then by Cauchy’s theorem the following integral representation holds $$f(p^2) = \frac{1}{2 \pi i} \int_{\gamma}\frac{ds \, f(s) }{s-p^2} \;,$$ provided that i) $p^2$ is inside the contour of ${\gamma}$ , ii) the contour of ${\gamma}$ does not cross any singularities.
![(left) Integration contour for 2-point function dispersion representation. (right) 2-point function in $\phi^3$ theory in perturbation theory.[]{data-label="fig:pacman"}](figs-pdf/fig4a.pdf "fig:"){width="2.6in"} ![(left) Integration contour for 2-point function dispersion representation. (right) 2-point function in $\phi^3$ theory in perturbation theory.[]{data-label="fig:pacman"}](figs-pdf/fig4b.pdf "fig:"){width="2.0in"}
Applying this techniques to the 2-point function in QFT one makes use of the knowledge of the analytic structure and chooses a contour ${\gamma}_R$ as in Fig. \[fig:pacman\] which does not cross any singularities. The radius is then taken to infinity, $R \to \infty$, which in the case where there are UV divergences results in subtraction which we generically parameterised $P(p^2)$ by a polynomial function (the $P = A$ in is a constant polynomial). The integral is then written as $$\begin{aligned}
\label{eq:disp}
\Gamma(p^2) &=& \frac{1}{2 \pi i} \int_{{\gamma}_R} \frac{ds \, \Gamma(s) }{s-p^2}
\stackrel{R \to \infty}{\to} \frac{1}{2 \pi i} \int_{s_1}^\infty \frac{ds \, ( \Gamma(s+i 0) - \Gamma(s-i 0)) }{s-p^2} + P(p^2) \nonumber \\[0.1cm]
&=& \frac{1}{2 \pi i} \int_{s_1}^\infty \frac{ds \, {\rm disc}[ \Gamma(s)] }{s-p^2 - i 0} + P(p^2) =
\frac{1}{ \pi } \int_{s_1}^\infty \frac{ds \, {\rm Im}[ \Gamma(s)] }{s-p^2 - i 0} + P(p^2) \;,\end{aligned}$$ where $s_1 < \text{singularities}$ (to the left of red part in Fig. \[fig:pacman\](left)) and the second line is the definition of what is called the discontinuity along the branch cut. The last equality follows from $2i {\rm Im}[ \Gamma(s)] = {\rm disc}[ \Gamma(s)]$. This formula can be verified in each order in PT but is also justified on general grounds by the Schwartz reflection principle (cf. appendix \[app:Schwartz\]). In summary we then have that the spectral function is related to the the imaginary part and the discontinuity by
$$\label{eq:main}
\rho(s) = \frac{1}{\pi} {\rm Im}[ \Gamma(s)] = \frac{1}{2 \pi i} {\rm disc}[ \Gamma(s)] \;.$$
This equation follows from equating and and the knowledge that the subtraction constant are the same in both cases since they originate from UV divergences. Hence eliminating the contributions from the arc may result in UV-divergences and subtraction constants.
Dispersion Relations in Perturbation Theory {#sec:disp-PT}
-------------------------------------------
This section aims to illustrate from the viewpoint of PT. In order to do PT one needs to specify a theory for which we choose ${\cal L}_{\rm int} = {\lambda}\phi^3 + {\rm h.c.}$. The pole contribution is then just the propagator and the first non-trivial interaction is generated by the diagram in Fig. \[fig:pacman\](right). The 1-loop graph is UV divergent and requires regularisation. Using dimensional regularisation $d = 4 - 2 {\epsilon}$ the result reads $$\label{eq:above}
\Gamma(p^2) = \frac{ Z({\lambda})}{ m^2 - p^2 - i 0} - {\lambda}^2 |A|\left(
\frac{1}{{\epsilon}} + 2 - {\beta}\ln \left( \frac{{\beta}+1}{{\beta}-1} \right) \right) + {\cal O}({\lambda}^4) \;,$$ with ${\beta}= \sqrt{1 - (4m^2-i0)/p^2}$. The corresponding imaginary part divided by $\pi$ must be the spectral function $$\label{eq:rho}
\rho(p^2) \stackrel{\eqref{eq:main}}{=} \frac{1}{\pi} {\rm Im}[ \Gamma(p^2)]
\stackrel{\eqref{eq:above}}{=} Z({\lambda}) \delta(p^2 - m^2) + {\lambda}^2 |A| {\beta}\theta(p^2 - 4 m^2) + {\cal O}({\lambda}^4) \;.$$ The UV divergence is not important for our purposes since it does not affect the imaginary part. Of course the UV-divergence means that the dispersion relation does not converge in the UV. This problem can be handled either by a subtracted dispersion relation or a polynomial counterterm and regularisation as discussed under point 6 in section \[sec:KL\].
Having resolved this technical issue we focus on the interpretation of the imaginary part. The propagator term is a pole singularity with a delta function in the spectral function and the logarithm corresponds to a branch cut singularity resulting in a $\theta$-function part. By the spectral representation this branch cut must correspond to some physical intermediate state. This state is a 2-particle state starting at the minimum centre of mass energy $4m^2$ ranging all the way up to infinity. The precise value depends on the corresponding momentum configuration. Let the two particle momenta be parameterised by $$p_{1,2} = ( \sqrt{m^2 + x^2},0,0,\pm x) \;, \quad x > 0 \;, \quad p_{1,2}^2 = m^2 \;, \quad p^2 = (p_1+p_2)^2 = 4 m^2 + 4 x^2$$ and therefore $4 x^2 = p^2 - 4m^2$ which can be satisfied for any (arbitrarily large) $p^2 \geq 4 m^2$.
Application of $2$-point Functions {#sec:app}
==================================
There are numerous applications of $2$-point functions and dispersion relations. For example deep-inelastic scattering, QCD sum rules which we have alluded to in and below , $e^+ e^- \to \text{hadrons}$ and inclusive $b \to X_{u,c} \ell \nu$ decays with the additional assumption of analytic continuation to Minkowski-space.[^9] We choose to present the Weinberg sum rules (WSR).
Weinberg Sum Rules
------------------
The Weinberg sum rules are an ingenious construction involving a variety of conceptual ideas. They were proposed in 1967 by Weinberg [@WSR] in the pre-QCD era but we are going to present them from the viewpoint of QCD e.g. [@deRafael; @Weinberg2]. One considers the correlation function of left and right-handed current with two massless quark flavours $$\label{eq:start}
i \int d^4 x e^{i q \cdot x} {\langle T J^{a,L}_\mu(x) J^{b,R}_\nu(x) \rangle} = (q_\mu q_\nu - q^2 g_{\mu \nu} )
\Pi^{a,b}_{\rm LR}(q^2) \;,$$ where $$J^{a,(L,R)}_\mu = \bar q T^a {\gamma}_\mu q_{L,R} \;,$$ with $q_{L,R} = P_{L,R} q $, $P_{L,R} = 1/2 (1 \mp {\gamma}_5)$, $T^a$ being an $SU(2)$-generator (Pauli-matrix). The Lorentz-decomposition in is valid in the limit of massless quarks. According to the previous sections the function $\Pi^{a,b}_{\rm LR}(-Q^2)$, with $-q^2 = Q^2 > 0$ satisfies a dispersion relation of the form $$\Pi^{a,b}_{\rm LR}(-Q^2) = \frac{1}{\pi} \int_{s_0}^\infty \frac{ds \, {\rm Im}[ \Pi^{a,b}_{\rm LR}(s)]}{s+Q^2}
+ A$$ where $A$ is a subtraction constant due to the potential logarithmic divergence which may arise since $\Pi^{a,b}_{\rm LR}$ is of mass dimension zero.
The peculiarity of the WSR relies on the absence of lower dimension corrections in the OPE. This can be seen in an elegant manner using $SU(2)$ representation theory. We denote by $1,F$ and $A$ the trivial, fundamental and adjoint representation of $SU(2)$ which are of dimension $1,2$ and $3$. The correlation function is in the (A,A)-representation of the $(SU(2)_L , SU(2)_R)$ global flavour symmetry. The individual OPE-contribution must be in the same global flavour symmetry representation or vanish otherwise.
One considers Wilson’s OPE in momentum space, valid for $Q^2 = - q^2 \gg \Lambda_{\rm QCD}^2$ $$\Pi^{a,b}_{\rm LR}(-Q^2) = C_{\mathbb{1}} (Q^2) {\langle \mathbb{1} \rangle} + C_{\bar q q}(Q^2)
\frac{ ( {\langle \bar q^{a}_L q^{b}_R \rangle} + {\rm h.c.})}{Q^3} + C_{JJ} (Q^2)
\frac{{\langle J^{L,a}_\mu J^{R,b}_\mu \rangle}}{Q^6} + {\cal O}\left( \frac{\Lambda_{\rm QCD}^8}{Q^8} \right) \;.$$ The functions $C(Q^2)$ are known as Wilson coefficients and carry logarithmic correction in QCD. As can be seen from the formula above the condensate terms of dimension $d$ are suppressed by $1/Q^d$ relative to the identity term. The ${\langle \mathbb{1} \rangle}$-term corresponds to PT and the condensates, i.e. VEVs of operators, are of non-perturbative nature. The former is in $(1,1)$-representation and therefore absent.[^10] A quark bilinear ${\langle \bar q^{a}_L q^{b}_R \rangle}$ is in the $(F,F)$-representation and absent for the same reason. The dimension six operator is, somewhat trivially, in the $(A,A)$-representation and therefore the leading term appears at ${\cal O}(1/Q^6)$. The vanishing of the $1/Q^2$- and $1/Q^4$-terms lead to constraints. The latter can be obtained by expanding the denominator in inverse powers of $Q^2$, $$\frac{1}{s+Q^2} = \frac{1}{Q^2}\frac{1}{1+s/Q^2} = \frac{1}{Q^2} - \frac{s}{Q^4} + \frac{s^2}{Q^6} + \dots \;.$$
The exact sum rules on the spectral function $$\begin{aligned}
\label{eq:WSR}
\int_{s_0}^\infty ds \, \Pi^{a,b}_{\rm LR}(s) = 0 \;, \qquad \int_{s_0}^\infty ds \, \Pi^{a,b}_{\rm LR}(s) s = 0\end{aligned}$$ known as the *first and second Weinberg sum rule* follow.
Note the absence of the perturbative term means in particular that there is no UV-divergence and hence $A=0$. Since the convergence is two powers in $s$ higher (1st and 2nd WSR) the dispersion relation is referred to a as superconvergent.
The WSR are a powerful non-perturbative constraint. We present the original application pursued by Weinberg. First we notice that the left-right correlator can be written as a difference of the vector and axial correlator $$\Pi^{a,b}_{\rm LR}(s) = \frac{1}{4} \left( \Pi^{a,b}_{\rm VV}(s) - \Pi^{a,b}_{\rm AA}(s) \right) \;,$$ where $J^{V(A),a}_\mu \equiv \bar q T^a {\gamma}_\mu ( {\gamma}_5) q$. Taking into account the lowest lying particles $\pi$, $\rho$ and $a_1$ in the narrow width approximation and assuming isospin symmetry (i.e. global $SU(2)_V$ -flavour symmetry) one arrives at[^11] $$\begin{aligned}
{2}
\label{eq:spec}
& \rho_{V}^{a,b}(s) &\;=\;& \frac{1}{\pi} {\rm Im}[ \Pi^{a,b}_{\rm VV}](s) = \delta^{ab}
(f_\rho^2 {\delta}(s- m_{\rho}^2) + \theta(s-s_0) \sigma_{V}) \;, \nonumber \\[0.1cm]
& \rho_{A}^{a,b}(s) &\;=\;& \frac{1}{\pi} {\rm Im}[ \Pi^{a,b}_{\rm AA}](s) = \delta^{ab}
(f_\pi^2 {\delta}(s) + f_{a_1}^2 {\delta}(s- m_{a_1}^2) + \theta(s-s_0) \sigma_{A}) \;.\end{aligned}$$ The functions $\sigma_{V,A}$ contain any higher states and multiparticle states. If one assumes that around $s_0$ perturbation theory is valid then $ \rho_{\rm LR}(s) = 0$ for $s > s_0$ which in turn implies $\sigma_{V} = \sigma_{A}$.
Hence using the two WSR read $$\begin{aligned}
{2}
f_\rho^2 = f_\pi^2 + f_{a_1}^2 \;, \qquad
m_\rho^2 f_\rho^2 = m_{a_1}^2 f_{a_1}^2 \;,\end{aligned}$$ where the decay constants are defined as $${\langle \rho[a_1]^b(p)| J^{V[A],a}_\mu |0\rangle} = {\delta}^{ab} \eta_\mu(p) m_{\rho[a_1]} f_{\rho[a_1]} \;, \quad
{\langle \pi^b(p)| J^{A,a}_\mu |0\rangle} = {\delta}^{ab} p_\mu f_\pi \;,$$ with $\eta$ being the polarisation vector.
In his original paper Weinberg used the experimentally motivated KSFR relation $f_\rho^2 = 2 f_\pi^2$ which then leads to $m_{a_1} = \sqrt{2} m_\rho$. This relation is reasonably satisfied by experiment: $m_{a_1}/m_\rho \simeq 1.63 \simeq 1.15 \sqrt{2}$. Let us end this section with mentioning two further applications of this reasoning.
- Being related to chirality the WSR, or the $\Pi_{LR}$ function, is a measure of contributions to electroweak precision measurement in the case of physics beyond the Standard Model coupling to new fermions. The WSR serve to estimate the contribution of strongly coupled extensions of the standard model such as technicolor and the composite Higgs model.
- The inverse moments of the spectral function, with pion pole subtracted, are related to the low energy constant $L_{10}$ of chiral perturbation theory. Note, chiral perturbation theory is an expansion in $Q^2$, and not $1/Q^2$ as the OPE, and thus leads to inverse moments rather than moments themselves. It is not the WSR per se which are important in this respect but the onset of the duality threshold of PT-QCD which allows to estimate $L_{10}$ in terms of $f_{\pi,\rho,a_1}$. The estimate of $L_{10}$ obtained is in reasonable agreement with experiment.
At this point the students were given the choice of continuing with applications (e.g. infrared interpretation of the chiral anomaly, positivity of low energy constants …) or a the more conceptual topic of higher point functions properties. Their choice was the latter.
Analyticity Properties of higher point Functions {#sec:higher}
================================================
There are many motivations to study higher point functions and their analytic structure amongst which we quote the following:
- As seen in the introduction they describe the scattering of $n$-particles.
- From the discussion in section \[sec:cauchy\] it is clear that to write down dispersion relations one needs to know first and foremost the analytic structure of the amplitude in question.
- 3-point functions are relevant for the study of form factors. Consider for example the $B \to \pi$ form factor, relevant for the determination of the CKM-element $|V_{\rm ub}|$, defined by $${\langle \pi(p)|V_\mu|B(p_B)\rangle} = (p_B)_\mu f^{B \to \pi}_+(q^2) + \dots \;,$$ where the dots stand for the other Lorentz structure and $V_\mu = \bar b {\gamma}_\mu u$ is the weak current (the axial part does not contribute in QCD by parity conservation). Then the form factor can be extracted from the following 3-point function, by using a double dispersion relation (dispersion relation in the $p_B^2$ and $p^2$-variable) $$\begin{aligned}
\label{eq:double}
\Gamma(p^2,p_B^2,q^2) &\;=\;& i^2 \int_{x,y} e^{ - i ( p_B \cdot x - p \cdot y)} {\langle T J_B(x) J_\pi(y) V_\mu(0) \rangle}
\;, \nonumber \\[0.1cm]
&\;=\;&
(p_B)_\mu \left( \frac{g_\pi \, f_B \, f_+^{B \to \pi}(q^2)}{(p_B^2-m_B^2)(p^2-m_\pi^2)} + \text{higher} \right) + \dots \;,\end{aligned}$$ where “higher" stands for higher contributions in the spectrum (the analogue of $\sigma(s)$ in ) and $$\begin{aligned}
{4}
& J_B &\;=\;& \bar q i {\gamma}_5 b \;, \quad & & {\langle B|J_B|0\rangle} &\;=\;& f_B \;, \nonumber \\[0.1cm]
& J_\pi &\;=\;& \bar q i {\gamma}_5 q \;, \quad & & {\langle 0|J_\pi|\pi\rangle} &\;=\;& g_\pi \;, \end{aligned}$$ play the role of the interpolating operators of the LSZ-formalism cf. footnote \[foot:inter\]. As previously mentioned the key idea is then to compute $\Gamma(p^2,p_B^2,q^2) $ in some formalism and to find ways to either estimate or suppress the higher states in order to extract the form factor where $g_\pi$ and $f_B$ are assumed to be known quantities.
In fact if we were able to compute $\Gamma(p^2,p_B^2,q^2)$ with arbitrary precision then the function would assume the form in and we could simply extract the form factor from the limiting expression $$f_+^{B \to \pi}(q^2) = \frac{1}{g_\pi f_B} \lim_{p_B^2 \to m_B^2, p^2 \to m_\pi^2} (p_B^2-m_B^2)(p^2-m_\pi^2)\Gamma(p^2,p_B^2,q^2) \;,$$ which makes the connection to the LSZ-formalism apparent. Unfortunately at present we cannot hope to do so and therefore we have to resort to the approximate techniques as alluded to above.
We have seen that for the 2-point function the analytic structure of the first sheet (physical sheet) is fully understood through the Källén-Lehmann representation. Moreover the singularities on the physical sheet are in one-to-one correspondence with the physical spectrum. For higher point function less is known in all generality. We refer the reader to the works of Källén-Wightman [@three] and Källén [@four] for some general studies of 3- and 4-point functions using first principles and the summary by Andrè Martin [@Martin] for a comparatively recent survey of rigorous results.[^12]
Hence one has to become immediately more modest! We are going to restrain ourselves to analysing singularities in PT for *physical momenta* (i.e. real momenta). This is done by the use of two major tools:
- *Landau equations*: which answer the question about the location of the singularities cf. section \[sec:LE\]. The question on which sheet the singularities are is a difficult question which we comment on.
- *Cutkosky rules*: are rules for computing the discontinuity of an amplitude cf. section \[sec:CR\].
Before analysing these matters in more details let us first consider the normal-thresholds for higher point functions.
Normal Thresholds: cutting Diagrams into two Pieces {#sec:normal}
---------------------------------------------------
The so-called normal thresholds are directly associated with unitarity. They originate from cutting (to be made more precise when discussing the Cutkosky rules) the diagram into two pieces and generalise the equal size optical theorems cuts. Cutting a diagram into two pieces is equivalent to the combinatorial problem of grouping the external momenta into two sets. Tab. \[tab:cuts\] provides the overview of the number of cuts versus number of independent kinematic variables. For the two lowest functions there are no constraints whereas for all higher point functions there are constraints due to momentum conservation. For the $4$-point functions this constraint is known as the famous *Mandelstam constraint* $$\label{eq:MSc}
\sum_{i=1}^4 p_i^2 = s + t + u \;, \quad s = (p_1+p_2)^2 \;, \quad t = (p_1+p_3)^2 \;,
\quad u = (p_1+p_4)^2 \;,$$ where we choose conventions such that all momenta are incoming (momentum conservation reads $\sum_{i=1}^4 (p_i)_\mu =0$).
![Sketch of generic $2,3,4$-point function. The number of unitarity cuts and independent kinematic variables are listed in Tab. \[tab:cuts\].[]{data-label="fig:three"}](figs-pdf/fig5.pdf){width="4in"}
The fact that the amplitudes become functions of several complex variables makes matters more difficult. For example:
- From the Mandelstam constraint it can be seen that the unitarity cuts from one channel do appear on the negative real axis in the complex plane of another channel. Say the $u$-channel cuts, with forward kinematics $t=0$, do appear on the negative real axis in the complex plane of $s =
\sum_i m_i^2 - u $ ($p_i^2 = m_i^2$ on shell). These $u$-channel cuts in the $s$-plane are sometimes referred to as left-hand cuts as opposed to the proper $s$-channel unitarity cuts on the right-hand side of the complex plane (cf. Fig. \[fig:pacman\]).
- Several complex variables allow for dispersion relations in multiple channels. An example of which is the conjectured Mandelstam representation, cf. [@IZ; @S-matrix], which is a double dispersion relation.
- There are cuts which are not directly related to unitarity, so-called anomalous thresholds, cutting the diagram into more than 2 pieces. We will return to the latter briefly when discussing the Landau equations and Cutkosky rules.
$$\begin{array}{c | c | c | c }
n\text{-point function} & \text{\#cuts} & \text{\#variables} & \text{\#constraints} \\ \hline
2 & 1 & 1 & 0 \\
3 & 3 & 3 & 0 \\
4 & 7 & 6 & 1 \\
5 & 15 & 10 & 5
\end{array}$$
Landau Equations {#sec:LE}
----------------
Before stating the Landau equations it is useful to look at singularities of a one-variable integral representation where the integrand has pole singularities as a function of external parameters. The Landau equations originate from analysing this problem for the integrals of several variables appearing in PT. The presentation in this section closely follows the original paper [@Landau] and the textbook[@S-matrix].
### Singularities of one-variable Integral Representations {#sec:singular}
Consider the following integral representation of a analytic function $f(z)$ $$\label{eq:fz}
f(z) = \int_{{\gamma}_{ab}} g(z,w) dw \;,$$ where the integrand $g(z,w)$ contains pole singularities $w_i(z)$ which depend on $z$. The path ${\gamma}_{ab}$ ranges from a point $a$ to $b$ and does not cross any singularities for some $z= z_0$ as shown in Fig. \[fig:oneC\](left). The analytic properties of $f(z)$ depend on whether or not the path ${\gamma}_{ab}$ can be smoothly deformed away from approaching pole singularities $\omega(z_i)$.
![Path ${\gamma}_{ab}$ between endpoints $a$ and $b$. (left) Poles $\omega_{1,2}(z_0)$ of the integrand $g(z,\omega)$ in for $z=z_0$ (middle) Deformation of poles $\omega_{1,2}$ by the parameter $z_0$ to $z_1$. This is not provide an analytic continuation of the function $f(z)$ around $z_0$ to $z_1$. (right) deformation of path ${\gamma}_{ab}$ to ${\gamma}_{ab}^{(1)}$ serves as a (legitimate) analytic continuation of the function $f(z)$ in around $z_0$ to $z_1$.[]{data-label="fig:oneC"}](figs-pdf/fig6.pdf){width="6in"}
For example, if we start from $z=z_0$ and go to $z=z_1$ with $\omega(z_1)$ crossing the ${\gamma}_{ab}$ (cf. Fig. \[fig:oneC\](middle)) then the path ${\gamma}_{ab}$ can be smoothly deformed as in Fig. \[fig:oneC\](right) and this constitutes an analytic continuation of the function $f(z)$. There are though instances when this is not possible:
- When a singularity $w_i(z)$ approaches one of the endpoints $a$ or $b$; e.g. $w_i(z_1) = a$ Fig. \[fig:sing\](left). This case is known as an *endpoint singularity*.
- When two singularities approach each other, $w_1(z_1) = w_2(z_1)$ from different direction of the integration path as depicted in Fig. \[fig:sing\](right). This case is known as a *pinch singularity*.
- When the path needs to be deformed to infinity (can be reduced to case b).
In PT it is the pinch singularity type that gives rise to the singularities.
![(left) endpoint singularity (right) pinch singularity[]{data-label="fig:sing"}](figs-pdf/fig7.pdf){width="5in"}
### Landau Equations = several-variable Case
Landau [@Landau] and others (cf. [@S-matrix] for further references ) have analysed the problem of singularities, discussed for a single integral above, for the case of several variables in the context of Feynman graphs. A generic Feynman graph of $L$-loop of momenta $k_i$ ($i=1 \dots L$), $N$-propagators, external momenta $p_i$ (cf. Fig. \[fig:rep\](left) for a representative graph) can be written as follows $$\label{eq:I}
I = \int Dk \frac{1}{\prod_{i=1}^N (q_i^2 -m_i^2 + i 0)} \;, \qquad Dk = \prod_{i=1}^L d^4 k_i \;,$$ where $q_i = q_i(p_j,k_l)$ are the momenta of the propagators. By the technique of Feynman parameters (generalisation of $(A B)^{-1} = \int_0^1 d{\alpha}({\alpha}A + (1-{\alpha}) B)^{-2}$) one may rewrite $I$ as follows $$\label{eq:D}
I = \int Dk \int_0^1 D{\alpha}\frac{1}{(F+ i0)^N} \;, \qquad D {\alpha}= \prod_{i=1}^N d {\alpha}_i {\delta}(1 - \sum_{i=1}^N {\alpha}_i) \;,$$ where the crucial denominator $F$ reads $$\label{eq:F}
F = \sum_{i=1}^N {\alpha}_i (q_i^2 - m_i^2 ) \;.$$ Is seems worthwhile to emphasise that even though these formulae look rather involved they are completely straightforward.
![(left) Generic Feynman diagram aimed to clarify notation used in text (right) bubble graph discussed in the text[]{data-label="fig:rep"}](figs-pdf/fig8a.pdf "fig:"){width="2.5in"} ![(left) Generic Feynman diagram aimed to clarify notation used in text (right) bubble graph discussed in the text[]{data-label="fig:rep"}](figs-pdf/fig8b.pdf "fig:"){width="2.2in"}
The key idea is that there are different types of singularities depending on how many of the $N$ propagators are on shell, i.e. $q_i^2 - m_i^2$. It is the number of on-shell propagators which serves as a classification of the singularities. The Landau equations in condensed form are:[^13]
#### Landau equations/conditions
There are singularities if and only if $$\begin{aligned}
\label{eq:LE1}
& & \text{[i] either } q_i^2 = m_i^2 \text{ or } {\alpha}_i = 0 \;, \\[0.1cm]
\label{eq:LE2}
& & \text{[ii] } \sum_{i \in \text{loop}(l)} {\alpha}_i (q_i)^\mu =0 \text{ for } l=1 \dots L.\end{aligned}$$ We have assumed $q_i = k + \dots$ with $k$ being the loop momentum as otherwise one needs to introduce spurious minus signs into the sum above.
We are not going to show a proof of the Landau equations but state the result and argue for its plausibility below. Let us emphasise that the Landau equations neither tell us on which sheet the singularities are (cf. section \[sec:example\] for the refinement in this direction) nor how to compute the discontinuity relevant for the dispersion relations (cf. Cutkosky rules section \[sec:CR\]). The first condition assures that $F=0$ by demanding that each summand is zero in . The interpretation of $q_i^2 = m_i^2$ is of course that the corresponding propagator is on-shell and contributes to the singularity. Correspondingly ${\alpha}_i = 0$ means that the corresponding line does not enter the singularity. In Fig. \[fig:three-cuts\] we give an example of a 3-point function cut. The second condition has a geometric interpretation. It means that the corresponding singularity surfaces are parallel to each other and that the hypercontour can therefore not be deformed away from the approaching singularity surfaces. This is the analogy of the pinch singularity discussed in section \[sec:singular\]. Eq. can be cast into a more convenient form by contracting the equation by a vector $(q_j)_\mu$ which leads to the matrix equatons $$\label{eq:LE2b}
\text{[ii']} \quad Q \vec{{\alpha}} = 0 \;, \qquad (Q)_{ij} = q_i \cdot q_j \;, \quad (\vec{{\alpha}})_i = {\alpha}_i \;.$$ For non-trivial $\vec{{\alpha}}$ the second Landau equation is then solved by demanding that the so-called Cayley-determinant $\rm det Q =0$ vanishes.
### Landau Equation exemplified: 1-loop 2-point Function (bubble graph) {#sec:example}
Consider the bubble graph depicted in Fig. \[fig:rep\](right) with external momenta $p$, loop momenta $k$ and momenta $q_1 = k$ and $q_2 = k-p$. The first Landau equation tells us that \[As a homework you could ask yourself why the case ${\alpha}_1 = 0$ and $q_2^2 = m_2^2$ is not an option for a singularity\] $q_1^2 = m_1^2$ and $q_2^2 = m_2^2$. The second Landau equation can be cast into the form $\det Q =0$ $$\det Q = \det \left( \begin{matrix} m_1^2 & q_1 \cdot q_2 \\ q_1 \cdot q_2 & m_2^2 \end{matrix} \right)
= 0 \quad \Leftrightarrow q_1 \cdot q_2 = \pm m_1 m_2$$ which we may reinsert back into $p= q_1 + q_2$ which yields the two singularities $p_{(+)}^2 = (m_1+m_2)^2$ *and* $p_{(-)}^2 = (m_1-m_2)^2$, $$p^2 = (q_1 - q_2)^2 = q_1^2 - 2 q_1 \cdot q_2 + q_2^2 = (m_1 \mp m_2)^2 \;.$$ This might surprise us at first since from unitarity we expect there to be a branch point at $p_{(+)}^2$ but the point $p_{(-)}^2 < p_{(+)}^2$ has no place in this picture. The resolution comes upon recalling that the Landau equations inform us about the singularities but do not tell us on which sheet they are! In order to learn more we may solve $Q \vec{{\alpha}} = 0$ with $\vec{{\alpha}} = ({\alpha}, (1-{\alpha}))^T$ which gives $${\alpha}_{\pm} = \frac{m_2}{m_2 \pm m_1} \Rightarrow \quad 0 < {\alpha}_+ < 1 \;, \quad {\alpha}_- > 1 \text{ or } {\alpha}_- <0
\;,$$ for $m_{1,2} > 0$. From this we learn that ${\alpha}_+$ is within the integration region (recall $\int_0^1 D{\alpha}$ ) and $p_{(+)}^2$ is therefore on the physical sheet, wheras ${\alpha}_-$ is outside the integration region necessitating the deformation of the ${\alpha}$-contour. This indicates that ${\alpha}_-$ may not be on the physical sheet as the contour may crosses singularities in the course of deformation. The singularity $p_{(-)}^2 =
(m_1-m_2)^2$ is sometimes referred to as a pseudo threshold. These findings suggest an important refinement of the Landau conditions.
#### Refinement of Landau equations
For physical configuration, by which we mean the real external momenta, the Landau singularities are
- on the first (physical) sheet when ${\alpha}_i \in [0,1]$
- may or may not be on the first (physicial) sheet when ${\alpha}_i \notin [0,1]$
For non physical configuration, complex momenta, the situation is far from straightforward to say the least. The method of choice is often deformation to a case of a physical configuration and then deform to complex momenta checking whether or not singularities are crossed in that process. Crossing a singularity correspond to changing the Riemann sheet. Alternatively one can deform the masses to complex values keeping the ${\alpha}_i \in [0,1]$ and then deform back.
#### Geometric interpretation and the forgotten second type singularity
We take a detour and give a geometrical interpretation of the singularities of the bubble graph. The first and second Landau equations (\[eq:LE1\],\[eq:LE2\]) read $$\label{eq:LEB}
(k-p)^2 = m_1^2 \;, \quad k^2 = m_2^2 \;, \quad
{\alpha}k_\mu + (1-{\alpha}) (k-p)_\mu = 0 \;.$$ The first Landau equation defines two hyperboloids with centres displaced by $p$ with respect to each other. The second Landau equation assures that $p$ and $k$ are parallel. First we discuss the previously found solutions and then uncover the forgotten second type singularity.
- *Normal and pseudo threshold*\
Equations are satisfied when the two hyperboloids touch each other at the symmetric point with displacement in the $k_0$-direction. The displacement is given by $p_{(\pm)}^2 = (m_1 \pm m_2,0,0,0)^2 = (m_1\pm m_2)^2$ in the case where the hyperboloids open up in opposite and the same direction respectively. We refer the reader to Fig. \[fig:hyper\] (left) and (middle) for an illustration and the relevant equations in the caption.
- *Second type singularity*\
Fig. \[fig:hyper\] (right) shows yet another type of singularity. For any light-like displacement $p$, the two hyperboloids meet at infinity. These type of singularities were first noted by Cutkosky [@CR] who named them non-Landauian singularities whereas nowadays they are known as *second type singularities* [@S-matrix]. The Cayley-determinant is zero since $q_1 $ and $q_2 $ are parallel with $k$ and $p$ being light-like. If we parameterise $p/k={\epsilon}$ then the second Landau equation is satisfied when $ {\alpha}= (1+{\epsilon})/{\epsilon}$ which diverges when ${\epsilon}\to 0$ and in particular outside the $[0,1]$-interval. Hence the singularity may therefore not be on the physical sheet. More directly the singularity $p_{(0)}^2 = 0$ has no interpretation in terms of the spectrum. By the one-to-one relation of the spectrum and the 2-point function singularities (cf. Källén-Lehmann representation in section \[sec:2pt\]) it is to be concluded that this singularity is not on the physical sheet.
More generally second type singularities are determined by the vanishing of the Gram-determinant $\det p_i \cdot p_j = 0 $ (where the $p_i$ are the external momenta) [@S-matrix]. The singularity $p_{(0)}^2 = 0$ for the bubble-graph is then just the solution of the one-dimensional Gram-determinant. In summary second-type singularities are therefore independent of the masses and not on the physical sheet.[^14]
![Choosen $m_1= 3$ and $m_2 = 2$ in arbitrary units. (left) *Normal threshold:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = (m_1+m_2) - \sqrt{k_3^2 +m_2^2}$ and $p_+ = (m_1+m_2,0,0,0)$. (middle) *Pseudo threshold:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = (m_1-m_2) + \sqrt{k_3^2 +m_2^2}$ and $p_- = (m_1-m_2,0,0,0)$. (right) *Second type singularity:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = \sqrt{k_3^2 +m_2^2}$ and $p_{(0)} = (0,0,0,0)$ the two curves meet at infinity and are independent of the masses. It is obvious that any light-like $p$, e.g. $p_{(0)} = (x,0,0,x)$ for $x$ real, is also a valid parametrisation.[]{data-label="fig:hyper"}](figs-pdf/fig11.pdf "fig:"){width="2.0in"} ![Choosen $m_1= 3$ and $m_2 = 2$ in arbitrary units. (left) *Normal threshold:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = (m_1+m_2) - \sqrt{k_3^2 +m_2^2}$ and $p_+ = (m_1+m_2,0,0,0)$. (middle) *Pseudo threshold:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = (m_1-m_2) + \sqrt{k_3^2 +m_2^2}$ and $p_- = (m_1-m_2,0,0,0)$. (right) *Second type singularity:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = \sqrt{k_3^2 +m_2^2}$ and $p_{(0)} = (0,0,0,0)$ the two curves meet at infinity and are independent of the masses. It is obvious that any light-like $p$, e.g. $p_{(0)} = (x,0,0,x)$ for $x$ real, is also a valid parametrisation.[]{data-label="fig:hyper"}](figs-pdf/fig12.pdf "fig:"){width="2.0in"} ![Choosen $m_1= 3$ and $m_2 = 2$ in arbitrary units. (left) *Normal threshold:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = (m_1+m_2) - \sqrt{k_3^2 +m_2^2}$ and $p_+ = (m_1+m_2,0,0,0)$. (middle) *Pseudo threshold:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = (m_1-m_2) + \sqrt{k_3^2 +m_2^2}$ and $p_- = (m_1-m_2,0,0,0)$. (right) *Second type singularity:* $k_0 = \sqrt{k_3^2 + m_1^2}$ and $k_0 = \sqrt{k_3^2 +m_2^2}$ and $p_{(0)} = (0,0,0,0)$ the two curves meet at infinity and are independent of the masses. It is obvious that any light-like $p$, e.g. $p_{(0)} = (x,0,0,x)$ for $x$ real, is also a valid parametrisation.[]{data-label="fig:hyper"}](figs-pdf/fig13.pdf "fig:"){width="2.0in"}
#### Discussion upon using explicit 1-loop result:
After these abstract considerations we consider it advantageous to illustrate these type of singularities explicitly. The bubble graph $$B_0(p^2, m_1^2,m_2^2) =
\frac{ 4 \pi^2}{ i}\int \frac{d^4 k}{(2 \pi)^4 } \frac{1}{(k-p)^2 - m_1^2+i0)(k^2 - m_2^2+i0)} \;,$$ is UV-divergent. In order to avoid regularisation we may use the same trick as for the subtracted dispersion relation and take the difference with a fixed value. The results, valid on the physical sheet, reads (e.g. [@IZ]) $$\label{eq:explicit}
B_0(p^2,m_1^2,m_2^2) - B_0(\bar{p}^2,m_1^2,m_2^2) =
\frac{1}{2 p^2}\left[ X(p^2,m_1,m_2) -X(0,m_1,m_2) \right] - [ p^2 \leftrightarrow \bar{p}^2]
\;,$$ where $\bar{p}^2$ is the arbitrary subtraction point, $$X(p^2,m_1,m_2) = \sqrt{{\lambda}} \ln \left( \frac{ \sqrt{(m_1+m_2)^2 -p^2 } +
\sqrt{(m_1-m_2)^2 -p^2 } }{ \sqrt{(m_1+m_2)^2 -p^2 } -
\sqrt{(m_1-m_2)^2 -p^2 } } \right)^2$$ and ${\lambda}= ( p^2- (m_1+m_2)^2) (p^2- (m_1-m_2)^2) $ the Källén-function and not to be confused with a coupling constant.[^15] It is noted that the expression above is consistent with in the limit $m= m_1 =m_2$. We can learn three things from the representation . On the physical sheet: (i) there is a branch cut starting at $p^2 \geq (m_1+m_2)^2$ (normal threshold) (ii) there is no branch cut at $p^2 = (m_1-m_2)^2$ (pseudo threshold) and (iii) there is no singularity for $ p^2 \to 0$ (second type singularity). To see the correctness of (ii) one has to note that $\sqrt{{\lambda}}$ is imaginary for $(m_1 -m_2)^2 < p^2 < (m_1+m_2)^2$. In summary it is confirmed from the explicit representation that the pseudo threshold and the second type singularity are, indeed, not present on the physical sheet.
Cutkosky rules {#sec:CR}
--------------
The question of how to compute the actual singularities for physical configurations is answered by the cutting rules stated by Cutkosky [@CR] shortly after the Landau equations were formulated. This is by no means accidental as they are closely related. The Landau equations tell us that there is a singularity if either $q_i^2 = m_i^2$ or ${\alpha}_i = 0$ and the Cutkosky rules state that the corresponding singularity can be computed by replacing each on-shell (or cut propagator) $$\label{eq:CR1}
\frac{1}{q_i^2 - m_i^2 - i0 } \to - 2 \pi i {\delta}^{(+)}( q_i^2 - m_i^2) \;,$$ with the $ {\delta}^{(+)}(p^2-m^2) \equiv \delta(p^2-m^2)\theta(p_0)$-distribution. Before we motivate this rather elegant and surprisingly simple prescription let us state the result more explicitly.
The discontinuity of $I$ , for real momenta, with propagators $i = 1 \dots r \leq N$ cut is given by $$\label{eq:CR}
{\rm disc} [I] = (- 2 \pi i )^r \int Dk \frac{ \prod_{i=1}^r {\delta}^{(+)}(q_i^2 - m_i^2)}{\prod_{j=1}^{N-r} (q_{r+j}^2 - m_{r+j}^2)}
\;,$$ and is known as the Cutkosky or cutting rule!
Before trying to make plausible the formula let us state the obvious. The rule certainly gives the discontinuity of the propagator. The somewhat surprising fact is that this seems to be the recipe in any diagram. In the book of Peskin and Schröder [@PS] one can find the bubble graph evaluated in this way.
![(left) pinch singularity (right) equivalent path ${\gamma}_{ab} \sim {\gamma}_{ab}^1 + {\gamma}_{ab}^2$ with isolated pinch singularity[]{data-label="fig:CR"}](figs-pdf/fig10.pdf){width="5.6in"}
In order to motivate the Cutkosky rules we are going to sketch an argument given in the original paper [@CR] which is also reproduced in [@S-matrix]. One considers an integral representation of the form $$I(z) = \int dk_1^2 \frac{ F(k_1^2,z)}{k_1^2 -m_1^2 - i 0} \;,$$ where the variable $z$ is a function of the other momenta external and internal. Let the integrand $F$ contain a pole $w_1(z)$ which approaches $m_1^2$ for some $z$ such that there is going to be a pinch singularity as shown in Fig.\[fig:CR\](left). One then switches to the equivalent configuration where the contour is deformed below the mass $m_1^2$ at the cost of encircling the singularity $m_1^2$. In the next step the integral is performed using Cauchy’s theorem which is equivalent to replacing the denominator by ${\delta}(k_1^2-m_1^2)$. This argument falls short in justifying the additional physical condition $\theta((k_1)_0)$. Repeated use of the argument above, for each propagator gives the celebrated Cutkosky rules. The Cutkosky rules have recently been proven more rigorously [@Kreimer] using ideas and methods by Pham. For the case of normal thresholds (unitarity cuts only) the method of the largest time equation [@tHV] provides an elegant derivation of the cutting rules.
Anomalous Thresholds & physical Interpretation of Landau Equations {#sec:fun}
------------------------------------------------------------------
### Brief remarks on anomalous Thresholds {#eq:termin}
By studying the $1$-loop bubble graph, in section \[sec:normal\], we have encountered singularities of the normal & pseudo and 2nd type; cf. Tab. \[tab:singularity\]. The most important singularities for practical purposes (e.g. dispersion relation, experiments) are the singularities which are on the physical sheet of which the normal one is the only type so far. This poses the questions whether there are any singularities on the physical sheet. By the work of Källén-Lehmann we know that this is not the case for $2$-point function and we must therefore look at higher point functions.
There are indeed new classes of singularities for $3$- and higher point functions. A $3$-point function Fig. \[fig:three-cuts\](right), for example, can be cut into more than two pieces which is going beyond the singularities discussed so far. Putting all three propagators on shell corresponds to ${\alpha}_i \neq 0$ at the level of the Landau equations. The corresponding singularities are known as *anomalous thresholds*.[^16] It is obvious that the singularities in say $p_1^2$ depend on the values of $p_2^2 $ and $p_3^2$ (provided the line between the two is not contracted ${\alpha}_1 \neq 0$ as otherwise one encounters the singularities discussed so far).
Whether or not anomalous thresholds appear on the physical sheet, returning to our original questions, depends on the external momenta $p_{1,2}^2$. Below we mention examples where the anomalous thresholds appear on the physical sheet.
- Consider the $1$-loop version of the $3$-point function with $p^2 \equiv p_2^2 =p_3^2$, $m \equiv m_2 =m_3$ with masses as indicated in Fig. \[fig:three-cuts\](left). For this configuration there is an anomalous threshold in $p_1^2 \geq X $ with $X= 4 m^2 - (p^2 - (m^2 +m_1^2))^2/m_1^2$ provided the condition $p^2 > m^2 +m_1^2$ holds [@IZ]. $X$ is a branch point and the higher values are branch cuts. This anomalous threshold is below the two particle threshold at $4 m^2$ and might be regarded as the very reason for calling these thresholds anomalous![^17]
- For an example of a momentum configuration where the anomalous threshold is complex and on the physical sheet we refer the reader to the appendix in [@DLZ]. This anomalous threshold has to be taken into account in the dispersion relations by choosing the contour accordingly.
The corresponding $3$-point function serves as an example where the Schwartz reflection principle does not apply since the amplitude is imaginary on the entire real axis. In some more detail: the $1$-loop $3$-point function evaluated in PT does obey Schwartz’s reflection principle with no anomalous threshold on the first sheet. It is though not the correct analytic continuation into the lower half-plane. Crucially, after elimination of an unphysical branch cut on the real line Schwartz’s reflection principle is not obeyed anymore allowing for the anomalous threshold to appear on the physical sheet in the lower half plane.
$$\begin{array}{l || l | l | l | l }
\text{singularity} & \text{normal} & \text{pseudo} & \text{2nd type} & \text{anomalous} \\ \hline \hline
\text{physical sheet} & \text{yes} & \text{no} & \text{no} & \text{$p_{2,3}^2$-dependent} \\
\text{remark} & \text{unitarity} & \text{ } & \text{mass-indep.} & \text{Leading Landau}
\end{array}$$
We end this section with a miscellaneous remarks on terminology and practicalities.
- Anomalous thresholds go beyond the concept of unitarity cuts in that they allow for cutting the diagram into more than two pieces. Cutkosky [@CR] was well aware of this and introduced the term *generalised unitarity* in the context of his equations.
- The number of propagators that are put on-shell (in a loop) give rise to a natural classification. The singularity with the maximal number of on-shell propagators (all ${\alpha}_i \neq 0 \Rightarrow \det Q =0$) is usually referred to as the *leading Landau singularity*. In the case of the triangle diagram the anomalous threshold is the leading Landau singularity.
- When all but one external momenta are kept below the thresholds there are only normal thresholds on the physical sheet in the corresponding four momentum squared. For analytic computations this constitutes a method for obtaining correlation functions in a certain kinematic range. The remaining domain is then obtained by analytic continuation.[^18]
For the assessment of dispersion relation, outside the range of concrete computations, this is not a practical method since one needs to know the location of all singularities on the physical sheet in order to choose a path $\gamma$ which does not cross any singularity.
![(left) normal threshold cut (non-leading singularity) (middle) the corresponding reduced graph for which ${\alpha}_1 = 0$ in the Landau equations. From the reduced graphs all lower singularities can be found. (right) the leading Landau singularity corresponding to a so called anomalous threshold (non-normal threshold)[]{data-label="fig:three-cuts"}](figs-pdf/fig9.pdf){width="5.6in"}
### Physical Interpretation of the second Landau Equations (\[eq:LE2\],\[eq:LE2b\])
For physical momenta Coleman and Norton have given an interpretation of the second Landau equation [@CN]. It is found that (\[eq:LE2\],\[eq:LE2b\]) ensures that the corresponding diagram can occur as a real process where the Feynman-parameter ${\alpha}_i \sim \tau_i/m_i$ has the interpretation of being the proper time $\tau_i$ divided by the mass of the propagating particle.
This means that ${\alpha}_i =0$ corresponds to the case where the $i^{\rm th}$ particle does not propagate at all and gives the reduced graphs (e.g. Fig. \[fig:three-cuts\](middle)) a more direct meaning. At last we note that the Coleman-Norton interpretation is a very reassuring result in view of the optical theorem’s statemant that the discontinuity (related to singularities), of the forward scattering amplitude, originates from physical intermediate states.
Outlook
=======
Even though dispersion relations are an old subject and a pure dispersive approach to particle physics has proven to be too complicated in practice, dispersion theory is and will remain a powerful tool in QFT as it follows from first principles and is intrinsically non-perturbative. This makes it particularly useful for hadronic physics which is not directly accessible by a perturbation expansion in the strong coupling constant. Dispersion relations are the most solid approach to quark hadron duality. Any approach to quark hadron duality should either start from or connect to dispersion relations.
In recent years dispersion relations and unitarity methods have also seen a major revival in evaluating perturbative diagrams (e.g. [@Dixon; @SAbook; @Henn; @RT] for reviews and applications). The bootstrapping programme has witnessed new exciting developments by limiting/fixing the target space functions of amplitudes. Old tools such as the Steinman relations, which are physical conditions on double discontinuities, have led to promising simplifications valid beyond perturbation theory [@boot].
Furthermore dispersion relations can serve to prove positivity, for example, when a physical quantity can be expressed as an unsubtracted dispersion integral with positive integrand (discontinuity). Examples are the so-called $c$- and $a$-theorems, which characterise the irreversibility of the renormalisation group flows in 2D and 4D. The dispersive proofs are given in [@Cappelli:1990yc; @KS11] in two and four dimensions by looking at two and four-point functions respectively. On another note, positivity of the Källén-Lehmann representation seemed to exclude the possibility of asymptotically free gauge theories in 1970 [@Wilson3]. The Faddeev-Popov ghosts (negative metric) proved to be the loophole in this argument as they give rise to the negative sign of the ${\beta}$-function [@AF1; @AF2].
I am grateful to James Gratrex for proofreading and to Einan Gardi for discussions on second type singularities. Apologies for all relevant references that were omitted. Last but not least I would like to thank the organisers of the “Strong Fields and Heavy Quarks" as well as the participants for a stimulating and pleasant atmosphere. I really did enjoy my trip to and around Dubna!
The Schwartz Reflection Principle {#app:Schwartz}
=================================
Consider an analytic function $f(z)$ with $f(z) \in \mathbb{R}$ for $z \in I_R$ where $I_R$ is an interval on the real line. Then the following relation holds $$\label{eq:2c}
f(z) = f(z^*)^* \;,$$ which can be analytically continued to the entire plane. Note that analytic continuation is unique from any set with an accumulation point for which an interval is a special case. Hence Eq. implies that $${\rm Re}[ f(z)] = {\rm Re}[ f(z^*)] \;, \quad {\rm Im}[ f(z)] = - {\rm Im}[ f(z^*)] \;.$$ Choosing $z = s + i0$ with $s \in \mathbb{R}$ it then follows that $$\label{eq:Schwartz}
{\rm disc}[ \Gamma(s)] = 2i {\rm Im}[ \Gamma(s)] \;,$$ which is a result known from experience with 2-point functions and intuitively in accordance with the optical theorem.
Conventions {#app:conventions}
===========
Here we summarise a few conventions. We are using the Minkowski metric of the form $$g_{\mu \nu} = {\rm diag}(1,-1,-1,-1) \;,$$ the following abbreviations for integrals over space and momentum space $$\int_x = \int d^4 x \;, \quad \int_k = \int d^4 k \;,$$ and the relativistic state normalisation $${\langle p|p' \rangle} =2 E_p (2 \pi)^3 {\delta}^{(3)}(\vec{p} - \vec{p'}) \;,$$ where $E_p = \sqrt{\vec{p}^2 + m^2}$ with $p = (E_p ,\vec{p})$.
[^1]: Based on a three-hours blackboard lecture given at the school in Dubna, Russia 18-20 July 2016 “Strong fields and Heavy Quarks”. Lectures to appear in upcoming proceedings of the school. Added discussion of anomalous thresholds and second type singularities.
[^2]: \[foot:inter\] The LSZ formalism, in its elegancy and efficiency, also allows for the description of composite particles. For example a pion of $SU(2)$-isospin quantum number $a$ may be described by $\phi \to \varphi^a = \phi \bar q T^a {\gamma}_5 q$ in the sense that ${\langle 0| \varphi^a |\pi^b\rangle} = g_\pi {\delta}^{ab}$. In such a case $\varphi^a$ is referred to as an interpolating field.
[^3]: It is crucial that this condition is only imposed on the matrix element (weak topology) as otherwise one runs into Haag’s theorem [@Haag] which states that any field which is unitarity equivalent to a free field is itself a free field.
[^4]: In perturbation theory (PT) the reality of the momenta is implicitly used when shifting momenta (e.g. completing squares for example).
[^5]: In it’s most standard formulation string theory is first quantised and does not allow this analytic continuation. String field theory does exist but is less developed than first quantised string theory for technical reasons.
[^6]: We will come back to these interchanges which are ill-defined when there are UV-divergences.
[^7]: More precisely the 2-point function is at first defined for real $p^2 + i0$ with $p^2 \in \mathbb{R}$. Analytic continuation which is unique from an interval proceeds through the upper half-plane to the left and passes below zero for real $p^2$ below the singlarities on the positive real line.
[^8]: When the particle becomes unstable and acquires a width then the pole wanders on the second sheet since the principle that there are no singularities on the physical sheet holds up e.g. [@S-matrix]. This would have been an interesting additional topic which we can unfortunately not cover in these short lectures.
[^9]: Without going into any details let us mention that it is in particular the inclusive decay rate and amplitudes of exclusive decays that are amenable to a dispersive treatment. It is the amplitude and not the rate that has the simple analytic properties. The inclusive case is special in that the rate can be written as an amplitude!
[^10]: The practitioner will notice the absence of from the orthogonality of the projectors $P_L P_R =0$ which necessarily arises in the a perturbative computation in the limit of massless quark.
[^11]: Note since we work in the massless limit the pion is massless as it is the goldstone boson of broken chiral symmetry $SU(2)_L \otimes
SU(2)_R \to SU(2)_V$. The spin parity quantum numbers $J^P$ of the $\pi$, $\rho$ and $a_1$ are $0^-,1^+,1^-$ respectively.
[^12]: Am important topic was the conjecture by Mandelstam of a double dispersion relation for $ 2 \to 2$ scattering (i.e. 4-point function) which was consistent with known results but never proven in all generality even in perturbation theory. From this the Froissart bound was derived which states that the cross section for the scattering of two particles cannot grow faster than $\ln^2 s$ (where $s$ is the centre of mass energy).
[^13]: The Landau equations enjoy an interpretation in terms of electric circuits since the equations are analoguous to Kirchoff’s equations cf. [@S-matrix] and references therein.
[^14]: There also exist mixed type singularities where some loop momenta are pinched at infinity and others not. The reader is referred to the book [@S-matrix] for examples and discussion.
[^15]: From the viewpoint of the optical theorem the Källén-function arises from the phase space integration. In the rest-frame of the particle associated with the four momentum squared $p^2$, the absolute velocity of one of the decaying particles $|v_1| = |v_2|$ is related to the Källén-function as follows $|v_{1,2}| = \sqrt{{\lambda}}/2p^2$. Hence at this singularity the velocity is infinite consistent with the hyperboloids meeting at infinity.
[^16]: Their existence can be deduced from hermitian analyticity [@S-matrix] which in our case corresponds to the property that the imaginary part is proportional to the discontinuity.
[^17]: This is of relevance for form factors which, as we have seen, can be related to $3$-point functions. The anomalous threshold does appear for the electromagnetic form factors of the hyperons whereas for for the pion and kaons they do not since global quantum numbers do not allow the condition to be satisfied [@IZ]. Appearing and not appearing stands for being on the physical sheet or not.
[^18]: This makes use of the analyticity postulate of the S-matrix theory in the that the domain of analyticity is the maximal one consistent with unitarity (normal thresholds) and crossing symmetry. Crossing symmetry means that if scattering $A + B \to C +D$ and the decay $A \to \bar{B} +C +D$ are both possible then they are described by the same amplitude through analytic continuation. These postulates are seen to be correct in concrete QFT computations and believed to be true in general. Some results are known for $3$- and $4$-point function through the work of Källén [@three; @four] and even more generally from tools like the Edge of the Wedge theorem [@IZ; @PCT].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Explicit expressions for anharmonic contributions to the thermodynamic properties with allowance for higher-order phonon-phonon interactions for closed-packed crystals are given, and the calculations for some fcc metals near the melting (Ir, Rh) and martensite phase transition (Ca, Sr) points are carried out. A detailed comparison of anharmonic and electron contributions to the heat capacity of these metals is carried out. The computational results for high-temperature heat capacity agree well with the available experimental data.'
author:
- 'M. I. Katsnelson, A. F. Maksyutov$^\dag$, A. V. Trefilov$^\dag$'
title: Peculiarities of anharmonic effects in the lattice thermodynamics of fcc metals
---
Institute of Metal Physics, 620219 Ekaterinburg, Russia\
$^\dag$ Russian Research Centre “Kurchatov Institute”, 123182 Moscow, Russia\
Corresponding author: M. I. Katsnelson, e-mail: [email protected]
PACS: 63.20.-e
Keywords: anharmonic thermodynamics, martensite phase transition, heat capacity, computation
Anharmonic effects (AE) in the lattice dynamics and thermodynamics have been extensively discussed in the literature \[1-5\]. However the quantitative information on AE, both experimental and theoretical, is rather poor so far. The experimental difficulties are due to the fact that at high temperatures ($T\sim T_m$, where $T_m$ is the melting temperature) when the AE in the thermodynamic properties become noticeable, it is usually difficult to identify their contribution to the heat capacity and thermal expansion from the contributions of vacancies and other thermally excited lattice defects \[2,5-8\]. The theoretical calculations, on the other hand, are based, as a rule, on rough models and, therefore, do not give any definite information on the AE in thermodynamic of real crystals. At the same time this information is needed for understanding the high temperature properties of solids. For example, it is important to clear up what is the role of self-anharmonic effects associated with phonon-phonon interactions, and what is the role of quasiharmonic ones which are only due to the change in the volume because of thermal expansion. Simple estimations [@pai] show that the AE contribution to heat capacity at high temperatures of the order of the electron contribution to heat capacity and has the same (linear) temperature dependence; the real ratio of these two contributions essentially depend on the features of the electron structure near the Fermi level $E_F$. However this problem has not been studied yet. All these problems are of special importance for refractory metals ($T_m>2000$ K) and for the metals with high-temperature martensite transitions which are usually accompanied by essential increase in the AE in the lattice dynamics [@KTKK]. In the present work the AE features (in the comparison with the electron contributions) in the thermodynamical properties of refractory fcc metals Ir and Rh and in the bcc and fcc phases of Ca and Sr near the points of structural transformation are studied. The calculations were made in terms of microscopic models used in [@KTKK; @JL] which describe a wide range of the lattice properties of these metals.
The initial Hamiltonian of phonon subsystem is $$H=\sum_\lambda \left (\frac{P_{\lambda}^2}{2M} + \frac{M\omega_\lambda^2
Q_\lambda^2}{2} \right ) +\sum_{n=3}^\infty H^{(n)}
\label {eq1}$$ $$H^{(n)}=\sum_{\lambda_1\dots\lambda_n}
\frac{
\Phi^{(n)} (\lambda_1,\dots,\lambda_n)
}
{n!}
Q_{\lambda_1}\dots Q_{\lambda_n}
\label{eq2}$$ where $M$ is the mass of atoms; $\omega_\lambda$ is the phonon frequency, $\lambda\equiv{\bf q}\xi$ where ${\bf q}$ – is the wave vector; $\xi$ is the branch number, $Q_\lambda$, $P_\lambda$ are the normal coordinates in the harmonic approximation and the corresponding pulses, $\Phi^{(n)}$ are amplitudes of intermode interactions. In the leading order of the parameter $\eta\equiv T/E_{at}$ where $E_{at}$ is the energy of the order of binding energy, the basic contribution of AE to free energy $F$ is determined by the second order by $H^{(3)}$ and the first order by $H^{(4)}$ [@Cow; @VKT] $$F_{an}=F^{(3)}+F^{(4)}
\label{eq3}$$ where at $T\gg\Theta_D$ ($\Theta_D$ is the Debye temperature), $$F^{(3)}=- \frac{T^2}{12 M^3} \sum_{\lambda\mu\nu} \frac{|\Phi^{(3)}(\lambda,
\mu,\nu)|^2}{\omega_\lambda^2\omega_\mu^2\omega_\nu^2},
\label{eq4}$$ $$F^{(4)}=\frac{T^2}{4 M^2} \sum_{\lambda\mu} \frac{\Phi^{(4)}(\lambda,
\lambda,\mu,\mu)}{\omega_\lambda^2\omega_\mu^2};
\label{eq5}$$ as usually, summing over phonon quasimomenta is fulfilled taking into account the conservation law for $\lambda\equiv(\xi,{\bf k})$, $\mu\equiv(\eta,{\bf q})$, $\nu\equiv(\zeta,{\bf k+q})$. The explicit microscopic expressions for $\Phi^{(3)}$ and $\Phi^{(4)}$ are given in [@VKT].
The heat capacity of the crystal at $T\gg\Theta_D$, without the standard harmonic contribution $C_V^{harm}=3R$ (where $R$ is the gaseous constant), has the form: $$\Delta C=C_P-C_V^{harm}=(C_P-C_V)+C_V^{an}+C_V^{e}
+C_V^{d}
\label{eq6}$$ where $C_V^{an}$ are the anharmonic contributions to the heat capacity $$C_V^e(T)=\frac{R}{T}\int\limits_{-\infty}^\infty {\rm d}E
\left [-\frac{\partial f(E)}{\partial E} \right ]
(E-\mu)^2 N(E)
\label{eq7}$$ is the electron heat capacity ($N(E)$ is the electron density of states, is the chemical potential, $f(E)$ is the Fermi distribution function), $C_V^d$ is the contribution of defects; the calculation formula for $C_P-C_V$ in the quasiharmonic approximation is presented in [@VKT]. The density of states $N(E)$ was calculated by the FP-LMTO method in the local approximation for the density functional (LDA) [@savras]. To estimate the contribution of defects to the heat capacity a common approximation of independent monovacancies was used with $$C_V^d(T)=R\left(\frac{E_v}{T}\right)^2
\exp\left(S_v-\frac{E_v}{T}\right)
\label{eq71}$$ where $E_v$, $S_v$ are the energy and entropy of monovacancy formation, respectively.
The calculations of AE in the Ir lattice dynamics, made in [@JL], show that near the melting temperature $T=T_m$ the temperature dependences of the phonon frequency and phonon damping drastically change over the Brillouin zone reaching the values of the order of 15-20%. This value is high enough so that the question on the role of higher orders in the anharmonic perturbation theory could be set. The theory of self-consistent phonons, often used for this purpose (see, e.g., [@betger]) is not adequate in this case as it only accounts for a part of $H^{(2n)}$ contributions to (\[eq1\]) and neglects the $H^{(2n+1)}$ contributions. At the same time, the calculations [@JL] show that in Ir the contributions of three-phonon processes to the temperature shifts of the phonon frequencies always dominate over the contribution of four-phonon processes. In the recent work [@shukla] another approximation is proposed which accounts both the three-phonon and four-phonon processes, and, as stated, reasonably describes the thermodynamics of crystals with the Lennard-Jones type interaction. The applicability of this approximation in the case considered seems somewhat doubtful as it is based on the replacement of anharmonic frequency shift by a certain value average over the Brillouin zone while the calculations in [@JL] show that the [**q**]{}-dependence of AE is essential. Moreover, it is known that the approximations adequate for the description of thermodynamics of systems with “hard” potentials of Lennard-Jones type are not usually applicable for metals as they have relatively “soft” potentials [@VKT].
To calculate the effects of higher orders by the Hamiltonian $H^{(3)}$ the approximation similar to that of Gell-Mann and Brueckner (“bubble approximation”) in the theory of electron liquid [@GB] was used. It can be shown that the accurate expression for the contribution of three-phonon processes to the free energy at $T\gg\Theta_D$ has the form: $$F^{(3)}=-\frac{T}{3} \int\limits_0^1 \frac{d\alpha}{\alpha}
\sum_\lambda \left [G_\lambda(\omega=0)\Sigma_\lambda(\omega=0)
\right ]\biggr|_{H^{(3)}\to \alpha H^{(3)}}
\label{eq8}$$ where $G_\lambda$, $\Sigma_\lambda$ are the Green function and self-energy part for phonon (see [@Cow]) connected by the Dyson equation
$$G_\lambda^{-1}(\omega)=\frac{\omega^2-\omega_\lambda^2}{2\omega_\lambda}-\Sigma_\lambda(\omega).
\label{eq9}$$
The calculations [@JL] show that the AE in fcc metals have pronounsed peaks as functions of ${\bf q}$ at points X and L. The largest contribution to the free energy comes from the “bubble” diagrams; this situation is similar to the electron gas of high density with the only difference that in the latter case the “dangerous” region is ${\bf q}\approx0$ [@GB]. Taking this into account, in (\[eq8\])-(\[eq9\]) we may restrict ourselves to the expression of the lowest (second) order of the perturbation theory for $\Sigma_\lambda$. Then we have $$F^{(3)}=\frac{T}{6}\sum_\lambda \ln \left [1-\frac{T}{2 M^3}
\sum_{\mu\nu}
\frac{|\Phi^{(3)}(\lambda\mu\nu)|^2}{\omega_\lambda^2\omega_\mu^2\omega_\nu^2}
\right ]
\label{eq10}$$ Unlike the bcc metals the fcc metals have no soft modes in the phonon spectrum, which makes the four-phonon processes relatively small comparing with the three-phonon ones [@KTKK; @JL]. Therefore for $F^{(4)}$ it is sufficient to use the simplest corresponding to the Hartry-Fock approximation (the simplest procedure of disengagement in averaging Hamiltonian $H^{(4)}$). Then for $T\gg\Theta_D$ we have $$F^{(4)}=T\sum_\lambda \left [\ln \left (1+\frac{\Delta_{4\lambda}}
{\omega_\lambda} \right ) -\frac{1}{2}\frac{\Delta_{4\lambda}}
{\omega_\lambda} \right ],
\label{eq11}$$ where $$\Delta_{4\lambda} =\frac{T}{4 \omega_\lambda} \sum_\mu \
\frac{\Phi^{(4)} (\lambda,\lambda,\mu,\mu)} {M^2 \omega_\mu^2}
\label{eq12}$$ is the corresponding correction to the phonon frequency.
{height="110mm" width="80mm"}
-3mm [Fig. 1. Anharmonic contributions to the free energy of Ir and Rh (in K) considered by the perturbation theory (\[eq4\]), (\[eq5\]) (“old”) and with taking into account the higher orders (\[eq10\]), (\[eq11\]) (“new”).]{}
{height="110mm" width="80mm"}
-3mm [Fig. 2. Electron contribution to the heat capacity (in $R$ units) and temperature dependence of chemical potential (to $10^{-4}$ Ry) in Ir; the dependencies for Rh are similar.]{}
{height="110mm" width="80mm"}
-3mm [Fig. 3. Comparison of the electron and anharmonic contributions to the heat capacity (in $R$ units) for the fcc phases of K and Ir; $t=T/T_m$.]{}
{height="110mm" width="80mm"}
-3mm [Fig. 4. Comparison of the calculated values of heat capacity $C_P$ (formula (\[eq6\])) with the experiment for Ir; the data of exp1, exp2, exp3 are taken from [@truh; @ramana; @kats], respectively.]{}
{height="110mm" width="80mm"}
-3mm [Fig. 5. Theoretical value of heat-capacity in Ca. The notation introduced $C_v=C_v^{harm}+C_v^{an}+C_v^{el}$, $C_p=C_v^{harm}+C_v^{an}+C_v^{el}+(C_p-C_v)$, $C_v^f=C_v^{harm}+C_v^{an}+C_v^{free\ el}$, $C_p^f=C_v^{harm}+C_v^{an}+C_v^{free\ el}+(C_p-C_v)$, $C_v^{free\ el}$ – the electron heat capacity in the approximation of free electrons.]{}
The computational results are shown in Figs. 1-5. One can see that the AE higher order contributions to the free energy of Ir and Rh are not large (do not exceed 20%). In Ir and Rh the anharmonic contribution to heat capacity is small comparing with the electron contribution while in the sp-metals (as the hypothetical fcc phase K shown in Fig. 3) these contributions are comparable. Among the metals studied only for Ir the detailed experimental information is available, the theoretical data being in good agreement with these results (Fig. 4). In the estimation of $C_V^d$ according to (\[eq71\]) the the first principles calculation result from $E_v=1.88$ eV [@GKM] was used for $E_v$, and $S_v=2.5$ for $S_v$ [@VKT]. It follows from Fig. 5 that in the difference of heat-capacity for bcc and fcc phases of Ca (and similarly Sr) a strong compensation for different contributions – electron, anharmonic ones – and the contribution of $C_P-C_V$ takes place. Thus, for the discussion of thermodynamics of high temperature structure transformations in metals the microscopic calculations of corresponding values including the anharmonic effects are needed.
The authors are grateful to K. Yu. Khromov for his assistance in performance of calculations and M. N. Khlopkin for helpful discussion of experimental results.
The work is sponsored by the Russian Basic Research Foundation, grant 01-02-16108, and the Netherlands Research Organization, grant NWO 047-008-16.
[99]{} R. A. Cowley, Adv. Phys. [**12**]{} (1963) 421. V. G. Vaks, S. P. Kravchuk and A. V. Trefilov, J. Phys. F [**10**]{} (1980) 2325. M. Zoli, Phys. Rev. B [**41**]{} (1990) 7497. E. R. Cowley and R. C. Shukla, Phys. Rev. B [**58**]{} (1998) 2596. M. I. Katsnelson and A. V. Trefilov, Phys. Metal. Metallogr. [**91**]{} (2001) 109. Ya. A. Kraftmakher, Fiz. Tverdogo Tela [**15**]{} (1966) 571. L. A. Reznitski, Calorimetry of solids, (MSU Publ., Moscow, 1981). S. A. Kats, V. Ya. Chekhovskii, N. L. Koryakovski et. al. In: “Rare metals and alloys with single-crystal structure”, (Nauka, Moscow, 1981), p. 203. R. Peierls, Quantum theory of solids, (Clarendon Press, Oxford, 1955). M. I. Katsnelson, A. V. Trefilov, M. N. Khlopkin and K. Yu. Khromov, Phil. Mag. B [**81**]{} (2001) 1893. M. I. Katsnelson, A. V. Trefilov and K. Yu. Khromov. JETP Lett. [**69**]{} (1999) 688. S. Y. Savrasov, Phys. Rev. B [**54**]{} (1996) 16470. H. Boettger, Principles of the theory of lattice dynamics, (Acad. Verlag, Berlin, 1983). M. Gell-Mann and K. A. Brueckner, Phys. Rev. [**106**]{}(1957) 364; P. Nozieres and D. Pines, ibid, [**111**]{} (1958) 442. L. N. Trukhanova and L. N. Filippov, Teplofiz. Visokikh Temp., [**8**]{} (1970) 919. G. R. Ramanauskas, V. Ya. Chekhovskoi, V. D. Tarasov et. al., Zhurn. Fiz. Khim., [**61**]{} (1987) 1420. Yu. N. Gornostyrev, M. I. Katsnelson, N. I. Medvedeva et. al., Phys. Rev. B [**62**]{} (2000) 7802.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Training deep neural networks with only a few labeled samples can lead to overfitting. This is problematic in where only a few labeled samples are available. In this paper, we show that a consequence of overfitting in is feature distribution misalignment between labeled and unlabeled samples. Hence, we propose two new feature distribution alignment methods to reduce overfitting. Our methods are particularly effective when using only a small amount of labeled samples. Furthermore, we add consistency regularization to our adversarial alignment method and demonstrate that we always outperform pure consistency regularization and achieve particularly high improvements when using only a small amount of labeled samples. We test our method on CIFAR-10 and SVHN. On SVHN we achieve a test error of 3.88% (250 labeled samples) and 3.39% (1000 labeled samples) which is close to the fully supervised model 2.89% (73k labeled samples). In comparison, the current state-of-the-art achieves only 4.29% and 3.74%.'
author:
- |
Christoph Mayer Matthieu Paul Radu Timofte\
Computer Vision Lab, ETH Z[ü]{}rich, Switzerland\
[{chmayer,paulma,timofter}@vision.ee.ethz.ch]{}
bibliography:
- 'references.bib'
title: 'Adversarial Feature Distribution Alignment for Semi-Supervised Learning'
---
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'It is well-known that the Artin-Mazur dynamical zeta function of a hyperbolic or quasi-hyperbolic toral automorphism is a rational function, which can be calculated in terms of the eigenvalues of the corresponding integer matrix. We give an elementary proof of this fact that extends to the case of general toral endomorphisms without change. The result is a closed formula that can be calculated by integer arithmetic only. We also address the functional equation and the relation between the Artin-Mazur and Lefschetz zeta functions.'
address: 'Fakultät für Mathematik, Universität Bielefeld, Postfach 100131, 33501 Bielefeld, Germany'
author:
- Michael Baake
- Eike Lau
- Vytautas Paskunas
title: |
A note on the dynamical zeta function\
of general toral endomorphisms
---
Introduction
============
Any $d$-dimensional toral endomorphism is represented by an integer matrix, $M \in \operatorname{Mat}(d,{\mathbb{Z}})$, with action mod $1$ on the $d$-torus ${\mathbb{T}}^{d}\simeq{\mathbb{R}{\hspace{0.5pt}}}^{d}/{\mathbb{Z}}^{d}$; see [@AP] and [@KH Ch. 1.8] for background and [@AA Ex. 1.16] for an illustration. Important aspects of the dynamical system $({\mathbb{T}}^{d},M)$ are related to its periodic orbits and their distribution over ${\mathbb{T}}^{d}$; compare [@DEI; @W]. The Artin-Mazur [@AM] dynamical zeta function provides a generating function for the orbit counts that is interesting both from an arithmetic and from a topological point of view [@Fel; @R]. The latter was also Smale’s approach [@Smale], who related the Artin-Mazur and Lefschetz zeta functions of a *hyperbolic* toral automorphism and calculated both in terms of eigenvalues.
In this note, we explain a different approach via elementary geometry and linear algebra, which bypasses more advanced topological methods as well as the need to calculate eigenvalues. A key observation is that all arguments apply to general toral endomorphisms without additional effort. We also treat the connection between the Artin-Mazur and the Lefschetz zeta function and their functional equations. As we have learned along the way, most arguments we use appear already in the literature, notably in [@Fel], but at least their combination seems to be new. Also, we make several steps explicit to facilitate their computational use.
For $M\in \operatorname{Mat}(d,{\mathbb{Z}})$ and $m\geq 1$, let $a_m$ be the number of *isolated* fixed points in ${\mathbb{T}}^d$ of the $m$-th iterate $M^m$. The starting point of our considerations is the identity $$\label{fixcount}
a^{}_{m} = \lvert {\hspace{0.5pt}}\det({\mathbbm{1}}- M^m) \rvert {\hspace{0.5pt}}.$$ This formula is well-known [@W; @BHP] when no eigenvalue of $M^m$ is $1$; it then follows from counting the number of points of ${\mathbb{Z}}^{d}$ in a fundamental domain of the lattice $({\mathbbm{1}}- M^{m}){\hspace{0.5pt}}{\mathbb{Z}}^d$. Otherwise, is true because both sides are zero. Indeed, since the fixed points of $M^m$ form a closed subgroup of ${\mathbb{T}}^{d}$, they are either all isolated, or they form entire subtori of positive dimension; see [@BHP Appendix] for a detailed discussion of the subtorus case. Incidentally, when no eigenvalue of $M^m$ is $1$, $a^{}_{m}$ is also the Reidemeister number of a toral endomorphism, see [@Fel Thm. 22 and p. 33], while, in general, $a^{}_m$ is its Nielsen number.
Following [@AM], the Artin-Mazur zeta function of a general $M
\in \operatorname{Mat}(d,{\mathbb{Z}})$ is defined as $$\label{zeta-1}
\zeta^{}_{M} (z) := {\hspace{0.5pt}}\exp
\Bigl(\sum_{m=1}^{\infty} \frac{a^{}_{m}}{m} z^m \Bigr) =
\prod_{m=1}^{\infty} (1-z^m)^{-c_m} .$$ Here, the exponents $c^{}_m$ of the Euler product representation are well-defined integers, see Proposition \[prop:euler\] below. An explicit representation of $\zeta^{}_{M} (z)$ as a rational function is given below in Theorem \[thm:main\]. We stress that at least for hyperbolic or quasihyperbolic toral endomorphisms, this result is well-known by [@Smale] or [@Fel]; our focus is the elementary method.
A matrix $M$ is called *hyperbolic* when it has no eigenvalue on the unit circle ${\mathbb{S}}^{1}$. Such toral automorphisms are expansive [@W p. 143]. Note that $M$ may possess eigenvalues on ${\mathbb{S}}^{1}$ other than roots of unity (for instance, if one eigenvalue of $M$ is a Salem number; see [@Wad; @BRtorus] for examples). Integer matrices without roots of unity in their spectrum constitute the *quasihyperbolic* cases, compare [@Wad] and references therein, where formula still counts all fixed points. For quasihyperbolic matrices $M$, the exponents $c_m$ are the cycle numbers, which are related to the fixed point counts via $$\label{mobius}
a^{}_m = \sum_{\ell | m} \ell{\hspace{0.5pt}}c^{}_{\ell}
\quad \text{and} \quad
c^{}_{m} = \frac{1}{m} \sum_{\ell | m}
\mu\bigl(\frac{m}{\ell}\bigr)\, a^{}_{\ell} \, .$$ This follows from a standard application of M[ö]{}bius inversion; compare [@PW; @BRW].
When roots of unity are among the eigenvalues of $M$, the Euler product still exists, with the same relation between the counts of (isolated) fixed points and the exponents $c^{}_m$, though the latter can now be negative. Let us briefly illustrate this phenomenon in one dimension. Endomorphisms of ${\hspace{0.5pt}}{\mathbb{T}}^{1}\!\simeq {\mathbb{S}}^{1}$ are represented by multiplication (mod $1$) with an integer $n$. The dynamical zeta function reads $\zeta^{}_{0}(z)=1/(1-z)$ and $$\label{dim-one}
\zeta^{}_{n} (z) =
\frac{1-\operatorname{sgn}(n){\hspace{0.5pt}}z}{1-\lvert n \rvert {\hspace{0.5pt}}z} {\hspace{0.5pt}}$$ for $n\neq 0$, due to our Theorem \[thm:main\] below (or a simple direct calculation). For $n=-1$, we get $\zeta^{}_{-1} (z) =
(1-z^2)/(1-z)^2$, thus $c^{}_{1} = 2$ and $c^{}_{2}=-1$, while $c_{m}=0$ for all $m\ge 3$. The negative $c^{}_2$ corresponds to the fact that the two isolated fixed points of the map fail to be *isolated* for any even iterate, which is the identity.
Finally, let us note that our arguments extend to the case of nilmanifolds $X=G/\varGamma$ considered in [@Fel Sec. 2.6], where $G$ is a simply connected nilpotent Lie group and $\varGamma$ a discrete subgroup such that $X$ is compact. Namely, any endomorphism $\varphi$ of $\varGamma$ extends to an endomorphism $\tilde\varphi$ of $X$, the isolated fixed points of which are counted by $\big\lvert
\det \bigl({\mathbbm{1}}-\operatorname{Lie}(\tilde{\varphi}) \bigr)\big \rvert$, analogously to .
A related zeta function
=======================
Let us start with the numbers $\widetilde{a}_{m} := \det({\mathbbm{1}}- M^m)$, which can be viewed as signed fixed point counts, and the corresponding zeta function $$\label{til-series}
\widetilde{\zeta}^{}_{M} (z) = \exp \Bigl( \sum_{m = 1}^{\infty}
\frac{\widetilde{a}_{m}}{m} z^m \Bigr).$$ In Section \[sec:lef\], we will see that this is actually a Lefschetz zeta function, see Eq. below.
For $A\in\operatorname{Mat}(d,{\mathbb{R}{\hspace{0.5pt}}})$, let ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^k(A)$ be the induced linear map on the exterior power ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^{k} ({\mathbb{R}{\hspace{0.5pt}}}^d)$. In terms of the standard basis of that space, ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^k(A)$ is represented by the matrix of all minors of $A$ of order $k$; see [@Gant Ch. 1.4] for details. This is an integer matrix of dimension $\binom{d}{k}$, with ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^0(A) = 1$, ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^1(A) = A$ and ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^d(A) = \det (A)$.
\[pr-zeta-torus\] For $M\in \operatorname{Mat}(d,{\mathbb{Z}})$, we have $\; \widetilde{\zeta}^{}_{M} (z) = \prod_{k=0}^{d}
\det \bigl( {\mathbbm{1}}- z {\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^k (M) \bigr)^{(-1)^{k+1}}$.
Since all ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^k(M)$ are integer matrices, $\widetilde{\zeta}^{}_{M}
(z)$ is a rational function with numerator and denominator in ${\mathbb{Z}}[z]$. It can be calculated by integer arithmetic alone (many algebraic program packages have the matrices of minors of arbitrary order $k$ as built-in functions). Also, since $0$ is never a root of the denominator, the series for $\widetilde{\zeta}^{}_{M}$ converges uniformly on sufficiently small disks around $0$.
This is analogous to [@Fel Lemma 17]: The assertion is immediate from the well-known formula in linear algebra $$\label{trace-formula-torus}
\det({\mathbbm{1}}-A)=\sum_{k=0}^{d} (-1)^{k}\operatorname{tr}({\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^{k}(A))$$ together with the power series identity $$\label{trace-identity}
\exp \Bigl(\sum_{m=1}^{\infty} \frac{\operatorname{tr}(A^{m})}{m}
\, z^m \Bigr) = \frac{1}{\det ({\mathbbm{1}}- z A)} {\hspace{0.5pt}},$$ which is omnipresent in connection with zeta functions of any kind (in particular, it appears in the calculation of dynamical zeta functions of shifts of finite type, see [@BL; @R]). We recall that is proved by evaluating the characteristic polynomial of $M$ at $1$, while is a simple consequence of the relation $\det \bigl( \exp (C)\bigr) = \exp\bigl(
\operatorname{tr}(C)\bigr)$ for square matrices $C$, together with the Taylor series for $-\log(1-z)$, which is the case $d=1$ of .
Let us also note that, in terms of the $d$ eigenvalues $\lambda^{}_{1}, \ldots ,\lambda^{}_{d}{\hspace{0.5pt}}$ of $M$, one has the relation $ \det \bigl( {\mathbbm{1}}- z {\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^k (M) \bigr) = P^{}_{k} (z)$ with the polynomials $P^{}_{0} (z) = 1-z$ and $$P^{}_{k} (z) = \prod_{1\le \ell^{}_{1} < \ell^{}_{2} <
\cdots < \ell^{}_{k} \le d} (1 - z {\hspace{0.5pt}}\lambda^{}_{\ell^{}_{1}}\!
\cdot\ldots\cdot \lambda^{}_{\ell^{}_{k}} )$$ for $1\le k\le d$. This version is useful for the derivation of the functional equation of $\widetilde{\zeta}^{}_{M}$.
\[functional\] If $M\in\operatorname{Mat}(d,{\mathbb{Z}})$ with $D:= \det (M) \ne 0$, one has $\, \widetilde{\zeta}^{}_{M} (1/Dz) = B {\hspace{0.5pt}}\bigl(
\widetilde{\zeta}^{}_{M} (z) \bigr)^{(-1)^{d}}$, where $B=D$ for $d=1$ and $B=1$ otherwise.
First, a direct calculation shows that $$P^{}_{k} \Bigl( \frac{1}{Dz} \Bigr) = \frac{1}{\beta^{}_{d-k}}
\Bigl( \frac{-1}{z} \Bigr)^{\binom{d}{k}} P^{}_{d-k} (z){\hspace{0.5pt}},$$ where $\beta^{}_{0} = 1$ and $\beta^{}_{k} = D^{\binom{d-1}{k-1}}$ for $1\le k\le d$. Note that each prefactor $\beta^{}_{k}$ involves products of eigenvalues, but is symmetric in them and thus simplifies to a power of the determinant.
Next, recall the binomial formula $\sum_{\ell=0}^{n} \binom{n}{\ell}
(-1)^{\ell} = \delta_{n,0}$ for $n\ge 0$, and insert the previous polynomial identities into the product expression of Proposition \[pr-zeta-torus\]. Our claim follows, because the prefactor that contains $z$ disappears by an application of the binomial formula, while the prefactor with the determinants simplifies to the factor $B$ by an analogous calculation; compare [@Fel Lemma 19] and its proof for a related argument.
The special situation for $d=1$ is also immediate from $\widetilde{\zeta}^{}_{n} (z) = (1-n z)/(1-z)$, as the determinant is $n$; compare the example in the introduction.
The Artin-Mazur zeta function
=============================
To derive a formula for the dynamical zeta function, we observe that $a_{m}= \widetilde{a}_{m} \operatorname{sgn}(\widetilde{a}_{m})$. Hence the signs of all nonzero $\widetilde{a}_{m}$ need to be determined. When $M$ is quasihyperbolic, this is done in [@Wad Lemma 2.1], see also the proof of [@Fel Lemma 15], but the argument works for general $M\in\operatorname{Mat}(d,{\mathbb{Z}})$, too: We employ the formula $$\label{eigen-1}
\widetilde{a}_{m} = \det({\mathbbm{1}}-M^{m}) =
\prod_{j=1}^{d} (1 - \lambda_{j}^{m})$$ with the $\lambda^{}_{j}$ as above. It is clear that neither complex eigenvalues play a role (as they come in complex conjugate pairs, and $(1 - \lambda^{m}) (1 - \bar{\lambda}^{m}) = \lvert 1 - \lambda^{m}
\rvert^2 \ge 0$), nor do eigenvalues $\lambda \in [-1,1]$ (because then $1 - \lambda^m \ge 0$). The remaining eigenvalues (evs) matter, and one finds $$\label{signs}
a_{m} \, = \, \widetilde{a}_{m} \,
\bigl( (-1)^{\# \text{ real evs } < -1}\bigr)^{m}
(-1)^{\# \text{ real evs outside } [-1,1] }
\, =: \, \widetilde{a}_{m} \, \delta^{{\hspace{0.5pt}}m} \, \varepsilon .$$
Inserting this into $\zeta^{}_{M} (z)$ and comparing with $\widetilde{\zeta}^{}_{M} (z)$ gives $$\label{zeta-4}
\zeta^{}_{M} (z) = \bigl({\hspace{0.5pt}}\widetilde{\zeta}^{}_{M}
(\delta{\hspace{0.5pt}}z) \bigr)^{\varepsilon}$$ for the dynamical zeta function of $M$. Though involves the eigenvalues of $M$, the signs $\delta$ and $\varepsilon$ can once again be obtained by integer arithmetic alone. When no eigenvalue of $M$ is $\pm 1$, they are simply given by $$\label{signs-1}
\delta = \operatorname{sgn}\bigl( \det ({\mathbbm{1}}+ M) \bigr)
\quad \text{and} \quad
\varepsilon = \delta \operatorname{sgn}\bigl( \det ({\mathbbm{1}}- M)\bigr) .$$ In general, the signs can be defined by the one-sided limits $$\delta = \lim_{\alpha\searrow {\hspace{0.5pt}}0}\operatorname{sgn}\bigl( \det ( (1+\alpha){\mathbbm{1}}+ M ) \bigr)
\quad \text{and} \quad
\varepsilon = \delta \lim_{\alpha\searrow {\hspace{0.5pt}}0} \operatorname{sgn}\bigl( \det ((1+\alpha){\mathbbm{1}}- M)\bigr) ,$$ which can be evaluated explicitly as follows. Factorise $\det (x{\mathbbm{1}}- M) = (x-1)^{\sigma} (x+1)^{\tau} Q(x)$ with $Q\in{\mathbb{Z}}[x]$ and $Q(\pm
1)\neq 0$, where the non-negative integers $\sigma,\tau$ are unique. This implies $\det (x{\mathbbm{1}}+M) = (x-1)^{\tau} R(x)$ with $R\in
{\mathbb{Z}}[x]$ and $R(1)\ne 0$. Consequently, one has $$\label{signs-2}
\delta = \operatorname{sgn}\biggl( \frac{\det (x{\mathbbm{1}}+ M)}
{(x-1)^{\tau}} \bigg|_{x=1} \biggr)
\quad \text{and} \quad
\varepsilon = \delta \operatorname{sgn}\biggl( \frac{\det (x {\mathbbm{1}}- M)}
{(x-1)^{\sigma}} \bigg|_{x=1} \biggr),$$ which is used in our sample program in the appendix.
Let us summarise the result of our derivation so far.
\[thm:main\] Consider a general toral endomorphism, represented by a matrix $M\in\operatorname{Mat}(d,{\mathbb{Z}})$. The associated Artin-Mazur zeta function, defined in terms of isolated fixed points, satisfies $$\zeta^{}_M(z)=\prod_{k=0}^d\det \bigl({\mathbbm{1}}- \delta
z{\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^k(M)\bigr)^{\varepsilon{\hspace{0.5pt}}(-1)^{k+1}},$$ where the signs $\delta$ and $\varepsilon$ are given by Eq. when $\pm 1$ is not an eigenvalue of $M$, and by Eq. in general. In particular, $\zeta^{}_{M} (z)$ is a rational function. When no eigenvalue of $M$ is a root of unity, all fixed points are covered this way.
For quasihyperbolic toral endomorphisms, this result follows from an analogous result for nilmanifolds [@Fel Thm. 45], which is proved by Reidemeister-Nielsen fixed point theory, while the case of hyperbolic toral automorphisms is already treated in [@Smale Prop. 4.5] in a slightly different formulation. Theorem \[thm:main\] covers the special cases of automorphisms for $d=2$ from [@KH; @DEI; @BRW]. Let us note that, in the hyperbolic case, the rationality of $\zeta^{}_{M}$ can also be seen as a consequence of the general rationality result [@Man] proved by Markov partitions. Of related interest is the approach of [@Miles], which connects the problem to an interesting class of ${\mathbb{Z}}^d$-actions.
Since $\delta$ and $\varepsilon$ are signs, in particular $\delta
=1/\delta$, the functional equation for $\zeta^{}_{M}$ is now immediate from Lemma \[functional\] and Theorem \[thm:main\].
When $D:= \det (M) \ne 0$, one has $\, \zeta^{}_{M} (1/Dz) =
B^{\varepsilon}_{} {\hspace{0.5pt}}\bigl( \zeta^{}_{M} (z) \bigr)^{(-1)^{d}}$ with $\varepsilon$ from Eq. , where $B=D$ for $d=1$ and $B=1$ otherwise.
Let us also mention that $$\label{genfun}
\sum_{m=1}^{\infty} a_{m} {\hspace{0.5pt}}z^m =
\frac{z \, \zeta^{{\hspace{0.5pt}}\prime}_{M} (z)}{\zeta^{}_{M} (z)}$$ is the ordinary power series generating function of the sequence $(a_{m})^{}_{m\in{\mathbb{N}}}$, which is still a rational function. Its radius of convergence $\varrho^{}_{M}$ is always positive (it is the absolute value of the smallest root of the denominator of in reduced form). Thus, when $\lim_{m\to\infty} \frac{a_{m+1}}{a_{m}}$ exists, $1/\varrho^{}_{M}$ is the asymptotic growth rate of the fixed point counts, which provides a simple alternative to the approach in [@Wad]. The limit exists precisely for hyperbolic (and hence expansive) endomorphisms, as follows from [@CEW Thm. 6.3]. The ratio as a growth measure is also employed in [@Lehmer Thm. 16], where the case of unimodular roots is briefly discussed, too.
Finally, we observe that the Artin-Mazur zeta function of a general toral endomorphism can be written as an Euler product.
\[prop:euler\] For any matrix $M\in\operatorname{Mat}(d,{\mathbb{Z}})$, the associated Artin-Mazur zeta function $\zeta^{}_M(z)$ has an Euler product representation with uniquely determined integers $c_n$.
The Euler product representation is equivalent to the relations . These define rational numbers $c_n$ which we must show to be integers. When $M$ is quasihyperbolic, this is true by their geometric interpretation as cycle numbers. The general case (including $d=1$, which also follows from ) can be proved by a deformation argument as follows.
Fix $n$ and recall that $c_n$ is linear in the $a_k$, while $a_k=\delta^k {\hspace{0.5pt}}\varepsilon\det({\mathbbm{1}}-M^k)$ from with the signs $\delta$ and $\varepsilon$. Hence, $c_n$ is a polynomial with rational coefficients in the entries of $M$, which can be written as $c_n=\varepsilon {\hspace{0.5pt}}P(M)$. Here, $P$ itself depends on $\delta$, but neither on $\varepsilon$ nor on $M$. For another matrix $M^{{\hspace{0.5pt}}\prime}$, let $\delta^{{\hspace{0.5pt}}\prime}$, $\varepsilon^{{\hspace{0.5pt}}\prime}$ and $c_{n}^{{\hspace{0.5pt}}\prime}$ denote the associated signs and numbers. When $\delta^{{\hspace{0.5pt}}\prime}=\delta$, we thus have $c^{{\hspace{0.5pt}}\prime}_{n}=\varepsilon^{{\hspace{0.5pt}}\prime}{\hspace{0.5pt}}P(M^{{\hspace{0.5pt}}\prime})$.
Let $\nu > 1$ be a common denominator of all coefficients of $P$ and define the diagonal $d\!\times\!d$-matrix $N=\operatorname{diag}(2\delta,3,4,\ldots)$, which shares the sign $\delta$ with $M$. Consider now $M^{{\hspace{0.5pt}}\prime} = M+\nu^{r} N= \nu^{r}
(N+\nu^{-r}M)$. Since $N$ has distinct real eigenvalues, the eigenvalues of $N+\nu^{-r}M$, for sufficiently large $r$, are real and close to those of $N$. Then, $M^{{\hspace{0.5pt}}\prime}$ is hyperbolic with $\delta^{{\hspace{0.5pt}}\prime} = \delta$ by construction; in particular, $c^{{\hspace{0.5pt}}\prime}_{n}=\varepsilon^{{\hspace{0.5pt}}\prime} P(M^{{\hspace{0.5pt}}\prime})$ is integral. Since the difference $P(M)-P(M^{{\hspace{0.5pt}}\prime})$ is integral as soon as $r\geq 1$, it follows that $c_n=\varepsilon {\hspace{0.5pt}}P(M)$ is integral, too.
Interpretation as Lefschetz zeta function {#sec:lef}
=========================================
Suppose that $X$ is a compact differentiable manifold, assumed orientable for simplicity, and $f:\, X\longrightarrow X$ is some differentiable map. In this situation (and also more generally), there is a fixed point index $I_f\in{\mathbb{Z}}$ which satisfies the Lefschetz trace formula $$\label{lefschetz-trace-formula}
I_f=\sum_k(-1)^k\operatorname{tr}\bigl(f_*|H_k(X;{\mathbb{Q}{\hspace{0.5pt}}})\bigr),$$ see [@Dold Prop. VII.6.6] or [@Bredon Thm. 12.9]. Here, $H_k(X;{\mathbb{Q}{\hspace{0.5pt}}})$ denotes singular homology with coefficients in ${\mathbb{Q}{\hspace{0.5pt}}}$. It is a finite-dimensional ${\mathbb{Q}{\hspace{0.5pt}}}$-vector space in our situation, on which $f$ acts by functoriality.
When all fixed points of $f$ are isolated, we have $I_f=\sum_{x\in\operatorname{Fix}(f)}i_f(x)$, where $i_f(x)\in{\mathbb{Z}}$ is the local index of $f$ at $x$. If $x$ is a regular fixed point, meaning that $1$ is not an eigenvalue of the tangential map $T_x(f)$, the local index is given by $$i_f(x)=\operatorname{sgn}\bigl(\det({\mathbbm{1}}-T_x(f))\bigr) \, \in \, \{\pm 1\}.$$
The Lefschetz zeta function associated to $f$ can be defined as $$\zeta^L_f(z)=\exp\Bigl(\sum_{n\geq 1}\frac{z^n}{n}\, I_{f^n}\Bigr).$$ This definition seems to appear first in [@Smale]; see also [@Fel]. By using the identity , the trace formula applied to all iterates of $f$ implies that $\zeta^L_f (z)$ is a rational function, $$\label{lefschetz-zeta-formula}
\zeta^L_f (z)=\prod_{k}
\det({\mathbbm{1}}-zf_*|H_{k}(X;{\mathbb{Q}{\hspace{0.5pt}}}))^{(-1)^{k+1}}.$$
Let us now assume that $X={\mathbb{T}}^d={\mathbb{R}{\hspace{0.5pt}}}^d/{\mathbb{Z}}^d$ as above, and that $f$ is given by an arbitrary $M\in\operatorname{Mat}(d,{\mathbb{Z}})$. In this case, with the zeta function $\widetilde{\zeta}^{}_{M}$ of Proposition \[pr-zeta-torus\], we have $$\label{rel-equals-lef}
\widetilde\zeta^{}_{M}(z)=\zeta^L_{f}(z){\hspace{0.5pt}},$$ including a correspondence of all related formulas (for the case of hyperbolic toral automorphisms, this was noted in [@Fel p. 86, lines 7 and 16]). Let us sketch a possible line of argument. First, it is well-known that the Künneth formula [@Bredon Thm. 3.2] gives an isomorphism $$\label{homology-torus}
H_{k}(X;{\mathbb{Q}{\hspace{0.5pt}}})\cong{\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^{k}({\mathbb{Q}{\hspace{0.5pt}}}^d)$$ such that the action of $f$ on $H_k$ corresponds to ${\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^k(M)$. This identifies the right hand sides of and , and similarly for Proposition \[pr-zeta-torus\] and Eq. . It follows that the left hand sides of the corresponding pairs of equations are equal as well, that is $$\label{fixpoint-index-torus}
I_f=\det({\mathbbm{1}}-M) ,$$ and similarly $I_{f^n}=\det({\mathbbm{1}}-M^n)$; this appears also in [@Smale Prop. 4.15] and in [@BBPT].
A direct proof of without using the Lefschetz trace formula can be done as follows. Assume first that $1$ is not an eigenvalue of $M$. All fixed points $x$ of $f$ are then regular, with the same local index $i_f(x)=\operatorname{sgn}\bigl(\det({\mathbbm{1}}-M)\bigr)$, because the tangent space $T_x(X)$ can be identified with ${\mathbb{R}{\hspace{0.5pt}}}^d$, where the action of $f$ is given by $M$. Thus is immediate. When no eigenvalue of $M$ is a root of unity, the same applies to all iterates of $f$.
For arbitrary $M$, we may use the following lemma.
\[le-index\] Assume $X$ to be a compact Lie group of dimension $d$ and $f \! : \,
X\to X$ a differentiable map. Let $g(x)=x\cdot f(x)^{-1}$. Then, $g$ acts on the $1$-dimensional ${\mathbb{Q}{\hspace{0.5pt}}}$-vector space $H_d(X;{\mathbb{Q}{\hspace{0.5pt}}})$ by the scalar $I_f$.
Granting the lemma, follows easily because, in the torus case, $g$ is given by ${\mathbbm{1}}-M$, and we have $H_d(X;{\mathbb{Q}{\hspace{0.5pt}}})\cong{\text{\raisebox{1.5pt}{$\scriptstyle {\hspace{0.5pt}}\bigwedge\!$}}}^d({\mathbb{Q}{\hspace{0.5pt}}}^d)$. We only sketch the proof of Lemma \[le-index\] and leave the details to the reader. The fixed point index can be defined as the homology intersection product in $X\!\times\! X$ of the graph of $f$ and the diagonal, $I_f=[\Gamma_X]\cdot[\Delta]$. Since the automorphism of $X\times X$ given by $(x,y)\mapsto(x,x\cdot y^{-1})$ acts on the orientation by $(-1)^{d}$, we get $I_f=(-1)^d[\Gamma_g]\cdot[X\times\{1\}]$. The assertion follows by a straightforward computation based on decomposing $[\Gamma_g]$ according to the Künneth formula for $X\!\times\! X$.
Appendix: A sample program for calculating $\zeta^{}_{M}$ {#appendixa-sample-program-for-calculating-zeta_m .unnumbered}
=========================================================
One can implement the explicit zeta function formulas of Proposition \[pr-zeta-torus\] and Theorem \[thm:main\] in a simple <span style="font-variant:small-caps;">Mathematica</span>${}^{\!\mbox{\tiny\textregistered}}$ program as follows.
\
\
\
\
\
\
\
\
The input is an integer matrix, in the standard format of a double list. The calculation is exact and reasonably fast for small dimensions, and can be used up to dimension $8$ or $10$ say.
Acknowledgements {#acknowledgements .unnumbered}
================
It is a pleasure to thank Alex Fel’shtyn and Tom Ward for various helpful suggestions and Doug Lind, John A.G. Roberts and Rudolf Scharlau for discussions. This work was supported by the German Research Council (DFG), within the CRC 701.
[99]{}
V.I. Arnold and A. Avez, *Ergodic Problems of Classical Mechanics*, reprint, Addison-Wesley, Redwood City, CA (1989).
R.L. Adler and R. Palais, Homeomorphic conjugacy of automorphisms of the torus, *Proc. AMS* **16** (1965) 1222–1225.
M. Artin and B. Mazur, On periodic points, *Ann. Math.* **81** (1965) 82–99.
M. Baake, J. Hermisson and P.A.B. Pleasants, The torus parametrization of quasiperiodic LI-classes, *J. Phys. A: Math. Gen.* **30** (1997) 3029–3056; `mp_arc/02-168`.
M. Baake and J.A.G. Roberts, Symmetries and reversing symmetries of toral automorphisms, *Nonlinearity* **14** (2001) R1–R24; `arXiv:math.DS/0006092`.
M. Baake, J.A.G. Roberts and A. Weiss, Periodic orbits of linear endomorphisms on the $2$-torus and its lattices, *Nonlinearity* **21** (2008) 2427–2446; `arXiv:0808.3489`.
R. Bowen and O.E. Lanford III, Zeta functions of restrictions of the shift transformation, in: *Global Analysis*, vol. 14 of *Proc. Sympos. Pure Math.*, AMS, Providence, RI (1970), pp. 43–49.
G. Bredon, *Topology and Geometry*, Springer, Berlin (1993).
R.B.S. Brooks, R.F. Brown, J. Pak and D.H. Taylor, Nielsen numbers of maps of tori, *Proc. AMS* **52** (1975) 398–400.
V. Chothi, G. Everest and T. Ward, S-integer dynamical systems: periodic points, *J. Reine Angew. Math. (Crelle)* **489** (1997) 99–132.
M. Degli Esposti and S. Isola, Distribution of closed orbits for linear automorphisms of tori, *Nonlinearity* **8** (1995) 827–842.
A. Dold, *Lectures on Algebraic Topology*, Springer, Berlin (1972).
A. Fel’shtyn, *Dynamical Zeta Functions, Nielsen Theory and Reidemeister Torsion*, Memoirs AMS vol. 147, no. 699, AMS, Providence, RI (2000).
F. Gantmacher, *The Theory of Matrices*, Chelsea, New York (1959).
A. Katok and B. Hasselblatt, *Introduction to the Modern Theory of Dynamical Systems*, Cambridge University Press, Cambridge (1995).
D.H. Lehmer, Factorization of certain cyclotomic functions, *Ann. Math.* (2) **34** (1933) 461–479.
A. Manning, Axiom $A$ diffeomorphisms have rational zeta functions, *Bull. London Math. Soc.* **3** (1971) 215–220.
R. Miles, Zeta functions for elements of entropy rank one actions, *Ergod. Th. & Dynam. Syst.* **27** (2007) 567–582.
Y. Puri and T. Ward, Arithmetic and growth of periodic orbits, *J. Integer Sequences* **4** (2001), paper 01.2.1.
D. Ruelle, *Dynamical Zeta Functions for Piecewise Monotone Maps of the Interval*, CRM Monograph Series, vol. 4, AMS, Providence, RI (1994).
D. Ruelle, Dynamical zeta functions and transfer operators, preprint IHES/M/02/66 (2002).
S. Smale, Differentiable dynamical systems, *Bull. AMS* **73** (1967) 747–817.
S. Waddington, The prime orbit theorem for quasihyperbolic toral automorphisms, *Monatsh. Math.* **112** (1991) 235–248.
P. Walters, *An Introduction to Ergodic Theory*, reprint, Springer, New York (2000).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Generative models for 3D geometric data arise in many important applications in 3D computer vision and graphics. In this paper, we focus on 3D deformable shapes that share a common topological structure, such as human faces and bodies. Morphable Models and their variants, despite their linear formulation, have been widely used for shape representation, while most of the recently proposed non-linear approaches resort to intermediate representations, such as 3D voxel grids or 2D views. In this work, we introduce a novel graph convolutional operator, acting directly on the 3D mesh, that explicitly models the inductive bias of the fixed underlying graph. This is achieved by enforcing consistent local orderings of the vertices of the graph, through the spiral operator, thus breaking the permutation invariance property that is adopted by all the prior work on Graph Neural Networks. Our operator comes by construction with desirable properties (anisotropic, topology-aware, lightweight, easy-to-optimise), and by using it as a building block for traditional deep generative architectures, we demonstrate state-of-the-art results on a variety of 3D shape datasets compared to the linear Morphable Model and other graph convolutional operators.'
author:
- |
Giorgos Bouritsas [^1] ^1^ Sergiy Bokhnyak ^\*\ 2^ Stylianos Ploumpis^1,3^\
Michael Bronstein^1,2,4^ Stefanos Zafeiriou^1,3^\
Imperial College London, UK^1^ Universita Svizzera Italiana, Switzerland^2^ FaceSoft.io^3^ Twitter^4^\
^1^[{g.bouritsas18, s.ploumpis, m.bronstein, s.zafeiriou}@imperial.ac.uk]{} ^2^[[email protected]]{}
bibliography:
- 'egbib.bib'
title: 'Neural 3D Morphable Models: Spiral Convolutional Networks for 3D Shape Representation Learning and Generation'
---
{width="\linewidth"}
Introduction
============
The success of deep learning in computer vision and image analysis, speech recognition, and natural language processing, has driven the recent interest in developing similar models for 3D geometric data. Generalisations of successful architectures such as convolutional neural networks (CNNs) to data with non-Euclidean structure (e.g. manifolds and graphs) is known under the umbrella term [*Geometric deep learning*]{} [@bronstein2017geometric]. In applications dealing with 3D data, the key challenge of geometric deep learning is a meaningful definition of intrinsic operations analogous to convolution and pooling on meshes or point clouds. Among numerous advantages of working directly on mesh or point cloud data is the fact that it is possible to build invariance to shape transformations (both rigid and nonrigid) into the architecture, as a result allowing to use significantly simpler models and much less training data. So far, the main focus of research in the field of geometric deep learning has been on [*analysis*]{} tasks, encompassing shape classification and segmentation [@qi2017pointnet; @qi2017pointnet++], local descriptor learning, correspondence, and retrieval [@monti2017geometric; @boscaini2016learning; @masci2015geodesic].
On the other hand, there has been limited progress in representation learning and generation of geometric data (*shape synthesis*). Obtaining descriptive and compact representations of meshes and point clouds is essential for downstream tasks such as classification and 3D reconstruction, when dealing with limited labelled training data. Additionally, geometric data synthesis is pivotal in applications such as 3D printing, computer graphics and animation, virtual reality, and game design, and can heavily assist graphics designers and speed-up production. Furthermore, given the high cost and time of acquiring quality 3D data, geometric generative models can be used as a cheap alternative for producing training data for geometric ML algorithms.
Most of the previous approaches in this direction rely on intermediate representations of 3D shapes, such as point clouds [@achlioptas2017learning], voxels [@wu2016learning] or mappings to a flat domain [@moschoglou20193dfacegan; @ben2018multi] instead of direct surface representations, such as meshes. Despite the success of such techniques, they either suffer from high computational complexity (voxels) or absence of smoothness of the data representation (point clouds), while usually pre- and post-processing steps are needed in order to obtain the output surface model. Learning directly on the mesh was only recently explored in [@litany2018deformable; @coma; @wang2018pixel2mesh; @kolotouros2019convolutional] for shape completion, non-linear facial morphable model construction and 3D reconstruction from single images, respectively. In this paper, we propose a novel representation learning and generative framework for fixed topology meshes. For this purpose, we formulate an ordering-based graph convolutional operator, contrary to the permutation invariant operators in the literature of Graph Neural Networks. In particular, similarly to image convolutions, for each vertex on the mesh, we enforce an explicit ordering of its neighbours, allowing a “1-1” mapping between the neighbours and the parameters of a learnable local filter. The order is obtained via a spiral scan, as proposed in [@spirals], hence the name of the operator, *Spiral Convolution*. This way we obtain anisotropic filters without sacrificing computational complexity, while simultaneously we explicitly encode the fixed graph connectivity. The operator can potentially be generalised to other domains that accept implicit local orderings, such as arbitrary mesh topologies and point clouds, while it is naturally equivalent to traditional grid convolutions. Via this equivalence, common CNN practices, such as dilated convolutions, can be easily formulated for meshes.
We use spiral convolution as a basic building block for hierarchical intrinsic mesh autoencoders, which we coin *Neural 3D Morphable Models*. We quantitatively evaluate our methods on several popular datasets: human faces with different expressions (COMA [@coma]) and identities (Mein3D [@booth2018large]) and human bodies with shape ad pose variation (DFAUST [@dfaust:CVPR:2017]). Our model achieves state-of-the-art reconstruction results, outperforming the widely used linear 3D Morphable Model [@vetterMM] and the COMA autoencoder [@coma], as well other graph convolutional operators, including the initial formulation of the spiral operator [@spirals]. We also qualitatively assess our framework showing ‘shape arithmetic’ in the latent space of the autoencoder and by synthesising facial identities via a spiral convolution Wasserstein GAN.
Related Work
============
**Generative models for arbitrary shapes:** Perhaps the most common approaches for generating arbitrary shapes are **volumetric CNNs** [@wu20153d; @qi2016volumetric; @maturana2015voxnet] acting on 3D voxels. For example, voxel regression from images [@girdhar2016learning], denoising autoencoders [@sharma2016vconv] and voxel-GANs [@wu2016learning] have been proposed. Among the key drawbacks of volumetric methods are their inherent high computational complexity and that they yield coarse and redundant representations. **Point clouds** are a simple and lightweight alternative to volumetric representation recently gaining popularity. Several methods have been proposed for representation learning of fixed-size point clouds [@achlioptas2017learning] using the PointNet [@qi2017pointnet] architecture. In [@yang2018foldingnet], point clouds of arbitrary size can be synthesised via a 2D grid deformation. Despite their compactness, point clouds are not popular for realistic and high-quality 3D geometry generation due to their lack of an underlying smooth structure. **Image-based** methods have also been proposed, such as multi-view [@arsalan2017synthesizing] and flat domain mappings such as UV maps [@moschoglou20193dfacegan; @ben2018multi], however they are computationally demanding, require pre- and post-processing steps and usually produce undesirable artefacts. It is also worth mentioning the recently introduced **implicit-surface** based approaches [@mescheder2018occupancy; @chen2018learning; @park2019deepsdf], that can yield accurate results, though with the disadvantage of slow inference (dense sampling of the 3D space followed by marching cubes).\
**Morphable models:** In the case of deformable shapes, such as faces, bodies, hands , where a fixed topology can be obtained by establishing dense correspondences with a template, the most popular methods are still statistical models given their simplicity. For **Faces**, the baseline is the PCA-based 3D Morphable Model (3DMM) [@vetterMM]. The Large Scale Face Model (LSFM) [@booth2018large] was proposed for facial identity and made publicly available, [@cao2014facewarehouse; @FLAME:SiggraphAsia2017] were proposed for facial expression, while for the entire head a large scale model was proposed in [@ploumpis2019combining]. For **Body & Hand**, the most well known models are the skinned vertex-based models SMPL [@SMPL:2015] and MANO [@MANO:SIGGRAPHASIA:2017], respectively. SMPL and MANO are non-linear and require (a) joint localisation and (b) solving special optimisation problems in order to project a new shape to the space of the models. In this paper, we take a different approach introducing a new family of differentiable Morphable Models, which can be applied on a variety of objects, with strong (body) and less strong (face) articulations. Our methods have better representational power and also do not require any additional supervision.\
**Geometric Deep Learning** is a set of recent methods trying to generalise neural networks to non-Euclidean domains such as graphs and manifolds [@bronstein2017geometric]. Such methods have achieved promising results in geometry processing and computer graphics [@masci2015geodesic; @boscaini2016learning], computational chemistry [@duvenaud2015convolutional; @gilmer2017neural], and network science [@kipfGCN; @monti2017geometric]. Multiple approaches have been proposed to construct convolution-like operations, including spectral methods [@bruna2013spectral; @defferrard2016convolutional; @kipfGCN; @yi2017syncspeccnn], local charting based [@masci2015geodesic; @boscaini2016learning; @monti2017geometric; @fey2018splinecnn; @spirals] and soft attention [@velivckovic2017graph; @verma2018feastnet]. Finally, graph or mesh coarsening techniques [@defferrard2016convolutional; @ying2018hierarchical] have been proposed, equivalent to image pooling.
Spiral Convolutional Networks
=============================
Spiral Convolution {#spiral_theory}
------------------
For the following discussion, we assume to be given a manifold, discretised as a triangular mesh $\mathcal{M} = (\mathcal{V,E,F})$ where $\mathcal{V}=\{1, \dots, n\}$, $\mathcal{E}$, and $\mathcal{F}$ denote the sets of vertices, edges, and faces respectively. Furthermore, let ${f:V\rightarrow\mathbb{R}}$, a function representing the vertex features.
One of the key challenges in developing convolution-like operators on graphs or manifolds is the lack of a global system of coordinates that can be associated with each point. The first intrinsic mesh convolutional architectures such as GCNN [@masci2015geodesic], ACNN [@boscaini2016learning] or MoNet [@monti2017geometric] overcame this problem by constructing a [*local*]{} system of coordinates $\mathbf{u}(x,y)$ around each vertex $x$ of the mesh, in which a set of local weighting functions $w_1,\hdots, w_L$ is applied to aggregate information from the vertices $y$ of the neighborhood $\mathcal{N}(x)$. This allows to define ‘patch operators’ generalising the sliding window filtering in images: $$\label{soft_attn}
(f\star g)_x = \sum_{\ell=1}^L g_\ell \sum_{y\in\mathcal{N}(x)} w_\ell(\mathbf{u}(x,y)) f(y)$$ where $\sum_{y\in\mathcal{N}(x)} w_\ell(\mathbf{u}(x,y)) f(y)$ are ‘soft pixels’ ($L$ in total), $f$ are akin to pixel intensity in images, and $g_\ell$ the filter weights. The problem of the absence of a global coordinate system is equivalent to the absence of canonical ordering of the vertices, and the patch-operator based approaches can be also interpreted as attention mechanisms, as in [@velivckovic2017graph] and [@verma2018feastnet]. In particular, the absence of ordering does not allow the construction of a “1-1” mapping between neighbouring features $f(y)$ and and filter weights $g_\ell$, thus a “all-to-all” mapping is performed via learnable *soft-attention* weights $w_\ell(\mathbf{u}(x,y))$. In the Euclidean setting, such operators boil down to the classical convolution, since an ordering can be obtained via the global coordinate system.
Besides the lack of a global coordinate system, another motivation for patch-operator based approaches when working on meshes, is the need for *insensitivity to meshing* of the continuous surface, ideally, each patch operator should be independent of the underlying graph topology.
[0.35]{} ![Spiral ordering on a mesh and an image patch[]{data-label="spiral_fig"}](Spiral_on_Mesh.png "fig:"){width="\linewidth"} \[spiral\_mesh\]
[0.33]{} ![Spiral ordering on a mesh and an image patch[]{data-label="spiral_fig"}](Spiral_on_Grid.png "fig:"){width="\linewidth"} \[spiral\_image\]
However, all the methods falling into this family, come at the cost of high computational complexity and parameter count and can be hard to optimise. Moreover, patch-operator based methods specifically designed for meshes, require hand-crafting and pre-computing the local systems of coordinates. To this end, in this paper we make a crucial observation in order to overcome the disadvantages of the aforementioned approaches: the issues of the absence of a global ordering and insensitivity to graph topology are irrelevant when dealing with [*fixed topology*]{} meshes. In particular, one can locally order the vertices and keep the order fixed. Then, graph convolution can be defined as follows: $$\label{ordering_based}
(f\star g)_x = \sum_{\ell=1}^L g_\ell f(x_\ell).$$ where $\{x_1, \hdots, x_L\}$ denote the neighbours of vertex $x$ ordered in a fixed way. Here, in analogy with the patch operators, each patch operator is a single neighbouring vertex.
In the Euclidean setting, the order is simply a raster scan of pixels in a patch. On meshes, we opt for a simple and intuitive ordering using spiral trajectories inspired by [@spirals]. Let $x\in \mathcal{V}$ be a mesh vertex, and let $R^d(x)$ be the $d$-[*ring*]{}, an ordered set of vertices whose shortest (graph) path to $x$ is exactly $d$ hops long; $R_j^d(x)$ denotes the $j$th element in the $d$-ring (trivially, $R_1^0(x) = x$). We define the [*spiral patch operator*]{} as the ordered sequence $$S(x) = \{x, R_1^1(x), R_2^1(x), \hdots, R^h_{|R^h|}\},$$ where $h$ denotes the patch radius, similar to the size of the kernel in classical CNNs. Then, [*spiral convolution*]{} is: $$(f * g)_x = \sum_{\ell=1}^{L} g_\ell \, f\big(S_\ell(x)\big). $$ The uniqueness of the ordering is given by fixing two degrees of freedom: the direction of the rings and the first vertex $R_1^1(x)$. The rest of the vertices of the spiral are ordered inductively. The direction is chosen by moving clockwise or counterclockwise, while the choice of the first vertex, the *reference point*, is based on the underlying geometry of the shape to ensure the robustness of the method. In particular, we fix a reference vertex $x_0$ on a template shape and choose the initial point for each spiral to be in the direction of the shortest geodesic path to $x_0$, . $$R_1^1(x) = \operatorname*{arg\,min}\limits_{y \in R^1(x)}{d_\mathcal{M}(x_0,y)},$$ where $d_{\mathcal{M}}$ is the geodesic distance between two vertices on the mesh $\mathcal{M}$. In order to allow for fixed-sized spirals, we choose a fixed length $L$ as a hyper-parameter and then either truncate or zero-pad each spiral depending on its size.
**Comparison to Lim [@spirals]**: The authors choose the starting point of each spiral at random, for every mesh sample, every vertex, and every epoch during training. This choice prevents us from explicitly encoding the fixed connectivity, since corresponding vertices in different meshes will not undergo the same transformation (as in image convolutions). Moreover, single vertices also undergo different transformations every time a new spiral is sampled. Thus, in order for the network to obtain robustness to different spiral samples, it inevitably has to become invariant to different rotations of the neighbourhoods, thus it has reduced capacity. To this end, we emphasise the need of consistent orderings across different meshes. Moreover, in [@spirals], the authors model the vertices on the spiral via a recurrent network, which has higher computational complexity, is harder to optimise and does not take advantage of the stationary properties of the 3D shape (local statistics are repeated across different patches), which are treated by our spiral kernel with weight sharing.
**Comparison to spectral filters:** Spectral convolutional operators developed in [@chebnet; @kipfGCN] for graphs and used in [@coma] for mesh autoencoders, suffer from the fact that the are inherently [*isotropic*]{}. This is a side-effect when one, under the absence of a canonical ordering, needs to design a permutation-invariant operator with small number of parameters. In particular, spectral filters rely on the [*Laplacian operator*]{}, which performs a weighted averaging of the neighbour vertices : $$(\Delta f)_x = \sum\nolimits_{y: (x,y) \in \mathcal{E}} w_{xy} \big(f(y) - f(x)\big),$$ where $w_{xy}$ denotes an edge weight. A polynomial of degree $r$ with learnable coefficients $\theta_0, \hdots, \theta_r$ is then applied to $\Delta$. Then, the graph convolution amounts to filtering the Laplacian eigenvalues, ${p(\Delta) = \Phi p(\Lambda) \Phi^\top}$. Equivalently: $$(f * g) = p(\Delta) f = \sum_{\ell = 0}^r \theta_\ell \Delta^\ell f,$$ While a necessary evil in general graphs, spectral filters on meshes are rather weak given that they are locally rotationally-invariant. On the other hand, spiral convolutional filters leverage the fact that on a mesh one can canonically order the neighbours. Thus, they are anisotropic by construction and as will be shown in the experimental section \[evaluation\] they are expressive by using just one-hop neighbourhoods, contrary to the large receptive fields used in [@coma]. In Fig \[isotropy\] we visualise the impulse response (centred on a vertex on the forehead) of a selected laplacian polynomial filter from the architecture of [@coma] (left) and from a spiral convolutional filter with $h=1$ (right).
Finally, the equivalence of spiral convolutions to image convolutions allows the use of long-studied practices in the computer vision community. For example, small patches can be used, leading to few parameters and fast computation. Furthermore, dilated convolutions [@yu2015multi] can also be adapted to the spiral operator by simply sub-sampling the spiral. Finally, we argue here that our operator could be applied to other domains, such as point clouds, where an ordering of the data points can be enforced.
Neural 3D Morphable Models
--------------------------
Let $\bm{F}=[\bm{f_0}| \bm{f_1}|..., \bm{f_N}]$, $\bm{f_i} \in \mathbb{R}^{d*m}$ the matrix of all the signals defined on a set of meshes in dense correspondence that are sampled from a distribution $\mathcal{D}$, where $d$ the dimensionality of the signal on the mesh (vertex position, texture etc.) and $m$ the number of vertices. A linear 3D Morphable Model [@vetterMM] represents arbitrary instances $\mathbf{y}\in \mathcal{D}$ as a linear combination of the k largest eigenvectors of the covariance matrix of $\bm{F}$ by making a gaussianity assumption: $$\mathbf{y} \approx \Bar{\bm{f}} + \sum_{i}^k \alpha_i \sqrt{d_i} \mathbf{v_i}$$
where $\Bar{\bm{f}}$ the mean shape, $\mathbf{v_i}$ is the $i$th principal component, $d_i$ the respective eigenvalue and $\alpha_i$ the linear weight coefficient. Given its linear formulation, the representational power of the 3DMM is constrained by the span of the eigenvectors, while its parameters scale linearly w.r.t the number of the eigencomponents used, leading to large parametrisations for meshes of high resolution.
In contrast, in this paper, we use spiral convolutions as a building block to build a fully differentiable non-linear Morphable Model. In essence, a Neural 3D Morphable Model is a deep convolutional mesh autoencoder, that learns hierarchical representations of a shape. An illustration of the architecture can be found in Fig \[fig:architecture\]. Leveraging the connectivity of the graph with spiral convolutional filters, we allow for local processing of each shape, while the hierarchical nature of the model allows learning in multiple scales. This way we manage to learn semantically meaningful representations and considerably reduce the number of parameters. Furthermore, we bypass the need to make assumptions about the distribution of the data.
Similar to traditional convolutional autoencoders, we make use of series of convolutional layers with small receptive fields followed by pooling and unpooling, for the encoder and the decoder respectively, where a decimated or upsampled version of the mesh is obtained each time and the features of the existing vertices are either aggregated or extrapolated. We follow [@coma] for the calculation of the features of the added vertices after upsampling, through interpolation by weighting the nearby vertices with barycentric coordinates. The network is trained by minimising the $L_1$ norm between the input and the predicted output.
Spiral Convolutional GAN
------------------------
In order to improve the synthesis of meshes of high resolution, thus increased detail, we extend our framework with a distribution matching scheme. In particular, we propose a mesh Wasserstein GAN [@pmlr-v70-arjovsky17a] with gradient penalty to enforce the Lipschitz constraint [@gulrajani2017improved], that is trained to minimise the wasserstein divergence between the real distribution of the meshes and the distribution of those produced by the generator network. The generator and discriminator architectures, have the same structure as the decoder and the encoder of the Neural3DMM respectively. Via this framework, we obtain two additional properties that are inherently absent from the autoencoder: high frequency detail and a straightforward way to sample from the latent space.
Evaluation
==========
In this section, we showcase the effectiveness of our proposed method on a variety of shape datasets. We conduct a series of ablation studies in order to compare our operator to other Graph Neural Networks, by using the same autoencoder architecture. Fist, we demonstrate the inherent higher capacity of spiral convolutions compared to ChebNet (spectral). Moreover, we discuss the advantages of our method compared to soft-attention based Graph Neural Networks, such as patch-operator based. Finally, we show the importance of the consistency of the ordering by comparing our method to different variants of the method proposed in [@spirals].
Furthermore, we quantitatively show that our method can yield better representations than the linear 3DMM and COMA, while maintaining a small parameter count and frequently allowing a more compact latent representation. Moreover, we proceed with a qualitative evaluation of our method by generating novel examples through vector space arithmetic. Finally, we assess our intrinsic GAN in terms of its ability to produce high resolution realistic examples.
For all the cases, we choose as signal on the mesh the normalised deformations from the mean shape, for every vertex we subtract its mean position and divide with the standard deviation. In this way, we encourage signal stationarity, thus facilitating optimisation. The code is available at <https://github.com/gbouritsas/neural3DMM>.
Datasets
--------
**COMA.** The facial expression dataset from Ranjan [@coma], consisting of 20K+ 3D scans (5023 vertices) of twelve unique identities performing twelve types of extreme facial expressions. We used the same data split as in [@coma]. **DFAUST.** The dynamic human body shape dataset from Bogo [@dfaust:CVPR:2017], consisting of 40K+ 3D scans (6890 vertices) of ten unique identities performing actions such as leg and arm raises, jumps, etc. We randomly split the data into a test set of 5000, 500 validation, and 34,5K+ train.
**MeIn3D.** The 3D large scale facial identity dataset from Booth [@booth20163d], consisting of more than 10,000 distinct identity scans with 28K vertices which cover a wide range of gender ethnicity and age. For the subsequent experiments, the MeIn3D dataset was randomly split within demographic constraints to ensure gender, ethnic and age diversity, into 9K train and 1K test meshes.
{width="0.33\linewidth"} {width="0.33\linewidth"} {width="0.33\linewidth"}
For the quantitative experiments of sections \[ablation\] and \[3dmm\_comparison\] the evaluation metric used is **generalisation**, which measures the ability of a model to represent novel shapes from the same distribution as it was trained on. More specifically we evaluate the average per sample and per vertex euclidean distance in the 3D space (in millimetres) between corresponding vertices in the input and its reconstruction.
Implementation Details {#implementation}
----------------------
We denote as $SC(h,w)$ a spiral convolution of $h$ hops and $w$ filters, $DS(p)$ and $US(p)$ a downsampling and an upsampling by a factor of $p$, respectively, $FC(d)$ a fully connected layer, $l$ the number of vertices after the last downsampling layer. The simple Neural3DMM for COMA and DFAUST datasets is the following:
$\mathit{Enc:} \, SC(1,16)\rightarrow DS(4)\rightarrow SC(1,16)\rightarrow DS(4)\rightarrow SC(1,16)\rightarrow DS(4)\rightarrow SC(1,32)\rightarrow DS(4)\rightarrow FC(d)$
$\mathit{Dec:} \,FC(l*32)\rightarrow US(4)\rightarrow SC(1,32)\rightarrow US(4)\rightarrow SC(1,16)\rightarrow US(4)\rightarrow SC(1,16)\rightarrow US(4)\rightarrow SC(1,3)$ For Mein3D, due to the high vertex count, we modified the COMA architecture for our simple Neural3DMM by adding an extra convolution and an extra downsampling/upsampling layer in the encoder and the decoder respectively (encoder filter sizes: \[8,16,16,32,32\], decoder: mirror of the encoder). The larger Neural3DMM follows the above architecture, but with an increased parameter space. For COMA, the convolutional filters of the encoder had sizes \[64,64,64,128\] and for Mein3D the sizes were \[8,16,32,64,128\], while the decoder is a mirror of the encoder. For DFAUST, the sizes were \[16,32,64,128\] and \[128,64,32,32,16\] and dilated convolutions with $h=2$ hops and dilation ratio $r=2$ were used for the first and the last two layers of the encoder and the decoder respectively. We observed that by adding an additional convolution at the very end (of size equal to the size of the input feature space), training was accelerated. All of our activation functions were ELUs [@ELU]. Our learning rate was $10^{-3}$ with a decay of $0.99$ after each epoch, and our weight decay was $5\times 10^{-5}$. All models were trained for 300 epochs.
![Spiral vs ChebNet (spectral) filters[]{data-label="fig:spectral_spiral_comparison"}](coma_isotropic_anisotropic.png "fig:"){width="0.6\linewidth"} \[fig:coma\_tropic\_comparison\]
![Spiral vs ChebNet (spectral) filters[]{data-label="fig:spectral_spiral_comparison"}](dfaust_isotropic_anisotropic.png "fig:"){width="0.6\linewidth"} \[fig:dfaust\_tropic\_comparison\]
Ablation Studies {#ablation}
----------------
### Isotropic vs Anisotropic Convolutions {#iso_comaparison}
For the purposes of this experiment we used the architecture deployed by the authors of [@coma]. The number of parameters in our case is slightly larger due to the fact that the immediate neighbours, that affect the size of the spiral, range from 7 to 10, while the polynomials used in [@coma] go up to the 6th power of the Laplacian. For both datasets, as clearly illustrated in Fig \[fig:spectral\_spiral\_comparison\], spiral convolution-based autoencoders consistently outperform the spectral ones for every latent dimension, in accordance with the analysis made in section \[spiral\_theory\]. Additionally, increasing the latent dimensions, our model’s performance increases at a higher rate than its counterpart. Notice that the number of parameters scales the same way as the latent size grows, but the spiral model makes better use of the added parameters especially looking at dimensions 16, 32, 64, and 128. Especially on the COMA dataset, the spectral model seems to be flattening between 64 and 128 while the spiral is still noticeably decreasing.
### Spiral vs Attention based Convolutions {#attention_comaparison}
In this experiment we compare our method with certain state-of-the-art soft-attention based Graph Neural Networks: **MoNet**: the patch-operator based model of [@monti2017geometric], where the attention weights are the learnable parameters of gaussian kernels defined on a pseudo-coordinate space[^2], **FeastNet** [@verma2018feastnet] and **Graph Attention** [@velivckovic2017graph], where the attention weights are learnable functions of the input features. In table \[soft\_attn\], we provide results on COMA dataset, using the simple Neural3DMM architecture with latent size 16. We choose the number of attention heads (gaussian kernels in [@monti2017geometric]) to be either 9 (equal to the size of the spiral in our method, for a fair comparison) or 25 (as in [@monti2017geometric], to showcase the effect of over-parametrisation). When it comes to similar number of parameters our method manages to outperform its counterparts, while compared to over-parametrised soft attention networks it either outperforms them, or achieves slightly worse performance. This shows that the spiral operator can make more efficient use of the available learnable parameters, thus being a lightweight alternative to attention-based methods without sacrificing performance. Also, its formulation allows for fast computation; in table \[soft\_attn\] we measure per mesh inference time in ms (on a GeForce RTX 2080 Ti GPU).
Ours
--------- ------- ------- ------- ------- ------- ----------- ------
kernels 9 25 9 25 9 25 -
error 0,762 0,732 0,750 0,623 0,708 **0,583** ****
params 101K 98K 95K ****
time 12,77 15,37 9,04 9,66 10,55 10,96 ****
: Spirals vs soft-attention operators[]{data-label="spiral_vs_attn"}
### Comparison to Lim et al. [@spirals] {#spiral_comaparison}
In order to showcase how the operator behaves when the ordering is not consistent, we perform experiments under four scenarios: the original formulation of [@spirals], where each spiral is randomly oriented for every mesh and every epoch (*rand mesh & epoch)*; choosing the same orientation across all the meshes randomly at every epoch (*rand epoch*); choosing different orientations for every mesh, but keeping them fixed across epochs (*rand mesh*); and fixed ordering (Ours). We compare the LSTM-based approach of [@spirals] and our linear projection formulation (Eq ). The experimental setting and architecture is the same as in the previous section. The proposed approach achieves over 28% improved performance compared to [@spirals], which substantiates the benefits of passing corresponding points through the same transformations.
operation rand mesh & epoch rand mesh rand epoch fixed ordering
------------ ------------------- ----------- ------------ ------------------
LSTM 0.888 [@spirals] 0.880 0,996 0.792
lin. proj. 0.829 0.825 0.951 **0.635 (Ours)**
: Importance of the ordering consistency
\[text\]
Neural 3D Morphable models {#3dmm_comparison}
--------------------------
![Colour coding of the per vertex euclidean error of the reconstructions produced by PCA (2nd), COMA (3rd), and our Neural3DMM (bottom). Top row is ground truth.[]{data-label="fig:dfaust_color_coding"}](dfaust_color_coding.jpg){width="0.6\linewidth"}
### Quantitative results
In this section, we compare the following methods for different dimensions of the latent space: **PCA**, the 3D Morphable Model [@vetterMM], **COMA**, the ChebNet-based Mesh Autoencoder, **Neural3DMM (small)**, ours spiral convolution autoencoder with the same architecture as in COMA, **Neural3DMM (ours)**, our proposed Neural3DMM framework, where we enhanced our model with a larger parameter space (see Sec. \[implementation\]). The latent sizes were chosen based on the variance explained by PCA (explained variance of roughly 85%, 95% and 99% of the total variance).
As can be seen from the graphs in Fig \[quantitative\], our Neural3DMM achieves smaller generalisation errors in every case it was tested on. For the COMA and DFAUST datasets all hierarchical intrinsic architectures outperform PCA for small latent sizes. That should probably be attributed to the fact that the localised filters used allow for effective reconstruction of smaller patches of the shape, such as arms and legs (for the DFAUST case), whilst PCA attempts a more global reconstruction, thus its error is distributed equally across the entire shape. This is well shown in Fig \[fig:dfaust\_color\_coding\], where we compare exemplar reconstructions of samples from the test set (latent size 16). It is clearly visible that PCA prioritises body shape over pose resulting to body parts in the wrong locations (for example see the right leg of the woman on the leftmost column). On the contrary COMA places the vertices in approximately correct locations, but struggles to recover the fine details of the shape leading to various artefacts and deformities; our model on the other hand seemingly balances these two difficult tasks resulting in quality reconstructions that preserve pose and shape.
Comparing to [@coma], it is again apparent here that our spiral-based autoencoder has increased capacity, which together with the increased parameter space, makes our larger Neural3DMM outperform the other methods by a considerably large margin in terms of both generalisation and compression. Despite the fact that for higher dimensions, PCA can explain more than 99% of the total variance, thus making it a tough-to-beat baseline, our larger model still manages to outperform it. The main advantage here is the substantially smaller number of parameters of which we make use. This is clearly seen in the comparison for the MeIn3D dataset, where the large vertex count makes non-local methods as PCA impractical. It is necessary to mention here, that larger latent space sizes are not necessarily desirable for an autoencoder because they might lead to less semantically meaningful and discriminative representation for downstream tasks.
### Qualitative results
Here, we assess the representational power of our models by the common practice of testing their ability to perform linear algebra in their latent spaces.
**Interpolation** Fig \[fig:interpolation\]: We choose two sufficiently different samples $\mathbf{x_1}$ and $\mathbf{x_2}$ from our test set, encode them in their latent representations $\mathbf{z_1}$ and $\mathbf{z_2}$ and then produce intermediate encodings by sampling the line that connects them $\mathbf{z} = a\mathbf{z_1}+(1-a)\mathbf{z_2}$, where $a\in(0,1)$.
**Extrapolation** Fig \[fig:extrapolation\]: Similarly, we decode latent representations that reside on the line defined by $\mathbf{z_1}$ and $\mathbf{z_2}$, but outside the respective line segment, $\mathbf{z} = a*\mathbf{z_1}+(1-a)*\mathbf{z_2}$, where $a\in(-\infty,0)\cup(1,+\infty)$. We choose $\mathbf{z_1}$ to be our neutral expression for COMA and neutral pose for DFAUST, in order to showcase the exaggeration of a specific characteristic on the shape.
![Extrapolation. Left: neutral expression/pose[]{data-label="fig:extrapolation"}](extrapolation_final.png){width="0.6\linewidth"}
**Shape Analogies** Fig \[fig:analogies\]: We choose three meshes $A$, $B$, $C$, and construct a $D$ such that it satisfies $A$:$B$::$C$:$D$ using linear algebra in the latent space as in [@mikolov]: $e(B)-e(A)=e(D)-e(C)$ ($e(*)$ the encoding), where we then solve for $e(D)$ and decode it. This way we transfer a specific characteristic using meshes from our dataset.
![Analogies in MeIn3D and DFAUST[]{data-label="fig:analogies"}](analogies.png){width="0.6\linewidth"}
GAN evaluation
--------------
In figure \[fig:gan\_samples\], we sampled several faces from the latent distribution of the trained generator. Notice that they are realistically looking and, following the statistics of the dataset, span a large proportion of the real distribution of the human faces, in terms of ethnicity, gender and age. Compared to the most popular approach for synthesizing faces, the 3DMM, our model learns to produce fine details on the facial structure, making them hard to distinguish from real 3D scans, whereas the 3DMM, although it produces smooth surfaces, frequently makes it easy to tell the difference between real and artificially produced samples. We direct the reader to the supplementary material to compare with samples drawn from the 3DMM’s latent space.
![Generated identities from our intrinsic 3D GAN[]{data-label="fig:gan_samples"}](gan_tile.png){width="0.7\columnwidth"}
Conclusion
==========
In this paper we introduced a representation learning and generative framework for fixed topology 3D deformable shapes, by using a mesh convolutional operator, spiral convolutions, that efficiently encodes the inductive bias of the fixed topology. We showcased the inherent representational power of the operator, as well as its reduced computational complexity, compared to prior work on graph convolutional operators and show that our mesh autoencoder achieves state-of-the-art results in mesh reconstruction. Finally, we present the generation capabilities of our models through vector space arithmetic, as well as by synthesising novel facial identities. Regarding future work, we plan to extend our framework to general graphs and 3D shapes of arbitrary topology, as well as to other domains that have capacity for an implicit ordering of their primitives, such as point clouds.
Acknowledgements
================
This research was partially supported by ERC Consolidator Grant No. 724228 (LEMAN), Google Research Faculty awards, and the Royal Society Wolfson Research Merit award. G. Bouritsas was funded by the Imperial College London, Department of Computing, PhD scholarship. Dr. Zafeiriou acknowledges support from a Google Faculty award and EPSRC fellowship Deform (EP/S010203/1). S. Ploumpis was suppored by EPSRC Project (EP/N007743/1) FACER2VM.
[^1]: Equal Contribution
[^2]: here we display the best obtained results when choosing the pseudo-coordinates to be local cartesian.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present two instances, L-GAE and L-VGAE, of the variational graph auto-encoding family (VGAE) based on separating feature propagation operations from graph convolution layers typically found in graph learning methods to a single linear matrix computation made prior to input in standard auto-encoder architectures. This decoupling enables the independent and fixed design of the auto-encoder without requiring additional GCN layers for every desired increase in the size of a node’s local receptive field. Fixing the auto-encoder enables a fairer assessment on the size of a nodes receptive field in building representations. Furthermore a by-product of fixing the auto-encoder design often results in substantially smaller networks than their VGAE counterparts especially as we increase the number of feature propagations. A comparative downstream evaluation on link prediction tasks show comparable state of the art performance to similar VGAE arrangements despite considerable simplification. We also show the simple application of our methodology to more challenging representation learning scenarios such as spatio-temporal graph representation learning.'
author:
- |
Paul M. Scherer\
Dept. of Computer Science and Technology\
University of Cambridge, UK Helena Andres-Terre\
Dept. of Computer Science and Technology\
University of Cambridge, UK Pietro Lió\
Dept. of Computer Science and Technology\
University of Cambridge, UK Mateja Jamnik\
Dept. of Computer Science and Technology\
University of Cambridge, UK
bibliography:
- 'neurips\_2019.bib'
title: 'Decoupling feature propagation from the design of graph auto-encoders'
---
Introduction
============
Learning useful and efficient data representations is an important facet of machine learning research as they can heavily sway the performance of downstream pattern recognition and inference algorithms [@representationlearning]. Machine learning methods on graph structured data rely on the fundamental assumption that nodes are better contextualised and thus better represented through the utilisation of local relational information [@goyal; @geometricdeeplearning; @relationalinductivebias; @representationlearninggraphs]. Hence, graph representation learning methods should also be designed to incorporate inductive biases to favour representations contextualising nodes based on its local receptive field [@HUBEL1959].
Driven by the successful interpretation of this inductive bias through convolutional methods, Kipf and Welling’s graph convolution layer (GCN) [@gcn] forms the basis of many current neural network designs on graph structured data. The variational graph auto-encoder (VGAE) and graph auto-encoder (GAE) [@vgae] also rely on GCN layers in its encoder followed by an inner product decoder of latent variables to reconstruct an adjacency matrix. A consequence of this design choice is that the number of layers in this network is directly linked to the size of a nodes local receptive field or the number of feature propagations performed on the features across the nodes in the graph.
We provide a simplification of the design by decoupling the feature propagation steps out of the auto-encoder into a single linear matrix multiplication performed on the dataset prior to input into the learning mechanism [@sgc]. This decouples the convolution operation from the layers in the auto-encoder allowing the design of the auto-encoder to be fixed regardless of the number of feature propagations we wish to perform. This allows the independent consideration of size of a nodes local receptive field and the application of standard off-the-shelf auto-encoder designs onto graph structured datasets. We call the resulting graph auto-encoder and variational graph auto-encoder L-GAE and L-VGAE respectively.
Graphs and graph methods
========================
A great number of real world datasets exhibit some form of graph structure on top of their observations such as social networks, citation networks, protein-protein interaction networks to name a few. For example, within the Cora citation dataset [@cora] a node represents a scientific publication associated with a bag-of-words feature vector describing some of its content. Publications which have cited each other would be connected by an edge thereby the entire dataset forms a graph. The standard task is to then classify each of the nodes into one of 7 classes. Whilst it is wholly possible to apply neural networks on the node features themselves, disregarding the relationships between the nodes to perform the classification, empirical results have shown that classification performance was significantly better when the relations between nodes were taken into account. Hence extending algorithms to appropriately incorporate relational information in its learning process has attracted a great deal of attention [@goyal; @geometricdeeplearning; @relationalinductivebias; @representationlearninggraphs].
We define $\mathcal{G} = (V, E)$ as a graph where $V$ is a set of nodes and $E \subseteq (V \times V)$ be a 2-tuple set of edges in the graph. Hence if $u$ and $v$ are nodes in $\mathcal{G}$, their relation would be recorded with an edge as $(u, v) \in E$. The *neighbours* of a node $v$ in graph $\mathcal{G}=(V,E)$, is the set of nodes which share an edge with $v$, denoted $\mathcal{N}(v) = \{u | (v,u) \in E \}$.
For a graph with $n$ nodes, $\mathbf{A} \in \mathbb{R}^{n \times n}$ is a symmetric adjacency matrix where element $a_{i,j}$ is the weight of the edge between nodes $v_i$ and $v_j$. If $(v_i, v_j) \notin E$ then $A_{i,j} = 0$. A diagonal degree matrix $\mathbf{D} \in \mathbb{R}^{n \times n}$ is defined as the matrix where each entry on the diagonal is the row-sum of the adjacency matrix. For graphs with node features, each node $v_i$ has an associated $d$-dimensional feature vector $\mathbf{x}_i \in \mathbb{R}^d$. The feature matrix $\mathbf{X} \in \mathbb{R}^{n \times d}$ represents feature vectors for every node in the graph.
Graph convolutional networks (GCN) {#section: gcn}
----------------------------------
Graph convolutional networks learn feature representations for nodes through several iterations of graph convolution layers in a composite manner. At each layer node representations $\mathbf{x}_i$ are updated through three stages: localized feature propagation, linear transformation, and a element-wise non-linear activation [@sgc]. The node representations fed into the $k$-th layer of a GCN can be denoted as $\mathbf{H}^{k-1}$ with its output being $\mathbf{H}^{k}$, initially $\mathbf{H}^{0} = X$ will be the input to the first GCN layer.
In the feature propagation stage of a GCN layer the features $\mathbf{h}_i$ of each node $v_i$ is averaged with the feature vectors of its neighbours $\mathcal{N}(v_i)$.
$$\label{eq: featureprop}
\bar{\mathbf{h}}_{i}^{(k)} = \frac{1}{d_i + 1}\mathbf{h}_{i}^{(k-1)} + \sum_{j=1}^{n} \frac{a_{i,j}}{\sqrt{(d_i + 1)(d_j + 1)}}\mathbf{h}_{k}^{(k-1)}$$
This can be expressed as the matrix operation by finding a normalized adjacency matrix $\mathbf{S}$ with added self loops defined as $\mathbf{S} = \tilde{\mathbf{D}}^{-\frac{1}{2}} \tilde{\mathbf{A}} \tilde{\mathbf{D}}^{-\frac{1}{2}}$ where $\tilde{\mathbf{A}} = \mathbf{A}+\mathbf{I}$ and $\tilde{\mathbf{D}} = \mathbf{D} + \mathbf{I}$ (the renormalization trick [@gcn]). Then equation \[eq: featureprop\] can be summarized as:
$$\label{eq: featurepropmatmul}
\bar{\mathbf{H}}^{(k)} = \mathbf{SH}^{(k-1)}$$
After the propagation of the node features in local neighbourhoods the rest of GCN layer is similar to a standard feed forward multi-layer perceptron network (MLP). The hidden representations $\bar{\mathbf{H}}^{(k)}$ at the $k$-th layer in multiplied with a weight matrix $\mathbf{W}^{(k)}$ and passed into a point-wise non-linear activation such as the ReLU function [@relu].
$$\label{eq: gcnactivation}
\mathbf{H}^{(k)} = \text{ReLU}(\bar{\mathbf{H}}^{(k)} \mathbf{W}^{(k)})$$
Variational graph auto-encoders {#sec: vgae}
-------------------------------
Kipf and Welling’s VGAE [@vgae] was born out of the combination of graph convolutional layers [@gcn] and the variational auto-encoder [@vae]. Focusing on link prediction as a target task, the VGAE is designed to capture the higher order dependencies between nodes using the adjacency matrix $\mathbf{A}$ as part of its input to the model and reconstructing the matrix at the decoder using an inner product of the latent variables.
The probabilistic encoder learns the distribution $Q(\mathbf{Z} | \mathbf{X,A})$ that is defined as two graph convolutional layers as defined in equation \[eq: featurepropmatmul\].
$$\label{eq: vgaeencoders}
Q(\mathbf{Z} | \mathbf{X,A}) = \prod_{i=1}^{N} Q(\mathbf{z}_i | \mathbf{X,A}), \thinspace \textrm{with} \thinspace \thinspace \thinspace Q(\mathbf{z}_i | \mathbf{X,A}) = \mathcal{N}(\mathbf{z}_i | \mathbf{\mu}_i, \textrm{diag}(\mathbf{\sigma}_{i}^{2}))$$
Here $\mathbf{\mu}_i \in \mathbf{M}$ and $\mathbf{M} = \textrm{GCN}_{\mathbf{\mu}}(\mathbf{X,A})$ is the matrix of mean vectors $\mathbf{\mu}_i$. Following on, $\log \mathbf{\sigma} = \textrm{GCN}_{\mathbf{\sigma}}(\mathbf{X,A})$. The two layer GCN encoder is thus defined as $\textrm{GCN} = \mathbf{S}(\textrm{ReLU}(\mathbf{SXW}_0))\mathbf{W}_1$. The probabilistic decoder is then defined as a single inner product of the latent variables
$$\label{eq: VGAE_DECODER}
p(\mathbf{A} | \mathbf{Z}) = \prod_{i=1}^{N} \prod_{j=1}^{N} p(A_{i,j} | \mathbf{z}_i, \mathbf{z}_j), \thinspace \textrm{with} \thinspace \thinspace \thinspace p(A_{i,j} | \mathbf{z}_i, \mathbf{z}_j) = \sigma(\mathbf{z}_{i}^{T} \mathbf{z}_j)$$
where $\sigma(\cdot)$ is an element-wise sigmoid function or more simply expressed as a matrix product $\hat{\mathbf{A}} = \sigma(\mathbf{Z}^{T}\mathbf{Z})$ wherein $\hat{\mathbf{A}}$ is interpreted as the networks’ reconstructed probabilistic adjacency matrix. The variational lower bound $\mathcal{L}$ is optimized with respect to the networks parameters $\mathbf{W}$
$$\label{eq: LossVGAE}
\mathcal{L} = \mathbb{E}_{Q(\mathbf{Z} | \mathbf{X,A})}[\log P(\mathbf{A}|\mathbf{Z})] - KL[Q(\mathbf{Z} | \mathbf{X,A}) || P(\mathbf{Z})]$$
One implication of using GCN layers in the design of these graph auto-encoders is that the number of layers used in the encoder is directly linked to the number of feature propagations performed upon the features of the graph. Within the VGAE encoder we see this happen twice, which is interpreted as two stages of feature propagation amongst the nodes of the graph through multiplication with $\mathbf{S}$. To integrate information from larger neighbourhoods per node, another GCN layer would have to be appended to the encoder changing the architecture of the model.
L-GAE and L-VGAE
================
The recursive application non-linear activation and learned weight matrices onto the propagated features shown in equation \[eq: gcnactivation\] creates considerable overhead for large graphs [@sgc]. This overhead may not be critical as previous works have shown that much of the power network based feature classifiers gain comes from local feature propagation [@labelprop; @labelprop2; @sgc]. Wu et al. [@sgc] have hence suggested removing all non-linear transitions except a final softmax layer for classification. We also remove this activation and weights to define a $k$-hop smoothed node feature matrix using $\mathbf{S}$ as defined in section \[section: gcn\].
$$\label{eq:xbar}
\begin{aligned}
\bar{\mathbf{X}} = \mathbf{SS...SX} \\
\bar{\mathbf{X}} = \mathbf{S}^{k}\mathbf{X}
\end{aligned}$$
$\bar{\mathbf{X}}$ can be interpreted as a parameter-free feature preprocessing step on the graph. This reduces any downstream node feature task such as classification, regression, clustering, etc. on training and evaluating standard vector models on the pre-processed $\bar{\mathbf{X}}$. Each application of $\mathbf{S}$ performs a localized feature propagation across the graph as defined in equation \[eq: featureprop\]. This considerably simplifies the construction of graph auto-encoders. As an example, a L-VGAE variational graph auto-encoder which considers a 3-hop local receptive field for each node can be constructed by feeding $\bar{\mathbf{X}} = \mathbf{S}^3\mathbf{X}$ into any variational auto-encoder.
Evaluation and discussion
=========================
---------- --------------------- --------------------- --------------------- --------------------- --------------------- ---------------------
SC 84.6 $\pm$ 0.01 88.5 $\pm$ 0.00 80.5 $\pm$ 0.01 85.0 $\pm$ 0.01 84.2 $\pm$ 0.02 87.8 $\pm$ 0.01
DW 83.1 $\pm$ 0.01 85.0 $\pm$ 0.00 80.5 $\pm$ 0.02 83.6 $\pm$ 0.01 84.4 $\pm$ 0.00 84.1 $\pm$ 0.00
GAE\* 84.3 $\pm$ 0.02 88.1 $\pm$ 0.01 78.7 $\pm$ 0.02 84.1 $\pm$ 0.02 82.2 $\pm$ 0.01 87.4 $\pm$ 0.00
VGAE\* 84.0 $\pm$ 0.02 87.7 $\pm$ 0.01 78.9 $\pm$ 0.03 84.1 $\pm$ 0.02 82.7 $\pm$ 0.01 87.5 $\pm$ 0.01
L-GAE\* **86.3 $\pm$ 0.01** **89.0 $\pm$ 0.01** **80.4 $\pm$ 0.01** **84.2 $\pm$ 0.01** 82.4 $\pm$ 0.00 87.5 $\pm$ 0.00
L-VGAE\* 85.9 $\pm$ 0.01 88.2 $\pm$ 0.01 79.8 $\pm$ 0.00 83.9 $\pm$ 0.00 **83.1 $\pm$ 0.01** **87.6 $\pm$ 0.01**
GAE 91.0 $\pm$ 0.02 92.0 $\pm$ 0.03 89.5 $\pm$ 0.04 89.9 $\pm$ 0.05 **96.4 $\pm$ 0.00** **96.5 $\pm$ 0.00**
VGAE 91.4 $\pm$ 0.01 92.6 $\pm$ 0.01 90.8 $\pm$ 0.02 92.0 $\pm$ 0.02 94.4 $\pm$ 0.02 94.7 $\pm$ 0.02
L-GAE 91.9 $\pm$ 0.00 93.0 $\pm$ 0.00 90.7 $\pm$ 0.01 91.3 $\pm$ 0.01 95.4 $\pm$ 0.00 95.6 $\pm$ 0.00
L-VGAE **92.9 $\pm$ 0.00** **93.9 $\pm$ 0.00** **92.5 $\pm$ 0.00** **93.6 $\pm$ 0.00** 92.6 $\pm$ 0.01 92.8 $\pm$ 0.01
---------- --------------------- --------------------- --------------------- --------------------- --------------------- ---------------------
: Link prediction task in citation networks recorded with mean and standard deviation. The (\*) signifies instances where the model was not given feature information for each node.
\[tab: results\]
A comparative evaluation was performed on the network reconstruction task set by VGAE [@vgae]. For a “fair” comparison between the GAE/VGAE and the proposed L-GAE/L-VGAE in the benchmark tests the design of the networks was made to be as similar as possible. For GAE/VGAE we adopt the definition made in their publication [@vgae]: a two GCN layer encoder with 32 and 16 hidden units incorporating 2 feature propagations and the inner product decoder (equation \[eq: VGAE\_DECODER\]). In our methodology this corresponds to 2 feature propagation on the feature matrix, i.e. $k=2$ in equation \[eq:xbar\], followed by an encoder of two linear layers with 32 and 16 hidden units and the same inner product decoder. Each of the presented models was trained for 200 epochs using the same Adam optimizer.
Experiments were run on two streams for the auto-encoder models. One stream did not use feature vector for each node in the graph, implemented by $\mathbf{X} = \mathbf{I}$, and one stream of results utilised the feature vectors for each node. The results shown in table \[tab: results\] show that L-GAE and L-VGAE perform on par and outperform their GAE and VGAE counterparts in both tasks. We discuss the by-product of smaller networks in the supplementary material.
------------------- --------------------- --------------------- ---------------------
****
**Method** 15 min 30 min 60 min
LSTM Baseline 3.67 $\pm$ 0.04 4.88 $\pm$ 0.02 6.53 $\pm$ 0.02
STDGI [@stdgi] 3.59 $\pm$ 0.00 **4.78 $\pm$ 0.00** **6.34 $\pm$ 0.01**
L-VGAE + Baseline **3.58 $\pm$ 0.00** 4.84 $\pm$ 0.01 6.43 $\pm$ 0.01
------------------- --------------------- --------------------- ---------------------
: Regression performance over time on METR-LA with mean absolute error score and standard deviation on the predictions of the baseline LSTM model applied on raw features as well the baseline LSTM model applied on features learned with L-VGAE and STDGI.[]{data-label="tab: resultsST"}
We have not proposed a model that is especially innovative. A fixed low-pass filter for graphs was studied very recently in Wu et al. [@sgc] but we presented a simple methodology to design graph auto-encoders by decoupling the feature propagations. This simplification allows us to quickly build methods for more challenging scenarios such as unsupervised learning of spatio-temporal node representations as covered in Opolka et al. [@stdgi] on the METR-LA dataset [@metr-la]. In this scenario we have constructed a simple non-forward looking model through simple application of $\bar{\mathbf{X}}$ into an Encoder-Decoder LSTM [@seq2seq] and achieved an improvement over the baseline supervised method which did not take relations into account. We are on par with the STDGI in short time horizons, however do worse as the later time predictions which makes sense as we do not incorporate future states as part of learning as in STDGI [@stdgi] and only relational information. We could potentially improve this by using a more sophisticated temporal model such as a flow-based generative model based around WaveNet [@wavenet] or a temporal auto-encoder. We will continue expanding on this work towards other challenging scenarios and hope that this could be seen as a “first step” or baseline method to consider when designing unsupervised graph representation techniques.
Appendix A: potentially smaller network in large receptive fields
=================================================================
An interesting aspect of auto-encoders is to look at the size and number of trainable parameters, or degrees of freedom, present in the model architectures. As the design of the auto-encoder in the GAE/VGAE is intimately coupled with the desired number of feature propagations, the resulting *size* of the network gets considerably larger with each additional feature propagation. In L-GAE/L-VGAE models the feature propagation of features is processed independently prior to input into the auto-encoder meaning that much smaller networks can be employed on the different levels of feature propagation. This enables the study of feature propagation depth as an independent variable on the same network and in model selection as an independent hyper-parameter. Tables \[cora\_weights\], \[citeseer\_weights\], \[pubmed\_weights\] show the number of degrees of freedom present in each auto-encoder based on number of feature propagation steps for Cora, Citeseer, and PubMed networks respectively. The network used in the L-GAE and L-VGAE are the same used in the evaluation in table 1, whilst the VGAE requires a new hidden layer with base 2 progression as implied in the original design of using 32 and 16 hidden units.
[@lllll@]{} &\
**Method** & k = 1 &
----------------------
k = 2
(Base Configuration)
----------------------
: Number of trainable parameters in the auto-encoder given the number of feature propagations for the Cora dataset. Note that the size of the networks for L-GAE and L-VGAE are fixed, we chose a two linear layer encoder (32 and 16 hidden units each) and single layer inner product decoder to match the architecture of the encoder model described in Kipf and Welling. The auto-encoder in the VGAE has to include a layer for each feature propagation step, sizes used are based on base 2 progression as implied in Kipf and Welling. ie for $k=1$ (16 hidden unit layer in auto-encoder), $k=2$ (32, 16), $k=3$ (64, 32, 16), and $k=7$ (1024, 512, 256, 128, 64, 32, 16)[]{data-label="cora_weights"}
& k = 3 & k = 7\
L-GAE & 46416 & **46416** & **46416** & **46416**\
L-VGAE & 46944 & **46944** & **46944** & **46944**\
VGAE & **45856** & 48928 & 94784 & 2166784\
[@lllll@]{} &\
**Method** & k = 1 &
----------------------
k = 2
(Base Configuration)
----------------------
: Number of trainable parameters in the auto-encoder given the number of feature propagations for the CiteSeer dataset.[]{data-label="citeseer_weights"}
& k = 3 & k = 7\
L-GAE & 119056 & **119056** & **119056** & **119056**\
L-VGAE & 119584 & **119584** & **119584** & **119584**\
VGAE & **118496** & 121568 & 240064 & 4491264\
[@lllll@]{} &\
**Method** & k = 1 &
----------------------
k = 2
(Base Configuration)
----------------------
: Number of trainable parameters in the auto-encoder given the number of feature propagations for the Pubmed dataset.[]{data-label="pubmed_weights"}
& k = 3 & k = 7\
L-GAE & 16560 & **16560** & **16560** & **16560**\
L-VGAE & 17088 & **17088** & **17088** & **17088**\
VGAE & **16000** & 19072 & 35072 & 1211392\
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
For output-symmetric DMCs at even moderately high rates, fixed-block-length communication systems show no improvements in their error exponents with feedback. In this paper, we study systems with fixed end-to-end delay and show that feedback generally provides dramatic gains in the error exponents.
A new upper bound (the uncertainty-focusing bound) is given on the probability of symbol error in a fixed-delay communication system with feedback. This bound turns out to have a similar form to Viterbi’s bound used for the block error probability of convolutional codes as a function of the fixed constraint length. The uncertainty-focusing bound is shown to be asymptotically achievable with noiseless feedback for erasure channels as well as any output-symmetric DMC that has strictly positive zero-error capacity. Furthermore, it can be achieved in a delay-universal (anytime) fashion even if the feedback itself is delayed by a small amount. Finally, it is shown that for end-to-end delay, it is generally possible at high rates to beat the sphere-packing bound for general DMCs — thereby providing a counterexample to a conjecture of Pinsker.
author:
- |
Anant Sahai[^1]\
[[email protected]]{}
bibliography:
- './IEEEabrv.bib'
- './MyMainBibliography.bib'
title: '**Why block length and delay behave differently if feedback is present**'
---
Feedback, delay, reliability functions, anytime reliability, sphere-packing bounds, random coding, hybrid ARQ, queuing, list decoding.
Introduction
============
The channel coding theorems studied in information theory are not just interesting as mathematical results, they also provide insights into the underlying tradeoffs in reliable communication systems. While in practice there are many different parameters of interest such as power, complexity, and robustness, perhaps the most fundamental two are end-to-end system delay and the probability of error. Error probability is fundamental because a low probability of bit error lies at the heart of the digital revolution justified by the source/channel separation theorem. Delay is important because it is the most basic cost that must be paid in exchange for reliability — it allows the laws of large numbers to be harnessed to smooth out the variability introduced by random communication channels.
In our entire discussion, the assumption is that information naturally arises as a stream generated in real time at the source (e.g. voice, video, or sensor measurements) and it is useful to the destination in finely grained increments (e.g. a few milliseconds of voice, a single video frame, etc.). The acceptable end-to-end delay is determined by the application and can often be much larger than the natural granularity of the information being communicated (e.g. voice may tolerate a delay of hundreds of milliseconds despite being useful in increments of a few milliseconds). This is different from cases in which information arises in large bursts with each burst needing to be received by the destination before the next burst even becomes available at the source.
Rather than worrying about what the appropriate granularity of information should be, the formal problem is specified at the individual bit level. (See Figure \[fig:timeline\].) If a bit is not delivered correctly by its deadline, it is considered to be erroneous. The upper and lower bounds of this paper turn out to not depend on the choice of information granularity, only on the fact that the granularity is much finer than the tolerable end-to-end delay.
\#1\#2\#3\#4\#5[ @font ]{}
(8124,3289)(439,-2594) [(451,-511)[( 1,-2)[150]{}]{} ]{}[(751,-511)[( 1,-2)[150]{}]{} ]{}[(1051,-511)[( 1,-2)[150]{}]{} ]{}[(1351,-511)[( 1,-2)[150]{}]{} ]{}[(1651,-511)[( 1,-2)[150]{}]{} ]{}[(1951,-511)[( 1,-2)[150]{}]{} ]{}[(2251,-511)[( 1,-2)[150]{}]{} ]{}[(2551,-511)[( 1,-2)[150]{}]{} ]{}[(2851,-511)[( 1,-2)[150]{}]{} ]{}[(3151,-511)[( 1,-2)[150]{}]{} ]{}[(3451,-511)[( 1,-2)[150]{}]{} ]{}[(3751,-511)[( 1,-2)[150]{}]{} ]{}[(4051,-511)[( 1,-2)[150]{}]{} ]{}[(4351,-511)[( 1,-2)[150]{}]{} ]{}[(4651,-511)[( 1,-2)[150]{}]{} ]{}[(4951,-511)[( 1,-2)[150]{}]{} ]{}[(5251,-511)[( 1,-2)[150]{}]{} ]{}[(5551,-511)[( 1,-2)[150]{}]{} ]{}[(5851,-511)[( 1,-2)[150]{}]{} ]{}[(6151,-511)[( 1,-2)[150]{}]{} ]{}[(6451,-511)[( 1,-2)[150]{}]{} ]{}[(6751,-511)[( 1,-2)[150]{}]{} ]{}[(7051,-511)[( 1,-2)[150]{}]{} ]{}[(7351,-511)[( 1,-2)[150]{}]{} ]{}[(7651,-511)[( 1,-2)[150]{}]{} ]{}[(7951,-511)[( 1,-2)[150]{}]{} ]{}[(1951,-2311)[( 1, 0)[2250]{}]{} ]{}[(3001,-1261)[( 0,-1)[450]{}]{} ]{}[(3601,-1261)[( 0,-1)[450]{}]{} ]{}[(4201,-1261)[( 0,-1)[450]{}]{} ]{}[(4801,-1261)[( 0,-1)[450]{}]{} ]{}[(5401,-1261)[( 0,-1)[450]{}]{} ]{}[(6001,-1261)[( 0,-1)[450]{}]{} ]{}[(6601,-1261)[( 0,-1)[450]{}]{} ]{}[(7201,-1261)[( 0,-1)[450]{}]{} ]{}[(7801,-1261)[( 0,-1)[450]{}]{} ]{}[(6751,464)[( 0,-1)[450]{}]{} ]{}[(7351,464)[( 0,-1)[450]{}]{} ]{}[(7951,464)[( 0,-1)[450]{}]{} ]{}[(751,464)[( 0,-1)[450]{}]{} ]{}[(1351,464)[( 0,-1)[450]{}]{} ]{}[(1951,464)[( 0,-1)[450]{}]{} ]{}[(2551,464)[( 0,-1)[450]{}]{} ]{}[(3151,464)[( 0,-1)[450]{}]{} ]{}[(3751,464)[( 0,-1)[450]{}]{} ]{}[(4351,464)[( 0,-1)[450]{}]{} ]{}[(4951,464)[( 0,-1)[450]{}]{} ]{}[(5551,464)[( 0,-1)[450]{}]{} ]{}[(6151,464)[( 0,-1)[450]{}]{} ]{}(451,-436)[(0,0)\[b\]]{} (751,-436)[(0,0)\[b\]]{} (1051,-436)[(0,0)\[b\]]{} (1351,-436)[(0,0)\[b\]]{} (1651,-436)[(0,0)\[b\]]{} (1951,-436)[(0,0)\[b\]]{} (2251,-436)[(0,0)\[b\]]{} (2551,-436)[(0,0)\[b\]]{} (2851,-436)[(0,0)\[b\]]{} (3151,-436)[(0,0)\[b\]]{} (3451,-436)[(0,0)\[b\]]{} (3751,-436)[(0,0)\[b\]]{} (4051,-436)[(0,0)\[b\]]{} (4351,-436)[(0,0)\[b\]]{} (4651,-436)[(0,0)\[b\]]{} (4951,-436)[(0,0)\[b\]]{} (5251,-436)[(0,0)\[b\]]{} (5551,-436)[(0,0)\[b\]]{} (5851,-436)[(0,0)\[b\]]{} (6151,-436)[(0,0)\[b\]]{} (6451,-436)[(0,0)\[b\]]{} (6751,-436)[(0,0)\[b\]]{} (7051,-436)[(0,0)\[b\]]{} (7351,-436)[(0,0)\[b\]]{} (7651,-436)[(0,0)\[b\]]{} (7951,-436)[(0,0)\[b\]]{} (601,-1111)[(0,0)\[b\]]{} (901,-1111)[(0,0)\[b\]]{} (1201,-1111)[(0,0)\[b\]]{} (1501,-1111)[(0,0)\[b\]]{} (1801,-1111)[(0,0)\[b\]]{} (2101,-1111)[(0,0)\[b\]]{} (2401,-1111)[(0,0)\[b\]]{} (2701,-1111)[(0,0)\[b\]]{} (3001,-1111)[(0,0)\[b\]]{} (3301,-1111)[(0,0)\[b\]]{} (3601,-1111)[(0,0)\[b\]]{} (3901,-1111)[(0,0)\[b\]]{} (4201,-1111)[(0,0)\[b\]]{} (4501,-1111)[(0,0)\[b\]]{} (4801,-1111)[(0,0)\[b\]]{} (5101,-1111)[(0,0)\[b\]]{} (5401,-1111)[(0,0)\[b\]]{} (5701,-1111)[(0,0)\[b\]]{} (6001,-1111)[(0,0)\[b\]]{} (6301,-1111)[(0,0)\[b\]]{} (6601,-1111)[(0,0)\[b\]]{} (6901,-1111)[(0,0)\[b\]]{} (7201,-1111)[(0,0)\[b\]]{} (7501,-1111)[(0,0)\[b\]]{} (7801,-1111)[(0,0)\[b\]]{} (8101,-1111)[(0,0)\[b\]]{} (3076,-2536)[(0,0)\[b\]]{} (6001,-2011)[(0,0)\[b\]]{} (7201,-2011)[(0,0)\[b\]]{} (3001,-2011)[(0,0)\[b\]]{} (3601,-2011)[(0,0)\[b\]]{} (4201,-2011)[(0,0)\[b\]]{} (4801,-2011)[(0,0)\[b\]]{} (5401,-2011)[(0,0)\[b\]]{} (6601,-2011)[(0,0)\[b\]]{} (7801,-2011)[(0,0)\[b\]]{} (751,539)[(0,0)\[b\]]{} (1351,539)[(0,0)\[b\]]{} (1951,539)[(0,0)\[b\]]{} (2551,539)[(0,0)\[b\]]{} (3151,539)[(0,0)\[b\]]{} (3751,539)[(0,0)\[b\]]{} (4351,539)[(0,0)\[b\]]{} (4951,539)[(0,0)\[b\]]{} (5551,539)[(0,0)\[b\]]{} (6151,539)[(0,0)\[b\]]{} (6751,539)[(0,0)\[b\]]{} (7351,539)[(0,0)\[b\]]{} (7951,539)[(0,0)\[b\]]{}
In the next section of this introduction, the example of the binary erasure channel at $R = \frac{1}{2}$ bits per channel use is used to constructively show how fixed-delay codes can dramatically outperform fixed-block-length codes at the same rates when feedback is present. Existing information-theoretic views of feedback and reliability are then reviewed in Section \[sec:background\]. Section \[sec:mainresults\] states the main results of the paper, with the constructions and proofs following in subsequent sections. Numerical examples and plots are also given in Section \[sec:mainresults\] to illustrate these results.
Section \[sec:nofeedback\] generalizes Pinsker’s result from [@PinskerNoFeedback] for non-block-code performance with fixed delay and also explains why, contrary to Pinsker’s assertion, this argument [*does not*]{} generalize to the case when feedback is present. The new upper bound (the “uncertainty-focusing bound”) on fixed-delay performance is proved in Section \[sec:feedbackbound\] by reviving Forney’s inverse concatenation construction to serve this new purpose. Asymptotic achievability of this new bound with noiseless feedback is shown in Section \[sec:erasurefeedback\] for erasure channels. These results are extended in Section \[sec:fortifiedfeedback\] to general DMCs. It turns out that for channels with strictly positive feedback-zero-error capacity, a low-rate error-free path can be constructed with very little overhead thereby attaining the performance of the uncertainty-focusing bound. For generic channels at high message rates, the overhead of this approach is non-negligible but the error probability still asymptotically beats that predicted by the sphere-packing bound for the same end-to-end delay.
A simple example using the BEC {#sec:becexample}
------------------------------
The natural question of end-to-end delay in situations with finely grained information was considered by Pinsker in [@PinskerNoFeedback]. He explicitly treats the BSC case, while asserting that the results hold for any DMC. The main result (Theorem 5 in [@PinskerNoFeedback]) is that the sphere-packing bound $E_{sp}(R)$ is an upper bound to the fixed-delay error exponent for any nonblock code. Theorem 8 in [@PinskerNoFeedback] asserts that the same bound continues to hold even with feedback. As reviewed in Section \[sec:fixedlengthreview\], these theorems parallel what is already known to hold for fixed-block-length codes.
The binary erasure channel (BEC) with erasure probability $\beta <
\frac{1}{2}$ used at rate $R' = \frac{1}{2}$ bits per channel use gives a counterexample to Pinsker’s generalized conjecture. The BEC is so simple that everything can be understood with a minimum of overhead. A counterexample that covers the BSC itself is given later in Section \[sec:comments\] (plotted in Figure \[fig:beatspherebsc\]) and others are given in [@TuncThesis; @SimsekJainVaraiya] using much more involved codes built around control-theoretic ideas.
The sphere-packing bound in the BEC case corresponds to the probability that the channel erases more than $\frac{1}{2}$ of the inputs during the block: $$\label{eqn:becspherebound}
E_{sp}(\frac{1}{2}) = D(\frac{1}{2}||\beta) =
- \frac{\ln(4\beta(1-\beta))}{2}.$$ For $\beta = 0.4$, this yields an error exponent of about $0.02$. Even with feedback, there is no way for a fixed-block-length code to beat this exponent. If the channel lets fewer than $\frac{n}{2}$ bits through, it is impossible to reliably communicate an $\frac{n}{2}$-bit message! Bit-error vs block-error considerations alone do not change the overall picture since they buy at most a factor of $\frac{2}{n}$ in the average probability of error — nothing on an exponential scale.
With noiseless feedback, the natural nonblock code just retransmits a bit over the BEC until it is correctly received. To be precise, as bits arrive steadily at the rate $R' = \frac{1}{2}$ bits per channel use, they enter a FIFO queue of bits awaiting transmission. At time $0$, both the encoder and decoder know that there are no bits waiting. From that time onward, the bit arrivals are modeled here as deterministic and come every other channel use. Since both the encoder and decoder know when a bit arrives as well as when a bit is successfully received, there is no ambiguity in how to interpret a channel output.
If the queue length is examined every two channel uses, exactly one new bit has arrived while the channel may have successfully served 0, 1, or 2 bits in this period. Thus, the length of the queue can either increase by one, stay the same, or decrease by one. The queue length can be modeled (see Figure \[fig:birthdeath\]) as a birth-death Markov chain with a $\beta^2$ probability of birth and a $(1-\beta)^2$ probability of death. The steady state distribution of the queue length is therefore $\pi_i = \kappa (\frac{\beta}{1-\beta})^{2i}$ where $\kappa$ is the normalization constant $(1 -
(\frac{\beta}{1-\beta})^2)$.
\#1\#2\#3\#4\#5[ @font ]{}
(6458,816)(518,-494) [(5626,-61) ]{}[(4051,-61) ]{}[(2476,-61) ]{}[(901,-61) ]{}[(6001, 89)[( 1, 0)[900]{}]{} ]{}[(6826,-211)[(-1, 0)[825]{}]{} ]{}[(4426, 89)[( 1, 0)[900]{}]{} ]{}[(5251,-211)[(-1, 0)[825]{}]{} ]{}[(2851, 89)[( 1, 0)[900]{}]{} ]{}[(3676,-211)[(-1, 0)[825]{}]{} ]{}[(1276, 89)[( 1, 0)[900]{}]{} ]{}[(2101,-211)[(-1, 0)[825]{}]{} ]{}(6376,164)[(0,0)\[b\]]{} (6376,-436)[(0,0)\[b\]]{} (4801,164)[(0,0)\[b\]]{} (4801,-436)[(0,0)\[b\]]{} (3226,164)[(0,0)\[b\]]{} (3226,-436)[(0,0)\[b\]]{} (1651,164)[(0,0)\[b\]]{} (1651,-436)[(0,0)\[b\]]{} (2476,-136)[(0,0)\[b\]]{} (901,-136)[(0,0)\[b\]]{} (4051,-136)[(0,0)\[b\]]{} (5626,-136)[(0,0)\[b\]]{} (6976,-136)[(0,0)\[b\]]{}
To understand the probability of error with end-to-end delay, just notice that the only way a bit can miss its deadline is if it is still waiting in the queue. If it was a bit from $d$ time steps ago, the queue must currently hold at least $\frac{d}{2}$ bits. The steady state distribution reveals that the asymptotic probability of this is: $$\kappa \sum_{i=\frac{d}{2}}^\infty (\frac{\beta}{1-\beta})^{2i} =
(\frac{\beta}{1-\beta})^d
\kappa \sum_{i=0}^\infty (\frac{\beta}{1-\beta})^{2i} =
(\frac{\beta}{1-\beta})^d.$$ Converting that into an error exponent with delay $d$ gives $$\label{eqn:halfbecfeedback}
E^{bec}_{a}(\frac{1}{2}) = \ln(1-\beta) - \ln(\beta).$$ Plugging in $\beta=0.4$ reveals an exponent of more than $0.40$. This is about twenty times higher than the sphere-packing bound! Simple computations can verify that the ratio of (\[eqn:halfbecfeedback\]) to (\[eqn:becspherebound\]) goes to infinity as $\beta \rightarrow
\frac{1}{2}$.
To help get an intuitive idea for why this happens, it is worthwhile to consider an idealized feedback-free code for erasure channels (the reader may find it helpful to think of packet erasure channels with large alphabets). Suppose that the encoder causally generated “parities” of all the message symbols so far with the property that symbols could be decoded whenever the receiver had as many unerased parities as there were undecoded symbols.[^2] The queue size can be reinterpreted in this setting as the number of additional parities required before the decoder could solve for the currently uncertain message symbols. The queue’s renewal times correspond to the times at which the decoder can solve for the current set of undecoded message symbols.
![A simulated run of an $0.4$ erasure channel using an idealized linear causal code without feedback. The red upper sawtooth represents the number of current message symbols that are still ambiguous at the decoder while the lower curve represents the number of additional parities that would enable it to resolve the current ambiguity. The lower curve is not coincidentally also the queue size for the natural FIFO-based code with feedback. The dotted line at 6 represents a potential delay deadline of 12 time units.[]{data-label="fig:backlogsimulation"}](backlogsimulation.edited.eps){width="4in" height="3in"}
![A zoomed-in look at the simulation of Figure \[fig:backlogsimulation\] showing the total number of decoded symbols as a function of time. The thin upper curve is the total number of symbols that have been received at the rate-$\frac{1}{2}$ encoder. The next lower line is the total number of symbols decoded by the code with feedback. The lowest curve corresponds to the code without feedback. The thin dotted line represents the deadline of $12$ time steps. Whenever the decoder curves are below this curve, they are missing the deadline.[]{data-label="fig:timedomainzoom"}](timedomainzoom.edited.eps){width="4in" height="3in"}
Figure \[fig:backlogsimulation\] illustrates the backlog of undecoded bits in a simulated run of a rate-$\frac{1}{2}$ code over a channel with erasure probability $0.4$. Figure \[fig:timedomainzoom\] zooms in on a particular segment of time corresponding to an “error event” and shows the differences between how the feedback-free code and feedback code make progress. During an error event in which the channel is erasing too many symbols, progress at the decoder seems to stop entirely in the feedback-free code, only catching up in a sudden burst when the error event ends. By contrast, the code with feedback makes visible, but slower, progress at the decoder even during these error events. As a result, it is able to meet the target delay deadline whereas the code without feedback misses it. This example also shows how the delays in the feedback-free code are related to the inter-renewal times of the queue, while the delays in the code with feedback are related to the length of the queue itself.
Stepping back, this example illustrates that Pinsker’s bound with delay does not generally apply when feedback is available. Instead, fixed-delay nonblock codes can dramatically outperform fixed-block-length codes with feedback. Moreover, it is possible to glimpse why this occurs. Reliable communication always takes place at message rates $R$ that are less than the capacity $C$. In a fixed-delay setting with feedback, the encoder has the flexibility to do flow control based on what the channel has been doing in the past. It can vary the short-term operational rate $R$ — in effect stealing channel uses from later bits to make sure that earlier bits meet their looming deadlines, while still hoping that the later bits will be able to meet their later deadlines. This flexibility is missing in the fixed-block-length setting because all the bits in the block are forced to share a common deadline.
This can also be seen by contrasting the total conditional entropy $H(B_{iR'}^{(i+d)R'}|Y_i^{i+d})$ of the message bits $B_{iR'}^{(i+d)R'}$ given the channel outputs $Y_i^{i+d}$ to the sum $\sum_{k=iR'}^{(i+d)R'}
H(B_{k}|Y_i^{i+d})$ of the marginal conditional entropies of the bits given the channel outputs. If the channel misbehaves slightly and makes it hard to distinguish only a single pair of bit strings, the marginal entropies $H(B_{k}|Y_i^{i+d})$ can become large even as the total conditional entropy is small. Such situations are common without feedback. From the decoder’s perspective, the feedback encoder’s strategy should be to focus the uncertainty $H(B_{iR'}^{(i+d)R'}|Y_i^{i+d})$ onto later bits $B_{(i+d)R' - \Delta}^{(i+d)R'}$ to pay for reducing it on earlier bits. The sum of the marginal conditional entropies can then be made the same as the total conditional entropy.
The total delay experienced by a bit can also be broken into two components: queuing delay and transmission delay. For the erasure channel, the transmission delay is just a geometric random variable governed by an exponent of $-\ln(\beta)$. This transmission exponent does not change with the message rate. The queuing delay is the dominant term, and its exponent does change with the message rate.
Finally, it is interesting to examine the computational burden of implementing this simple code. At the encoder, all that is needed is a FIFO queue that costs a constant (assuming memory is free) per unit time to operate. The decoder has similar complexity since it too just tracks how many bits it has received so far in comparison with the number of bits known to have arrived at the encoder. The computational burden does not change with either the target delay or the quality of the channel!
Background {#sec:background}
==========
Fixed-length codes {#sec:fixedlengthreview}
------------------
Traditionally, reliable communication was first explored in the context of block codes [@ShannonOriginalPaper]. If physical information sources are considered to produce bits steadily at $R'$ bits per second, then the use of a block code of length $n$ channel uses (with channel uses assumed to occur once per second) contributes to end-to-end delay in two ways.
- Enough bits must first be buffered up to even compute the codeword. This takes no more than $n$ seconds and can take less if the block code is systematic in nature.
- The decoder must wait for $n$ seconds to get the $n$ channel outputs needed to decode the block. This second delay would be present even if the source bits were realized entirely in advance of the use of the channel.
In this context, the fundamental lower bound on error probability comes from the sphere-packing bound. To understand this bound, it is helpful to think about the message block as representing a certain [*volume*]{} of entropic uncertainty that the decoder has about the message. The objective of using the channel is to reduce this uncertainty. Let $P$ be the transition matrix ($p_{y|x}$ is the probability of seeing output $y$ given input $x$) for the DMC. Each channel use can reduce the uncertainty on average by no more than the capacity $$\label{eqn:capacitydefinition}
C(P) = \max_{\vec{q}} I(\vec{q},P)$$ where $I(\vec{q},P)$ is the mutual information between input and output of channel $P$ when $\vec{q}$ is the input distribution and is defined by $$\label{eqn:mutualinformationdef}
I(\vec{q},P) =
\sum_x q_x \sum_y p_{y|x} \ln \frac{p_{y|x}}{\sum_{k}q_k p_{y|k}}.$$
With or without feedback, successful communication is not possible if during the block, the memoryless channel acts like one whose capacity is less than the target message rate. Following [@csiszarkorner; @Haroutunian], for fixed-block-length codes this idea immediately gives the following upper bound (referred to as the [*Haroutunian bound*]{} throughout this paper) on the block-coding error exponent ($\limsup_{n \rightarrow \infty} \frac{-\ln P_e}{n}$): $$\begin{aligned}
E^+(R) & = & \inf_{G: C(G) < R}
\sup_{\vec{r}} D\left(G||P |
\vec{r}\right) \label{eqn:primitiveupperbound} \\
& = & \min_{G: C(G) \leq R}
\max_x \sum_y g_{y|x} \ln
\frac{g_{y|x}}{p_{y|x}} \label{eqn:Haroutunianbound}\end{aligned}$$ where $D(G || P | \vec{r})$ is the divergence term that governs the exponentially small probability of the true channel $P$ behaving like channel $G$ when facing the input distribution $\vec{r}$. The divergence is defined as $$\label{eqn:divergencedef}
D(G||P | \vec{r}) =
\sum_x r_x \sum_y g_{y|x} \ln \frac{g_{y|x}}{p_{y|x}}.$$
Without feedback, the encoder does not have the flexibility to change the input distribution in response to the channel’s behavior. The optimization can take this into account to get the bound traditionally known as the sphere-packing bound $$\label{eqn:spherepackingbound}
E_{sp}(R) = \max_{\vec{r}} \min_{G: I(\vec{r},G) \leq R } D\left(G||P|\vec{r}\right).$$ It is clear that $E_{sp}(R) \leq E^+(R)$ and Figure \[fig:zchannelbounds\] illustrates that the inequality can be strict.
![The sphere-packing and Haroutunian bounds for the Z-channel with nulling probability $0.5$. The upper curve is the Haroutunian upper bound for the error exponent of block codes with feedback and the lower curve is the classical sphere-packing bound. Both approach zero very rapidly around the capacity of $0.223$ nats per channel use. Due to the asymmetry of the Z-channel, the capacity-achieving distribution is not the same as the sphere-packing-bound-achieving distribution.[]{data-label="fig:zchannelbounds"}](zchannelbounds.edited.eps){width="4in" height="3in"}
It is often useful to use an alternate form for $E_{sp}(R)$ given by [@gallager] $$\label{eqn:rhospherepack}
E_{sp}(R) = \max_{\rho \geq 0} \big[ E_0(\rho) - \rho R \big]$$ with the Gallager function $E_0(\rho)$ defined as: $$\begin{aligned}
E_0(\rho) &=& \max_{\vec{q}} E_0(\rho,\vec{q}), \nonumber \\
E_0(\rho, \vec{q}) &=& -\ln \sum_y \bigg[ \sum_x q_x
p_{y|x}^\frac{1}{1+\rho} \bigg]^{(1+\rho)}. \label{eqn:enought}\end{aligned}$$ Since the random-coding error exponent is given by $$\label{eqn:rhorandom}
E_{r}(R) = \max_{0 \leq \rho \leq 1} \big[ E_0(\rho) - \rho R \big],$$ it is clear that the sphere-packing bound is achievable, even without feedback, at message rates close to $C$ since for those rates, $\rho <
1$ optimizes both expressions [@gallager].
It is less well appreciated that the points on the sphere-packing bound where $\rho > 1$ are also achievable by random coding if the sense of “correct decoding” is relaxed. Rather than forcing the decoder to emit a single estimated codeword, list decoding allows the decoder to emit a small list of guessed codewords. The decoding is considered correct if the true codeword is on the list. For list decoding with list size $\ell$ in the context of random codes, Problem 5.20 in [@gallager] reveals that $$\label{eqn:listrhorandom}
E_{r,\ell}(R) = \max_{0 \leq \rho \leq \ell} \big[ E_0(\rho) - \rho R \big]$$ is achievable. At high message rates (where the maximizing $\rho$ is small), there is no benefit from relaxing to list decoding, but it makes a difference at low rates.
![The sphere-packing bound divided up into two sections: a blue segment where list decoding is needed for random codebooks to achieve it, and a red segment where lists are not needed. The tangents represent list sizes of $8$, $4$, $2$, and $1$. []{data-label="fig:listnolist"}](listnolist.edited.eps){width="4in" height="3in"}
Figure \[fig:listnolist\] illustrates the range of exponents for which list decoding is required for a BSC. The blue part of the sphere-packing curve shows where list decoding is important and the red part shows where lists are not required. Four tangents are illustrated corresponding to list sizes of $8$, $4$, $2$, and $1$. The y-intercepts of these tangents represent the maximum error exponents possible using those list sizes and random codes.
For output-symmetric channels (see Definition \[def:channels\]), it is clear that $E^+(R) = E_{sp}(R)$ since the input distribution $\vec{r}$ can always be chosen to be uniform [@DobrushinReliability]. Thus, for fixed-block-length codes and output-symmetric DMCs, not only does causal feedback not improve capacity, it does not improve reliability either, at least at high rates.[^3]
The extreme limit of reliability in the fixed-block-length setting is given by the study of zero-error capacity, in which the probability of decoding error is required to be exactly zero. As pointed out in [@ShannonZeroError], this can be different with and without feedback. For zero-error capacity, the details of the channel matrix $P$ are not important as it clearly only depends on which entries are zero. The true zero-error capacity without feedback $C_0$ is very hard to evaluate, but the zero-error capacity with feedback $C_{0,f}$ can be easily evaluated when it is greater than zero [@ZeroErrorSurvey].
Although there is an explicit expression for $C_{0,f}$ in [@ShannonZeroError], the interpretation is more straightforward in the context of (\[eqn:rhospherepack\]). $$\label{eqn:zeroerrorexpression}
C_{0,f} = \lim_{\rho \rightarrow \infty} \frac{E_0(\rho)}{\rho}$$ was established in [@ShannonGallagerBerlekampI] by evaluating the limit and showing that it is identical to the expression for $C_{0,f}$ from [@ShannonZeroError]. If $C_{0,f}$ is nonzero, both the sphere-packing bound (\[eqn:rhospherepack\]) and Haroutunian bound (\[eqn:primitiveupperbound\]) are infinite at message rates below $C_{0,f}$ and finite above it.
Variable-length codes {#sec:varlength}
---------------------
Since feedback neither improves the capacity nor significantly improves the fixed-block-length reliability function, it seemed that this particular reliability somehow represented the wrong technical question to ask. After all, it was unable to answer why feedback seemed to be so useful in practice. The traditional response to this was to fall back to the issue of complexity.
Because classical decoding of fixed-block-length codes has a complexity that is not linear in the block length, the block length was viewed as a proxy for implementation complexity rather than only for end-to-end delay. Just as in variable-length source-coding, the idea in variable-block-length channel-coding is to extend use of the channel when the channel is behaving atypically. This way, the presumed complexity of increased block lengths is only experienced rarely and on average, the system can be simpler to operate.
Without feedback, a variable-length mode of operation is impossible since the encoder has no way to know if the channel is behaving typically or atypically. With noiseless feedback, the length of the codeword can be made to vary based on what the channel has done so far — as long as this variation depends only on the received channel symbols. This is the counterpart to the unique decodability requirement in source coding in that both are needed to prevent an irrecoverable loss of synchronization between the encoder and decoder.
One proposed error exponent for variable-length channel codes divides the negative log of the probability of block error $\epsilon$ by the expected block length $E[N_\epsilon]$ of an average rate-$\bar{R}$ variable-length code [@burnashev].
$$E_{vl}(\bar{R}) = \limsup_{\epsilon \rightarrow 0} -\frac{\ln(\epsilon)}{E[N_\epsilon]}.$$
Burnashev gave an upper bound to this exponent by using martingale arguments treating the ending of a block as a stopping time and studying the rate of decrease in the conditional entropy of the message at the receiver [@burnashev]. This gives $$\label{eqn:burnashevbound}
E_{v}(\bar{R}) = C_1 \bigg(1-\frac{\bar{R}}{C}\bigg)$$ where $C$ is the Shannon capacity of the channel and $$\label{eqn:c1def}
C_1 = \max_{x,x'} D(P(\cdot|x)||P(\cdot|x'))
= \max_{x,x'} \sum_{y} p_{y|x} \ln \frac{p_{y|x}}{p_{y|x'}}$$ represents the maximum divergence possible between channel output distributions given choice of two input letters.
While Burnashev gives an explicit variable-length scheme in [@burnashev] that asymptotically attains the exponent of (\[eqn:burnashevbound\]), the scheme of Yamamoto and Itoh in [@yamamoto] is simpler and makes clear the idea of separating reliability from efficiency. Suppose there is a single message of $nR$ nats to send:
1. Transmit the message using any reliable block code at a rate $\widetilde{R} < C$ close to capacity but larger than the target average rate $\bar{R}$. This will consume $n \frac{R}{\widetilde{R}}$ channel uses.
2. Use the noiseless feedback to decide at the encoder whether the message was received correctly or incorrectly.
3. If the message was received correctly, send a “confirm” signal by sending input $x$ from (\[eqn:c1def\]) repeated $n(1-\frac{R}{\widetilde{R}})$ times. Otherwise, use the channel to send a “deny” signal by repeating input $x'$ the same number of times.
This part can be interpreted as a sort of punctuation: a “deny” is a backspace telling the decoder to erase what it has seen so far while a “confirm” is a comma telling the decoder that this block is finished.
4. The decoder performs a simple binary hypothesis test on the received confirm/deny channel outputs to decide whether to accept the current message block. If it rejects the block, then the encoder will retransmit it until it is accepted. Since errors only occur when the message is falsely accepted, the decoder minimizes the probability of false alarm while holding the probability of missed detection to some acceptably low level.
Since retransmissions can be made as rare as desired as long as $\widetilde{R} < C$, the overall average rate $\bar{R}$ of the scheme approaches $R$. Since the number of slots for the “confirm/deny” message can be made to approach $n(1-\frac{R}{C})$, the reliability approaches (\[eqn:burnashevbound\]) by Stein’s Lemma [@coverthomas]. [ *Our approach to generic channels in Section \[sec:comments\] can be considered as using variable-block-length codes to achieve good fixed-delay performance by combining an alternative approach to punctuation with a softer sense of retransmission.*]{}
The Burnashev exponent is dramatically higher than the fixed-block-length exponents (see Figure \[fig:straightbsc\]) and thus seems to demonstrate the advantage of feedback. However, it is unclear what the significance of average delay or block length really is in a system. The block length under the Yamamoto and Itoh scheme is distributed like a scaled geometric random variable. Consequently, the block length will exceed a target deadline (like an underlying channel’s coherence time or an application-specific latency requirement) far more often than the scheme makes an undetected error. There are also no known nontrivial separation theorems involving either average block length or average delay.
Nonblock codes {#sec:intrononblock}
--------------
Another classical approach to the problem of reliable communication is to consider codes without any block structure. Convolutional and tree codes represent the prototypical examples. It was realized early on that in an infinite-constraint-length convolutional code under ML decoding, all bits will eventually be decoded correctly [@gallager]. Given that this asymptotic probability of error is zero, there are two possible ways to try to understand the underlying tradeoffs: look at complexity or look at the delay.
The traditional approach was to focus on complexity by examining the case of finite constraint lengths. The per-symbol encoding complexity of a convolutional code is linear in the constraint length, and if sequential decoding algorithms are used and the message rate is below the cutoff rate $E_0(1)$, so is the average decoding complexity [@ForneySeq]. With a fixed constraint length $\nu$, the probability of error cannot go to zero and so it is natural to consider the tradeoff between the error probability and constraint length $\nu$. Viterbi used a genie-aided argument to map the sphere-packing bound for block codes into an upper bound for fixed-constraint-length convolutional codes. (A variant of this argument is used in Section \[sec:feedbackbound\] to bound performance with delay.) This gives the following parametric upper bound for the exponent governing how fast the bit error probability can improve with the constraint length:[@viterbi] $$\label{eqn:viterbibound}
E_{c}(R) = E_0(\rho)\,\,;\,\, R = \frac{E_0(\rho)}{\rho}$$ where $\rho \geq 0$. The “inverse concatenation construction” (illustrated in Figure \[fig:erasurebounds\]) is the graphical representation of the above curve — it is the envelope of the $(R,E)$ intercepts traced out by the tangents to the sphere-packing bound. Thus, this upper bound can be tightened in the low-rate regime by using the “straight-line bound” from [@ShannonGallagerBerlekampII]. The bound (\[eqn:viterbibound\]) is also achievable in the high-rate regime ($R > E_0(1)$) [@ForneySeq].
The $E_c(R)$ from (\[eqn:viterbibound\]) for fixed constraint lengths is substantially higher than $E_{sp}(R)$ from (\[eqn:rhospherepack\]) for fixed block-lengths. This was used to argue for the superiority of convolutional codes over block codes from an implementation point of view. However, it is important to remember that this favorable comparison does not hold when end-to-end delay, rather than complexity, is considered.
If the end-to-end delay is forced to be bounded, then the bit-error probability with delay is governed by $E_{r}(R)$ for random convolutional codes, even when the constraint lengths are unbounded [@ForneyML]. This performance with delay is also achievable using an appropriately biased sequential decoder [@JelinekSequential]. A nice feature of sequential decoders is that they are not tuned to any target delay — they can be prompted for estimates at any time and they will give the best estimate that they have. Thus an infinite-constraint-length convolutional code with appropriate sequential decoding achieves the exponent $E_r(R)$ delay universally over all (sufficiently long) delays. This property turns out to be important for this paper since such codes are used in place of two-point block codes to encode punctuation information in Section \[sec:comments\].
The role of feedback in nonblock codes has also been investigated considerably by considering a variety of different schemes [@Horstein; @Fang; @HashimotoARQ; @Veugen; @SchalkwijkPost; @KudryashovPPI; @KudryashovIT], each with an idiosyncratic way of defining a relevant error exponent. The simplest approach is to consider a variable-constraint-length model in which complexity is counted by the expected number of multiply-accumulate operations that are required to encode a new channel symbol. This is done in Appendix \[app:feedbackconvolution\]. The result is that for all rates below the computational cutoff rate, a finite amount of expected computation per input bit is enough to get an arbitrarily low probability of error — that the computational error exponent is infinite.
At first glance, this infinite exponent seems to show the superiority of variable-constraint-length codes over variable-block-length codes with feedback. After all, the Burnashev bound (\[eqn:burnashevbound\]) is only infinite for channels whose probability matrices $P$ contains a zero. However, this is not a fair comparison since it is comparing expected per-channel-use computational complexity here with expected block length in the variable-block-length case.
The variable-block-length schemes of Ooi and Wornell [@ooiwornell; @Ooi] achieve linear complexity in the block length for the message-communication part. Once complexity is linear in the expected length, it is constant on an average per-symbol basis. Thus block codes can also achieve any desired probability of error by adjusting the length of the confirm/deny phase in the same way that a large enough terminator $d$ can be chosen for the variable-constraint-length convolutional codes of Appendix \[app:feedbackconvolution\]. So both have infinite computational error exponents with feedback.
An infinite exponent just means that the asymptotic tradeoff of probability of error with expected per-symbol computation is uninteresting when noiseless feedback is allowed. As a result, it is very natural to consider the tradeoff with end-to-end delay instead. The open questions that are addressed in this paper are whether the end-to-end delay performance can generally be improved using feedback, and if so, what are the limits to such improvements.
Main results and examples {#sec:mainresults}
=========================
First, some basic definitions are needed. Vector notation $\vec{x}$ is used to denote sequences $x_1^n$ where the indices are obvious from the context.
\[def:channels\] A [*discrete time discrete memoryless channel*]{} (DMC) is a probabilistic system with an input and an output. At every time step $t$, it takes an input $x_t \in {\cal X}$ and produces an output $y_t \in {\cal Y}$ with probability ${\cal P}(Y_t = y|X_t = x) =
p_{y|x}$. Both ${\cal X},{\cal Y}$ are finite sets and the transition probability matrix $P$ containing the $p_{y|x}$ entries is a stochastic matrix. The current channel output is independent of all past random variables in the system conditioned on the current channel input.
Following [@gallager page 94], a DMC is called [ *output-symmetric*]{} if the set of outputs ${\cal Y}$ can be partitioned into disjoint subsets[^4] in such a way that for each subset, the matrix of transition probabilities has the property that each row is a permutation of each other row and each column is a permutation of each other column.
\[def:nofeedbackencoder\] A [*rate-$R$ encoder ${\cal E}$ without feedback*]{} is a sequence of maps $\{{\cal E}_t\}$. Each ${\cal E}_t: \{0,1\}^{\lfloor R't \rfloor}
\rightarrow {\cal X}$ where the range is the finite set of channel inputs $\cal X$. The $t$-th map takes as input the available message bits $B_1^{\lfloor R't \rfloor}$ where $R' = \frac{R}{\ln 2}$ is the encoder’s rate in bits rather than nats per channel use.
For a [*rate-$R$ encoder with noiseless feedback*]{}, the maps ${\cal E}_t: {\cal Y}^{t-1} \times \{0,1\}^{\lfloor R't \rfloor}
\rightarrow {\cal X} $ also get access to all the past channel outputs $Y_1^{t-1}$.
A [*delay-$d$ rate-$R$ decoder*]{} is a sequence of maps $\{ {\cal
D}_i \}$. Each ${\cal D}_i: {\cal Y}^{\lceil \frac{i}{R'} \rceil + d}
\rightarrow \{0,1\}$ where the output of each map is the estimate $\widehat{B}_i$ for the $i$-th bit. The $i$-th map takes as input the available channel outputs $Y_1^{\lceil \frac{i}{R'} \rceil + d}$. This means that it can see $d$ time units (channel uses) beyond when the bit to be estimated first had the potential to influence the channel inputs.
[*Randomized encoders and decoders*]{} also have access to random variables $W_t$ denoting common randomness available in the system.
\[def:achievable\] The fixed-delay error exponent $\alpha$ is asymptotically [*achievable*]{} at message rate $R$ across a noisy channel if for every delay $d_j$ in some strictly increasing sequence indexed by $j$ there exist rate-$R$ encoders ${\cal E}^{j}$ and delay-$d_j$ rate-$R$ decoders ${\cal D}^{j}$ that satisfy the following properties when used with input bits $B_i$ drawn from iid fair coin tosses.
1. For the $j$-th code, there exists an $\epsilon_j < 1$ so that ${\cal P}(B_i
\neq \widehat{B}_i(d_j)) \leq \epsilon_j$ for every bit position $i \geq 1$. The $\widehat{B}_i(d_j)$ represents the delay-$d_j$ estimate of $B_i$ produced by the $({\cal E}^j,
{\cal D}^j)$ pair connected through the channel in question.
2. $\lim_{j \rightarrow \infty} \frac{-\ln \epsilon_j}{d_j} \geq
\alpha$
The exponent $\alpha$ is asymptotically [*achievable universally over delay*]{} or in an [*anytime fashion*]{} if a single encoder ${\cal
E}$ can be used simultaneously for all sufficiently long delays $d$.
Main results
------------
With these definitions, the five main results of this paper can be stated:
\[thm:nofeedbackbound\] For a DMC, no fixed-delay exponent greater than the Haroutunian bound ($\alpha > E^+(R)$ from (\[eqn:Haroutunianbound\])) is asymptotically achievable without feedback.
[*Uncertainty-focusing bound:*]{} \[thm:generalfeedbackbound\] For a DMC, no delay exponent $\alpha > E_{a}(R)$ is asymptotically achievable even if the encoders are allowed access to noiseless feedback. $$\label{eqn:generalfeedbackbound}
E_a(R) = \inf_{0 \leq \lambda < 1} \frac{E^+(\lambda R)}{1-\lambda}$$ where $E^+$ is the Haroutunian bound from (\[eqn:Haroutunianbound\]). Whenever $E^+(R) = E_{sp}(R)$ (e.g. the DMC is output-symmetric), $E_a(R) = E_{a,s}(R)$ where the latter is expressed parametrically as $$\begin{aligned}
\label{eqn:symmetricfeedbackbound}
E_{a,s}(R) & = & E_0(\eta),\\
R & = & \frac{E_0(\eta)}{\eta} \nonumber\end{aligned}$$ where $E_0(\eta)$ is the Gallager function from (\[eqn:enought\]), and $\eta$ ranges from $0$ to $\infty$.
The curve (\[eqn:symmetricfeedbackbound\]) has negative slope of at least $2C/ \frac{\partial^2 E_0(0)}{\partial \eta^2}$ in the vicinity of the $(C,0)$ point where the derivatives of $E_0$ are taken fixing the capacity-achieving distribution.
\[thm:erasurecase\] For the binary erasure channel with erasure probability $\beta > 0$, there exists a code using noiseless feedback with a delay error exponent that asymptotically approaches the uncertainty-focusing bound $E_a(R)$ for all message rates $R < C$. Viewed as a reliability-dependent capacity, the tradeoff is given by $$\label{eqn:becanytimecapacity}
C'(\alpha) = \frac{\alpha}{\alpha +
\log_2 \left(\frac{1-\beta}{1-2^\alpha \beta}\right)}$$ where $\alpha$ is the desired reliability (in base 2) with fixed delay and $C'(\alpha)$ is the supremal rate (in bits per channel use) at which reliable communication can be sustained with fixed-delay reliability $\alpha$.
Furthermore, for every $r \geq \frac{2 - \log_2 \log_2 \beta^{-1}}{\log_2
\beta^{-1}}$ (in particular: any $r \geq 0$ as long as $\beta \leq
\frac{1}{16}$), at all rates $R' < \frac{1}{1+2r}$ bits per channel use, the error exponent (in base 2) with respect to delay is $\geq
\log_2 \beta^{-1} - 2\beta^r$.
\[thm:withzeroerror\] For any DMC with strictly positive zero-error capacity $C_{0,f} > 0$, it is possible to asymptotically approach all delay exponents within the region $\alpha < E_{a,s}(R)$ defined by (\[eqn:symmetricfeedbackbound\]) using noiseless feedback and randomized encoders, even if the feedback is delayed by a constant $\phi$ channel uses.
This rate/reliability region can also be asymptotically achieved for any DMC by an encoder/decoder pair that has access to noiseless feedback if it also has access to an error-free forward communication channel with any strictly positive rate.
Furthermore, the delay exponents can be achieved in a delay-universal or “anytime” sense.
As is shown in Section \[sec:fortifiedfeedback\], the scheme that approaches the uncertainty-focusing bound is built around a variable-length channel code with the zero-error aspects used to convey unambiguous “punctuation” information that allows the decoder to stay synchronized with the encoder. Without any zero-error capacity, this punctuation information can be encoded in a separate parallel stream of channel uses to give the following result.
\[thm:genericachieve\] For any DMC, it is possible with noiseless feedback and randomized encoders to asymptotically achieve all delay exponents $\alpha < E'(R)$ where the tradeoff curve is given parametrically by varying $\rho \in (0,\infty)$: $$\begin{aligned}
\label{eqn:timesharingregion}
E'(\rho) & = & \left(\frac{1}{E_0(\rho)} + \frac{1}{E_0(1)}\right)^{-1}, \\
R(\rho) & = & \frac{E'(\rho)}{\rho}. \nonumber\end{aligned}$$
The curve (\[eqn:timesharingregion\]) has strictly negative slope $-E_0(1) / (C - \frac{E_0(1)}{2C} \left(\frac{\partial^2
E_0(0)}{\partial \rho^2} \right))$ in the vicinity of the $(C,0)$ point.
Furthermore, these delay exponents are also achievable in a delay-universal or “anytime” sense.
The fact that this achievable region (\[eqn:timesharingregion\]) generically has strictly negative slope in the vicinity of $(C,0)$ while the Haroutunian bound $E^+$ and sphere-packing bound $E_{sp}$ both generically approach $(C,0)$ only quadratically with zero slope establishes that noiseless feedback generally improves the tradeoff between end-to-end delay and the probability of error.
The above results relate to the strict interior of the region defined by $E_{a,s}(R)$ or $E'(R)$ for achievability and the strict exterior region corresponding to $E_a(R)$ for the converse. Unlike the case of fixed-block-length codes where the sphere-packing bound is known to be achievable at high rates, the results above do not cover points on the $E_{a,s}(R)$ curve itself at any rates.
The results of Theorems \[thm:erasurecase\], \[thm:withzeroerror\] and \[thm:genericachieve\] are also stated using asymptotic language — they apply in the limit of large end-to-end delays. In the case of Theorems \[thm:withzeroerror\] and \[thm:genericachieve\], the parameters defining the randomized codes are also allowed to get asymptotically large in order to approach the delay-error-exponent frontier. However, the proofs use techniques that make it possible to evaluate the performance of schemes with finite parameters.
Numerical examples {#sec:regularexamples}
------------------
![The binary erasure channel with $\beta=0.4$. The vertical dashed line represents the rate of $\frac{1}{2}$ bits per channel use while the horizontal dashed line is the ultimate limit of $-\ln(0.4)$ for the reliability function. Notice how the uncertainty-focusing bound gets very close to that ultimate bound even at moderately small rates. The triangle illustrates the “inverse concatenation construction” connecting the two bounds to each other. []{data-label="fig:erasurebounds"}](erasurebounds.edited.eps){width="4in" height="3in"}
The erasure channel is the simplest channel for understanding the asymptotic tradeoffs between message rate, end-to-end delay, and probability of error when noiseless feedback is allowed. Figure \[fig:erasurebounds\] illustrates how when the erasure probability $\beta$ is small, even moderately low rates achieve spectacular reliabilities with respect to fixed delay.
![The sphere-packing, uncertainty-focusing, and Burnashev bounds for a BSC with crossover probability $0.02$. The thin lines represent the parametric bounds in (\[eqn:generalfeedbackbound\]) setting $\lambda = \frac{1}{8}, \frac{1}{2}, \frac{7}{8}$ and the thick middle curve is the uncertainty-focusing bound — the lower envelope of the parametric bounds over all $\lambda$. []{data-label="fig:straightbsc"}](threebounds.edited.eps){width="4in" height="6in"}
Now, consider a binary symmetric channel with crossover probability $0.02$. The capacity of this channel is about $0.60$ nats per channel use. Figure \[fig:straightbsc\] shows how the different choices of $\lambda$ used in the bound (\[eqn:generalfeedbackbound\]) kiss the uncertainty-focusing bound for the BSC. It also shows the Burnashev bound for variable-block-length coding for comparison. In this particular plot, the Burnashev bound appears to always be higher than the uncertainty-focusing bound. Figure \[fig:focusvsburnashev\] illustrates that this is not always the case by plotting both bounds in the high-rate regime for a BSC with crossover probability $0.003$. It is unknown whether any scheme can actually achieve fixed-delay reliabilities above the Burnashev bound since the scheme of Theorem \[thm:genericachieve\] does not do so.
![The uncertainty-focusing and Burnashev bounds for a BSC with crossover probability $0.003$.[]{data-label="fig:focusvsburnashev"}](focusvsburnashev.edited.eps){width="4in" height="3in"}
![The delay-reliability bounds for the BSC used with noiseless feedback. The sphere-packing bound approaches capacity in a quadratically flat manner while the uncertainty-focusing bound and the scheme of Theorem \[thm:genericachieve\] both approach the capacity point linearly, albeit with different slopes. The random-coding error exponent is also plotted for convenience since it beats the scheme of Theorem \[thm:genericachieve\] at low rates and can be attained without feedback.[]{data-label="fig:beatspherebsc"}](twoboundsplusrandomplustimesharing.edited.eps){width="4in" height="3in"}
The gap between the uncertainty-focusing bound and the scheme of Theorem \[thm:genericachieve\] is illustrated in Figure \[fig:beatspherebsc\] for the BSC. This also shows how the sphere-packing bound is significantly beaten at high rates even when the channel has no zero-error capacity and thus provides an explicit counterexample to Pinsker’s Theorem 8 in [@PinskerNoFeedback]. Examples showing how the uncertainty-focusing bound is met for communication systems with strictly positive zero-error capacity are deferred to Section \[sec:fortifiedexamples\].
Upper-bounding the fixed-delay reliability function without feedback {#sec:nofeedback}
====================================================================
This section proves Theorem \[thm:nofeedbackbound\] giving a generalization of Pinsker’s BSC argument from [@PinskerNoFeedback] to the case of general DMCs. The Haroutunian exponent $E^+(R)$ from (\[eqn:Haroutunianbound\]) is shown to upper bound the reliability function with delay if feedback is not available. For output-symmetric DMCs, this is the same as the sphere-packing bound $E_{sp}(R)$. Furthermore, we discuss why this proof [*does not*]{} go through when feedback is present.
The complete proof spans the next few sections with some technical details in the Appendices.
Feedforward decoders and their equivalent forms
-----------------------------------------------
For notational convenience, assume that $R' < 1$ so that at least one channel use comes between each message bit’s arrival. If $R' > 1$, the same argument will work (at the cost of uglier notation) by considering the incoming bits to arrive in pairs, triples, etc. Theorem \[thm:nofeedbackbound\] is proven by considering a more powerful class of decoders that have access to extra information that can only improve their performance.
\[def:feedforwarddecoder\] A [*delay-$d$ rate-$R$ decoder ${\cal D}$ with feedforward information*]{} is a decoder ${\cal D}_i: \{0,1\}^{i-1}\times {\cal
Y}^{\lceil \frac{i}{R'} \rceil +
d} \rightarrow \{0,1\} $ that has noiseless access to the past message bits $B_1^{i-1}$ in addition to the available channel outputs $Y_1^{\lceil \frac{i}{R'} \rceil + d}$.
The first property is that with access to the feedforward information, it suffices to ignore very old channel outputs.
\[lem:windowdependence\] For a memoryless channel, given a rate-$R$ encoder ${\cal E}$ without feedback and a delay-$d$ rate-$R$ decoder ${\cal D}_i$ with feedforward for bit $i$, there exists a decoder ${\cal D}^f_i: \{0,1\}^{i-1}\times {\cal Y}^{d+1}
\rightarrow \{0,1\}$ for bit $i$ that only depends on all the past message bits $B_1^{i-1}$ and the recent channel outputs $Y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil
+ d}$. The bit error probability ${\cal P}(B_i \neq {\cal
D}^f_i(B_1^{i-1},Y_{\lceil \frac{i}{R'} \rceil}^{\lceil
\frac{i}{R'} \rceil + d})) \leq {\cal P}(B_i \neq {\cal
D}_i(B_1^{i-1},Y_1^{\lceil \frac{i}{R'} \rceil + d}))$ assuming that the message bits $B$ are all iid fair coin tosses.
[*Proof:*]{} The result follows immediately from the following Markov chain that holds since there is no feedback. $$\label{eqn:basicmarkovchain}
Y_1^{\lceil \frac{i}{R'} \rceil -
1}~-~B_1^{i-1}X_1^{\lceil \frac{i}{R'} \rceil - 1}~-~B_i X_{\lceil \frac{i}{R'}
\rceil}^{\lceil \frac{i}{R'} \rceil + d} Y_{\lceil
\frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}.$$
To see the result explicitly, let ${\cal D}^{map}_i$ be the MAP decoder for bit $i$ based on feedforward information $B_1^{i-1}$ and observations $Y_1^{\lceil \frac{i}{R'} \rceil + d}$.
$$\begin{aligned}
& & {\cal D}^{map}_i(b_1^{i-1},y_1^{\lceil \frac{i}{R'} \rceil + d}) \\
& = &
\operatorname*{argmax}_{b_i} {\cal P}(B_i = b_i|B_1^{i-1} = b_1^{i-1}, Y_1^{\lceil \frac{i}{R'}
\rceil + d} = y_1^{\lceil \frac{i}{R'} \rceil + d}) \\
& = &
\operatorname*{argmax}_{b_i} {\cal P}(B_i = b_i, B_1^{i-1} = b_1^{i-1}, Y_1^{\lceil \frac{i}{R'}
\rceil + d} = y_1^{\lceil \frac{i}{R'} \rceil + d}) \\
& = &
\operatorname*{argmax}_{b_i} \sum_{b_{i+1}^{i + \lceil dR' \rceil}} {\cal
P}(B_i = b_i, B_1^{i-1} = b_1^{i-1}, B_{i+1}^{i + \lceil dR' \rceil} = b_{i+1}^{i + \lceil dR' \rceil}, Y_1^{\lceil \frac{i}{R'}
\rceil + d} = y_1^{\lceil \frac{i}{R'} \rceil + d}) \\
& = &
\operatorname*{argmax}_{b_i} {\cal P}(B_1^{i-1} = b_1^{i-1}, Y_1^{\lceil \frac{i}{R'} \rceil - 1} = y_1^{\lceil
\frac{i}{R'} \rceil - 1}) \\
& & ~~~~ \cdot \sum_{b_{i+1}^{i + \lceil dR' \rceil}}
{\cal P}(B_{i}^{i + \lceil dR' \rceil} = b_{i}^{i + \lceil dR' \rceil})
{\cal P}(Y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'}
\rceil + d} = y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}
\big| B_1^{i + \lceil dR' \rceil} = b_1^{i + \lceil dR' \rceil},
Y_1^{\lceil \frac{i}{R'} \rceil - 1} = y_1^{\lceil \frac{i}{R'} \rceil
- 1}) \\
& =_{(a)} &
\operatorname*{argmax}_{b_i} \sum_{b_{i+1}^{i + \lceil dR' \rceil}}
{\cal P}(Y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'}
\rceil + d} = y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}
\big| B_1^{i + \lceil dR' \rceil} = b_1^{i + \lceil dR' \rceil},
Y_1^{\lceil \frac{i}{R'} \rceil - 1} = y_1^{\lceil \frac{i}{R'} \rceil - 1}) \\
& =_{(b)} &
\operatorname*{argmax}_{b_i} \sum_{b_{i+1}^{i + \lceil dR' \rceil}}
{\cal P}(Y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'}
\rceil + d} = y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}
\big| X_1^{\lceil \frac{i}{R'} \rceil + d} = {\cal E}(b_1^{i + \lceil dR' \rceil}),
B_1^{i + \lceil dR' \rceil} = b_1^{i + \lceil dR' \rceil},
Y_1^{\lceil \frac{i}{R'} \rceil - 1} = y_1^{\lceil \frac{i}{R'} \rceil - 1}) \\
& =_{(c)} &
\operatorname*{argmax}_{b_i} \sum_{b_{i+1}^{i + \lceil dR' \rceil}}
{\cal P}(Y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'}
\rceil + d} = y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}
\big| X_1^{\lceil \frac{i}{R'} \rceil + d} = {\cal E}(b_1^{i + \lceil dR' \rceil})). \end{aligned}$$
The first few lines above are standard expansions of probability in the MAP context and use the fact that the message bits are drawn iid. (a) holds by dropping terms that do not depend on the exact values for $b_i^{i + \lceil dR' \rceil}$ and thus do not impact the $\operatorname*{argmax}$. (b) uses the fact that the channel input $X$ is entirely determined[^5] by the message bits $B$ for an encoder without feedback. (c) is due to the memoryless nature of the channel.
Define ${\cal D}^f_i$ directly as $$\begin{aligned}
& & {\cal D}^f_i(b_1^{i-1},y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}) \\
& = &
\operatorname*{argmax}_{b_i} \sum_{b_{i+1}^{i + \lceil dR' \rceil}}
{\cal P}(Y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'}
\rceil + d} = y_{\lceil \frac{i}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}
\big| X_1^{\lceil \frac{i}{R'} \rceil + d} = {\cal E}(b_1^{i + \lceil dR' \rceil})). \end{aligned}$$
This decoder only depends on the recent channel outputs in addition to the feedforward information and achieves MAP performance. Since MAP is optimal, the probability of bit error would be the same or better than any other decoder. $\Diamond$
The second property is that it suffices to feedforward the error sequence $\widetilde{B}_i = \widehat{B}_i + B_i \bmod 2$ rather than the past message bits themselves.
\[lem:forwarderrors\] Given a rate-$R$ encoder ${\cal E}$ and delay-$d$ rate-$R$ decoder ${\cal D}_i$ for bit $i$ with feedforward. There exists another decoder $\widetilde{\cal
D}_i: \{0,1\}^{i-1}\times {\cal Y}^{\lceil \frac{i}{R'} \rceil +
d} \rightarrow \{0,1\}$ that only depends on the error sequence $\widetilde{B}_1^{i-1}$ in addition to the channel outputs $Y_{1}^{\lceil \frac{i}{R'} \rceil + d}$. If $\widetilde{B}_i =
\widehat{B}_i + B_i \bmod 2$, then the outputs of the two decoders are identical $\widetilde{\cal
D}_i(\widetilde{B}_1^{i-1},Y_{1}^{\lceil \frac{i}{R'} \rceil + d})
= {\cal D}_i(B_1^{i-1},Y_{1}^{\lceil \frac{i}{R'} \rceil + d})$.
[*Proof:*]{} This holds very generally by induction. Neither memorylessness nor even the absence of feedback is required. It clearly holds for $i=1$ since there are no prior bits and so the same $\widehat{B}_1$ results. Assume now that it holds for all $j<k$ and consider $i=k$. By the induction hypothesis, the action of all the prior decoders $j$ can be simulated since the decoder has access to $\widetilde{B}_1^{j-1}$ and $Y_{1}^{\lceil \frac{j}{R'} \rceil + d}$. The resulting estimates $\widehat{B}_j$ for $j<k$ can be XORed with $\widetilde{B}_j$ to recover $B_j$ itself. Since $B_1^{k-1}$ can be recovered from the given information, the original ${\cal D}_k$ decoder can be run as a subroutine to give $\widehat{B}_k$. $\Diamond$
Lemmas \[lem:windowdependence\] and \[lem:forwarderrors\] tell us that feedforward decoders can be thought in three ways: having access to all past message bits and all past channel outputs, having access to all past message bits and only a recent window of past channel outputs, or having access to all past decoding errors and all past channel outputs.
Constructing a rate-$(R - \delta_1)$ block code
-----------------------------------------------
Consider the system illustrated in Figure \[fig:newdataprocessing\]. The message bitstream consisting of fair coin tosses is encoded using the given rate-$R$ encoder. The channel outputs are decoded using the delay-$d$ rate-$R$ decoders with feedforward, with the feedforward in the form of the error signals $\widetilde{B}$ by Lemma \[lem:forwarderrors\]. These error signals are generated by XORing the message bits with the output of an equivalent feedforward decoder. Finally, the feedforward error signals are used one more time and combined with the estimates $\widehat{B}$ to recover the message bits $B$ exactly. It is immediately clear that this hypothetical system never makes an error from end to end.
\#1\#2\#3\#4\#5[ @font ]{}
(6997,2594)(1929,-3683) (3451,-2611)[(0,0)\[b\]]{} (3451,-2836)[(0,0)\[b\]]{} [(4951,-3136)[(900,900)]{} ]{}(5401,-2536)[(0,0)\[b\]]{} (5401,-2761)[(0,0)\[b\]]{} (5401,-2986)[(0,0)\[b\]]{} [(6901,-2011)[(900,900)]{} ]{}(7351,-1411)[(0,0)\[b\]]{} (7351,-1636)[(0,0)\[b\]]{} (7351,-1861)[(0,0)\[b\]]{} [(4801,-1561) ]{}[(2401,-2011)[(900,900)]{} ]{}[(1951,-1561)[( 1, 0)[450]{}]{} ]{}[(3301,-1561)[( 1, 0)[1050]{}]{} ]{}[(5251,-1561)[( 1, 0)[1650]{}]{} ]{}[(2101,-1561)[( 0,-1)[1125]{}]{} (2101,-2686)[( 1, 0)[900]{}]{} ]{}[(3901,-2686)[( 1, 0)[1050]{}]{} ]{}[(5401,-1561)[( 0,-1)[675]{}]{} ]{}[(5851,-2686)[( 1, 0)[450]{}]{} ]{}[(6301,-2911)[(600,450)]{} ]{}[(2401,-2686)[( 0,-1)[975]{}]{} (2401,-3661)[( 1, 0)[4200]{}]{} (6601,-3661)[( 0, 1)[750]{}]{} ]{}[(6901,-2686)[( 1, 0)[450]{}]{} (7351,-2686)[( 0, 1)[675]{}]{} ]{}[(7801,-1561)[( 1, 0)[300]{}]{} (8101,-1561)[( 0,-1)[900]{}]{} ]{}[(7801,-2911)[(600,450)]{} ]{}[(7351,-2686)[( 1, 0)[450]{}]{} ]{}[(8401,-2686)[( 1, 0)[450]{}]{} ]{}[(5701,-1111)(12.718,-12.718)[201]{}[(1.6667,11.6667)[.]{}]{} ]{}(4801,-1561)[(0,0)\[b\]]{} (4801,-1786)[(0,0)\[b\]]{} (2851,-1561)[(0,0)\[b\]]{} (2851,-1786)[(0,0)\[b\]]{} (6601,-2761)[(0,0)\[b\]]{} (7951,-1411)[(0,0)\[lb\]]{} (8101,-2761)[(0,0)\[b\]]{} (8926,-2761)[(0,0)\[lb\]]{} (1951,-1636)[(0,0)\[rb\]]{} (3751,-1486)[(0,0)\[b\]]{} (5701,-1486)[(0,0)\[b\]]{} (5551,-2011)[(0,0)\[lb\]]{} (7051,-2911)[(0,0)\[lb\]]{} (7401,-2311)[(0,0)\[lb\]]{} (4476,-2611)[(0,0)\[b\]]{} (5951,-2436)[(0,0)\[lb\]]{}
Now, this system will be interpreted as a block code. Pick an arbitrarily small $\delta_1 > 0$. To avoid cumbersome notation, some integer effects will be neglected. For every delay $d$, pick a block length $n = \frac{d R}{\delta_1}$. For notational convenience, let $\delta_1'$ be such that $\frac{(R - \delta_1)}{\ln
2} = R' - \delta_1'$ so that $n (R' - \delta_1') = n R' - dR'$.
The data processing inequality implies:
\[lem:newdataprocessing\] Suppose $n$ is the block length, the block rate is $R - \delta_1$ nats per channel use, the $X_1^{n}$ are the channel inputs, the $Y_1^{n}$ are the channel outputs, and the $\widetilde{B}_1^{n (R' -
\delta_1')}$ are the error signals coming from the underlying rate-$R$ delay-$d$ encoding and decoding system. Then $$\label{eqn:newdataprocessing}
H(\widetilde{B}_1^{n(R' - \delta_1')}) \geq n (R-\delta_1)- I(X_1^{n};
Y_1^{n}).$$
[*Proof:*]{} See Appendix \[app:lemdataprocessing\].
Lower-bounding the error probability
------------------------------------
Now, suppose this system of Figure \[fig:newdataprocessing\] were to be run over the noisy channel $G$ that minimizes (\[eqn:Haroutunianbound\]) at $R - 2\delta_1$ nats per channel use. Since the capacity of $G$ is at most $R - 2\delta_1$ nats per channel use and there is no feedback to the encoder, the mutual information between the channel inputs and outputs is upper-bounded by $$\label{eqn:goofychannelinformation}
I(X_1^{n};Y_1^{n}) \leq n(R - 2\delta_1) = n(R - \delta_1) - n\delta_1.$$ Plugging (\[eqn:goofychannelinformation\]) into (\[eqn:newdataprocessing\]) from Lemma \[lem:newdataprocessing\] gives $$\label{eqn:toomuchentropy}
H(\widetilde{B}_1^{n(R' - \delta_1')}) \geq n \delta_1.$$ Since the sum of marginal entropies $\sum_{i=1}^{n(R' -
\delta_1')}H(\widetilde{B}_i) \geq H(\widetilde{B}_1^{n(R' -
\delta_1')})$, the average entropy of the error bits $\widetilde{B}_i$ is at least $\frac{\delta_1}{R' - \delta_1'} >
0$. Consider $i^*$ whose individual entropy $H(\widetilde{B}_{i^*}) \geq \frac{\delta_1}{R' - \delta_1'}$.
By the strict monotonicity of the binary entropy function for probabilities less than $\frac{1}{2}$, there exists a $\delta_2 > 0$ so that the probability of bit error ${\cal P}(\widetilde{B}_{i^*} = 1) =
{\cal P}(\widehat{B}_{i^*} \neq B_{i^*}) \geq \delta_2$. While the specific positions $i^*$ might vary for different delays $d$, the lower bound $\delta_2$ on minimum error probability does not vary.
At this point, Lemma \[lem:windowdependence\] implies that even if the channel $G$ were used only for the $d + 1$ time steps from $[\lceil\frac{i^*}{R'} \rceil, \lceil \frac{i^*}{R'} \rceil + d]$, the same minimum error probability $\delta_2$ must hold, regardless of how large $d$ is. For each possible message prefix $b_1^{i^*}$, there is an error event $A(b_1^{i^*})$ corresponding to the channel outputs that would cause erroneous decoding of the $i^*$-th bit. Formally, $A(b_1^{i^*}) := \{y_{\lceil\frac{i^*}{R'}\rceil}^{\lceil
\frac{i^*}{R'} \rceil + d} | {\cal
D}^f_{i^*}(b_1^{i^*-1},y_{\lceil\frac{i^*}{R'}\rceil}^{\lceil
\frac{i^*}{R'} \rceil + d}) \neq b_{i^*}\}$.
Averaging out the probability of error over message prefixes gives $$\begin{aligned}
\delta_2
& \leq &\sum_{b_1^{i^*}} \frac{1}{2^{i^*}}
{\cal P}(A(b_1^{i^*})|B_1^{i^*} = b_1^{i^*}) \\
& = & \sum_{b_1^{\lfloor (\lceil \frac{i^*}{R'} \rceil + d_j)R'
\rfloor}} 2^{-\lfloor (\lceil \frac{i^*}{R'} \rceil + d_j)R' \rfloor}
{\cal P}(A(b_1^{i^*})|\vec{X} = {\cal E}(b_1^{\lfloor (\lceil
\frac{i^*}{R'} \rceil + d_j)R'\rfloor})).\end{aligned}$$ Since the average over messages $b_1^{\lfloor (\lceil \frac{i^*}{R'}
\rceil + d_j)R' \rfloor}$ is at least $\delta_2$, and the probabilities can be no bigger than $1$ and no smaller than $0$, at least a $\frac{\delta_2}{2}$ proportion of messages result in the $A(b_1^{i^*})$ having a conditional probability of at least $\frac{\delta_2}{2}$ if channel $G$ is used.
All that remains is to show that the probability of this event under the true channel $P$ cannot be too small. To distinguish between the probability of an event when using channel $P$ or channel $G$, subscripts are used with ${\cal P}_P$ used to refer to the probability of an event when the channel is $P$ and ${\cal P}_G$ used for when the channel is $G$.
This simple lemma is useful:
\[lem:measurechange\] If under channel $G$ and input sequence $\vec{x}$, the probability ${\cal P}_G(Y_1^d \in A|X_1^d = \vec{x}) \geq \delta > 0$, then for any $\epsilon > 0$, there exists $d_0(\epsilon, \delta, G, P)$ so that as long as $d > d_0(\epsilon, \delta, G, P)$, the $A$ event’s conditional probability using channel $P$ must satisfy ${\cal
P}_P(Y_1^d \in A|X_1^d = \vec{x}) \geq \frac{\delta}{2} \exp(-d
(D(G||P | \vec{r}) + \epsilon)$ where $\vec{r}$ is the type of $\vec{x}$.
[*Proof:* ]{} See Appendix \[app:lemmeasurechange\].
Given $\frac{\delta_2}{2}, G$ and an arbitrary $\epsilon > 0$, apply Lemma \[lem:measurechange\] to consider delays $d > d_0$. This reveals that $$\begin{aligned}
{\cal P}_P(B_{i^*} \neq \widehat{B}_{i^*})
& = &
\sum_{b_1^{\lfloor (\lceil \frac{i^*}{R'} \rceil + d_j)R' \rfloor}}
2^{-\lfloor (\lceil \frac{i^*}{R'} \rceil + d_j)R' \rfloor}
{\cal P}_P\left(A(b_1^{i^*})|\vec{X} =
{\cal E}(b_1^{\lfloor (\lceil \frac{i^*}{R'} \rceil + d_j)R'\rfloor})\right) \\
& \geq_{(a)} & \frac{(\delta_2)^2}{8} \exp(-(d_j+1)(\max_{\vec{r}}
D(G||P | \vec{r}) + \epsilon)) \\
& =_{(b)} & \frac{(\delta_2)^2}{8} \exp\left(-(d_j+1)(E^+(R - 2\delta_1) + \epsilon)\right).\end{aligned}$$ (a) follows from the fact that a proportion $\frac{\delta_2}{2}$ of the messages must have probability of bit error of at least $\frac{\delta_2}{2}$ with the final factor of $2$ coming from Lemma \[lem:measurechange\]. Since the local type of the channel input is unknown, the maximum is taken over the channel input type $\vec{r}$. (b) is using the definition of $G$ and the Haroutunian bound.
Since $\epsilon > 0$ is an arbitrary choice and $\delta_2$ does not depend on the delay $d$, taking logs quickly reveals that the error exponent with delay cannot be any larger than $E^+(R -
2\delta_1)$. For any $\alpha > E^+(R)$, it is always possible to pick a $0 < \delta_1 < \frac{R - C_{0,f}}{2}$ so that $\alpha > E^+(R -
2\delta_1)$ as well since the Haroutunian bound $E^+$ is continuous in the rate for all rates strictly below Shannon capacity and above the feedback zero-error capacity $C_{0,f}$. Thus, no exponent $\alpha >
E^+(R)$ can be asymptotically achieved and Theorem \[thm:nofeedbackbound\] is proved. $\Box$
Comments
--------
For output-symmetric channels, $E^+(R) = E_{sp}(R)$ and so the usual sphere-packing bound is recovered in the fixed-delay context. Since $E_{sp}(R)$ is achieved universally with delay at high rates by using infinite-length random time-varying convolutional codes, this means that such codes achieve the best possible asymptotic tradeoff between probability of bit error and end-to-end delay. However, the proof in the previous section does not get to the sphere-packing bound for asymmetric channels like the Z-channel plotted in Figure \[fig:zchannelbounds\].
We could apply the sphere-packing bound to the $n$-length block-code by trying the $G$ channel that optimizes $E_{sp}(R - 2\delta_1 -
\gamma)$ for one of the block codeword compositions $\vec{r}$ that contains at least $\exp(n (R - \delta_1 - \gamma))$ codewords for some small $\gamma > 0$ that can be chosen after $\delta_1$. As a result, there would be weak bits whose probabilities of error are at least $\delta_2$ when used with the $G$ channel. The problem arises when we attempt to translate this back to the original channel $P$. Because the local $(d+1)$-length input-type is unknown in the vicinity of these weak bits, we would only be able to prove an exponent of $$\label{eqn:betterHaroutunianbound}
\widetilde{E}^+(R) = \inf_{G: \max_{\vec{r}: I(\vec{r},P) \geq R}
I(\vec{r},G) < R} \max_x D(G(\cdot|x)||P(\cdot|x)).$$ This is formally better than (\[eqn:Haroutunianbound\]) since there is slightly more flexibility in choosing the mimicking channel $G$. It now just has to have a mutual information across it lower than $R$ when driven with an input distribution that is good enough for the original channel. But, it seems unlikely that (\[eqn:betterHaroutunianbound\]) is tight the way that $E_{sp}$ is since for the Z-channel, it can evaluate to the same thing as (\[eqn:Haroutunianbound\]).
It is more interesting to reflect upon why this proof does not go through when feedback is available. This reveals why Pinsker’s assertion of Theorem 8 in [@PinskerNoFeedback] is incorrect. Although the lack of feedback was used in many places, the most critical point is Lemma \[lem:windowdependence\] which corresponds to [@PinskerNoFeedback Eqn. (39)]. When feedback is present, the current channel inputs can depend on the past channel [*outputs*]{}, even if we condition on the past channel [ *inputs*]{}. Thus it is not possible to take a block error and then focus attention on the channel behavior only during the delay period. It could be that the atypical channel behavior has to begin well before the bit in question even arrived at the encoder. This is seen clearly in the BEC case with feedback discussed in Section \[sec:becexample\] — the most common failure mode is for a bit to enter finding a large queue of senior bits already waiting and then finding that service continues to be so slow that the senior bits are not all able to leave the queue before the bit’s own deadline expires.
Upper-bounding the fixed-delay reliability function with feedback {#sec:feedbackbound}
=================================================================
To prove Theorem \[thm:generalfeedbackbound\] and get a proper upper bound to the fixed-delay reliability function when feedback is allowed, we need to account for the fact that the dominant error event might begin before the bit in question even arrives at the encoder. To do this, Viterbi’s argument from [@viterbi] is repurposed to address delay rather than constraint length. We call this upper bound the “uncertainty-focusing bound” because it is based on the idea of focusing the decoder’s uncertainty about the message bits given the channel outputs onto bits whose deadlines are not pending.
To bound what is possible, a fixed-delay code is translated into a fixed-block-length code. A lower bound on error probability for block codes is then pulled back to give a bound on the probability of error for the original fixed-delay code. The key difference from the previous section is that the block-length $n$ is not automatically made large compared to the delay. Rather, each different block length provides its own bound at all rates, with the final bound at any given rate and delay coming from optimizing over the block length.
[*Proof:* ]{} Given a code with fixed delay $d$, pick an arbitrary $0
< \lambda < 1$ and set the block length $n = \frac{d}{1-\lambda}$. As illustrated in Figure \[fig:lambdameaning\], this implies that $n =
\lambda n + d = \frac{\lambda}{1 - \lambda} d + d$. To avoid cumbersome notation, integer effects are ignored here. When $d$ is small, the fact that the block length must be an integer limits our choices for $\lambda$ in an insignificant way.
The block decoder operates by running the delay-$d$ decoder. This decodes the first $\lambda nR'$ bits, thus making the effective rate for the block code $\lambda R'$ bits per channel use or $\lambda R$ nats per channel use. The encoder just applies the given causal encoders with feedback using the actual message bits as the first $\lambda nR'$ bits. Random coin tosses can be used for the final $(1-\lambda) n R'$ inputs to the encoders since these will not be decoded anyway.
\#1\#2\#3\#4\#5[ @font ]{}
(3624,2547)(589,-2023) [(3001, 89)[( 0,-1)[300]{}]{} ]{}[(601,-961)[( 1, 0)[3600]{}]{} ]{}[(3001,-811)[( 0,-1)[300]{}]{} ]{}(3601,-361)[(0,0)\[b\]]{} (1801,-361)[(0,0)\[b\]]{} (1801,-1261)[(0,0)\[b\]]{} (3601,-1261)[(0,0)\[b\]]{} (1801, 14)[(0,0)\[b\]]{} (3601, 14)[(0,0)\[b\]]{} (1801,389)[(0,0)\[b\]]{} (3601,389)[(0,0)\[b\]]{} (1801,-886)[(0,0)\[b\]]{}
Let $B_1^{\lambda nR'}$ be the original message consisting entirely of independent fair coin tosses. The Haroutunian bound reveals that given any $\delta_1,\epsilon > 0$ there exists a sufficiently large block length $n_1$ and a constant $K$, so that as long as $n > n_1$, this fixed-block-length code with feedback must have a probability of block error that is lower bounded by [@Haroutunian] $$\label{eqn:haroutunianerror}
{\cal P}_P(B_{1}^{\lambda nR'} \neq \widehat{B}_{1}^{\lambda nR'}) \geq
K \exp\left(-n[E^+(\lambda R - \delta_1) + \epsilon] \right).$$
Substitute in $n = \frac{d}{1 - \lambda}$ and then notice that there must be at least one message bit position $i^*$ whose probability of bit error is no worse than $\frac{1}{n \lambda R'}$ times the probability of block error. This gives $${\cal P}_P(B_{i^*} \neq \widehat{B}_{i^*}) \geq
\frac{(1-\lambda)K}{\lambda R' d}
\exp\left(-d[\frac{E^+(\lambda R - \delta_1)}{1-\lambda} +
\frac{\epsilon}{1-\lambda}] \right).$$
Since the $\frac{1}{d}$ term in front is dominated by the exponential and $\delta_1,\epsilon$ are arbitrarily small and $\lambda$ was arbitrary, taking logs and the limit $d \rightarrow \infty$ proves (\[eqn:generalfeedbackbound\]).
Whenever $E^+(R) = E_{sp}(R)$, by using (\[eqn:rhospherepack\]) and following arguments identical to those used in the analysis of convolutional codes, (\[eqn:generalfeedbackbound\]) turns into (\[eqn:symmetricfeedbackbound\]). These arguments are given in Appendix \[app:focussymmetric\] for completeness.
Expanding (\[eqn:symmetricfeedbackbound\]) by Taylor expansion in the vicinity of $\eta=0$, noticing that the first derivative of $E_0$ there is the capacity $C$, and applying simple algebra leads to the negative slope of $2C/\frac{\partial^2 E_0(0)}{\partial \eta^2}$ in the vicinity of the $(C,0)$ point. When the second derivative term is equal to zero, then [@gallager] reveals that the channel’s sphere-packing bound hits $(C,0)$ at a positive slope of at least $-1$ and thus (\[eqn:symmetricfeedbackbound\]) evaluated at $\eta=1$ already has hit the capacity. There is no need to consider lower values of $\eta$. The uncertainty-focusing bound in such cases jumps discontinuously down to zero at rates above capacity. $\Box$
It is also important to notice that the core idea driving the proof is the inverse-concatenation construction from [@viterbi] and [@ForneyML]. This allows us to map an upper bound on the fixed-block-length reliability function into an upper bound on the fixed-delay reliability. As a result, the uncertainty-focusing bound can also be used for channels [*without*]{} feedback.
For a DMC, no fixed-delay exponent greater than the expurgated bound at rate 0 ($\alpha >
E_{ex}(0)$ from [@gallager]) is asymptotically achievable without feedback.
[*Proof:*]{} Because the straight-line bound [@gallager] can tighten the low-rate exponent for block-codes without feedback, this means that it can also be used to tighten the bound for fixed-delay codes in the low-rate regime. The inverse concatenation construction immediately turns the straight-line bound for fixed-block-length codes turns into a horizontal line at $E_{ex}(0)$ for fixed-delay codes. $\Box$
Thus the best upper bound we have for the reliability function for end-to-end delay in a system without feedback is $\min(E_{ex}(0),
E^+(R))$.
For the case of output-symmetric channels with feedback (or whenever the $E_{a,s}$ bound is tight), it is also possible to explicitly calculate the worst case $\lambda^*$ in parametric form using the arguments of Appendix \[app:focussymmetric\]: $$\label{eqn:optimumlambda}
\lambda^* =
\frac{(\frac{\partial E_0(\rho = \eta)}{\partial \rho})}{R(\eta)}
=
\frac{\eta}{E_0(\eta)}
(\frac{\partial E_0(\rho = \eta)}{\partial \rho}).$$ The exponentially dominating error event involves $\frac{\lambda^*}{1
- \lambda^*} d $ of the past channel outputs as well as the $d$ time steps in the future — for a error event length of $\frac{d}{1 -
\lambda^*}$. Thus $\lambda^*$ captures the critical balance between how badly the channel must misbehave and how long it must misbehave for. In general, when $R$ is near $C$, the $\eta$ will be near zero. Since $\frac{\partial E_0(\rho = 0)}{\partial \rho} = C$, this implies $\lambda^*$ there will be near $1$, and the dominant error events will be much longer than the desired end-to-end delay.
Achievability of the fixed-delay reliability with feedback for erasure channels {#sec:erasurefeedback}
===============================================================================
This section proves Theorem \[thm:erasurecase\] and thereby demonstrates the asymptotic achievability of $E_a(R) = E_{a,s}(R)$ everywhere for erasure channels with noiseless feedback.
The optimal code and its reliability
------------------------------------
The optimal scheme for the binary erasure channel with instantaneous[^6] causal noiseless feedback is intuitively obvious — buffer up message bits as they arrive and attempt to transmit the oldest message bit that has not yet been received correctly by the receiver. What is not immediately obvious is how well this scheme actually performs with end-to-end delay.
The Markov-chain analysis in Section \[sec:becexample\] becomes unwieldy at rates that are not simple rational numbers like $\frac{1}{2}$. In [@ACC00Paper; @SahaiThesis], an analysis of this scheme is given by translating the communication problem into a problem of stabilization of an unstable scalar plant over a noisy feedback link using techniques from [@ControlPartI]. The stabilization problem can then be studied explicitly in terms of its $\eta$-th moments, which can be understood using certain infinite sums. The dominant terms in these sums are found using heuristic arguments (rigorous only for $\eta=2,3$) and the convergence of those reveals which $\eta$-moments are finite. This in turn implicitly gives a lower bound to the reliability function with delay. It turns out that this calculation agrees with the uncertainty-focusing bound. In the following section, a direct and rigorous proof is given for Theorem \[thm:erasurecase\] at all rates.
A BEC with erasure probability $\beta$ is output-symmetric and so the Haroutunian bound and the sphere-packing bound are identical. Evaluating the symmetric uncertainty-focusing bound (\[eqn:symmetricfeedbackbound\]) gives the following parametric expression: (in units of bits and power of two reliability exponents since the computation is simpler in that base) $$\label{eqn:becreliability}
E_{a}^{bec}(R') = \eta-\log_2 (1+\beta(2^\eta-1)) ~~\,,\,~~ R' =
\frac{\eta-\log_2(1+\beta(2^\eta-1))}{\eta}$$ where $\eta$ ranges from $0$ to $\infty$.
Simple algebraic manipulation allows the parameter $\eta$ to be eliminated and this results in the rate-reliability tradeoff of (\[eqn:becanytimecapacity\]). The calculations for this and the simple low-rate bound are in Appendix \[app:erasurelowrate\].
Direct proof of achievability {#sec:becproof}
-----------------------------
In this section, the asymptotic achievability of the BEC’s fixed-delay reliability function (\[eqn:becreliability\]) is proven directly using a technique that parallels the bounding technique used for Theorem \[thm:generalfeedbackbound\].
The key idea is to use the first-in-first-out property of the “repeat until received” strategy, treating the system as a D/M/1 queue. The only way the $i$-th bit would not be received by the deadline is if there were too few successes. It is easy to see that this could happen if there were zero successes after it enters the system. But it could also happen if there were only one success since the previous bit entered the system, and so on. This is captured in the following:
\[lem:notthereyet\] The probability that bit $i$ is unable to meet deadline $\lceil
\frac{i}{R'} \rceil + d$ can be upper-bounded by: $$\label{eqn:sumbound}
{\cal P}(\widehat{B}_i(\lceil \frac{i}{R'} \rceil + d) \neq B_i)
\leq
\sum_{k=1}^i {\cal P}\left(\frac{1}{d + \lceil \frac{i}{R'} \rceil - \lceil
\frac{k}{R'} \rceil} \sum_{t=\lceil \frac{k}{R'} \rceil}^{\lceil \frac{i}{R'}
\rceil + d} Z_t \leq \frac{i-k}{d + \lceil \frac{i}{R'} \rceil - \lceil
\frac{k}{R'} \rceil}
\right)$$ where the $\{Z_t\}$ are the iid random variables that are $1$ if the $t$-th channel use is successful and $0$ if it is erased.
[*Proof:*]{} See Appendix \[app:lemnotthereyet\].
\#1\#2\#3\#4\#5[ @font ]{}
(5274,1935)(139,-1090) [(4951,539)[( 0,-1)[300]{}]{} ]{}[(4201,614)[( 0,-1)[450]{}]{} ]{}[(4201,-61)[( 1, 0)[1200]{}]{} ]{}[(1951,-511)(4.49775,0.00000)[668]{}[(1.6667,11.6667)[.]{}]{} ]{}[(1951,614)(0.00000,-4.50000)[101]{}[(1.6667,11.6667)[.]{}]{} ]{}(5176,164)[(0,0)\[b\]]{} (4576,164)[(0,0)\[b\]]{} (4801,-211)[(0,0)\[b\]]{} (4801,-361)[(0,0)\[b\]]{} (1951,689)[(0,0)\[b\]]{} (4951,689)[(0,0)\[b\]]{} (4201,689)[(0,0)\[b\]]{} (3451,-736)[(0,0)\[b\]]{} (1951,-211)[(0,0)\[rb\]]{} (1951,-361)[(0,0)\[rb\]]{} (3451,-886)[(0,0)\[b\]]{} (3451,-1036)[(0,0)\[b\]]{} (1951,-61)[(0,0)\[rb\]]{}
The next idea is to isolate the dominant term in the sum (\[eqn:sumbound\]) and to bound the whole sum explicitly in terms of this. The idea is depicted in Figure \[fig:erasurebounding\]. The potentially unbounded-length sum (since $i$ is arbitrary) is broken into two parts. One part has a finite number of terms and each term is upper-bounded by the dominant term. The other part has an unbounded number of terms but that sum is bounded using a convergent geometric series. This is done explicitly rather than relying on asymptotic large-deviations theorems so that the resulting constants are available to us to calculate plots for finite delays. The details are in Appendix \[app:erasuredetails\], but result in $$\begin{aligned}
& & {\cal P}(\widehat{B}_i(\lceil \frac{i}{R'} \rceil + d) \neq B_i)
\nonumber \\
& \leq &
\exp(-d \frac{D(\lambda^* R' || 1 - \beta)}{1-\lambda^*}
(\frac{D(R' + \frac{2}{\bar{n}} || 1 - \beta) - \epsilon_1}
{D(R' || 1 - \beta)}
))
\left[ \lceil \frac{1}{R'} \rceil
\sum_{l=0}^{\infty}
\exp(-l (D(R' + 2\bar{n}^{-1} || 1 - \beta) - \epsilon_1)) \right]
\nonumber \\
& &
+ d(\frac{D(\lambda^*R' || 1 - \beta)}
{(1-\lambda^*)D(R' || 1 - \beta)})
\exp(-d \bigg((1-\epsilon_2) \frac{D(\lambda^*R' || 1 - \beta)}
{1 - \lambda^*}
- \frac{\epsilon_1}{1-\lambda^*} \bigg)) \label{eqn:erasureintermediate}\end{aligned}$$ where $\lambda^*$ is coming from (\[eqn:optimumlambda\]), $\bar{n} =
d \frac{D(\lambda^*R' || 1 - \beta)}{(1-\lambda^*)D(R' || 1 -
\beta)}$, and $\epsilon_1, \epsilon_2$ are constants that can be made arbitrarily small as $d$ gets large. The term in the brackets $[
\cdots ]$ is a convergent geometric series while $(\frac{D(R' +
2\bar{n}^{-1} || 1 - \beta)}{D(R' || 1 - \beta)})$ approaches $1$ as $d$ and hence $\bar{n}$ gets large.
Since $\lambda^*$ does not depend on $d$, just notice that $E_a(R) =
\frac{D(\lambda^*R' || 1 - \beta)}{1-\lambda^*}$ for the binary erasure channel to get: $$\label{eqn:becachieve}
{\cal P}(\widehat{B}_i(\lceil \frac{i}{R'} \rceil + d) \neq B_i)
\leq
(\gamma + \xi d)\exp(-d (1-\epsilon_3) E_a(R))$$ for all $d > d_3(\epsilon_3, \beta, R)$ where $\epsilon_3 > 0$ is arbitrary and $\gamma, \xi > 0$ are constants depending on $\epsilon_3, \beta$ and $R$. Since the linear term is dominated by the exponential, it is clear that the bound of (\[eqn:symmetricfeedbackbound\]) is asymptotically achievable for the BEC with large delays at all rates $R$. For non-binary erasure channels, it is obvious that the same proof holds.
Furthermore, since the FIFO-based encoder does not need to know what the target delay is, the code is clearly delay-universal or anytime in nature. $\Box$
The transmission delay view
---------------------------
An alternative view of the communication problem over the binary erasure channel is useful when considering more general cases. Each bit’s delay can be viewed as the sum of a queuing delay (that can be correlated across different bits) and its transmission delay $T_j$ which is a geometric $(1-\beta)$ random variable that is iid over different bits $j$. The event $\{\frac{1}{d + \lceil \frac{i}{R'} \rceil
- \lceil \frac{k}{R'} \rceil} \sum_{t=\lceil \frac{k}{R'}
\rceil}^{\lceil \frac{i}{R'} \rceil + d} Z_t \leq \frac{i-k}{d +
\lceil \frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil}\}$ from (\[eqn:sumbound\]) can alternatively be expressed in this language as: $\{\sum_{j=k}^i T_j > d + \lceil \frac{i}{R'} \rceil - \lceil
\frac{k}{R'} \rceil\}$. This expresses the event that even if any backlog before $k$ is ignored, the unlucky transmission delays alone are too much for the bit $i$ to meet its deadline. (\[eqn:sumbound\]) then becomes $$\label{eqn:sumbound2}
{\cal P}(\widehat{B}_i(\lceil \frac{i}{R'} \rceil + d) \neq B_i)
\leq
\sum_{k=1}^i
{\cal P}(\sum_{j=k}^i T_j > d + \lceil \frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil).$$ With this interpretation, Theorem \[thm:erasurecase\] about the binary erasure channel implies the following result about large delays in certain D/G/1 queues:
\[cor:delaybasedbound\] Consider a communication system in which point messages arrive deterministically at a steady rate of $R'$ messages per unit time, are FIFO queued up until ready to be served, and are then independently served using geometric $(1-\beta)$ service times $T_j$. Given any $\epsilon_4 > 0$, there exists a $d_4(\epsilon_4, \beta, R)$ so that for all $d \geq d_4$, the probability that point message $i$ has not completed service by time $d+\lceil \frac{i}{R'} \rceil$ is upper bounded by $\exp(-d
(1-\epsilon_4) E_a^{bec}(R')\ln 2)$ where $E_a^{bec}$ from (\[eqn:becreliability\]) is the fixed-delay error exponent for the binary erasure channel with erasure probability $\beta$ and rate $R'$ in bits per channel use. This fixed delay exponent is attained universally over all sufficiently long delays $d \geq d_3$.
Furthermore, this result continues to hold even if the independent service times $T_j$ merely have complementary CDFs that are bounded by: ${\cal P}(T_j > k) \leq \beta^{k}$. The service times do not need to be identically distributed.
Finally, suppose the point message rate $R' = \frac{1}{m}$ where $m >
\widetilde{m}$ is a positive integer and the independent service times $T_j$ satisfy ${\cal P}(T_j > \widetilde{m} + k) \leq \beta^{k}$. Then the probability that point message $i$ (which arrived at time $im$) has not completed service by time $d+ i m$ is upper bounded by $\exp(-d (1-\epsilon_4) E_a^{bec}(R'') \ln 2)$ where $R'' = (m -
\widetilde{m})^{-1}$. This fixed delay exponent is also attained universally over all sufficiently long delays $d \geq d_4 + m$.
[*Proof:* ]{}In place of bits, there are messages. The geometric random variables can be interpreted as the interarrival times for the Bernoulli process of successful transmissions. The rate $R'$ bits per channel use turns into $R' \ln 2$ nats per channel use. Finally, the $\gamma + \xi d$ polynomial from (\[eqn:becachieve\]) can be absorbed into the exponential by just making $d_4$ and $\epsilon_4$ a little bigger than the original $d_3$ and $\epsilon_3$. This establishes the result for independent geometric service times.
For the case of general service times whose complementary CDF is bounded by the geometric’s complementary CDF, the reason is that the errors all come from large deviations events of the form $\sum_{j=k}^i
T_j \geq l$. For each $j$, start with an independent continuous uniform$[0,1]$ random variable $V_j$ and obtain both $T_j$ and $T'_j$ from $V_j$ through the inverse of their respective CDFs. This way, each of the $T_j$ can be paired with a geometric $T'_j$ random variable such that $\forall \omega, T_j(\omega) \leq T'_j(\omega)$ where $\omega$ represents an element from the sample space. Since $$\begin{aligned}
\{\omega | \sum_{j=k}^i T_j(\omega) \geq l \}
& = &
\bigcup_{\vec{l}| l_n \in {\cal N}, \sum_n l_n = l}
\bigcap_{n} \{\omega | T_n(\omega) \geq l_n \} \\
& \subseteq &
\bigcup_{\vec{l}| l_n \in {\cal N}, \sum_n l_n = l}
\bigcap_{n} \{\omega | T'_n(\omega) \geq l_n \} \\
& = &
\{\omega | \sum_{j=k}^i T'_j(\omega) \geq l \},\end{aligned}$$ it is clear that ${\cal P}(\sum_{j=k}^i T_j(\omega) \geq l) \leq
{\cal P}(\sum_{j=k}^i T'_j(\omega) \geq l)$ and so the same error probability bounds can be achieved.
Finally, consider the case of $R' =
\frac{1}{m}$ and independent service times bounded by those of a constant plus geometrics. (\[eqn:sumbound2\]) simplifies to $$\begin{aligned}
\label{eqn:sumbound3}
{\cal P}(\widehat{M}_i(im + d) \neq M_i)
& \leq &
\sum_{k=1}^i {\cal P}\left(\sum_{j=k}^i T_j > d + im -
km\right) \label{eqn:sumbound3} \\
& = &
\sum_{k=1}^i {\cal P}\left(\sum_{j=k}^i T_j > d + (i-k)m \right) \nonumber \\
& = &
\sum_{k=1}^i {\cal P}\left(\sum_{j=k}^i (T_j - \widetilde{m}) > d +
(i-k)(m-\widetilde{m}) \right) \nonumber \\
& = &
\sum_{k=1}^i {\cal P}\left(\sum_{j=k}^i (T_j - \widetilde{m}) > d +
i(m-\widetilde{m}) -k(m-\widetilde{m}) \right). \label{eqn:sumbound4}\end{aligned}$$ Notice that in (\[eqn:sumbound4\]), the random variable $\tilde{T}_j
= T_j - \widetilde{m}$ has a complementary CDF bounded by a geometric and corresponds to (\[eqn:sumbound2\]) with a point-message rate of $R'' = (m-\widetilde{m})^{-1}$. Thus, the error exponent with delay is at least as good as $E_a^{bec}(R'') \ln 2$ for the point messages. $\Box$
Achievability for general channels {#sec:fortifiedfeedback}
==================================
The goal of this section is to prove Theorems \[thm:withzeroerror\] and \[thm:genericachieve\]. Rather than starting with channels with strictly positive zero-error capacity, it is conceptually easier to start with generic DMCs but add a low-rate error-free side channel that can be used to carry “control” information. This information is interpreted as a kind of punctuation used to make the channel output stream unambiguously understandable to the decoder. The idea is that the rate of this error-free control channel is much lower than the message rate that needs to be communicated. This allows the result to extend immediately to channels with strictly positive zero-error capacity. For general channels, the control channel is synthesized and its own errors must be taken into account.
The scheme for fortified systems with noiseless feedback {#sec:nclscheme}
--------------------------------------------------------
A “fortified” model is an idealization (depicted in Figure \[fig:fortification\]) that makes an error-free control channel explicit:
\[def:fortify\] Given a DMC $P$ for the forward link, a $\frac{1}{k}$-[*fortified*]{} communication system built around it is one in which every $k$-th use of $P$ is supplemented with the ability to transmit a single error-free bit to the receiver.
\#1\#2\#3\#4\#5[ @font ]{}
(6774,1627)(-11,-1469) [( 76,-136)[( 0,-1)[150]{}]{} ]{}[(151,-136)[( 0,-1)[150]{}]{} ]{}[(226,-136)[( 0,-1)[150]{}]{} ]{}[(301,-136)[( 0,-1)[150]{}]{} ]{}[(376,-136)[( 0,-1)[150]{}]{} ]{}[(451,-136)[( 0,-1)[150]{}]{} ]{}[(526,-136)[( 0,-1)[150]{}]{} ]{}[(601,-136)[( 0,-1)[150]{}]{} ]{}[(676,-136)[( 0,-1)[150]{}]{} ]{}[(751,-136)[( 0,-1)[150]{}]{} ]{}[(826,-136)[( 0,-1)[150]{}]{} ]{}[(901,-136)[( 0,-1)[150]{}]{} ]{}[(976,-136)[( 0,-1)[150]{}]{} ]{}[(1051,-136)[( 0,-1)[150]{}]{} ]{}[(1126,-136)[( 0,-1)[150]{}]{} ]{}[(1201,-136)[( 0,-1)[150]{}]{} ]{}[(1276,-136)[( 0,-1)[150]{}]{} ]{}[(1351,-136)[( 0,-1)[150]{}]{} ]{}[(1426,-136)[( 0,-1)[150]{}]{} ]{}[(1501,-136)[( 0,-1)[150]{}]{} ]{}[(1576,-136)[( 0,-1)[150]{}]{} ]{}[(1651,-136)[( 0,-1)[150]{}]{} ]{}[(1726,-136)[( 0,-1)[150]{}]{} ]{}[(1801,-136)[( 0,-1)[150]{}]{} ]{}[(1876,-136)[( 0,-1)[150]{}]{} ]{}[(1951,-136)[( 0,-1)[150]{}]{} ]{}[(2026,-136)[( 0,-1)[150]{}]{} ]{}[(2101,-136)[( 0,-1)[150]{}]{} ]{}[(2176,-136)[( 0,-1)[150]{}]{} ]{}[(2251,-136)[( 0,-1)[150]{}]{} ]{}[(2326,-136)[( 0,-1)[150]{}]{} ]{}[(2401,-136)[( 0,-1)[150]{}]{} ]{}[(2476,-136)[( 0,-1)[150]{}]{} ]{}[(2551,-136)[( 0,-1)[150]{}]{} ]{}[(2626,-136)[( 0,-1)[150]{}]{} ]{}[(2701,-136)[( 0,-1)[150]{}]{} ]{}[(2776,-136)[( 0,-1)[150]{}]{} ]{}[(2851,-136)[( 0,-1)[150]{}]{} ]{}[(2926,-136)[( 0,-1)[150]{}]{} ]{}[(3001,-136)[( 0,-1)[150]{}]{} ]{}[(3076,-136)[( 0,-1)[150]{}]{} ]{}[(3151,-136)[( 0,-1)[150]{}]{} ]{}[(3226,-136)[( 0,-1)[150]{}]{} ]{}[(3301,-136)[( 0,-1)[150]{}]{} ]{}[(3376,-136)[( 0,-1)[150]{}]{} ]{}[(3451,-136)[( 0,-1)[150]{}]{} ]{}[(3526,-136)[( 0,-1)[150]{}]{} ]{}[(3601,-136)[( 0,-1)[150]{}]{} ]{}[(3676,-136)[( 0,-1)[150]{}]{} ]{}[(3751,-136)[( 0,-1)[150]{}]{} ]{}[(3826,-136)[( 0,-1)[150]{}]{} ]{}[(3901,-136)[( 0,-1)[150]{}]{} ]{}[(3976,-136)[( 0,-1)[150]{}]{} ]{}[(4051,-136)[( 0,-1)[150]{}]{} ]{}[(4126,-136)[( 0,-1)[150]{}]{} ]{}[(4201,-136)[( 0,-1)[150]{}]{} ]{}[(4276,-136)[( 0,-1)[150]{}]{} ]{}[(4351,-136)[( 0,-1)[150]{}]{} ]{}[(4426,-136)[( 0,-1)[150]{}]{} ]{}[(4501,-136)[( 0,-1)[150]{}]{} ]{}[(4576,-136)[( 0,-1)[150]{}]{} ]{}[(4651,-136)[( 0,-1)[150]{}]{} ]{}[(4726,-136)[( 0,-1)[150]{}]{} ]{}[(4801,-136)[( 0,-1)[150]{}]{} ]{}[(4876,-136)[( 0,-1)[150]{}]{} ]{}[(4951,-136)[( 0,-1)[150]{}]{} ]{}[(5026,-136)[( 0,-1)[150]{}]{} ]{}[(5101,-136)[( 0,-1)[150]{}]{} ]{}[(5176,-136)[( 0,-1)[150]{}]{} ]{}[(5251,-136)[( 0,-1)[150]{}]{} ]{}[(5326,-136)[( 0,-1)[150]{}]{} ]{}[(5401,-136)[( 0,-1)[150]{}]{} ]{}[(5476,-136)[( 0,-1)[150]{}]{} ]{}[(5551,-136)[( 0,-1)[150]{}]{} ]{}[(5626,-136)[( 0,-1)[150]{}]{} ]{}[(5701,-136)[( 0,-1)[150]{}]{} ]{}[(5776,-136)[( 0,-1)[150]{}]{} ]{}[(5851,-136)[( 0,-1)[150]{}]{} ]{}[(5926,-136)[( 0,-1)[150]{}]{} ]{}[(6001,-136)[( 0,-1)[150]{}]{} ]{}[(6076,-136)[( 0,-1)[150]{}]{} ]{}[(6151,-136)[( 0,-1)[150]{}]{} ]{}[(6226,-136)[( 0,-1)[150]{}]{} ]{}[(6301,-136)[( 0,-1)[150]{}]{} ]{}[(6376,-136)[( 0,-1)[150]{}]{} ]{}[(6451,-136)[( 0,-1)[150]{}]{} ]{}[(6526,-136)[( 0,-1)[150]{}]{} ]{}[( 1,-1036)[( 1, 0)[6750]{}]{} ]{}[(1051,-886)[( 0,-1)[300]{}]{} ]{}[(2101,-886)[( 0,-1)[300]{}]{} ]{}[(3151,-886)[( 0,-1)[300]{}]{} ]{}[(4201,-886)[( 0,-1)[300]{}]{} ]{}[(5251,-886)[( 0,-1)[300]{}]{} ]{}[(6301,-886)[( 0,-1)[300]{}]{} ]{}[( 1,-211)[( 1, 0)[6750]{}]{} ]{}( 1, 14)[(0,0)\[lb\]]{} (1051,-1411)[(0,0)\[b\]]{} (2101,-1411)[(0,0)\[b\]]{} (3151,-1411)[(0,0)\[b\]]{} (4201,-1411)[(0,0)\[b\]]{} (5251,-1411)[(0,0)\[b\]]{} (6301,-1411)[(0,0)\[b\]]{} ( 1,-811)[(0,0)\[lb\]]{}
In comparison to the encoders with feedback from Definition \[def:nofeedbackencoder\], fortified encoders get to send an additional error-free bit $S_{\frac{t}{k}}$ at times $t$ that are integer multiples of $k$. The decoders are naturally modified to get causal access to the error-free bits as well.
The idea is to generalize the repeat-until-received strategy used for the erasure channel in Theorem \[thm:erasurecase\]. A family of schemes indexed by three parameters $(n,c,l)$ is described first, and the asymptotic achievability of $E_{a,s}(R)$ is shown by taking an appropriate limit over such schemes.
\#1\#2\#3\#4\#5[ @font ]{}
(7704,2310)(-41,-2161) [(113,-136)[( 0,-1)[150]{}]{} ]{}[(188,-136)[( 0,-1)[150]{}]{} ]{}[(263,-136)[( 0,-1)[150]{}]{} ]{}[(338,-136)[( 0,-1)[150]{}]{} ]{}[(413,-136)[( 0,-1)[150]{}]{} ]{}[(488,-136)[( 0,-1)[150]{}]{} ]{}[(563,-136)[( 0,-1)[150]{}]{} ]{}[(638,-136)[( 0,-1)[150]{}]{} ]{}[(713,-136)[( 0,-1)[150]{}]{} ]{}[(788,-136)[( 0,-1)[150]{}]{} ]{}[(863,-136)[( 0,-1)[150]{}]{} ]{}[(938,-136)[( 0,-1)[150]{}]{} ]{}[(1013,-136)[( 0,-1)[150]{}]{} ]{}[(1088,-136)[( 0,-1)[150]{}]{} ]{}[(1163,-136)[( 0,-1)[150]{}]{} ]{}[(1238,-136)[( 0,-1)[150]{}]{} ]{}[(1313,-136)[( 0,-1)[150]{}]{} ]{}[(1388,-136)[( 0,-1)[150]{}]{} ]{}[(1463,-136)[( 0,-1)[150]{}]{} ]{}[(1538,-136)[( 0,-1)[150]{}]{} ]{}[(1613,-136)[( 0,-1)[150]{}]{} ]{}[(1688,-136)[( 0,-1)[150]{}]{} ]{}[(1763,-136)[( 0,-1)[150]{}]{} ]{}[(1838,-136)[( 0,-1)[150]{}]{} ]{}[(1913,-136)[( 0,-1)[150]{}]{} ]{}[(1988,-136)[( 0,-1)[150]{}]{} ]{}[(2063,-136)[( 0,-1)[150]{}]{} ]{}[(2138,-136)[( 0,-1)[150]{}]{} ]{}[(2213,-136)[( 0,-1)[150]{}]{} ]{}[(2288,-136)[( 0,-1)[150]{}]{} ]{}[(2363,-136)[( 0,-1)[150]{}]{} ]{}[(2438,-136)[( 0,-1)[150]{}]{} ]{}[(2513,-136)[( 0,-1)[150]{}]{} ]{}[(2588,-136)[( 0,-1)[150]{}]{} ]{}[(2663,-136)[( 0,-1)[150]{}]{} ]{}[(2738,-136)[( 0,-1)[150]{}]{} ]{}[(2813,-136)[( 0,-1)[150]{}]{} ]{}[(2888,-136)[( 0,-1)[150]{}]{} ]{}[(2963,-136)[( 0,-1)[150]{}]{} ]{}[(3038,-136)[( 0,-1)[150]{}]{} ]{}[(3113,-136)[( 0,-1)[150]{}]{} ]{}[(3188,-136)[( 0,-1)[150]{}]{} ]{}[(3263,-136)[( 0,-1)[150]{}]{} ]{}[(3338,-136)[( 0,-1)[150]{}]{} ]{}[(3413,-136)[( 0,-1)[150]{}]{} ]{}[(3488,-136)[( 0,-1)[150]{}]{} ]{}[(3563,-136)[( 0,-1)[150]{}]{} ]{}[(3638,-136)[( 0,-1)[150]{}]{} ]{}[(3713,-136)[( 0,-1)[150]{}]{} ]{}[(3788,-136)[( 0,-1)[150]{}]{} ]{}[(3863,-136)[( 0,-1)[150]{}]{} ]{}[(3938,-136)[( 0,-1)[150]{}]{} ]{}[(4013,-136)[( 0,-1)[150]{}]{} ]{}[(4088,-136)[( 0,-1)[150]{}]{} ]{}[(4163,-136)[( 0,-1)[150]{}]{} ]{}[(4238,-136)[( 0,-1)[150]{}]{} ]{}[(4313,-136)[( 0,-1)[150]{}]{} ]{}[(4388,-136)[( 0,-1)[150]{}]{} ]{}[(4463,-136)[( 0,-1)[150]{}]{} ]{}[(4538,-136)[( 0,-1)[150]{}]{} ]{}[(4613,-136)[( 0,-1)[150]{}]{} ]{}[(4688,-136)[( 0,-1)[150]{}]{} ]{}[(4763,-136)[( 0,-1)[150]{}]{} ]{}[(4838,-136)[( 0,-1)[150]{}]{} ]{}[(4913,-136)[( 0,-1)[150]{}]{} ]{}[(4988,-136)[( 0,-1)[150]{}]{} ]{}[(5063,-136)[( 0,-1)[150]{}]{} ]{}[(5138,-136)[( 0,-1)[150]{}]{} ]{}[(5213,-136)[( 0,-1)[150]{}]{} ]{}[(5288,-136)[( 0,-1)[150]{}]{} ]{}[(5363,-136)[( 0,-1)[150]{}]{} ]{}[(5438,-136)[( 0,-1)[150]{}]{} ]{}[(5513,-136)[( 0,-1)[150]{}]{} ]{}[(5588,-136)[( 0,-1)[150]{}]{} ]{}[(5663,-136)[( 0,-1)[150]{}]{} ]{}[(5738,-136)[( 0,-1)[150]{}]{} ]{}[(5813,-136)[( 0,-1)[150]{}]{} ]{}[(5888,-136)[( 0,-1)[150]{}]{} ]{}[(5963,-136)[( 0,-1)[150]{}]{} ]{}[(6038,-136)[( 0,-1)[150]{}]{} ]{}[(6113,-136)[( 0,-1)[150]{}]{} ]{}[(6188,-136)[( 0,-1)[150]{}]{} ]{}[(6263,-136)[( 0,-1)[150]{}]{} ]{}[(4201,-586)[( 1, 0)[2100]{}]{} ]{}[(4201,-511)[( 0,-1)[150]{}]{} ]{}[(6301,-511)[( 0,-1)[150]{}]{} ]{}(5251,-736)[(0,0)\[b\]]{} [(526,-1786)[( 0, 1)[525]{}]{} ]{}[(526,-1486)[( 1, 0)[525]{}]{} (1051,-1486)[( 0, 1)[225]{}]{} ]{}(526,-1936)[(0,0)\[lb\]]{} [( 1,-136)[( 0,-1)[150]{}]{} ]{}[( 76,-136)[( 0,-1)[150]{}]{} ]{}[(151,-136)[( 0,-1)[150]{}]{} ]{}[(226,-136)[( 0,-1)[150]{}]{} ]{}[(301,-136)[( 0,-1)[150]{}]{} ]{}[(376,-136)[( 0,-1)[150]{}]{} ]{}[(451,-136)[( 0,-1)[150]{}]{} ]{}[(526,-136)[( 0,-1)[150]{}]{} ]{}[(601,-136)[( 0,-1)[150]{}]{} ]{}[(676,-136)[( 0,-1)[150]{}]{} ]{}[(751,-136)[( 0,-1)[150]{}]{} ]{}[(826,-136)[( 0,-1)[150]{}]{} ]{}[(901,-136)[( 0,-1)[150]{}]{} ]{}[(976,-136)[( 0,-1)[150]{}]{} ]{}[(1051,-136)[( 0,-1)[150]{}]{} ]{}[(1126,-136)[( 0,-1)[150]{}]{} ]{}[(1201,-136)[( 0,-1)[150]{}]{} ]{}[(1276,-136)[( 0,-1)[150]{}]{} ]{}[(1351,-136)[( 0,-1)[150]{}]{} ]{}[(1426,-136)[( 0,-1)[150]{}]{} ]{}[(1501,-136)[( 0,-1)[150]{}]{} ]{}[(1576,-136)[( 0,-1)[150]{}]{} ]{}[(1651,-136)[( 0,-1)[150]{}]{} ]{}[(1726,-136)[( 0,-1)[150]{}]{} ]{}[(1801,-136)[( 0,-1)[150]{}]{} ]{}[(1876,-136)[( 0,-1)[150]{}]{} ]{}[(1951,-136)[( 0,-1)[150]{}]{} ]{}[(2026,-136)[( 0,-1)[150]{}]{} ]{}[(2101,-136)[( 0,-1)[150]{}]{} ]{}[(2176,-136)[( 0,-1)[150]{}]{} ]{}[(2251,-136)[( 0,-1)[150]{}]{} ]{}[(2326,-136)[( 0,-1)[150]{}]{} ]{}[(2401,-136)[( 0,-1)[150]{}]{} ]{}[(2476,-136)[( 0,-1)[150]{}]{} ]{}[(2551,-136)[( 0,-1)[150]{}]{} ]{}[(2626,-136)[( 0,-1)[150]{}]{} ]{}[(2701,-136)[( 0,-1)[150]{}]{} ]{}[(2776,-136)[( 0,-1)[150]{}]{} ]{}[(2851,-136)[( 0,-1)[150]{}]{} ]{}[(2926,-136)[( 0,-1)[150]{}]{} ]{}[(3001,-136)[( 0,-1)[150]{}]{} ]{}[(3076,-136)[( 0,-1)[150]{}]{} ]{}[(3151,-136)[( 0,-1)[150]{}]{} ]{}[(3226,-136)[( 0,-1)[150]{}]{} ]{}[(3301,-136)[( 0,-1)[150]{}]{} ]{}[(3376,-136)[( 0,-1)[150]{}]{} ]{}[(3451,-136)[( 0,-1)[150]{}]{} ]{}[(3526,-136)[( 0,-1)[150]{}]{} ]{}[(3601,-136)[( 0,-1)[150]{}]{} ]{}[(3676,-136)[( 0,-1)[150]{}]{} ]{}[(3751,-136)[( 0,-1)[150]{}]{} ]{}[(3826,-136)[( 0,-1)[150]{}]{} ]{}[(3901,-136)[( 0,-1)[150]{}]{} ]{}[(3976,-136)[( 0,-1)[150]{}]{} ]{}[(4051,-136)[( 0,-1)[150]{}]{} ]{}[(4126,-136)[( 0,-1)[150]{}]{} ]{}[(4201,-136)[( 0,-1)[150]{}]{} ]{}[(4276,-136)[( 0,-1)[150]{}]{} ]{}[(4351,-136)[( 0,-1)[150]{}]{} ]{}[(4426,-136)[( 0,-1)[150]{}]{} ]{}[(4501,-136)[( 0,-1)[150]{}]{} ]{}[(4576,-136)[( 0,-1)[150]{}]{} ]{}[(4651,-136)[( 0,-1)[150]{}]{} ]{}[(4726,-136)[( 0,-1)[150]{}]{} ]{}[(4801,-136)[( 0,-1)[150]{}]{} ]{}[(4876,-136)[( 0,-1)[150]{}]{} ]{}[(4951,-136)[( 0,-1)[150]{}]{} ]{}[(5026,-136)[( 0,-1)[150]{}]{} ]{}[(5101,-136)[( 0,-1)[150]{}]{} ]{}[(5176,-136)[( 0,-1)[150]{}]{} ]{}[(5251,-136)[( 0,-1)[150]{}]{} ]{}[(5326,-136)[( 0,-1)[150]{}]{} ]{}[(5401,-136)[( 0,-1)[150]{}]{} ]{}[(5476,-136)[( 0,-1)[150]{}]{} ]{}[(5551,-136)[( 0,-1)[150]{}]{} ]{}[(5626,-136)[( 0,-1)[150]{}]{} ]{}[(5701,-136)[( 0,-1)[150]{}]{} ]{}[(5776,-136)[( 0,-1)[150]{}]{} ]{}[(5851,-136)[( 0,-1)[150]{}]{} ]{}[(5926,-136)[( 0,-1)[150]{}]{} ]{}[(6001,-136)[( 0,-1)[150]{}]{} ]{}[(6076,-136)[( 0,-1)[150]{}]{} ]{}[(6151,-136)[( 0,-1)[150]{}]{} ]{}[(6226,-136)[( 0,-1)[150]{}]{} ]{}[( 1,-211)[( 1, 0)[7650]{}]{} ]{}[( 1,-1036)[( 1, 0)[7650]{}]{} ]{}[(2101,-886)[( 0,-1)[300]{}]{} ]{}[(3151,-886)[( 0,-1)[300]{}]{} ]{}[(4201,-886)[( 0,-1)[300]{}]{} ]{}[(5251,-886)[( 0,-1)[300]{}]{} ]{}[(6301,-886)[( 0,-1)[300]{}]{} ]{}[(1576,-886)[( 0,-1)[300]{}]{} ]{}[(2626,-886)[( 0,-1)[300]{}]{} ]{}[(3676,-886)[( 0,-1)[300]{}]{} ]{}[(4726,-886)[( 0,-1)[300]{}]{} ]{}[(5776,-886)[( 0,-1)[300]{}]{} ]{}[( 1,-2011)[( 0, 1)[750]{}]{} ]{}[(3676,-1786)[( 0, 1)[525]{}]{} ]{}[(3676,-1486)[(-1, 0)[2100]{}]{} (1576,-1486)[( 0, 1)[225]{}]{} ]{}[(2626,-1486)[( 0, 1)[225]{}]{} ]{}[(3151,-1486)[( 0, 1)[225]{}]{} ]{}[(3676,-1486)[( 1, 0)[2100]{}]{} (5776,-1486)[( 0, 1)[225]{}]{} ]{}[(5251,-1486)[( 0, 1)[225]{}]{} ]{}[(4726,-1486)[( 0, 1)[225]{}]{} ]{}[(6826,-886)[( 0,-1)[300]{}]{} ]{}[(7351,-886)[( 0,-1)[300]{}]{} ]{}[(7351,-1711)[( 0, 1)[525]{}]{} ]{}[(7351,-1486)[(-1, 0)[525]{}]{} (6826,-1486)[( 0, 1)[225]{}]{} ]{}[(6301,-136)[( 0,-1)[150]{}]{} ]{}[(7463,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7426,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7388,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7351,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7313,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7276,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7238,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7201,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7163,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7126,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7088,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7051,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7013,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6976,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6938,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6901,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6863,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6826,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6788,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6751,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6713,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6676,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6638,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6601,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6563,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6526,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6488,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6451,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6413,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6376,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6338,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(1051,-886)(0.00000,-4.47761)[68]{}[(1.6667,11.6667)[.]{}]{} ]{}[(526,-886)(0.00000,-4.47761)[68]{}[(1.6667,11.6667)[.]{}]{} ]{}[( 1,-886)(0.00000,-4.47761)[68]{}[(1.6667,11.6667)[.]{}]{} ]{}( 1, 14)[(0,0)\[lb\]]{} ( 1,-811)[(0,0)\[lb\]]{} ( 1,-2161)[(0,0)\[lb\]]{} (3676,-1936)[(0,0)\[lb\]]{} (2101,-1336)[(0,0)\[b\]]{} (4201,-1336)[(0,0)\[b\]]{} (6301,-1336)[(0,0)\[b\]]{} (7351,-1936)[(0,0)\[rb\]]{}
Call $c \geq 1$ the chunk length in terms of how many control bits are associated with each chunk, $2^l$ the list length (with $l \leq c
-1$), and $n > l$ the message block length in units of chunks. The $(n,
c, l)$ randomized communication scheme (illustrated in Figure \[fig:nclcode\]) is:
1. The encoder queues up incoming message bits and assembles them into message blocks of size $\frac{n c k R}{\ln 2}$ bits. One such message block arrives deterministically every $nck$ channel uses.
2. At every noisy channel use, the encoder sends the channel input corresponding to the next position in an infinite-length random codeword associated with the current message block.
Formally, the codewords are $X_{i}(j,t)$ where $i > 0$ represents the current block number, $t > 0$ is the current channel-use time, and $0 \leq j < \exp(n c k R)$ is the value of the current message block. Each $X_{i}(j,t)$ is drawn iid from ${\cal X}$ using the $E_0(\eta)$ maximizing distribution $\vec{q}$. An $\eta$ is chosen such that the desired rate $R < \frac{E_0(\eta,\vec{q})}{\eta}$, while the target reliability is also $\alpha < E_0(\eta,\vec{q})$.
If there is no message block to send, the encoder just idles by transmitting the next letter in the past message block.
3. If the time is an integer multiple of $ck$, the encoder uses the noiselessly fedback channel outputs to simulate the decoder’s attempt to ML-decode the current codeword to within a list of size $2^l$.
If the true codeword is one of the $2^l$ entries on the list, the encoder sends a $1$ ([*confirm*]{}) over the noiseless forward link. The encoder places into a control queue $l$ bits representing the true codeword’s index within the decoder’s list. The encoder then removes the current message block from the message queue.
If the true block is not in the decoder’s list, the encoder sends a $0$ ([*deny*]{}) over the error-free forward link.
The $0$ can be viewed as a null punctuation mark while the $1$ corresponds to a comma delimiting one variable-length block from another. When the list-disambiguation information is sent, it can be interpreted as a specific type of comma. There are thus just $l+1$ different kinds of punctuation in the system.
4. If the time is an integer multiple of $k$ but not an integer multiple of $ck$, then the encoder looks in the control queue and transmits one of these bits over the error-free link, removing it from this second queue. If there are no control bits waiting, then the error-free link is ignored.
Since $c > l$, all $l$ of the control bits will be communicated within one chunk.
5. At the decoder, the encoder’s message queue length is known perfectly since it can only change by the deterministic arrival of message blocks or when an error-free confirm or deny bit has been sent over the noise-free link. Thus the decoder can correctly parse the received channel uses and always knows which message block a given channel output $Y_t$ or fortification symbol $S_t$ corresponds to.
6. If the time is an integer multiple of $ck$ and the decoder receives a $1$ noiselessly, then it decodes what it has seen to a list of the top $2^l$ possibilities for this message block. It uses the next $l$ error-free bits to disambiguate this list and commits to the result as its estimate for the message block.
Analysis of end-to-end delay and probability of error
-----------------------------------------------------
It is clear that this hybrid-ARQ scheme does not commit any errors at the decoder. Some blocks just take longer to make it across than others do. Furthermore, notice that the delay experienced by any message [*bit*]{} can be divided into four parts:
1. [*Assembly delay*]{}: How long it takes before the rest of the message block has arrived at the encoder. This is bounded by a constant $nck$ channel uses.
2. [*Queuing delay*]{}: How long the message block must wait before it begins to be transmitted.
3. [*Transmission delay*]{}: How many channel uses it takes before the codeword can be correctly decoded to within a list of $2^l$. This is a random quantity $T_j$ that must be an integer multiple of $ck$ channel uses. The $T_j$ are iid since the channel is memoryless and the random codebooks are also iid.
4. [*Termination delay*]{}: How long the decoder must wait before the block is disambiguated by the error-free control signals. This is bounded by a constant $lk$ channel uses.
Since the assembly and termination delays are constants that do not depend on the target end-to-end delay, they can be ignored and the focus kept on the queuing and transmission delays. This is because our interest is in the fixed-delay behavior for asymptotically large delays much longer than $nck$. Since the transmission delays are iid, the approach is to apply Corollary \[cor:delaybasedbound\] and this requires a bound in terms of a constant plus a geometric.
\[lem:transmissiontimes\] The $(n,c,l)$ transmission scheme using input-distribution $\vec{q}$ at rate $R$ for a $\frac{1}{k}$-fortified communication system over a DMC has iid transmission times $T_j$ satisfying $$\label{eqn:bettertransmissionbound}
{\cal P}(T_j - \lceil \widetilde{t}(\rho,R,n,\vec{q}) \rceil c k > t ck) \leq
[\exp(-ck E_0(\rho,\vec{q}) )
]^{t}$$ for all $0 \leq \rho \leq 2^l$ and positive integer $t$ where $\widetilde{t}(\rho,R,n,\vec{q}) = \frac{R}{\widetilde{C}(\rho,\vec{q})}n$ and $\widetilde{C}(\rho,\vec{q}) = \frac{E_0(\rho,\vec{q})}{\rho}$.
[*Proof:* ]{}See Appendix \[app:lemtransmissiondelay\].
\#1\#2\#3\#4\#5[ @font ]{}
(4824,2319)(1189,-1972) [(1201,-61)[( 0,-1)[600]{}]{} ]{}[(1801,-61)[( 0,-1)[600]{}]{} ]{}[(2401,-61)[( 0,-1)[600]{}]{} ]{}[(3001,-61)[( 0,-1)[600]{}]{} ]{}[(3601,-61)[( 0,-1)[600]{}]{} ]{}[(4201,-61)[( 0,-1)[600]{}]{} ]{}[(4801,-61)[( 0,-1)[600]{}]{} ]{}[(5401,-61)[( 0,-1)[600]{}]{} ]{}[(6001,-61)[( 0,-1)[600]{}]{} ]{}[(3601,-1111)[( 0, 1)[375]{}]{} ]{}[(4801,-1561)[( 0, 1)[825]{}]{} ]{}[(4801, 89)[(-1, 0)[1200]{}]{} ]{}[(4801, 89)[( 1, 0)[1200]{}]{} ]{}[(5401,-961)[(-1, 0)[600]{}]{} ]{}[(5401,-961)[( 1, 0)[600]{}]{} ]{}[(1501,-61)[( 0,-1)[600]{}]{} ]{}[(2101,-61)[( 0,-1)[600]{}]{} ]{}[(2701,-61)[( 0,-1)[600]{}]{} ]{}[(3301,-61)[( 0,-1)[600]{}]{} ]{}[(3901,-61)[( 0,-1)[600]{}]{} ]{}[(4501,-61)[( 0,-1)[600]{}]{} ]{}[(5101,-61)[( 0,-1)[600]{}]{} ]{}[(5701,-61)[( 0,-1)[600]{}]{} ]{}[(1201,-661)(18.00000,0.00000)[201]{}[(1.6667,11.6667)[.]{}]{} ]{}[(1201,-511)(18.04511,0.00000)[134]{}[(1.6667,11.6667)[.]{}]{} ]{}[(1201,-211)(18.04511,0.00000)[134]{}[(1.6667,11.6667)[.]{}]{} ]{}[(1201,-61)(18.00000,0.00000)[201]{}[(1.6667,11.6667)[.]{}]{} ]{}(4801,164)[(0,0)\[b\]]{} (3601,-1411)[(0,0)\[rb\]]{} (5401,-1186)[(0,0)\[b\]]{} (1351,-1261)[(0,0)\[b\]]{} (3601,-1261)[(0,0)\[rb\]]{} (4801,-1711)[(0,0)\[rb\]]{} (1201,239)[(0,0)\[lb\]]{} (5401,-1411)[(0,0)\[b\]]{} (5401,-1561)[(0,0)\[b\]]{} (4801,-1936)[(0,0)\[rb\]]{}
![Why at least a $0.44$ fixed-delay exponent is achievable at rate $0.37$ nats per channel use. The thick curve is the sphere-packing bound and the thin curve on top is the uncertainty-focusing bound. The thick tangent represents using a list size of $1$. The gap between $0.44$ and $0.37$ on the rate axis depicts the fraction of “slack” channel uses that are available. The thin tangent is the one used in the inverse-concatenation construction for the bound at rate $0.37$ nats per channel use.[]{data-label="fig:slackillustrated"}](slackillustrated.edited.eps){width="4in" height="3in"}
Lemma \[lem:transmissiontimes\] is illustrated in Figure \[fig:blockslack\] and then the application of Corollary \[cor:delaybasedbound\] is illustrated in Figures \[fig:slackillustrated\] and \[fig:chunkslackerror\]. These illustrate the achievability of the fixed-delay exponent $0.44$ at a rate of $0.37$ nats. The gap between $0.44$ and $0.37$ on the rate axis in Figure \[fig:slackillustrated\] depicts the fraction of “slack” channel uses that are available to communicate a message block with reliability $0.44$ while still draining the queue faster than it is being filled. The block length $n$ must be long enough so that the slack represents at least a few channel uses. As the block length $n$ becomes longer, it is possible to move up to the reliability limit illustrated by the inverse concatenation construction.
\#1\#2\#3\#4\#5[ @font ]{}
(9204,7470)(289,-7111) [(1201,-2611)[( 0,-1)[300]{}]{} ]{}[(1351,-2611)[( 0,-1)[300]{}]{} ]{}[(1501,-2611)[( 0,-1)[300]{}]{} ]{}[(1651,-2611)[( 0,-1)[300]{}]{} ]{}[(1801,-2611)[( 0,-1)[300]{}]{} ]{}[(1951,-2611)[( 0,-1)[300]{}]{} ]{}[(2251,-2611)[( 0,-1)[300]{}]{} ]{}[(2401,-2611)[( 0,-1)[300]{}]{} ]{}[(2551,-2611)[( 0,-1)[300]{}]{} ]{}[(2701,-2611)[( 0,-1)[300]{}]{} ]{}[(2851,-2611)[( 0,-1)[300]{}]{} ]{}[(3001,-2611)[( 0,-1)[300]{}]{} ]{}[(3151,-2611)[( 0,-1)[300]{}]{} ]{}[(3451,-2611)[( 0,-1)[300]{}]{} ]{}[(3601,-2611)[( 0,-1)[300]{}]{} ]{}[(3751,-2611)[( 0,-1)[300]{}]{} ]{}[(3901,-2611)[( 0,-1)[300]{}]{} ]{}[(4051,-2611)[( 0,-1)[300]{}]{} ]{}[(4201,-2611)[( 0,-1)[300]{}]{} ]{}[(4351,-2611)[( 0,-1)[300]{}]{} ]{}[(4501,-2611)[( 0,-1)[300]{}]{} ]{}[(4651,-2611)[( 0,-1)[300]{}]{} ]{}[(4801,-2611)[( 0,-1)[300]{}]{} ]{}[(4951,-2611)[( 0,-1)[300]{}]{} ]{}[(5101,-2611)[( 0,-1)[300]{}]{} ]{}[(5251,-2611)[( 0,-1)[300]{}]{} ]{}[(5401,-2611)[( 0,-1)[300]{}]{} ]{}[(5551,-2611)[( 0,-1)[300]{}]{} ]{}[(5851,-2611)[( 0,-1)[300]{}]{} ]{}[(6001,-2611)[( 0,-1)[300]{}]{} ]{}[(6151,-2611)[( 0,-1)[300]{}]{} ]{}[(6301,-2611)[( 0,-1)[300]{}]{} ]{}[(6451,-2611)[( 0,-1)[300]{}]{} ]{}[(6601,-2611)[( 0,-1)[300]{}]{} ]{}[(6751,-2611)[( 0,-1)[300]{}]{} ]{}[(6901,-2611)[( 0,-1)[300]{}]{} ]{}[(7051,-2611)[( 0,-1)[300]{}]{} ]{}[(7201,-2611)[( 0,-1)[300]{}]{} ]{}[(7351,-2611)[( 0,-1)[300]{}]{} ]{}[(7501,-2611)[( 0,-1)[300]{}]{} ]{}[(7651,-2611)[( 0,-1)[300]{}]{} ]{}[(7801,-2611)[( 0,-1)[300]{}]{} ]{}[(7951,-2611)[( 0,-1)[300]{}]{} ]{}[(2101,-2611)[( 0,-1)[300]{}]{} ]{}[(5701,-2311)[( 0,-1)[450]{}]{} ]{}[(8101,-2311)[( 0,-1)[300]{}]{} ]{}[(8401,-2611)[( 0,-1)[300]{}]{} ]{}[(8701,-2611)[( 0,-1)[300]{}]{} ]{}[(8851,-2611)[( 0,-1)[300]{}]{} ]{}[(9001,-2611)[( 0,-1)[300]{}]{} ]{}[(9301,-2611)[( 0,-1)[300]{}]{} ]{}[(8551,-2611)[( 0,-1)[300]{}]{} ]{}[(9151,-2611)[( 0,-1)[300]{}]{} ]{}[(9451,-2761)[(-1, 0)[9150]{}]{} ]{}[(751,-2611)[( 0,-1)[300]{}]{} ]{}[(601,-2611)[( 0,-1)[300]{}]{} ]{}[(1201,-661)[( 0,-1)[300]{}]{} ]{}[(1351,-661)[( 0,-1)[300]{}]{} ]{}[(1501,-661)[( 0,-1)[300]{}]{} ]{}[(1651,-661)[( 0,-1)[300]{}]{} ]{}[(1801,-661)[( 0,-1)[300]{}]{} ]{}[(1951,-661)[( 0,-1)[300]{}]{} ]{}[(2251,-661)[( 0,-1)[300]{}]{} ]{}[(2401,-661)[( 0,-1)[300]{}]{} ]{}[(2551,-661)[( 0,-1)[300]{}]{} ]{}[(2701,-661)[( 0,-1)[300]{}]{} ]{}[(2851,-661)[( 0,-1)[300]{}]{} ]{}[(3001,-661)[( 0,-1)[300]{}]{} ]{}[(3151,-661)[( 0,-1)[300]{}]{} ]{}[(3451,-661)[( 0,-1)[300]{}]{} ]{}[(3601,-661)[( 0,-1)[300]{}]{} ]{}[(3751,-661)[( 0,-1)[300]{}]{} ]{}[(3901,-661)[( 0,-1)[300]{}]{} ]{}[(4051,-661)[( 0,-1)[300]{}]{} ]{}[(4201,-661)[( 0,-1)[300]{}]{} ]{}[(4351,-661)[( 0,-1)[300]{}]{} ]{}[(4501,-661)[( 0,-1)[300]{}]{} ]{}[(4651,-661)[( 0,-1)[300]{}]{} ]{}[(4801,-661)[( 0,-1)[300]{}]{} ]{}[(4951,-661)[( 0,-1)[300]{}]{} ]{}[(5101,-661)[( 0,-1)[300]{}]{} ]{}[(5251,-661)[( 0,-1)[300]{}]{} ]{}[(5401,-661)[( 0,-1)[300]{}]{} ]{}[(5551,-661)[( 0,-1)[300]{}]{} ]{}[(5851,-661)[( 0,-1)[300]{}]{} ]{}[(6001,-661)[( 0,-1)[300]{}]{} ]{}[(6151,-661)[( 0,-1)[300]{}]{} ]{}[(6301,-661)[( 0,-1)[300]{}]{} ]{}[(6451,-661)[( 0,-1)[300]{}]{} ]{}[(6601,-661)[( 0,-1)[300]{}]{} ]{}[(6751,-661)[( 0,-1)[300]{}]{} ]{}[(6901,-661)[( 0,-1)[300]{}]{} ]{}[(7051,-661)[( 0,-1)[300]{}]{} ]{}[(7201,-661)[( 0,-1)[300]{}]{} ]{}[(7351,-661)[( 0,-1)[300]{}]{} ]{}[(7501,-661)[( 0,-1)[300]{}]{} ]{}[(7651,-661)[( 0,-1)[300]{}]{} ]{}[(7801,-661)[( 0,-1)[300]{}]{} ]{}[(7951,-661)[( 0,-1)[300]{}]{} ]{}[(2101,-661)[( 0,-1)[300]{}]{} ]{}[(5701,-361)[( 0,-1)[450]{}]{} ]{}[(8101,-361)[( 0,-1)[450]{}]{} ]{}[(8401,-661)[( 0,-1)[300]{}]{} ]{}[(8701,-661)[( 0,-1)[300]{}]{} ]{}[(8851,-661)[( 0,-1)[300]{}]{} ]{}[(9001,-661)[( 0,-1)[300]{}]{} ]{}[(9301,-661)[( 0,-1)[300]{}]{} ]{}[(8551,-661)[( 0,-1)[300]{}]{} ]{}[(9151,-661)[( 0,-1)[300]{}]{} ]{}[(9301,-811)[(-1, 0)[9000]{}]{} ]{}[(751,-661)[( 0,-1)[300]{}]{} ]{}[(601,-661)[( 0,-1)[300]{}]{} ]{}[(826,-361)[(1.6667,11.6667)[.]{}]{} ]{}[(3301,-61)[( 0,-1)[250]{}]{} ]{}[(3601,-61)[(-1, 0)[2700]{}]{} (901,-61)[( 0,-1)[250]{}]{} ]{}[(5701,-61)[( 1, 0)[2400]{}]{} (8101,-61)[( 0,-1)[250]{}]{} ]{}[(5701,-1561)[( 1, 0)[3600]{}]{} ]{}[(901,-661)[( 0,-1)[300]{}]{} ]{}[(3301,-661)[( 0,-1)[300]{}]{} ]{}[(5701,-661)[( 0,-1)[300]{}]{} ]{}[(8101,-661)[( 0,-1)[300]{}]{} ]{}[(3301,-2611)[( 0,-1)[300]{}]{} ]{}[(5701,-2611)[( 0,-1)[300]{}]{} ]{}[(8101,-2611)[( 0,-1)[300]{}]{} ]{}[(5251,-3811)[( 0, 1)[300]{}]{} (5251,-3511)[(-1, 0)[4650]{}]{} (601,-3511)[( 0, 1)[375]{}]{} ]{}[(6151,-2761)[( 0,-1)[450]{}]{} ]{}[(9451,-2761)[( 0,-1)[450]{}]{} ]{}[(6151,-3511)[( 0, 1)[250]{}]{} ]{}[(3601,-3511)[( 0, 1)[250]{}]{} ]{}[(3601,164)[( 0,-1)[225]{}]{} (3601,-61)[( 1, 0)[2100]{}]{} (5701,-61)[( 0,-1)[250]{}]{} ]{}[(901,-4111)(0.00000,9.00000)[126]{}[(1.6667,11.6667)[.]{}]{} ]{}[(5701,-1861)[( 1, 0)[1775]{}]{} ]{}[(7501,-1861)[( 1, 0)[1800]{}]{} ]{}[(7501,-5086)[( 1, 0)[1800]{}]{} ]{}[(6901,-5386)[( 0,-1)[450]{}]{} ]{}[(6301,-5386)[( 0,-1)[450]{}]{} ]{}[(6901,-5686)[( 0,-1)[300]{}]{} ]{}[(6751,-5686)[( 0,-1)[300]{}]{} ]{}[(6601,-5686)[( 0,-1)[300]{}]{} ]{}[(6451,-5686)[( 0,-1)[300]{}]{} ]{}[(6301,-5686)[( 0,-1)[300]{}]{} ]{}[(6151,-5686)[( 0,-1)[300]{}]{} ]{}[(8101,-5686)[( 0,-1)[300]{}]{} ]{}[(7951,-5686)[( 0,-1)[300]{}]{} ]{}[(7801,-5686)[( 0,-1)[300]{}]{} ]{}[(7651,-5686)[( 0,-1)[300]{}]{} ]{}[(7501,-5686)[( 0,-1)[300]{}]{} ]{}[(7351,-5686)[( 0,-1)[300]{}]{} ]{}[(7201,-5686)[( 0,-1)[300]{}]{} ]{}[(9151,-5686)[( 0,-1)[300]{}]{} ]{}[(9001,-5686)[( 0,-1)[300]{}]{} ]{}[(8851,-5686)[( 0,-1)[300]{}]{} ]{}[(8701,-5686)[( 0,-1)[300]{}]{} ]{}[(8551,-5686)[( 0,-1)[300]{}]{} ]{}[(8401,-5686)[( 0,-1)[300]{}]{} ]{}[(8251,-5686)[( 0,-1)[300]{}]{} ]{}[(6001,-5686)[( 0,-1)[300]{}]{} ]{}[(5851,-5686)[( 0,-1)[300]{}]{} ]{}[(5701,-5686)[( 0,-1)[300]{}]{} ]{}[(5551,-5686)[( 0,-1)[300]{}]{} ]{}[(5401,-5686)[( 0,-1)[300]{}]{} ]{}[(9451,-5836)[(-1, 0)[4350]{}]{} ]{}[(6001,-5836)[( 0,-1)[300]{}]{} ]{}[(7201,-5836)[( 0,-1)[450]{}]{} ]{}[(8101,-5386)[( 0,-1)[300]{}]{} ]{}[(8701,-5386)[( 0,-1)[300]{}]{} ]{}[(7951,-5836)[( 0,-1)[450]{}]{} ]{}[(3601,-3061)[( 1, 0)[2525]{}]{} ]{}[(6151,-3061)[( 1, 0)[3275]{}]{} ]{}[(6301,-6136)[( 1, 0)[875]{}]{} ]{}[(7201,-6136)[( 1, 0)[725]{}]{} ]{}[(7951,-6136)[( 1, 0)[1475]{}]{} ]{}[(2701,-2911)(-18.00000,0.00000)[101]{}[(1.6667,11.6667)[.]{}]{} ]{}[(901,-2611)(18.00000,0.00000)[101]{}[(1.6667,11.6667)[.]{}]{} ]{}[(3301,-2611)(18.00000,0.00000)[101]{}[(1.6667,11.6667)[.]{}]{} ]{}[(3601,-2911)(18.00000,0.00000)[101]{}[(1.6667,11.6667)[.]{}]{} ]{}[(5701,-2611)(18.00000,0.00000)[101]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6151,-2911)(18.00000,0.00000)[101]{}[(1.6667,11.6667)[.]{}]{} ]{}(7501,-1411)[(0,0)\[b\]]{} (5701,-2236)[(0,0)\[b\]]{} (8101,-2236)[(0,0)\[b\]]{} (3301,-2236)[(0,0)\[b\]]{} (5701,-1186)[(0,0)\[b\]]{} (8101,-1186)[(0,0)\[b\]]{} (3301,-1186)[(0,0)\[b\]]{} (901,-1186)[(0,0)\[b\]]{} (3601,-3661)[(0,0)\[b\]]{} (901,-2236)[(0,0)\[b\]]{} (9451,-3661)[(0,0)\[b\]]{} (6151,-3661)[(0,0)\[b\]]{} (601,-3661)[(0,0)\[b\]]{} (3601,239)[(0,0)\[b\]]{} (8401,-2041)[(0,0)\[b\]]{} (6601,-1786)[(0,0)\[b\]]{} (901,-4261)[(0,0)\[lb\]]{} (8401,-5011)[(0,0)\[b\]]{} (7501,-5311)[(0,0)\[b\]]{} (6901,-5311)[(0,0)\[b\]]{} (6301,-5311)[(0,0)\[b\]]{} (5701,-5311)[(0,0)\[b\]]{} (8101,-5311)[(0,0)\[b\]]{} (8701,-5311)[(0,0)\[b\]]{} (3151,-5536)[(0,0)\[rb\]]{} (3151,-5761)[(0,0)\[rb\]]{} (5251,-3961)[(0,0)\[b\]]{} (2251,-3211)[(0,0)\[b\]]{} (4876,-3211)[(0,0)\[b\]]{} (7801,-3211)[(0,0)\[b\]]{} (6751,-6361)[(0,0)\[b\]]{} (7576,-6361)[(0,0)\[b\]]{} (8701,-6361)[(0,0)\[b\]]{} (7201,-6511)[(0,0)\[b\]]{} (7951,-6511)[(0,0)\[b\]]{} (6301,-7111)[(0,0)\[lb\]]{} (6001,-6286)[(0,0)\[b\]]{} (9451,-6511)[(0,0)\[b\]]{}
Consider time in $ck$ units. Let $R' = \frac{1}{n}$ be the rate at which message blocks are generated in terms of blocks generated per $ck$ channel uses. $R'' = \frac{1}{n - \lceil
\widetilde{t}(\rho,R,n,\vec{q}) \rceil}$ is the rate at which we evaluate the BEC’s fixed-delay reliability in the application of Corollary \[cor:delaybasedbound\]. The effective “erasure probability” is $\beta = \exp(-ck E_0(\rho,\vec{q}))$.
Recall that $R < \frac{E_0(\rho)}{\rho}$ where the $\vec{q}$ distribution is chosen as the $E_0(\rho)$ achieving distribution. The quantity $n - \lceil \widetilde{t}(\rho,R,n,\vec{q}) \rceil$ has a special significance since it captures the amount of slack in the system when viewed with parameter $\rho$. This slack term is positive for large enough $n >
\frac{\widetilde{C}(\rho,\vec{q})}{\widetilde{C}(\rho,\vec{q}) - R}$ since
$$\begin{aligned}
n - \lceil \widetilde{t}(\rho,R,n,\vec{q}) \rceil
& \geq &
\left(\frac{\widetilde{C}(\rho,\vec{q}) -
R}{\widetilde{C}(\rho,\vec{q})} \right) n - 1. \label{eqn:slackbound}\end{aligned}$$
Thus $$\begin{aligned}
R'' & = & \frac{1}{n - \lceil \widetilde{t}(\rho,R,n,\vec{q}) \rceil} \\
& \leq & (\left(\frac{\widetilde{C}(\rho,\vec{q}) - R}{\widetilde{C}(\rho,\vec{q})} \right) n - 1 )^{-1}.\end{aligned}$$ Notice that $R''$ can be made as small as desired by choosing $n$ large while $\beta$ can be made extremely small by choosing $c$ large. Applying Theorem \[thm:erasurecase\] tells us to set $$\begin{aligned}
1 + 2r & = &
\frac{\widetilde{C}(\rho,\vec{q}) - R}{\widetilde{C}(\rho,\vec{q})} n
- 1 \nonumber \\
n(\rho, c, k, l, r)
& = & \frac{\widetilde{C}(\rho,\vec{q})}{\widetilde{C}(\rho,\vec{q}) -
R}(2 + 2r)
\label{eqn:ndefinition}\end{aligned}$$ in order to get to within $(2 \ln 2)\exp(-rck E_0(\rho, \vec{q}))$ of the exponent $ck E_0(\rho, \vec{q})$ in terms of delays measured in $ck$ time units, or to within $\frac{2 \ln 2}{ck}\exp(-rck E_0(\rho,
\vec{q}))$ of the exponent of $E_0(\rho,\vec{q})$ in terms of delays measured in channel uses.
Putting it all together, for any small $\Delta > 0$, and $\rho \geq 0$ such that $R < \frac{E_0(\rho)}{\rho}$ a delay-exponent of $E_0(\rho)
- (3 \ln 2) \Delta$ is clearly achievable by setting $l = \max(0, \lceil
\log_2 \rho \rceil)$, choosing chunk size $$\label{eqn:cselection}
c = \max(l+1, \lceil \frac{\ln 16}{k E_0(\rho)} \rceil)$$ and then choosing $r$ big enough using $$\label{eqn:rselection}
r \geq \max(0,\frac{\ln(\Delta c k)}{ck E_0(\rho)}).$$ With $c$ and $r$ defined, $n$ can be obtained from (\[eqn:ndefinition\]).
Notice that $k$ is arbitrary here and can thus be made as large as desired. This corresponds to the fact that the amount of “punctuation” information can be made as small as desired, assuming that the target end-to-end delay is large enough.[^7]
Each $(n,c,l)$ code is also delay universal since it is not designed with a maximum $d$ in mind. The longer the decoder is willing to wait, the lower the probability of error becomes. This property is inherited from the repeat-until-success code for the erasure channel through Corollary \[cor:delaybasedbound\]. $\Box$
Channels with strictly positive zero-error capacity {#sec:strictzero}
---------------------------------------------------
The above communication scheme is easily adapted to channels with strictly positive zero-error capacity by just using a zero-error code to carry the punctuation information. There is no $k$. Instead, let $\theta$ be the block length required to realize feedback zero-error transmission of at least $l+1$ bits. As illustrated in Figure \[fig:zeroerrorcode\], terminate each chunk with a block-length-$\theta$ feedback zero-error code and use it to transmit the punctuation information. If the chunk size is $c$ channel uses, then it is as though we are operating with only a fraction $(1 -
\frac{\theta}{c})$ of the channel uses. This effectively increases the rate to $R/(1 - \frac{\theta}{c})$ and reduces the achieved delay exponent to $\alpha(1 - \frac{\theta}{c})$ as well. This overhead becomes negligible by making the chunk size $c$ large giving us the desired result.
\#1\#2\#3\#4\#5[ @font ]{}
(7674,2284)(-11,-1760) [(1651,-286)[( 1, 0)[450]{}]{} ]{}[( 1,-211)[( 1, 0)[7650]{}]{} ]{}[(6301,-136)[( 0,-1)[150]{}]{} ]{}[(7463,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7426,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7388,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7351,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7313,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7276,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7238,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7201,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7163,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7126,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7088,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7051,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(7013,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6976,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6938,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6901,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6863,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6826,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6788,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6751,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6713,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6676,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6638,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6601,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6563,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6526,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6488,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6451,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6413,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6376,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6338,-136)(0.00000,-4.54545)[34]{}[(1.6667,11.6667)[.]{}]{} ]{}[(451,-136)[( 0,-1)[150]{}]{} ]{}[(526,-136)[( 0,-1)[150]{}]{} ]{}[(301,-136)[( 0,-1)[150]{}]{} ]{}[(376,-136)[( 0,-1)[150]{}]{} ]{}[( 1,-136)[( 0,-1)[150]{}]{} ]{}[( 76,-136)[( 0,-1)[150]{}]{} ]{}[(151,-136)[( 0,-1)[150]{}]{} ]{}[(226,-136)[( 0,-1)[150]{}]{} ]{}[(188,-136)[( 0,-1)[150]{}]{} ]{}[(263,-136)[( 0,-1)[150]{}]{} ]{}[( 38,-136)[( 0,-1)[150]{}]{} ]{}[(113,-136)[( 0,-1)[150]{}]{} ]{}[(338,-136)[( 0,-1)[150]{}]{} ]{}[(413,-136)[( 0,-1)[150]{}]{} ]{}[(488,-136)[( 0,-1)[150]{}]{} ]{}[(563,-136)[( 0,-1)[150]{}]{} ]{}[(601,-136)[( 0,-1)[150]{}]{} ]{}[(676,-136)[( 0,-1)[150]{}]{} ]{}[(638,-136)[( 0,-1)[150]{}]{} ]{}[(713,-136)[( 0,-1)[150]{}]{} ]{}[(751,-136)[( 0,-1)[150]{}]{} ]{}[(826,-136)[( 0,-1)[150]{}]{} ]{}[(788,-136)[( 0,-1)[150]{}]{} ]{}[(863,-136)[( 0,-1)[150]{}]{} ]{}[(901,-136)[( 0,-1)[150]{}]{} ]{}[(976,-136)[( 0,-1)[150]{}]{} ]{}[(938,-136)[( 0,-1)[150]{}]{} ]{}[(1013,-136)[( 0,-1)[150]{}]{} ]{}[(1088,-136)[( 0,-1)[150]{}]{} ]{}[(1163,-136)[( 0,-1)[150]{}]{} ]{}[(1051,-136)[( 0,-1)[150]{}]{} ]{}[(1126,-136)[( 0,-1)[150]{}]{} ]{}[(1238,-136)[( 0,-1)[150]{}]{} ]{}[(1313,-136)[( 0,-1)[150]{}]{} ]{}[(1388,-136)[( 0,-1)[150]{}]{} ]{}[(1463,-136)[( 0,-1)[150]{}]{} ]{}[(1201,-136)[( 0,-1)[150]{}]{} ]{}[(1276,-136)[( 0,-1)[150]{}]{} ]{}[(1351,-136)[( 0,-1)[150]{}]{} ]{}[(1426,-136)[( 0,-1)[150]{}]{} ]{}[(1501,-136)[( 0,-1)[150]{}]{} ]{}[(1576,-136)[( 0,-1)[150]{}]{} ]{}[(1651,-136)[( 0,-1)[150]{}]{} ]{}[(1726,-136)[( 0,-1)[150]{}]{} ]{}[(1688,-136)[( 0,-1)[150]{}]{} ]{}[(1763,-136)[( 0,-1)[150]{}]{} ]{}[(1538,-136)[( 0,-1)[150]{}]{} ]{}[(1613,-136)[( 0,-1)[150]{}]{} ]{}[(1838,-136)[( 0,-1)[150]{}]{} ]{}[(1913,-136)[( 0,-1)[150]{}]{} ]{}[(1988,-136)[( 0,-1)[150]{}]{} ]{}[(2063,-136)[( 0,-1)[150]{}]{} ]{}[(1801,-136)[( 0,-1)[150]{}]{} ]{}[(1876,-136)[( 0,-1)[150]{}]{} ]{}[(1951,-136)[( 0,-1)[150]{}]{} ]{}[(2026,-136)[( 0,-1)[150]{}]{} ]{}[(2101,-136)[( 0,-1)[150]{}]{} ]{}[(2176,-136)[( 0,-1)[150]{}]{} ]{}[(2251,-136)[( 0,-1)[150]{}]{} ]{}[(2326,-136)[( 0,-1)[150]{}]{} ]{}[(2138,-136)[( 0,-1)[150]{}]{} ]{}[(2213,-136)[( 0,-1)[150]{}]{} ]{}[(2288,-136)[( 0,-1)[150]{}]{} ]{}[(2363,-136)[( 0,-1)[150]{}]{} ]{}[(2438,-136)[( 0,-1)[150]{}]{} ]{}[(2513,-136)[( 0,-1)[150]{}]{} ]{}[(2588,-136)[( 0,-1)[150]{}]{} ]{}[(2663,-136)[( 0,-1)[150]{}]{} ]{}[(2738,-136)[( 0,-1)[150]{}]{} ]{}[(2813,-136)[( 0,-1)[150]{}]{} ]{}[(2888,-136)[( 0,-1)[150]{}]{} ]{}[(2963,-136)[( 0,-1)[150]{}]{} ]{}[(2401,-136)[( 0,-1)[150]{}]{} ]{}[(2476,-136)[( 0,-1)[150]{}]{} ]{}[(2551,-136)[( 0,-1)[150]{}]{} ]{}[(2626,-136)[( 0,-1)[150]{}]{} ]{}[(2701,-136)[( 0,-1)[150]{}]{} ]{}[(2776,-136)[( 0,-1)[150]{}]{} ]{}[(2851,-136)[( 0,-1)[150]{}]{} ]{}[(2926,-136)[( 0,-1)[150]{}]{} ]{}[(3151,-136)[( 0,-1)[150]{}]{} ]{}[(3226,-136)[( 0,-1)[150]{}]{} ]{}[(3038,-136)[( 0,-1)[150]{}]{} ]{}[(3113,-136)[( 0,-1)[150]{}]{} ]{}[(3188,-136)[( 0,-1)[150]{}]{} ]{}[(3263,-136)[( 0,-1)[150]{}]{} ]{}[(3301,-136)[( 0,-1)[150]{}]{} ]{}[(3376,-136)[( 0,-1)[150]{}]{} ]{}[(3001,-136)[( 0,-1)[150]{}]{} ]{}[(3076,-136)[( 0,-1)[150]{}]{} ]{}[(3338,-136)[( 0,-1)[150]{}]{} ]{}[(3413,-136)[( 0,-1)[150]{}]{} ]{}[(3451,-136)[( 0,-1)[150]{}]{} ]{}[(3526,-136)[( 0,-1)[150]{}]{} ]{}[(3488,-136)[( 0,-1)[150]{}]{} ]{}[(3563,-136)[( 0,-1)[150]{}]{} ]{}[(3751,-136)[( 0,-1)[150]{}]{} ]{}[(3826,-136)[( 0,-1)[150]{}]{} ]{}[(3601,-136)[( 0,-1)[150]{}]{} ]{}[(3676,-136)[( 0,-1)[150]{}]{} ]{}[(4088,-136)[( 0,-1)[150]{}]{} ]{}[(4163,-136)[( 0,-1)[150]{}]{} ]{}[(4238,-136)[( 0,-1)[150]{}]{} ]{}[(4313,-136)[( 0,-1)[150]{}]{} ]{}[(4201,-136)[( 0,-1)[150]{}]{} ]{}[(4276,-136)[( 0,-1)[150]{}]{} ]{}[(3938,-136)[( 0,-1)[150]{}]{} ]{}[(4013,-136)[( 0,-1)[150]{}]{} ]{}[(3638,-136)[( 0,-1)[150]{}]{} ]{}[(3713,-136)[( 0,-1)[150]{}]{} ]{}[(3788,-136)[( 0,-1)[150]{}]{} ]{}[(3863,-136)[( 0,-1)[150]{}]{} ]{}[(3901,-136)[( 0,-1)[150]{}]{} ]{}[(3976,-136)[( 0,-1)[150]{}]{} ]{}[(4051,-136)[( 0,-1)[150]{}]{} ]{}[(4126,-136)[( 0,-1)[150]{}]{} ]{}[(4351,-136)[( 0,-1)[150]{}]{} ]{}[(4426,-136)[( 0,-1)[150]{}]{} ]{}[(4388,-136)[( 0,-1)[150]{}]{} ]{}[(4463,-136)[( 0,-1)[150]{}]{} ]{}[(4501,-136)[( 0,-1)[150]{}]{} ]{}[(4576,-136)[( 0,-1)[150]{}]{} ]{}[(4651,-136)[( 0,-1)[150]{}]{} ]{}[(4726,-136)[( 0,-1)[150]{}]{} ]{}[(4538,-136)[( 0,-1)[150]{}]{} ]{}[(4613,-136)[( 0,-1)[150]{}]{} ]{}[(4688,-136)[( 0,-1)[150]{}]{} ]{}[(4763,-136)[( 0,-1)[150]{}]{} ]{}[(5138,-136)[( 0,-1)[150]{}]{} ]{}[(5213,-136)[( 0,-1)[150]{}]{} ]{}[(4988,-136)[( 0,-1)[150]{}]{} ]{}[(5063,-136)[( 0,-1)[150]{}]{} ]{}[(4838,-136)[( 0,-1)[150]{}]{} ]{}[(4913,-136)[( 0,-1)[150]{}]{} ]{}[(4951,-136)[( 0,-1)[150]{}]{} ]{}[(5026,-136)[( 0,-1)[150]{}]{} ]{}[(4801,-136)[( 0,-1)[150]{}]{} ]{}[(4876,-136)[( 0,-1)[150]{}]{} ]{}[(5101,-136)[( 0,-1)[150]{}]{} ]{}[(5176,-136)[( 0,-1)[150]{}]{} ]{}[(5251,-136)[( 0,-1)[150]{}]{} ]{}[(5326,-136)[( 0,-1)[150]{}]{} ]{}[(5288,-136)[( 0,-1)[150]{}]{} ]{}[(5363,-136)[( 0,-1)[150]{}]{} ]{}[(5438,-136)[( 0,-1)[150]{}]{} ]{}[(5513,-136)[( 0,-1)[150]{}]{} ]{}[(5551,-136)[( 0,-1)[150]{}]{} ]{}[(5626,-136)[( 0,-1)[150]{}]{} ]{}[(5588,-136)[( 0,-1)[150]{}]{} ]{}[(5663,-136)[( 0,-1)[150]{}]{} ]{}[(5401,-136)[( 0,-1)[150]{}]{} ]{}[(5476,-136)[( 0,-1)[150]{}]{} ]{}[(5701,-136)[( 0,-1)[150]{}]{} ]{}[(5776,-136)[( 0,-1)[150]{}]{} ]{}[(6151,-136)[( 0,-1)[150]{}]{} ]{}[(6226,-136)[( 0,-1)[150]{}]{} ]{}[(6001,-136)[( 0,-1)[150]{}]{} ]{}[(6076,-136)[( 0,-1)[150]{}]{} ]{}[(5851,-136)[( 0,-1)[150]{}]{} ]{}[(5926,-136)[( 0,-1)[150]{}]{} ]{}[(5738,-136)[( 0,-1)[150]{}]{} ]{}[(5813,-136)[( 0,-1)[150]{}]{} ]{}[(6188,-136)[( 0,-1)[150]{}]{} ]{}[(6263,-136)[( 0,-1)[150]{}]{} ]{}[(6038,-136)[( 0,-1)[150]{}]{} ]{}[(6113,-136)[( 0,-1)[150]{}]{} ]{}[(5888,-136)[( 0,-1)[150]{}]{} ]{}[(5963,-136)[( 0,-1)[150]{}]{} ]{}[(4201,-136)[(-1, 0)[450]{}]{} ]{}[(4201,-286)[(-1, 0)[450]{}]{} ]{}[(5851,-136)[( 1, 0)[450]{}]{} ]{}[(5851,-286)[( 1, 0)[450]{}]{} ]{}[(1651,-136)[( 3,-1)[450]{}]{} ]{}[(1651,-286)[( 3, 1)[450]{}]{} ]{}[(3751,-286)[( 3, 1)[450]{}]{} ]{}[(3751,-136)[( 3,-1)[450]{}]{} ]{}[(5851,-136)[( 3,-1)[450]{}]{} ]{}[(5851,-286)[( 3, 1)[450]{}]{} ]{}[(1651,-286)(0.00000,-4.50000)[151]{}[(1.6667,11.6667)[.]{}]{} ]{}[(2101,-286)(0.00000,-4.50000)[151]{}[(1.6667,11.6667)[.]{}]{} ]{}[(3751,-286)(0.00000,-4.50000)[151]{}[(1.6667,11.6667)[.]{}]{} ]{}[(4201,-286)(0.00000,-4.50000)[151]{}[(1.6667,11.6667)[.]{}]{} ]{}[(5851,-286)(0.00000,-4.50000)[151]{}[(1.6667,11.6667)[.]{}]{} ]{}[(6301,-286)(0.00000,-4.50000)[151]{}[(1.6667,11.6667)[.]{}]{} ]{}[(3976,-1561)[( 0, 1)[675]{}]{} ]{}[(3976,-1486)[( 1, 0)[2100]{}]{} (6076,-1486)[( 0, 1)[225]{}]{} ]{}[(3976,-1486)[(-1, 0)[2100]{}]{} (1876,-1486)[( 0, 1)[600]{}]{} ]{}[( 1,-511)[( 0,-1)[150]{}]{} ]{}[(2101,-511)[( 0,-1)[150]{}]{} ]{}[(2101,-586)[( 1, 0)[2100]{}]{} ]{}[(2101,-511)[( 0,-1)[150]{}]{} ]{}[(4201,-511)[( 0,-1)[150]{}]{} ]{}[(4201,-586)[( 1, 0)[2100]{}]{} ]{}[(4201,-511)[( 0,-1)[150]{}]{} ]{}[(6301,-511)[( 0,-1)[150]{}]{} ]{}[( 1,-586)[( 1, 0)[2100]{}]{} ]{}(6076,-1186)[(0,0)\[b\]]{} (3976,-1711)[(0,0)\[b\]]{} ( 1,389)[(0,0)\[lb\]]{} (1876, 14)[(0,0)\[b\]]{} (3976, 14)[(0,0)\[b\]]{} (6076, 14)[(0,0)\[b\]]{} (1876,-811)[(0,0)\[b\]]{} (3976,-811)[(0,0)\[b\]]{} (6076,-961)[(0,0)\[b\]]{} (751, 14)[(0,0)\[b\]]{} (2851, 14)[(0,0)\[b\]]{} (4951, 14)[(0,0)\[b\]]{} (1051,-511)[(0,0)\[b\]]{} (3151,-511)[(0,0)\[b\]]{} (5251,-511)[(0,0)\[b\]]{} (751,-811)[(0,0)\[b\]]{} (2851,-811)[(0,0)\[b\]]{} (4951,-811)[(0,0)\[b\]]{}
Delayed feedback
----------------
Let $\phi$ be the delay in the noiseless feedback. So the encoders now know only $Y_1^{t-\phi}$ in addition to the message bits. Everything continues to work because the chunks $c$ can be made much longer than $\phi$. The last $\phi - 1$ channel uses in a chunk can then be discarded without any significant overhead.
Thus, Theorem \[thm:withzeroerror\] holds for any communication system in the asymptotic limit of large end-to-end delays even if there are small round-trip delays in the feedback. All that is required is some way to provide infrequent, but unmistakable, punctuation information from the encoder to the decoder. $\Box$
Channels without zero-error capacity: paying for punctuation {#sec:comments}
------------------------------------------------------------
All that remains is to prove Theorem \[thm:genericachieve\]. When the channel has no zero-error capacity, then it is still possible to follow the Section-\[sec:strictzero\] approach of allocating $\theta$ channel uses per chunk to carry punctuation information. The channel uses are partitioned as before into two streams assigned to two sub-encoders. The first is exactly as it was in the Section \[sec:strictzero\] and carries the message itself using a variable-length channel code with the dynamic length chosen to ensure correct list-decoding. This first encoder generates punctuation messages at the end of every chunk and these are the input to the second encoder. The second encoder’s role is to convey this punctuation information consisting of $l+1$ bits for every chunk.
Instead of using a zero-error code, the second encoder is implemented using an infinite-constraint-length time-varying random convolutional code. The trick of Appendix \[app:feedbackconvolution\] can be used to reduce the expected computational burden for encoding/decoding by using feedback, but essentially this sub-code operates without feedback.
The decoder also runs with two subsystems. One subsystem is responsible for decoding the punctuation stream. This can be implemented using either an ML decoder or a sequential decoder from [@JelinekSequential]. Either way, it is responsible for giving its current best estimate for all punctuation so far. By the properties of random infinite-constraint-length convolutional codes, this attains the random-coding error exponent with respect to delay for every piece of punctuation in the stream. The earlier punctuation marks are almost certainly decoded correctly while more recent punctuation marks are more likely to be subject to error.
This current estimate for all the punctuation so far is then used by the subsystem responsible for decoding the message bits themselves. The decoded punctuation is used to tentatively parse the channel outputs into variable-length blocks and then tentatively decode those blocks under the assumption that the punctuation is correct. Any bits that have reached their deadlines are then emitted. Although the decisions for those bits are now committed from the destination’s point of view, this does not prevent the system from re-parsing them in the future when considering estimates for other bits.
### Analysis
An error can occur at the decoder in two different ways. As before, the message-carrying stream could be delayed due to channel atypicality in its own channel slots. The new source of errors is that the punctuation stream could also become corrupted through atypicality in these other channel slots. As a result, the punctuation overhead $\theta$ must be kept proportional to the chunk length $c$ to avoid having punctuation errors cause too many decoding errors.
Set $\theta = \psi c$ for a constant $\psi$ to be optimized. The rate of the punctuation information is $\frac{l+1}{\theta} =
\frac{l+1}{\psi c} $ and goes to zero as $c \rightarrow \infty$. Since the random-coding error exponent at rate $0$ approaches $E_0(1)$, this is the relevant error exponent for the second stream relative to the channel uses that it gets. But there are only $\psi$ punctuation-code channel uses per second and so the delay-exponent for the punctuation stream is actually $\psi E_0(1)$ with respect to true delay.
Meanwhile, the chunk size in the message stream is $c' = c
(1-\psi)$. The effective rate of the message stream is thereby increased to $\frac{R}{1-\psi}$. Assuming that the punctuation information is correct, the fixed-delay error-exponent is as close as we would like to $E_{a,s}(\frac{R}{1-\psi})$ with respect to the delay in terms of message-code channel uses. But there are only $(1-\psi)$ message-code channel uses per second and so the delay exponent approaches $(1-\psi)E_{a,s}(\frac{R}{1-\psi})$ with respect to true end-to-end delay.
Consider a large fixed delay $d$. It can be written as $d = d_f + d_m$ in $d$ different ways. Let $d_f$ be the part of the end-to-end delay that is burned by errors in the punctuation stream. That is, with probability exponentially small in $d_f$, this suffix of time has possibly incorrect punctuation information and so cannot be trusted to be interpreted correctly. If the bit did not make it out correctly in the $d_m$ time steps (corresponding to $(1-\psi)d_m$ channel uses for the message-code) where the punctuation is correct, we assume that it will not come out correctly.
Since the channel uses are disjoint between the punctuation and message streams, the two error events are independent. The probability of an error with delay $d$ can thus be union-bounded as $$\begin{aligned}
& & {\cal P}(\widehat{B}_i(d) \neq B_i) \\
& \leq & \sum_{d_m = 1}^d {\cal P}(\mbox{message error with delay }d_m)
{\cal P}(\mbox{punctuation error with delay }d-d_m) \\
& \leq & \sum_{d_m = 1}^d K_1 \exp(-d_m((1-\psi)E_a(\frac{R}{1-\psi}) +
\epsilon_1)) K_2 \exp(-(d-d_m)(\psi E_0(1) +
\epsilon_2)) \\
& \leq & d K_1 K_2 \exp(-d (\min\left(\psi
E_0(1),(1-\psi)E_{a,s}(\frac{R}{1-\psi})\right) - \epsilon_1
-\epsilon_2)\end{aligned}$$ where $\epsilon_1,\epsilon_2$ are arbitrarily tiny constants and $K_1
K_2$ are large constants that together capture the nonasymptotic terms in the earlier analysis.
Since the focus here is on the asymptotic error exponent with delay, the polynomial and $\epsilon$ terms can be ignored and an achievable exponent is found by choosing $\psi$ so that the two exponents are balanced: $$E' = \psi E_0(1) = (1-\psi)E_{a,s}(\frac{R}{1-\psi}).$$
Evaluating the parametric forms (\[eqn:symmetricfeedbackbound\]) using $\eta = \rho$ for $E_{a,s}$, we get a pair of equations $$\begin{aligned}
\psi E_0(1) & = & (1-\psi)E_0(\rho), \\
\frac{E_0(\rho)}{\rho} & = & \frac{R}{1-\psi}.\end{aligned}$$ The first thing to notice is that simple substitution gives $$R = \frac{(1-\psi)E_0(\rho)}{\rho} = \frac{\psi E_0(1)}{\rho} = \frac{E'}{\rho}.$$
Solving for $\psi$ shows (after a little algebra) that $$\label{eqn:psidef}
\psi = \frac{E_0(\rho)}{E_0(1) + E_0(\rho)}.$$ This way $1-\psi = \frac{E_0(1)}{E_0(1) + E_0(\rho)}$ and the first equation is clearly true. Similarly $\frac{1}{1 - \psi} = 1 +
\frac{E_0(\rho)}{E_0(1)}$ and $\frac{\psi}{1 - \psi} =
\frac{E_0(\rho)}{E_0(1)}$ and thus the second equation is also true. Evaluating,
$$\begin{aligned}
E' & = & \psi E_0(1) \\
& = & \frac{E_0(\rho) E_0(1)}{E_0(1) + E_0(\rho)} \\
& = & \left(\frac{1}{E_0(\rho)} + \frac{1}{E_0(1)}\right)^{-1}.\end{aligned}$$
Simple (but mildly tedious) Taylor series expansion around the $\rho =
0$ point gives $E'(\rho) = 0 + C\rho + \frac{1}{2}(\frac{\partial^2
E_0(0)}{\partial \rho^2} - 2\frac{C^2}{E_0(1)})\rho^2 +
o(\rho^2)$ and thus $R(\rho) = C - (\frac{C^2}{E_0(1)} - \frac{1}{2}
\frac{\partial^2 E_0(0)}{\partial \rho^2})\rho + o(\rho)$. Taking the ratio of the first order terms gives the desired slope in the vicinity of the $(C,0)$ point. The fact that this slope is strictly negative is clear from the fact that $\frac{\partial^2
E_0(0)}{\partial \rho^2} \leq 0$. $\Box$
### Computation
As in the rate-$\frac{1}{2}$ erasure case discussed in Section \[sec:becexample\], the computational burden for the $(n,c,l)$ schemes is a constant that depends only on the particular scheme (and hence indirectly on the target rate-reliability pair) and not on the target end-to-end delay. As described, the complexity is exponential in the block length $nc$ since both the encoder and decoder must do list decoding among the codewords. The computational burden of the punctuation code is light since by Appendix \[app:feedbackconvolution\] it is like running a sequential decoder for a very-low-rate convolutional code.
More examples {#sec:fortifiedexamples}
-------------
![The sphere-packing and uncertainty-focusing bounds, with and without a noiseless side-channel of rate $\frac{1}{50}$ for a BSC with crossover probability $0.02$. The lower curves are the sphere-packing bounds and the upper curves are the uncertainty-focusing bounds. The thin lines represent the fortified cases with the added noiseless side-channel.[]{data-label="fig:fortifiedbsc"}](twoboundsfortified.edited.eps){width="4in" height="3in"}
Rather than considering an example using a DMC with strictly positive zero-error capacity, it is more instructive to consider a BSC with a fortification side-channel of rate $\frac{1}{50}$ bits per channel use. The capacity of the BSC with crossover probability $0.02$ increases to $0.61$ nats with such fortification and the Burnashev bound becomes infinite. Figure \[fig:fortifiedbsc\] shows the effect of zero-error capacity on the sphere-packing and uncertainty-focusing bounds. At high rates, the fortified uncertainty-focusing bound looks like it has just been shifted in rate by $0.01$ nats, just like the fortified sphere-packing bound. However, because of the flatness of the classical sphere-packing bound at high rates, the sphere-packing bound visually appears unchanged by fortification on a plot. At very low rates, the two behave differently. The fortified uncertainty-focusing bound tends smoothly to infinity at $0.01$ nats while the fortified sphere-packing bound jumps abruptly to infinity, reflecting the typical behavior of the error exponent curves for channels with strictly positive zero-error capacity.
![The dominant error events illustrated by plotting the ratio of future to past in dB scale. The horizontal axis is rate and the vertical axis is $10\log_{10}\frac{1-\lambda^*}{\lambda^*}$ where $\lambda^*$ is from (\[eqn:optimumlambda\]). The thicker red curve represents the unfortified channel while the thin black curve is the $\frac{1}{50}$-fortified system.[]{data-label="fig:pastfuturedb"}](pastfuturedb.edited.eps){width="4in" height="3in"}
Looking at a deeper level of detail, Figure \[fig:pastfuturedb\] illustrates the time-nature of the dominant error events at different rates. The question is for how long does the channel behave atypically for a bit to miss its deadline. In fixed-block-length coding, the usual source of errors is slightly atypical behavior across the entire block. As shown in Section \[sec:nofeedback\], when feedback is not available, the usual errors mainly involve the channel behaving atypically [*after*]{} the bit in question arrived at the encoder.
By contrast, in the fixed-delay context with feedback, the dominant error events involve more and more of the past as the rates get large. This means that the typical way for a bit to miss its deadline is for the channel to have been behaving atypically for some time before the bit even arrived at the encoder, and for this atypical behavior to continue till the deadline. At intermediate rates, the future behavior (after the bit has arrived at the encoder) becomes more important since it is more likely for the channel to become very bad for a shorter period.
At very low rates, the fortified and unfortified systems exhibit qualitatively different behavior. For unfortified systems, the dominant error events soon involve essentially only the future. The dominant event approaches the channel going into complete “outage” (e.g. the channel flipping half the inputs of a BSC) after the bit arrives at the encoder. For systems with positive zero-error capacity, such a complete outage is not possible as the message bits can always dribble across the error-free part. For an error to occur, it is essential to build up a large enough backlog in the queue and thus the past behavior starts to become dominant again. The curves diverge for the same rates at which the fortified case’s uncertainty-focusing bound is much better than the unfortified case.
![The fixed-delay error exponents of different schemes for a $\frac{1}{50}$-fortified BSC with crossover probability $0.02$ used with noiseless feedback. The lowest curve is the sphere-packing bound limiting feedback-free performance. The three new curves represent what is attained by the $(10,3,2), (20,4,3), (50,8,6)$ schemes described in Section \[sec:nclscheme\] and vary by block length, granularity, and the size of the lists used for list decoding. The uncertainty-focusing bound with and without fortification is plotted for reference.[]{data-label="fig:fortifiedperformance"}](performance.edited.eps){width="4in" height="3in"}
Figure \[fig:fortifiedperformance\] shows the fixed-delay reliabilities achieved by the $(n,c,l)$ schemes[^8] of Section \[sec:nclscheme\] for the specific cases of $(10,3,2)$, $(20,4,3)$, $(50,8,6)$. These are delay universal since they hold with all sufficiently long delays. Increasing $l$ increases the list size and helps the low rate performance while large block lengths $n$ are needed to perform well at higher rates. It is interesting to see how how the $(10,3,2)$ scheme is already spectacularly better than the feedback-free case for all low to moderate rates. In this case, there are $10*3*50 = 1500$ BSC uses and only $10*3 = 30$ error-free control bits corresponding to a typical message block.
Conclusions
===========
This paper has shown that fixed-block-length and fixed-delay systems behave very differently when feedback is allowed. While fixed-block-length systems do not usually gain substantially in reliability with noiseless feedback, fixed-delay systems can achieve very substantial gains for any generic DMC. The uncertainty-focusing bound complements the classical sphere-packing bound and gives limits to what is possible. Furthermore, these limits can be approached in a delay-universal fashion for erasure channels and any channel with positive feedback zero-error capacity if the encoders have access to noiseless channel output feedback, even if that feedback is slightly delayed. The computational requirements in doing so do not scale with the desired probability of error and only depend on the target rate and delay exponent. The details of this work establish a connection between queuing and communication over noisy channels with feedback. For the constructions given here, the end-to-end delay is asymptotically dominated by time spent waiting in a queue.
Given that complete noiseless feedback now has unambiguously clear value for reliable communication, it is important for the community to explore the required quality of feedback. This paper only shows that slightly delayed feedback can be tolerated. The case of noisy or rate-constrained feedback in the fixed-delay context is almost entirely open (see [@BalancedForwardFeedback] for the case of erasure channels on both the forward and feedback links). In addition, both the upper and lower bounds here only cover the case of a single message stream. The multistream rate/reliability region is still unknown even for the BEC case [@ControlPartII].
Stepping back, these results are also interesting because they show how feedback changes the qualitative nature of the dominant error events. Without feedback, errors are dominated by future channel behavior, but when feedback is available, the dominant event involves a mixture of the past and future. When the rate is low, the future tends to be more important but when the rate is high, the past starts to dominate. This brings to mind Shannon’s intriguing comment at the close of [@ShannonLossy]:
> \[The duality between source and channel coding\] can be pursued further and is related to a duality between past and future and the notions of control and knowledge. Thus we may have knowledge of the past and cannot control it; we may control the future but have no knowledge of it.
In [@OurSourceCodingBound], we explore the source-coding analogs of the results given here. In particular, feedback is found to be irrelevant in point-to-point lossless source coding and the dominant error events involve only the past! That makes precise the duality hinted at by Shannon.
Finally, in [@WaterslidePaper], the techniques developed here are extended to lower-bound the complexity of decoding based on iterative message-passing for general codes. The linear concept of time here is generalized to the message-passing graph. The role of delay is thus played by the decoding neighborhood within the graph and the corresponding bounds reveal the complexity cost of approaching capacity with such decoding algorithms.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks his student Tunc Simsek for many productive discussions. This work builds on the line of investigation that we opened up in Tunc’s doctoral thesis [@TuncThesis], although the proofs are different. I also thank Pravin Varaiya for his support, Sanjoy Mitter and Nicola Elia for many discussions over a long period of time which influenced this work in important ways, and the Berkeley students in the Fall 2004 advanced information theory course who forced me to simplify the presentation considerably. The anonymous reviewers are also thanked for their very helpful comments.
Feedback, convolutional codes, and complexity {#app:feedbackconvolution}
=============================================
The encoder is allowed to “look over the shoulder” of the decoder and have access to noiseless feedback of the channel outputs. This appendix gives[^9] the convolutional parallel to the Burnashev problem of variable-block-length codes. For ease of exposition, suppose the channel is binary input and that the uniform distribution is an optimal input distribution. If another input distribution is desired, mappings in the style of Figure 6.2.1 of [@gallager] can be used to approximate the desired channel input distribution. Use $R' = \frac{R}{\ln 2}$ to refer to the input rate in bits per channel use rather than nats per channel use. Apply the “encode the error signals” advice of [@ooiwornell] to get the following simple construction of a random code:
- Start with an infinite-constraint-length random time-varying convolutional code. The $j$-th channel input $X_j = \sum_k
H_k(j)B_k \bmod 2$ is generated by correlating the input bits $B_1^{jR'}$ with a random binary string $H_1^{jR'}(j)$.
- Use the noiseless feedback to run a sequential decoder at the encoder. This gives the encoder access to $\widehat{B}_1^{(j-1)R'}(j-1)$ — the tentative estimates of the past input bits based on the channel outputs so far. Set $\widehat{B}_{jR'}(j-1) = 0$ since there is no estimate for the new bit, and then compute $\widetilde{B}_k(j) = B_k +
\widehat{B}_k(j-1) \bmod 2$ to represent the current error sequence. Since the probability of bit error is exponentially decreasing in delay [@JelinekSequential], only a small number of the $\widetilde{B}_k(j)$ are nonzero, and furthermore, these are all around the more recent bits. The expected number of nonzero error bits is therefore upper bounded by some constant.
- Run the infinite-constraint-length convolutional code using the error sequence rather than the input bits. $\widetilde{X}_j =
\sum_k H_k(j) \widetilde{B}_k(j) \bmod 2 = X_j + [\sum_k H_k(j)
\widehat{B}_k(j-1)] \bmod 2$. Input the resulting $\widetilde{X}_j$ into the channel.
Since the additional term $[\cdots]$ is entirely known at the receiver and modulo $2$ addition is invertible, this feedback code has exactly the same distance properties as the original code without feedback. Furthermore, since there are only a finite random number of nonzero error bits and the encoder knows where these are, the encoding complexity is a random variable with finite expectation.
If a block-code is desired, then pick an arbitrary length $d$ to terminate a block, and choose an overall block length $n$ so that $d$ is insignificant in comparison.
The expected per-channel-input constraint-length used by the code is a finite constant that only depends on the rate, while the overall probability of block error dies exponentially with the terminator length $d$. Consequently, the expected-constraint-length error-exponent for variable-constraint-length convolutional codes is infinite with noiseless feedback. If we also count the expected number of computations required to run the encoder’s copy of the decoder, then this result holds for all rates strictly below[^10] the computational cutoff rate $E_0(1)$. Even though noiseless feedback is used by the encoder to generate the channel inputs, the decoding is “sequential” in the sense of Jacobs and Berlekamp [@JacobsBerlekamp] and suffers from the resulting computational limitation of having a search-effort distribution with certain unbounded moments.
At rates above $E_0(1)$, the same flavor of result can be preserved in principle by using the concatenated-coding transformations of Pinsker [@PinskerComplexity] (as well as others described more recently by Arikan [@Arikan05]) to bring the computational-cutoff rate $E_0(1)$ as close to $C$ as desired. Thus, [*the expected-computation error exponent for convolutional-style codes with noiseless output feedback can be made essentially infinite at all rates below capacity.*]{} The expected complexity is a constant that depends only on the desired rate, not on the target probability of error.
Extended Proofs
===============
Lemma \[lem:newdataprocessing\] {#app:lemdataprocessing}
-------------------------------
$$\begin{aligned}
n (R - \delta_1) & = & H(B_1^{n (R' - \delta_1')}) \\
& = & I(B_1^{n (R' - \delta_1')} ; B_1^{n(R' - \delta_1')}) \\
& \leq_{(a)} & I(B_1^{n(R' - \delta_1')} ; \widetilde{B}_1^{n(R' - \delta_1')},
Y_1^{n}) \\
& = & H(Y_1^{n}) + H(\widetilde{B}_1^{n(R' - \delta_1')}|Y_1^{n})
-H(Y_1^{n}|B_1^{n(R' - \delta_1')})
-H(\widetilde{B}_1^{n(R' - \delta_1')}|Y_1^{n},B_1^{n(R' -
\delta_1')}) \\
& =_{(b)} & H(Y_1^{n}) + H(\widetilde{B}_1^{n(R' - \delta_1')}|Y_1^{n})
-H(Y_1^{n}|B_1^{n(R' - \delta_1')}) \\
& = & H(\widetilde{B}_1^{n(R' - \delta_1')}|Y_1^{n})
+ I(Y_1^{n};B_1^{n(R' - \delta_1')}) \\
& \leq_{(c)} & H(\widetilde{B}_1^{n(R' - \delta_1')}) + I(Y_1^{n};B_1^{n(R' -
\delta_1')}) \\
& \leq_{(d)} & H(\widetilde{B}_1^{n(R' - \delta_1')}) + I(X_1^{n};Y_1^{n}).\end{aligned}$$
The first equality holds because the message bits are fair coin tosses. (a) comes from the data processing inequality when considering the following trivial Markov chain: $B_1^{n(R'-\delta_1')}~-~(\widetilde{B}_1^{n(R' - \delta_1')},
Y_1^n)~-~B_1^{n(R'-\delta_1')}$ that comes from the fact that the channel outputs and the error signals are enough to reconstruct the original bits. After expanding in terms of entropies, the $H(\widetilde{B}_1^{n(R' - \delta_1')}|Y_1^{n},B_1^{n(R' -
\delta_1')})$ term can be dropped to give (b) since this conditional entropy is zero because the error signal can be reconstructed from the message bits and the channel outputs. (c) comes from dropping conditioning, while the final inequality (d) comes from applying the data processing inequality to the Markov chain $B_1^{n(R'-\delta_1')}~-~
X_1^{n} ~-~ Y_1^n$ capturing the lack of feedback in the system. $\Diamond$
Lemma \[lem:measurechange\] {#app:lemmeasurechange}
---------------------------
Before proving Lemma \[lem:measurechange\], it is useful to establish a result involving typical sets.
### Typical set lemma
\[lem:typicalsequenceprob\] For every finite DMC $G$ and $\epsilon_1, \epsilon_2 > 0$, there exists a constant $K$ such that for every $\vec{x}$ $$\label{eqn:smallprob}
{\cal P}_G(\vec{Y} \in J_{\vec{x}}^{\epsilon_1, \epsilon_2} | \vec{X} =
\vec{x}) \geq 1 - |{\cal X}||{\cal Y}| \exp(-K d)$$ where $d$ is the length of the vectors $\vec{x}, \vec{Y}$, and the appropriate typical set is $$\label{eqn:Jdef}
J_{\vec{x}}^{\epsilon_1, \epsilon_2} = \left\{\vec{y} | \forall x \in {\cal
X} \mbox{ either }\frac{n_x(\vec{x})}{d} < \epsilon_2 \mbox{ or } \forall y \in {\cal Y},
\frac{n_{x,y}(\vec{x},\vec{y})}{n_x(\vec{x})} \in (g_{y|x} -
\epsilon_1, g_{y|x} + \epsilon_1) \right\}$$ where $n_{x,y}(\vec{x},\vec{y})$ is the count of how many times $(x,y)$ occurs in the sequence $(x_1, y_1), (x_2, y_2),
\ldots, (x_d, y_d)$, and $n_{x}(\vec{x})$ is the count of how many $x$ are present in the $d$ length vector $\vec{x}$.
Furthermore, for any $\vec{y} \in J_{\vec{x}}^{\epsilon_1,
\epsilon_2}$, the probability of the sequence $\vec{y}$ under a different channel $P$ satisfies $$\label{eqn:typicalratio}
\frac{{\cal P}_P(Y_1^d = \vec{y} | X_1^d = \vec{x})}{{\cal P}_G(Y_1^d
= \vec{y} | X_1^d = \vec{x})}
\geq
\exp(-d \left[D\bigg(G || P | \vec{r}(\vec{x})\bigg) +
(2\epsilon_2 + \epsilon_1) \sum_{x,y : n_x(\vec{x}) \neq 0, g_{y|x} \neq 0}
|\ln\frac{g_{y|x}}{p_{y|x}}|\right])$$ where $\vec{r}(\vec{x})$ is the type of $\vec{x}$. In particular, $$\label{eqn:typicalratioworst}
\frac{{\cal P}_P(Y_1^d = \vec{y} | X_1^d = \vec{x})}{{\cal P}_G(Y_1^d = \vec{y} | X_1^d
= \vec{x})}
\geq
\exp(-d [\max_{\vec{r}} D(G || P | \vec{r}) +
(2\epsilon_2 + \epsilon_1) \sum_{x,y : n_x(\vec{x}) \neq 0, g_{y|x} \neq 0}
|\ln\frac{g_{y|x}}{p_{y|x}}|]).$$
[*Proof:* ]{} The first goal is to establish (\[eqn:smallprob\]). For every $x \in {\cal X}$, the weak law of large numbers for iid finite random variables says that the relative frequency of $y$’s will concentrate around $g_{y|x} \pm
\epsilon_1$. Simple Chernoff bounds for the Bernoulli random variables representing the indicator functions tell us that this convergence is exponentially fast in that $\forall (x,y) \exists \zeta_{x,y} > 0$ so that if the channel input is always $x$ for a length $d$, the random number $N_y$ of times the channel output is $y$ satisfies $${\cal P}(|\frac{N_y}{d} - g_{y|x}| \geq \epsilon_1 |X_1^d = x_1^d)
\leq \exp(-\zeta_{x,y} d).$$
Let $K' = \min_{x \in {\cal X}, y \in {\cal Y}} \zeta_{x,y} > 0$. Set $K = \epsilon_2 K'$ since there are at least $\epsilon_2 d$ occurrences of the relevant $x$ values. Finally, apply the union bound over all $|{\cal X}||{\cal Y}|$ possible pairs to get (\[eqn:smallprob\]).
To show (\[eqn:typicalratio\]), first note that those $(x,y)$ pairs for which $g_{y|x} = 0$ can be ignored since these cannot occur in any sequence with nonzero probability under $G$. Then $$\begin{aligned}
{\cal P}_G(Y_1^d = \vec{y} | X_1^d = \vec{x})
& = &
\prod_{i=1}^d g_{y_i|x_i} \\
& = &
\prod_{x \in {\cal X}, y \in {\cal Y}} g_{y|x}^{n_{x,y}(\vec{x},\vec{y})} \\
& = &
(\prod_{x \in {\cal X}, y \in {\cal Y}}
g_{y|x}^{\frac{n_{x,y}(\vec{x},\vec{y})}{d}})^d \\
& = & \exp(d \sum_{x \in {\cal X}, y \in {\cal Y}} \frac{n_{x,y}(\vec{x},\vec{y})}{d}
\ln g_{y|x}) \\
& = & \exp(d \sum_{x \in {\cal X}} \frac{n_x(\vec{x})}{d} \sum_{y \in {\cal Y}}
\frac{n_{x,y}(\vec{x},\vec{y})}{n_x(\vec{x})}\ln g_{y|x}).\end{aligned}$$ Similarly $${\cal P}_P(Y_1^d = \vec{y} | X_1^d = \vec{x}) =
\exp(d \sum_{x \in {\cal X}} \frac{n_x(\vec{x})}{d} \sum_{y \in {\cal Y}}
\frac{n_{x,y}(\vec{x},\vec{y})}{n_x(\vec{x})}\ln p_{y|x}).$$ The ratio of the two probabilities is thus $$\frac{{\cal P}_P(Y_1^d = \vec{y} | X_1^d = \vec{x})}{{\cal P}_G(Y_1^d = \vec{y} |
X_1^d = \vec{x})} = \exp(-d \sum_{x \in {\cal X}} r_x(\vec{x}) \sum_{y
\in {\cal Y}} \frac{n_{x,y}(\vec{x},\vec{y})}{n_x(\vec{x})}\ln
\frac{g_{y|x}}{p_{y|x}}).$$
Now apply the definition of $J_{\vec{x}}^{\epsilon_1,
\epsilon_2}$ and first bound the contribution to the exponent by those inputs $x \in {\cal X}^{rare}$ that occur too rarely: $n_x(\vec{x}) < \epsilon_2 d$. We drop the arguments of $(\vec{x},
\vec{y})$ when they are obvious from context. $$\begin{aligned}
\sum_{x \in {\cal X}^{rare}} r_x \sum_{y \in {\cal Y}}
\frac{n_{x,y}}{r_x d}\ln \frac{g_{y|x}}{p_{y|x}}
& \leq &
\sum_{x \in {\cal X}^{rare}} \left[r_x \sum_{y \in {\cal Y}}
g_{y|x} \ln \frac{g_{y|x}}{p_{y|x}}
+ 2\epsilon_2 \sum_{y \in {\cal Y},
g_{y|x} \neq 0} |\ln\frac{g_{y|x}}{p_{y|x}}|\right] \\
& \leq &
2\epsilon_2 \sum_{x,y : r_x \neq 0, g_{y|x} \neq 0} |\ln\frac{g_{y|x}}{p_{y|x}}| +
\sum_{x \in {\cal X}^{rare}} r_x \sum_{y \in {\cal Y}}
g_{y|x} \ln \frac{g_{y|x}}{p_{y|x}}.\end{aligned}$$ For the non-rare $x$, the $n_{x,y}$ are already within $\epsilon_1$ of $g_{y|x}$ and thus, for $y_1^d \in J_{\vec{x}}^{\epsilon_1,
\epsilon_2}$, $$\begin{aligned}
\frac{{\cal P}_P(Y_1^d = y_1^d | X_1^d = x_1^d)}{{\cal P}_G(Y_1^d = y_1^d | X_1^d
= x_1^d)}
& = & \exp(-d \sum_{x \in {\cal X}} r_x \sum_{y \in {\cal Y}}
\frac{n_{x,y}}{r_x d}\ln \frac{g_{y|x}}{p_{y|x}}) \\
& \geq &
\exp(-d [\sum_{x \in {\cal X}} r_x \sum_{y \in {\cal Y}}
g_{y|x} \ln \frac{g_{y|x}}{p_{y|x}} +
(2\epsilon_2 + \epsilon_1) \sum_{x,y : r_x \neq 0, g_{y|x} \neq 0}
|\ln\frac{g_{y|x}}{p_{y|x}}|] \\
& = &
\exp(-d [D(G || P | \vec{r}) +
(2\epsilon_2 + \epsilon_1) \sum_{x,y : r_x \neq 0, g_{y|x} \neq 0}
|\ln\frac{g_{y|x}}{p_{y|x}}|]\end{aligned}$$ which establishes (\[eqn:typicalratio\]). To get (\[eqn:typicalratioworst\]), just bound by the worst possible $\vec{r}$. $\Diamond$
### Proof of Lemma \[lem:measurechange\] itself
If $g_{y|x} \neq 0$, then it is safe to assume $p_{y|x}
\neq 0$ as well since otherwise the divergence is infinite and the Lemma is trivially true.
The finite sum $\sum_{x,y : r_x \neq 0, g_{y|x} \neq 0}
|\ln\frac{g_{y|x}}{p_{y|x}}|$ is thus just some finite constant $K'$ that depends only on $G$ and $P$. By choosing $\epsilon_1, \epsilon_2$ small enough, it is possible to satisfy $(2\epsilon_2 + \epsilon_1)
\sum_{x,y : r_x \neq 0, g_{y|x} \neq 0} |\ln\frac{g_{y|x}}{p_{y|x}}| <
\epsilon$.
The event $A$ has a substantial conditional probability $\delta$ when channel $G$ is used and this probability does not diminish with $d$. Consequently, Lemma \[lem:typicalsequenceprob\] implies that for the chosen $\epsilon_1, \epsilon_2 > 0$, there exists a constant $K$ so that ${\cal P}_G(\vec{Y} \in J_{\vec{x}}^{\epsilon_1,
\epsilon_2} | \vec{X} = \vec{x}) \geq 1 - |{\cal X}||{\cal Y}|
\exp(-K d)$.
Pick a $d_0(G,\delta, \epsilon_1,\epsilon_2) > 0$ large enough so that $|{\cal X}||{\cal Y}| \exp(-K d_0) < \frac{\delta}{2}$. Thus ${\cal
P}_G(A \cap J_{\vec{x}}^{\epsilon_1, \epsilon_2} | \vec{X} =
\vec{x}) \geq \frac{\delta}{2}$. The immediate application of the second part of Lemma \[lem:typicalsequenceprob\] gives $$\begin{aligned}
{\cal P}_P(A | \vec{X} = \vec{x}) & \geq & {\cal P}_P(A \cap
J_{\vec{x}}^{\epsilon_1, \epsilon_2} | \vec{X} = \vec{x}) \\
& \geq &
\frac{\delta}{2} \exp\left(-d [D(G || P | \vec{r}(\vec{x})) +
(2\epsilon_2 + \epsilon_1) \sum_{x,y : n_x(\vec{x}) \neq 0, g_{y|x} \neq 0}
|\ln\frac{g_{y|x}}{p_{y|x}}|]\right) \\
& \geq &
\frac{\delta}{2} \exp(-d [D(G || P | \vec{r}(\vec{x})) + \epsilon]) \end{aligned}$$ which is the desired result. $\Diamond$
Expressing the symmetric uncertainty-focusing bound in parametric form {#app:focussymmetric}
----------------------------------------------------------------------
$$\begin{aligned}
E_{a,s}(R) & = &
\inf_{0 \leq \lambda < 1} \frac{E^+(\lambda R)}{1-\lambda} \\
& = & \inf_{0 \leq \lambda < 1} \max_{\rho \geq 0} \frac{E_0(\rho) - \rho
\lambda R}{1-\lambda}.\end{aligned}$$
To find the minimizing $\lambda$, first observe that given $\lambda$, the maximizing $\rho$ is the solution to $$\label{eqn:rhosolve}
\frac{\partial E_0(\rho)}{\partial \rho} = \lambda R.$$ A solution exists because $E_0$ is concave $\cap$ [@gallager]. If the solution is not unique, just pick the smallest solution. Call this solution to (\[eqn:rhosolve\]) as $\rho(\lambda,R)$. Let $$g(\lambda,R) = E_0(\rho(\lambda,R)) - \rho(\lambda,R)\lambda R.$$ Now, the goal is to minimize $\frac{g(\lambda,R)}{1 - \lambda}$ with respect to $\lambda$. Take a derivative and set it to zero: $$g(\lambda,R) + (1 - \lambda) \frac{\partial g(\lambda,R)}{\partial
\lambda} = 0.$$ But $$\begin{aligned}
\frac{\partial g(\lambda,R)}{\partial \lambda}
& = &
\frac{\partial E_0(\rho(\lambda,R))}{\partial \rho}
\frac{\partial \rho(\lambda,R)}{\partial \lambda}
-
\lambda R
\frac{\partial \rho(\lambda,R)}{\partial \lambda}
-
\rho(\lambda,R) R \\
& = & \frac{\partial \rho(\lambda,R)}{\partial \lambda} (
\frac{\partial E_0(\rho(\lambda,R))}{\partial \rho} -
\lambda R) - \rho(\lambda,R) R \\
& = & - \rho(\lambda,R) R.\end{aligned}$$ So, solve for $\lambda^*$ in $$g(\lambda^*,R) = \rho(\lambda^*,R) R (1 - \lambda^*).$$ Plugging in the definition of $g$ gives $$\frac{E_0(\rho(\lambda^*,R))}{\rho(\lambda^*,R) R} - \lambda^* =
1 - \lambda^*.$$ Which implies $$\frac{E_0(\rho(\lambda^*,R))}{\rho(\lambda^*,R) R} = 1$$ or $R = \frac{E_0(\rho(\lambda^*,R))}{\rho(\lambda^*,R)}$. For the other part, just notice $$\begin{aligned}
E_a(R) & = & \frac{1}{1 - \lambda^*} g(\lambda^*,R) \\
& = & \frac{1}{1 - \lambda^*} \rho(\lambda^*,R) R (1 - \lambda^*) \\
& = & \rho(\lambda^*,R) R \\
& = & E_0(\rho(\lambda^*,R)).\end{aligned}$$ Setting $\eta = \rho(\lambda^*,R)$ gives (\[eqn:symmetricfeedbackbound\]). $\Box$
Proof of the low-rate approximation in Theorem \[thm:erasurecase\] {#app:erasurelowrate}
------------------------------------------------------------------
First, solve for $2^\eta$ in (\[eqn:becreliability\]) in terms of the reliability $E_{a}^{bec}(R') = \alpha$. This gives $2^{\eta} =
\frac{1-\beta}{2^{-\alpha} - \beta}$ and so $\eta = \alpha +
\log_2({1-\beta}{1 - 2^\alpha \beta})$. Plugging into the $R'$ expression (\[eqn:becreliability\]) gives the desired $C'(\alpha)$ tradeoff.
It is worthwhile to investigate the behavior of this $C'(\alpha)$ for values of reliability $\alpha$ close (within a factor of 2) to the fundamental upper limit of $-\log_2 \beta$. Consider $0 \leq \epsilon
\leq -\frac{\log_2 \beta}{2}$. When $\alpha = (-\log_2 \beta -
\epsilon)$, $$\begin{aligned}
C'(-\log_2 \beta - \epsilon)
& = & \frac{-\log_2 \beta - \epsilon}{-\log_2 \beta - \epsilon +
\log_2 (1-\beta) - \log_2(1-\beta 2^{-\epsilon - \log_2 \beta})} \\
& = & \left(1 + \frac{\log_2 (1-\beta) - \log_2(1-
2^{-\epsilon})}{-\log_2 \beta - \epsilon}\right)^{-1}.\end{aligned}$$ $(1-2^{-\epsilon})$ is a concave $\cap$ function of $\epsilon\in[0,1]$ and can thus be lower-bounded by $\frac{1}{2} \epsilon$. This gives $$\begin{aligned}
C'(-\log_2 \beta - \epsilon)
& \geq & \left(1 +
\frac{\log_2 (1-\beta) + \log_2(2 \epsilon^{-1})}{-\log_2 \beta - \epsilon}\right)^{-1} \\
& > & \left(1 +
\frac{\log_2(2 \epsilon^{-1})}{\log_2(\beta^{-1}) -
\epsilon}\right)^{-1} \\
& \geq & \left(1 +
2\frac{\log_2(2 \epsilon^{-1})}{\log_2(\beta^{-1})}\right)^{-1} .\end{aligned}$$ Plugging in $\epsilon = 2 \beta^r$ is valid as long as $r \geq \frac{2 - \log_2 \log_2 \beta^{-1}}{\log_2
\beta^{-1}}$. This gives $$\label{eqn:smallRbehaviour}
C'\left((-\log_2 \beta) - 2\beta^r\right) \geq \frac{1}{1+2r}.$$
Proof of Lemma \[lem:notthereyet\] {#app:lemnotthereyet}
----------------------------------
An error can occur only when there have not been enough successful transmissions to get the $i$-th bit out in time. Applying the union bound to such events gives $$\begin{aligned}
{\cal P}(\widehat{B}_i(\lceil \frac{i}{R'} \rceil + d) \neq B_i)
& \leq &
\sum_{k=1}^i {\cal P}\left(\sum_{t=\lceil \frac{k}{R'} \rceil}^{\lceil \frac{i}{R'}
\rceil + d} Z_t \leq i-k\right) \\
& = & \sum_{k=1}^i {\cal P}(\frac{\sum_{t=\lceil \frac{k}{R'} \rceil}^{\lceil
\frac{i}{R'} \rceil + d} Z_t}{d + \lceil \frac{i}{R'} \rceil - \lceil
\frac{k}{R'} \rceil} \leq \frac{i-k}{d + \lceil \frac{i}{R'} \rceil - \lceil
\frac{k}{R'} \rceil}).\end{aligned}$$ This establishes the desired result.$\Diamond$
The details in the proof of Theorem \[thm:erasurecase\] {#app:erasuredetails}
-------------------------------------------------------
Notice that the event $\{\frac{\sum_{t=\lceil \frac{k}{R'}
\rceil}^{\lceil \frac{i}{R'} \rceil + d} Z_t}{d + \lceil \frac{i}{R'}
\rceil - \lceil \frac{k}{R'} \rceil} \leq \frac{i-k}{d + \lceil
\frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil} \}$ is just the error event for an ideal erasure-channel block code with block length $n(k) = d + \lceil \frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil$ and a bit rate of $R'(k) = \frac{i-k + 1}{d + \lceil \frac{i}{R'} \rceil - \lceil
\frac{k}{R'} \rceil}$. This is because it represents the event that the channel erases too many symbols. Let $\lambda(k) = 1 - \frac{d}{d + \lceil
\frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil}$. Then $n(k) = \lambda(k)
n(k) + d$ and $R'(k) \in (\lambda(k) R', \lambda(k) R' +
\frac{2}{n(k)})$. Thus, for every $\epsilon_1 > 0$, there exists a $d_1(\epsilon_1)$ so that for all $d > d_1(\epsilon_1)$, $$\begin{aligned}
{\cal P}(\frac{\sum_{t=\lceil \frac{k}{R'}
\rceil}^{\lceil \frac{i}{R'} \rceil + d} Z_t}{d + \lceil \frac{i}{R'}
\rceil - \lceil \frac{k}{R'} \rceil} \leq \frac{i-k}{d + \lceil
\frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil})
& \leq &
\exp(-n(k) \left(D(R'(k) || 1 - \beta) - \epsilon_1\right)) \\
& < &
\exp(-n(k) \left(D(\lambda(k)R' + \frac{2}{n(k)} || 1 - \beta) - \epsilon_1\right)) \\
& = &
\exp(-d \left[\frac{D(\lambda(k)R' + \frac{2}{n(k)} || 1 - \beta) - \epsilon_1}{1-\lambda(k)}\right]).\end{aligned}$$
Now, divide the events in (\[eqn:sumbound\]) into two categories (illustrated in Figure \[fig:erasurebounding\]) based on a critical value for $\lambda(k)$ and $k$. Let $\lambda^*$ from (\[eqn:optimumlambda\]) be the $\lambda$ that minimizes the exponent $\frac{D(\lambda R' || 1 - \beta)}{1 - \lambda}$. Set $\bar{n} = d
\frac{D(\lambda^*R' || 1 - \beta)}{(1-\lambda^*)D(R' || 1 -
\beta)}$. Let $\bar{k}$ be the largest $k$ for which $n(k) >
\bar{n}$. For all $k < \bar{k}$, $$\begin{aligned}
& & {\cal P}(\frac{\sum_{t=\lceil \frac{k}{R'}
\rceil}^{\lceil \frac{i}{R'} \rceil + d} Z_t}{d + \lceil \frac{i}{R'}
\rceil - \lceil \frac{k}{R'} \rceil} \leq \frac{i-k}{d + \lceil
\frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil}) \nonumber\\
& < &
\exp(-n(k)
(D\bigg(\lambda(k)R' + \frac{2}{n(k)} || 1 - \beta \bigg) - \epsilon_1)) \nonumber\\
& < &
\exp(-n(k)
(D\bigg(R' + \frac{2}{\bar{n}} || 1 - \beta \bigg) - \epsilon_1)) \nonumber\\
& = &
\exp(-(\bar{n} + (n(k) - \bar{n})) (D(R' + \frac{2}{\bar{n}} || 1 - \beta) -
\epsilon_1)) \nonumber\\
& = &
\exp\left(-d \frac{D(\lambda^*R' || 1 - \beta)}{1-\lambda^*}(\frac{D(R' +
\frac{2}{\bar{n}} || 1 - \beta) - \epsilon_1}{D(R' || 1 - \beta)})\right)
\exp\left(-(n(k) - \bar{n})(D(R' + \frac{2}{\bar{n}} || 1 - \beta) -
\epsilon_1)\right) \label{eqn:caseksmall}.\end{aligned}$$ Meanwhile, for $k \geq \bar{k}$, $$\begin{aligned}
{\cal P}(\frac{\sum_{t=\lceil \frac{k}{R'}
\rceil}^{\lceil \frac{i}{R'} \rceil + d} Z_t}{d + \lceil \frac{i}{R'}
\rceil - \lceil \frac{k}{R'} \rceil} \leq \frac{i-k}{d + \lceil
\frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil})
& < &
\exp(-d \frac{D(\lambda(k)R' + \frac{2}{n(k)} || 1 - \beta) -
\epsilon_1}{1-\lambda(k)}) \nonumber\\
& \leq &
\exp(-d [\frac{D(\lambda(k)R' + \frac{2}{d} || 1 - \beta)}{1 - \lambda(k)} -
\frac{\epsilon_1}{1-\lambda^*}]). \label{eqn:caseklarge}\end{aligned}$$ If there were no $\frac{2}{d}$ term above, then the terms (\[eqn:caseklarge\]) could be bounded by using $\lambda^*$ in place of $\lambda(k)$ since $\lambda^*$ is the worst possible $\lambda$. But since the divergence is continuous in its first argument and $\frac{2}{d}$ is small, we bound them all by allowing for a small slop $\epsilon_2$. Explicitly, for every $\epsilon_2 > 0$, it is clear there exists a $d_2(\epsilon_2) > 0$ so that for all $d > d_2(\epsilon_2)$ and $k$ such that $n(k) \leq \bar{n}$, we have $$\begin{aligned}
{\cal P}(\frac{\sum_{t=\lceil \frac{k}{R'}
\rceil}^{\lceil \frac{i}{R'} \rceil + d} Z_t}{d + \lceil \frac{i}{R'}
\rceil - \lceil \frac{k}{R'} \rceil} \leq \frac{i-k}{d + \lceil
\frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil})
& < &
\exp(-d [(1-\epsilon_2) \frac{D(\lambda^*R' || 1 - \beta)}{1 - \lambda^*} -
\frac{\epsilon_1}{1-\lambda^*}]). \label{eqn:caseklarge2}\end{aligned}$$ Putting the two bounds (\[eqn:caseksmall\]) and (\[eqn:caseklarge2\]) together for $d > \max(d_1(\epsilon_1),
d_2(\epsilon_2))$ gives $$\begin{aligned}
& & {\cal P}(\widehat{B}_i(\lceil \frac{i}{R'} \rceil + d) \neq B_i) \\
& \leq &
\sum_{k=1}^i {\cal P}\left(\frac{1}{d + \lceil \frac{i}{R'} \rceil -
\lceil \frac{k}{R'} \rceil}
\sum_{t=\lceil \frac{k}{R'} \rceil}^{\lceil \frac{i}{R'} \rceil + d}
Z_t \leq \frac{i-k}
{d + \lceil \frac{i}{R'} \rceil - \lceil \frac{k}{R'} \rceil}
\right) \\
& < &
[\sum_{k=1}^{\bar{k}} \exp(-d \frac{D(\lambda^*R' || 1 -
\beta)}{1-\lambda^*}(\frac{D(R' + \frac{2}{\bar{n}} || 1 - \beta) -
\epsilon_1}{D(R' || 1 - \beta)}))
\exp(-(n(k) - \bar{n})(D(R' + \frac{2}{\bar{n}} || 1 -
\beta) - \epsilon_1))] \\
& &
+ (i - \bar{k} + 1) \exp(-d [(1-\epsilon_2) \frac{D(\lambda^*R' || 1 -
\beta)}{1 - \lambda^*} - \frac{\epsilon_1}{1-\lambda^*}]) \\
& < &
\exp(-d \frac{D(\lambda^* R' || 1 - \beta)}{1-\lambda^*}
(\frac{D(R' + \frac{2}{\bar{n}} || 1 - \beta) - \epsilon_1}
{D(R' || 1 - \beta)}
))
\left[ \lceil \frac{1}{R'} \rceil
\sum_{l=0}^{\infty}
\exp(-l (D(R' + 2\bar{n}^{-1} || 1 - \beta) - \epsilon_1)) \right] \\
& &
+ d(\frac{D(\lambda^*R' || 1 - \beta)}
{(1-\lambda^*)D(R' || 1 - \beta)})
\exp(-d \bigg((1-\epsilon_2) \frac{D(\lambda^*R' || 1 - \beta)}
{1 - \lambda^*}
- \frac{\epsilon_1}{1-\lambda^*} \bigg)).\end{aligned}$$
Proof of Lemma \[lem:transmissiontimes\] {#app:lemtransmissiondelay}
----------------------------------------
That the transmission times $T_j$ are iid is obvious since each depends on disjoint channel uses and the channel is memoryless and stationary. Before proving (\[eqn:bettertransmissionbound\]), it is useful to first establish $$\label{eqn:transmissionbound}
{\cal P}(T_j > t ck) \leq \exp(\rho R nck)
[\exp(-ck E_0(\rho,\vec{q}))]^t.$$ The only way that the transmission time can be longer than $tck$ for some integer $t \geq 1$ is if the block-length-$tck$ code cannot be correctly decoded to within a list of size $l$. The effective rate of the block code in nats is thus $$\begin{aligned}
\frac{\frac{nck}{R}}{tck}
& = & R\frac{n}{t}.\end{aligned}$$ Applying the list-decoding upper-bound (\[eqn:listrhorandom\]) on the probability of error for random block coding gives $$\begin{aligned}
{\cal P}(T_j > t ck)
& \leq &
\exp(-tck[E_0(\rho,\vec{q}) - \rho R\frac{n}{t}]) \\
& = &
\exp(\rho R nck)
\exp(-tck E_0(\rho,\vec{q})) \\
& = &
\exp(\rho R nck)
[\exp(-ck E_0(\rho,\vec{q}))]^t\end{aligned}$$ where this holds for all $0 \leq \rho \leq 2^l$. Pulling the constant into the exponent gives $$\begin{aligned}
{\cal P}(T_j > t ck)
& \leq &
[\exp(-ck E_0(\rho,\vec{q}))]^{t - \frac{\rho R nck}{ck
E_0(\rho,\vec{q})}} \\
& = &
[\exp(-ck E_0(\rho,\vec{q}))]^{t - \frac{\rho R n}{E_0(\rho,\vec{q})}}
\\
& \leq &
[\exp(-ck E_0(\rho,\vec{q}))]^{t - \lceil \frac{\rho R
n}{E_0(\rho,\vec{q})} \rceil} \\
& = &
[\exp(-ck E_0(\rho,\vec{q}))]^{t - \lceil \widetilde{t}(\rho,R,n,\vec{q}) \rceil}\end{aligned}$$ and this proves the desired result. $\Diamond$
[^1]: Wireless Foundations, Department of Electrical Engineering and Computer Science at the University of California at Berkeley.
[^2]: This is in the style of rateless block coding [@LTCodes], except that the message bits are revealed to the encoder in time rather than being known all at the beginning.
[^3]: Notice how the situation for unconstrained DMCs is dramatically different from the behavior of the AWGN channel with noiseless feedback for which Schalkwijk and Kailath showed double-exponential reliability with block length [@Kailath66; @Schalkwijk66]. However, those results rely crucially on the [ *variable*]{} nature of an input constraint that only has to hold on average. An unconstrained DMC is more like an AWGN channel with just a hard amplitude constraint on the channel inputs [@WynerSKLimit].
[^4]: Notice how Gallager’s definition of output-symmetric channels slightly generalizes the symmetric channel definitions of Dobrushin [@DobrushinReliability] and Csiszár and Körner [@csiszarkorner page 114]. Such output-symmetric channels can be understood as convex combinations of symmetric channels, each with its own distinct output alphabet. Knowledge of the partition the output lands in just tells the decoder which of the symmetric channels it happens to be encountering, but does not reveal anything about the channel input itself.
[^5]: Note that the same argument would also work if the encoder and decoder are allowed to share common randomness.
[^6]: If the feedback is not instantaneous, then there is no obvious scheme. Asymptotically optimal schemes for such cases are given in [@BalancedForwardFeedback].
[^7]: The target end-to-end delay must at least be large enough to absorb the roughly $2nck$ channel uses corresponding to the sum of assembly delay and essential service time for the message block. It is beyond that point that the delay exponent analysis here kicks in.
[^8]: The schemes plotted here have been slightly modified to use the noiseless side-channel to carry codeword information whenever it is not needed to carry punctuation information. This more accurately reflects the typical behavior of channels with strictly positive zero-error capacity.
[^9]: The scheme we describe in this subsection is too obvious to be original to us, but we are unaware of who might have come up with it earlier.
[^10]: At depth $\tau$ within a false path, each node expansion for a sequential decoder requires $O(\tau)$ multiply-accumulate operations to evaluate. This polynomial-order term is insignificant when compared to the rate-dependent exponential increase in the number of false nodes with increasing search depth. Thus, the polynomial term can be bounded away by just treating it as slight increase in the rate.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We investigate the formation of fingered flow in dry granular media under simulated rainfall using a quasi-2D experimental set-up composed of a random close packing of mono-disperse glass beads. Using controlled experiments, we analyze the finger instabilities that develop from the wetting front as a function of fundamental granular (particle size) and fluid properties (rainfall, viscosity).These finger instabilities act as precursors for water channels, which serve as outlets for water drainage. We look into the characteristics of the homogeneous wetting front and channel size as well as estimate relevant time scales involved in the instability formation and the velocity of the channel finger tip. We compare our experimental results with that of the well-known prediction developed by Parlange and Hill \[1976\]. This model is based on linear stability analysis of the growth of perturbations arising at the interface between two immiscible fluids. Results show that in terms of morphology, experiments agree with the proposed model. However, in terms of kinetics we nevertheless account for another term that describes the homogenization of the wetting front. This result shows that the manner we introduce the fluid to a porous medium can also influence the formation of finger instabilities.'
author:
- 'Cesare M. Cejas$^{1}$, Yuli Wei$^{1,2}$, Remi Barrois$^{1}$, Christian Fr$\acute{e}$tigny$^{3}$, Douglas J. Durian$^2$, and R$\acute{e}$mi Dreyfus$^{1}$'
bibliography:
- 'Kinetics\_References.bib'
title: 'Kinetics of Gravity-Driven Water Channels Under Steady Rainfall'
---
Water infiltration in soil is a long-standing research topic due to a wealth of interesting physical phenomena, such as fluid-granular interactions, as well as also having a wide variety of industrial applications. For example, rainwater can induce solute leaching as it drives contaminants from the unsaturated zone just below the soil surface to deeper areas underground such as the water table [@DeRooij00]. This affects the quality of groundwater and thus such infiltration studies have aimed to limit the adverse effects of groundwater contamination [@Glass91; @Wang98; @Hillel03]. Both laboratory [@Hill72; @Diment85] and real field experiments [@Ritsema98; @Hendrickx93] have confirmed the existence of preferential drainage paths in sandy soils under uniform flow via rainfall or irrigation water. In agricultural applications, when water drains through preferential channels, drainage greatly reduces the quantity of water around the root zone that could otherwise be absorbed by the plants. Understanding the physical mechanisms involved in water infiltration during rain can help in developing novel techniques that could potentially have direct applications in soil remediation and water retention. Since infiltration is an example of multiphase flow, basic interest on the subject have initially focused on the dynamics of the interface between two immiscible fluids.
Infiltration proceeds via the formation of preferential paths. Extensive experiments have shown that apart from soil structural heterogeneities like macropores [@Davidson84], preferential paths may also occur in homogeneous dry sand. This is due to the fingering instabilities developing from the interface of a wetting front that occur during initially uniform and gravity-driven fluid flow [@Hill72; @Diment85; @Glass89; @Baker90]. This has been observed in homogeneous sandy soil [@Raats73; @Vanommen89; @Ritsema93; @Yao93] but is nevertheless also proven in materials of varying wettability [@Bond69; @Wang00; @Cejas13].
Over the years, infiltration studies have employed empirical [@Green11; @Horton40; @Chen06], numerical [@Brutsaert77; @Pachepsky03], and theoretical [@Parlange76] solutions to describe the phenomenon observed in both real soil fields and laboratory simulations. Recent studies [@Cueto-Felgueroso08; @Cueto-Felgueroso09] have brought additional insight into existing equations in modelling gravity-driven flow. Such equations are fundamentally based on Richards$^{\prime}$ equation for unsaturated flow, which couples Darcy$^{\prime}$s law and mass conservation law. However, Richards$^{\prime}$ equation is unable to simulate fingering phenomenon [@Cueto-Felgueroso08; @Cueto-Felgueroso09; @Nieber05], thus extensions are normally added to account for certain aspects of multiphase flow [@Eliassi01]. Previous studies [@Glass91; @Diment85; @Raats73; @Parlange76; @Saffman58; @Chuoke59] have proposed models to explain experiments based on parameters that condition wetting front instability, such as water repellency [@Ritsema93; @Ritsema98; @Wang98] and water redistribution [@Philip69]. However, to our knowledge, most studies on infiltration have focused on morphology of the water channels that form during infiltration. Also, not much has been performed with regards to understanding kinetics. Some studies have focused on the change in the pressure jump that accompanies flow velocity through the unsaturated zone [@DiCarlo00]. Others have focused on how flow velocity is affected by hydraulic properties such as conductivity and saturation [@Raats73]. Still, not much has been brought to light regarding the influence of the water source, which is normally introduced to a porous medium in a homogeneous manner. Acquiring a full grasp of the dynamics of the phenomenon first requires comprehensive analysis of the fundamental physical features that arise from the infiltration process. This means that understanding how channel size and channel velocity are affected by granular and fluid properties remain to be key pieces in mapping out the entire puzzle of the phenomenon of finger instability.
In this paper, we present an experimental kinetic study on the dynamics of the formation of water channels during steady rainfall. Using a quasi-two dimensional (2D) set-up that simulates different rainfall rates and at the same time provides good visualization of water channel formation, we determine systematically the influence of physical parameters on the formation of the wetting front, instability, and propagation of water channels. These physical parameters include granular properties, such as particle size, which have been commonly studied. We also vary fluid viscosity, an important parameter whose effect on channel formation has not yet been sufficiently surveyed. Moreover, we also estimate relevant time scales involved in water channel formation, thereby providing additional information on the kinetics of the instability.
Experiment
==========
We use a quasi-two-dimensional ($2$D) cell, of cross-sectional area $A$ (length, $l$ = $30$ cm, cell thickness, $e$ = $0.8$ cm). To reduce wetting effects on the glass wall, the sample cell is made hydrophobic by washing with hydrophobic silane solution (OMS Chemicals). We attach screen meshes at the bottom portion of the cell to freely circulate air and to freely drain water while preventing glass beads from emptying out of the 2D cell.
This cell is filled with random close packing of monodisperse glass beads (A-series, Potters Industries, Inc.) as our model soil system. The glass beads are hydrophilic. We clean them by burning them in a furnace for $72$ hours at high temperature. Then we soak the glass beads in $1$ M HCl, rinse with deionized water, and then bake them in a vacuum oven for $12$ hours at $110^{\circ}$C and then left to cool at room temperature. Contact angle measurements on the clean hydrophilic beads reveal a contact angle, $\theta^*$, of $\theta^*$ $=$ $16\pm 2^{\circ}$. The packing porosity is $\epsilon$ = $0.36 - 0.40$ and is measured using the imbibition method. The cell is first filled with dry glass beads then the glass beads are taken out of the cell. Water is then slowly poured into the glass beads until they are fully saturated. The glass beads are weighed before and after the imbibition and since the density is known, the difference gives the pore volume. Calculating pore volume with respect to bulk volume of the cell gives the porosity.
We build a rain source with equally spaced glass capillary tubes (borosilicate micropipettes, VWR). The spacing between the tubes is $1$ cm. The rain source provides a constant rain rate $Q$ and we control the distance, $h$, between the tip of the capillary and the soil surface to control the droplet impinging speed. From the average masses of the raindrops, we estimate the droplet diameter to be $3$ mm. We suspend the sample cell beneath the rain source as shown in Fig. 1.
![(Color online) Diagram of the quasi-$2$D experimental set-up used to visualize the formation of water channels during steady rain. A rain source built with equally spaced capillary tubes provides a constant rainfall rate on the sample cell of cross-section $A$. The sample cell is filled with monodisperse glass beads as model soil and is suspended under the rain source. We can vary the distance between the capillary and model soil surface to vary the free-fall height of the raindrops. []{data-label="Samplecell"}](Samplecell){width="3.7in"}
We measure the rain rate by determining the volume of water per time per cross-sectional area. As expected, experiments show that the rain rate is proportional to the water level in the rain source. Because also of the design limitation of the size of the rain source, extremely high flow rates can only be achieved when the size of the capillary tubes is also modified. Control of the flow rate is set according to the water level height in the rain source and the size of the capillary tubes. Thus, to achieve higher flow rates, we vary the capillary tubes using readily available capillary tubes ($\pm0.5\%$) in the market: $5$ ${\mu}$L, $10$ ${\mu}$L, $25$ ${\mu}$L, and $50$ ${\mu}$L. The full lengths of all these commercial capillary tubes are $12.70 \pm 0.05$ cm and the outer diameters (OD) of all the tubes are measured to be within the range of approximately $1.6 -1.8$ $\pm$ $0.5$ mm. Since these tubes are in fact micropipettes, the volumes are calibrated only up to a certain effective length, which is $7.30 \pm 0.05$ cm. The inner diameters (ID) vary according to its volumetric capacity and can be calculated from the effective length. The values are $0.295 \pm 0.001$ mm, $0.418 \pm 0.002$ mm, $0.660 \pm 0.003$ mm, $0.934 \pm 0.004$ mm for the four aforementioned tubes respectively. But even though the capacity of the capillary and their ID values change, the OD values are roughly constant. Hence, the size of the droplet also roughly remains the same.
The presence of a light box behind the sample cell illuminates it from behind when taking images at $5$-second intervals using an SLR camera (D$90$, Nikon and Canon-SLR, Canon) that is automatically pre-set by a corresponding computer software.
To further probe the kinetics of the infiltration process, we modify the viscosity of the primary fluid (water) by adding a concentration of glycerol (Sigma Aldrich) to create a water/glycerol solution. The addition of glycerol increases the viscosity of the fluid. The properties of the water/glycerol solutions are presented in Table 1. The density and viscosity values are calculated from Cheng $et$ $al$ [@Cheng00]. The densities of the water/glycerol solutions do not significantly change whereas the viscosities are increased by a factor up to $8$ times. The interfacial tension values are obtained from a study of interfacial tensions as a function of the volume of glycerol fraction performed by Shchekotov [@Shchekotov10]. The interfacial tension values also do not significantly change. In addition, it has been determined from previous studies [@Drelich96] that glycerol is hydrophilic and thus the contact angle of the water/glycerol solutions is essentially hydrophilic and is similar to water.
For some infiltration experiments of more viscous fluids, we use a camera (Phantom) with a frame rate of $24$ fps to capture the infiltration and instability formation in slightly greater detail.
Before proceeding on how the infiltration phenomenon is influenced by physical parameters of the model soil, we probe the effects of the droplet impinging speed, $U_T$, on our system and we do not see any significant effect [@Wei13]. For the experiments described in this paper, we keep the droplet impinging speed constant at $U_T$ = $1.0$ m/s but vary the rain rat, $Q$, accordingly.
{width="6in"}
Experimental Results
====================
Experimental Observations - Infiltration Process
------------------------------------------------
In our experiments, infiltration under steady rain proceeds with rainwater initially wetting the soil surface as shown in Fig. $2$a.
As rain is continuously supplied, a homogeneous wetting front begins to penetrate and develop inside the porous medium. As water continues to infiltrate vertically in the direction of gravity, the interface of the front eventually becomes unstable. Some of these instabilities fully develop into water channels while others do not. As soon as the water channels form, the wetting front ceases to infiltrate further deeper into the medium and these water channels serve as preferential paths for the drainage of water. We study the infiltration process in initially dry and hydrophilic granular beads. First, we keep rain rate $Q$ constant but vary the diameter of the glass beads, $D$ = $2R$, which is proportional to the characteristic size of the pore [@Culligan05]. Fig. $2$b shows a representative experimental image sequence for infiltration of water at $\mu$ = $1$ mPa$\cdot$s, $Q$ = $14.5$ cm/hr, and at $D$ = $300~{\mu}$m. In Fig. $2$b, we also observe a second finger instability, which results to a second water channel.
Next, we keep the diameter $D$ constant but vary viscosity $\mu$ and rain rate $Q$. Fig. $2$c shows another representative image sequence for infiltration of water/glycerol mixture at $\mu$ = $4$ mPa$\cdot$s, $Q$ = $96.0$ cm/hr, and at $D$ = $1~$mm.
In all these experiments, we measure the extent of the physical observations, such as the maximum width of the homogenous wetting front, $z_{wet}$, the average width or diameter of the channels, $d$, and the distance between two channels, $d^{\prime}$. These results concerning morphology are discussed in our companion paper [@Wei13].
Experimental Observations - Kinetics
------------------------------------
We can observe features common to all the performed experiments regardless of viscosity, flow rate, impinging speed, or bead diameters. Once rain begins to reach the soil, the first thing we observe is the establishment of a homogeneous wet front or wet zone, Fig. $3$. The development of the wet front may be slow or fast depending on the both properties of the fluid and the granular medium. The front gradually increases in size due to the presence of a continuous rain source, which supplies water to the top of the model soil surface. When it sufficiently forms, results suggest at first glance that the front is completely saturated. However, it has been shown that in the direction of gravity, a gradient of water saturation actually exists between the surface of the model soil and the area immediately just below it. This is to say that saturation levels are higher in the bottom of the front than at the top [@Rezanezhad07; @Cueto09]. Recent studies have suggested the role of saturation levels in the wetting front on instability formation although currently in literature, the wetting front is still a subject of ongoing investigations.
From experiments, the formation of the wetting front is particularly recognizable for fluids of low viscosities ($1$ mPa$\cdot$s). At larger viscosities, the homogeneous wetting front is easier to identify at higher flow rates. At larger viscosities but at lower flow rates, the wetting front appears faint because experiments give an impression that the viscous droplets do not spread enough to sufficiently coalesce with neighboring droplets.
We plot the average velocity of the wetting front, $v_{wet}$, as a function of bead diameter, $D$, in Fig. $4$a. The plot shows that the wetting front propagates at higher velocities in larger bead diameters than in smaller ones.
![(Color online). (a) Experimental images showing the evolution of the homogeneous wet zone with time during the early moments of rainfall. This is for the following experiment: droplet impinging speed, $U_T = 1.0$ m/s, particle size $D=300~\mu$m, viscosity $\mu = 1$ mPa$\cdot$s, and rain rate $Q=14.5$ cm/hr. []{data-label="Fig3"}](Fig3){width="3.5in"}
We also observe that the formation of the homogeneous wetting front takes time, particularly for low flow rates. In the case of low rainfall rate, this time scale can be measured by simple image analysis. Rain falling unto the soil surface allows the wetting front to expand downward in a homogeneous fashion, while moving at a certain velocity, $v_{wet}$. Eventually at a certain time, $t_C$, an instability occurs at the interface with the development of a finger. Fingers can appear either successively or simultaneously depending on the experimental conditions. The parameter $t_C$ refers to the time of formation of the channels and is the second main kinetics observation. This parameter will be discussed further in the next section. We will mainly focus on the time it takes for the first set of fingers to appear though data will be presented for the succeeding fingers.
Finally, once a finger is formed, the finger propagates deep in the soil until it reaches the bottom of the cell, where water drains. The propagation of the finger exhibits a certain velocity, $v$, making it the third kinetics observation. The plot of the velocities of the water channels as a function of time is seen in Fig. $4$b (at constant $\mu$, $U_T$, $Q$, but varying $D$) and Fig. $4$c (at constant $U_T$, $D$ but varying $\mu$, $Q$). In Fig. $4$b, we compare the finger channel velocities of both the first and second channel that develop. Results suggest that the velocities of the first channel are always greater than the velocities of the second channel.
Fig. $4$c shows the average velocities of the channel fronts as a function of rainfall flow rate, $Q$, and fluid viscosity, $\mu$. These data are taken using hydrophilic beads at constant bead diameter, $D$ = $1$ mm, and at constant $U_T$. In this figure, there is a clear trend pertaining to channel front (or finger) velocities with respect to rainfall flow rate and fluid viscosity. First, at constant fluid viscosity, $\mu$, the channel finger velocities increase with flow rate. The larger volume of fluid entering the soil results to faster propagation of channels downward. Second, at constant flow rate, $Q$, channel fingers in less viscous fluids ($1$ mPa$\cdot$s) propagate faster than in more viscous fluids.
![(Color online). (a) Plot of the velocity of the wetting front, $v_{wet}$ as a function of bead diameter, $D$, at constant $U_T$, $Q$, and $\mu$. The graph shows that water moves faster in larger pores than in smaller ones. (b) Average channel front velocity of the first and second channels as function of bead diameter, $D$, at constant $U_T$, $Q$, and $\mu$. (c) Experimental data on the average channel front velocity, $v$, as a function of flow rate $Q$ also for different fluid viscosities, $\mu$, at constant $D$ and $U_T$.[]{data-label="Fig4"}](Fig4){width="3in"}
Now that we have presented a general description of the experimental observations, we look into them more closely in the next section. Among these observations, we first look into the onset of finger instabilities when the wetting front has fully formed. These instabilities serve as precursors for the formation of fluid channels.
Time of formation of water channels
-----------------------------------
In the infiltration process, the onset of the instability in the homogeneous wetting front corresponds to the transition towards channel formation. Capillary forces dominate the formation of the wetting front. As the front develops inside the medium, capillary forces stabilize the interface while gravity has a destabilizing effect. At the onset of the instability, certain areas of the homogeneous front develop relatively faster than others. Many of these proto-fingers [@Cho02] develop in the front but only one or a few mature and grow into a full water channel. Once a finger fully grows into a channel, the other proto-fingers cease to develop.
During the growth of the instability, the entire wetting front still continues to move in the direction of gravity as water is still continuously supplied at the surface. When the finger instability grows into a water channel, only then will the front plane stop growing. This is because the water channel serves as a preferential path for water drainage as it provides an outlet for water.
Expectedly for different bead sizes, the time of appearance of the water channel also varies as water flow through a porous medium is limited by the size of the pore. It takes a longer time for water to flow through smaller pores than it takes through larger pores. Thus, it takes a longer time before the water channel appears. Fig. $5$a shows different time scales observed in our experiments. In this graph, we consequently plot the time of appearance of the first and second channel, $t_{C_1}$ and $t_{C_2}$, respectively. Using water ($\mu$ = $1$ mPa$\cdot$s) and also at constant $Q$, $U_T$, we consistently observe two channels regardless of $D$. There is also clearly a trend for the formation of the first water channel as a function of particle size. However, the second channel does not exhibit such a clear trend. We also plot in the same graph the quantity $t_h$, which is the characteristic time when the front becomes homogeneous. This is due to the experimental design, in which the front during the first few seconds of rain is not homogeneous as previously shown in Fig. $3$.
In the infiltration of more viscous fluids using constant $D$ and $U_T$ but at varying $Q$, the time of appearance of the formation of water channels seems to generally decrease with increasing flow rate as shown in Fig. $5$b. This physically means that as more volume of water enters the medium at high flow rates, water immediately requires a drainage outlet and thus channels form rather quickly. The rightmost point in this graph, however, corresponds to a more viscous fluid ($8$ mPa$\cdot$s) infiltrating at extremely high flow rate. Experiments show that instead of forming distinct water channels having widths considerably less than the length of the $2$D cell, a massive front is generated that covers the entire length of the cell.
With these observations, we use a proposed model from literature to explain the physics of the phenomenon.
![(Color online). (a) Plot of the time of formation of the first channel, $t_{C_1}$, and second channel, $t_{C_2}$, as well as the time it takes for the front to become homogeneous, $t_h$, all as a function of bead size diameter, $D$, at constant $Q$, $\mu$, and $U_T$. A clear trend exists during the formation of the first channel, meaning water channels form much later in smaller particles where liquid flow is much slower. However, the formation of the second channel seems to be conditioned not just by particle size but also by other parameters, which are yet to be fully determined. A third quantity $t_h$ is the time when the front becomes homogeneous. Owing to the unique design of the rain source, it takes time for neighboring droplet impact sites to coalesce and form a continuous front. (b) Experimental data of average time of formation of the first water channel, $t_{C_1}$, as a function of flow rate $Q$ for different fluid viscosities, $\mu$, but at constant particle size diameter $D$ and droplet impinging speed, $U_T$. There appears to be a decreasing trend with respect to flow rate $Q$ initially at low flow rates but this trend slowly increases at higher flow rates and higher viscosities. []{data-label="Fig5"}](Fig5){width="3.3in"}
Model Discussion
================
The linear stability approach has been used in numerous studies. Through experimental results, it became more apparent that infiltration is a form of immiscible fluid displacement between a wetting phase (liquid) and a non-wetting phase (air). To study such a phenomenon, Saffman and Taylor [@Saffman58] performed one of the pioneering approaches on the subject using experiments in Hele-Shaw cells filled with two fluids of different viscosities using the fact that flows in porous media and in Hele-Shaw cells are formally analogous. The different properties of both fluids result to perturbations occurring in the interface. These perturbations develop into instabilities. Crucial to the analysis of the formation of the instability is the definition of the pressure at the interface of these two fluids. Saffman and Taylor [@Saffman58] notes that a sharp interface is nonexistent but nevertheless assumed that there is no pressure jump across the interface since the characteristic width of the perturbations in the interface is smaller than the length scale of the motion. Thus, for Saffman and Taylor [@Saffman58], pressure is continuous. This results to an equation where any perturbation, whether large or small, can grow into a water channel. This contradicts our observation where the finger size is clearly defined from a characteristic perturbation that develops the fastest. Chuoke $et$ $al$ [@Chuoke59] incorporated this limitation in the modification of the original analysis of Saffman and Taylor. In their assessment, the pressure is in fact not continuous and the discontinuities are defined by a Young-Laplace relationship. Hence, the interfacial pressure jump was described by an effective macroscopic surface tension. However, it is often difficult to determine exactly the effective macroscopic surface tension and thus Parlange and Hill [@Parlange76] later argues that this might only be valid for fluid displacements in parallel plates and not for porous media such as soil. The analysis of Saffman and Taylor [@Saffman58] and Chuoke $et$ $al$ [@Chuoke59] were performed for a less viscous fluid driving a more viscous one. While the opposite of the infiltration process described in this paper where the more viscous fluid, water, displaces the less viscous fluid, air, its principles are certainly analogous to water infiltration in a dry porous medium. Nevertheless, Parlange and Hill [@Parlange76] further proposed another approach for calculating channel width taking into account the influence of soil-water diffusivity when the curved front propagates. While using the same basic principles of linear stability analysis initially described by Saffman and Taylor [@Saffman58], Parlange and Hill [@Parlange76] considered the front as a discontinuity and assumed that if $u_1$ is the velocity of the relatively flatter front, then the velocity of the curved front, $u_2$, is decelerated proportionally to its curvature, ($r_1^{-1}$ + $r_2^{-1}$) according to Eq. 1, where $r_1$ and $r_2$ are the front$^{\prime}$s two principal radii of curvature:
$$u_2 = u_1 - \xi \left( \frac{1}{r_1}+\frac{1}{r_2} \right),
\label{Eq1}$$
where $\xi$ is a function describing soil properties. In other words, Parlange and Hill [@Parlange76] described the interfacial pressure as a function of front velocity.
Assuming that the fluid is incompressible and the porosity of the granular material is uniform, the velocity potential satisfies Laplace$^{\prime}$s equation, ${\nabla^2}{\phi}$ . Darcy$^{\prime}$s law is then used to describe the velocity of the front in the $z$ direction, where $z$ is pointing downward:
$$q_z = - \frac{\kappa_s}{ \left(S_s - S_0 \right)} \nabla \phi,
\label{Eq2}$$
where $\kappa_s$ is the hydraulic conductivity, $S_s$ is the saturated water content, and $S_0$ is the initial water content. The hydraulic conductivity $\kappa_s$ measures the ease in which a fluid flows through pore spaces [@Hillel03]. As $\kappa_s$ will appear in succeeding equations, it is worthwhile to note its definition [@Verneuil11].
$$K = \frac{\rho g K_0 D^2}{\mu},
\label{Eq3}$$
where $\rho$ is the fluid density (water), $\mu$ is the dynamic viscosity, $g$ is the acceleration due to gravity, $D$ is the particle diameter, and $K_0$ is the intrinsic permeability. For a random close packing of spheres having porosity, $\epsilon$ = $0.36 - 0.40$, $K_0$ can be determined using the Karman-Cozeny equation [@Verneuil11]. From this approach, $K_0$ = $6.3$x$10^{-4}$. From Eq. 3, conductivity is proportional to the square of the particle size, so we expect that water infiltration proceeds extremely faster in larger bead sizes as indeed observed from our experiments.
From an initial condition of $z = 0$, taking the derivative of the front position with time results to the velocity of the curved interface (see Parlange and Hill [@Parlange76] for more details on the linear stability analysis): $$u_2 = u_1 + {a}{\lambda^2}{\xi} \exp \left({i}{\lambda}{y}+{\omega}{t}\right),
\label{Eq4}$$ where $a$ is the amplitude, $\lambda$ is the wavelength, $\omega$ is the growth rate of the instability, and $\xi$ is a function describing soil properties defined as:
$$\xi = \int_{S_0}^{S_s}{ \frac{D_f}{S_s - S_0}} \mathrm{d}\theta,
\label{Eq5}$$
where $D_f$ is hydraulic diffusivity, which varies with water content, $S$, in this equation. The hydraulic diffusivity is defined as the ratio of the flux to the soil-water content gradient [@Hillel03]. It is to note that water movement in soil is not actually described as diffusion, in the strictest sense, but of mass flow or convection [@Hillel03], although the term diffusivity has been used for historical reasons.
Nevertheless, the solution to the Laplace equation such that $\omega$ $>$ $0$, gives:
$$\omega = \lambda \left( \frac{\kappa_s - u_1\left( S_s - S_0 \right) }{ S_s - S_0 } \right) - {\xi}{\lambda^2},
\label{Eq6}$$
If pressure is continuous across the front, Parlange and Hill [@Parlange76] notes that the instability that grows the fastest and results to a channel satisfies d$\omega$/d$\lambda = 0$, where $\lambda$ is given by:
$$\lambda = {\frac{1}{2}}\left( \frac{\kappa_s - u_1(S_s - S_0)}{\int_{S_0}^{S_s}{D_f}\mathrm{d}\theta}\right),
\label{Eq7}$$
Using substitution of Eq. 7 to Eq. 6, we obtain the growth rate of the unstable wavelength, $\omega$ = $\xi$$\lambda^2$, where $\lambda$ is related to the equation for determining finger width or diameter, $d$. Parlange and Hill [@Parlange76] argues that the finger width is roughly of this dimension, $d$ = $\pi$/$\lambda$, and the soil diffusivity can be expressed in terms of soil sorptivity, written as:
$$s_w^2 = {2\left(S_s - S_0\right)}{\int_{S_0}^{S_s}{D_f}\mathrm{d}\theta}
\label{Eq8}$$
Sorptivity is the measure of the capacity of a medium to absorb or desorb liquid through capillary forces [@Philip69]. Culligan $et$ $al$ [@Culligan05] states that the sorptivity depends on the properties of both the fluid and the porous material. Using scaling analysis, Culligan $et$ $al$ [@Culligan05] used experiments in real sandy soil to arrive at the following relationship for sorptivity:
$$s_w = {s^*} \left( {\frac{\epsilon l^* \sigma cos\theta^*}{\mu S_{av}^{c - 1}}}\right)^{1/2},
\label{Eq9}$$
where $s^*$ is the dimensionless intrinsic sorptivity with a value equivalent to $s^* = 0.133$ as experimentally determined for sandy-type soil [@Culligan05], $\epsilon$ is the porosity, $\rho$ is the fluid density, $\mu$ is the dynamic viscosity, $l^*$ is a microscopic characteristic length scale of the medium, $\sigma$ is the surface tension, $\theta^*$ is the effective contact angle, $S_{av}$ is the average saturation of the infiltrating fluid at the inlet of the porous medium, and $c$ is an empirical coefficient determined from the Brooks-Corey pore size distribution index. We can expect $l^*$ to be proportional with the particle size diameter, $D$. However, it is difficult to determine the value of $S_{av}^{c - 1}$; thus, we have assigned $\frac{l^*}{S_{av}^{c - 1}}$ in Eq. 9 to be equivalent to $\beta D$, where $\beta$ is a fitting parameter equal to $\beta$ = $0.015\pm0.002$. The value of $\beta$ is kept constant for all equations where this parameter appears. Thus, Eq. 9 is now simplified into the following equation:
$$s_w = {s^*} \left( \frac{\epsilon \beta D \sigma cos\theta^*}{\mu}\right)^{1/2}.
\label{Eq10}$$
Further substitutions result to the equation for determining channel width, which is written as:
$$d = \frac{\pi s_w^2}{\kappa_s \left(S_s-S_0\right) \left( 1- Q/\kappa_s \right) } .
\label{Eq11}$$
The parameter $S$ is the water content defined as the ratio of the volume of water in the soil and the total volume of the soil. The subscripts $s$ and $0$ respectively represent the saturated state and initial state of the soil. Since these experiments have been performed from an initially dry and random close-packed glass beads, $S_0$ = $0$ and $S_s$ = $\epsilon$, where $\epsilon$ is the porosity, representing the maximum amount of water that can be contained within the pore spaces. Eq. 11 has shown good agreement with experimental results obtained from sand [@Glass91; @Glass89; @Selker92].
The measure of the capillary forces is manifested in the surface tension factor in the sorptivity. As capillary forces are increased, so does the sorptivity. This consequently leads to an increase in channel width or size. Capillary forces stabilize the wetting front. As the instability develops, the characteristic size of the perturbations that can develop also increases with increasing capillary forces. This is the fundamental reason why soil with smaller bead diameters shows larger channel widths at constant viscosity and flow rate.
We can also use the same analyses as a starting point to calculate the characteristic time it takes for a wavelength to become unstable. The linear stability analysis also provides the period, $\tau$, which is given by $\tau$ = $2\pi$/$\omega$, where $\omega$ = $\xi$$\lambda^2$. We can calculate for $\lambda$ via substitution of Eq. 5 to Eq. 8 to yield the following relationship:
$$\xi = \frac{s_w^2}{{2\left(S_s-S_0\right)}^2}.
\label{Eq12}$$
Furthermore, $\tau$ can be further simplified into:
$$\tau = \frac{4 \epsilon^2 d^2}{s_w^2 \pi},
\label{Eq13}$$
where ($S_s$ - $S_0$) $\approx$ $\epsilon$ and $\epsilon$ is porosity, $d$ is the channel finger width (Eq. 11) and $s_w$ is sorptivity (Eq. 10). This time scale reflects the time of appearance of the first channel that develops from the instability of the homogeneous wetting front. We put focus on the formation of the first water channel because experimentally a clear trend is observed with respect to bead size as shown in Fig. $5$a. Combining Eq. 3, 10, 11, and 13 gives the following scaling for $\tau$:
$$\tau \sim \frac{\mu}{D^3 \left(1-\frac{\mu Q}{\rho g K_0 D^2}\right)^2} ,
\label{Eq14}$$
Parlange and Hill$^{\prime}$s main contribution has been to describe the beginning of the instability and the morphology of the fingers that develop from such an instability. The model is the basis of subsequent analysis from Glass $et$ $al$ [@Glass89] in describing the channel finger propagation velocity, $v$, to arrive at the following relationship:
$$v = \frac{\kappa_s}{\left(S_s-S_0\right)}f(Q/\kappa_s).
\label{Eq15}$$
Further analysis by Glass $et$ $al$ [@Glass89] and Wang $et$ $al$ [@Wang98] shows that Eq. 15 can be written as follows:
$$v = \frac{\kappa_s}{\epsilon} \left(C + (1 - C) \sqrt{Q/\kappa_s}\right),
\label{Eq16}$$
where $C$ is the projected zero flow velocity for fingers [@Wang98], which is dependent on the dimensionality (whether $2$D or $3$D systems) of the granular system.
Further expansion of Eq. 16, using substitution of the definition of hydraulic conductivity (Eq. 3) yields the following scaling:
$$\mu v \sim (\mu Q)^{1/2}.
\label{Eq17}$$
In Eq. 17, the dependence of the channel finger velocity, $v$, on $\sqrt{Q}$ is demonstrated for all fluid viscosities.
We have so far discussed the evolution of the instability formation. Instability precedes the formation of the water channel and we have shown that the drainage of water channels from an initial wetting front can be described by linear stability analysis. More recently, numerical simulations performed by Cueto-Felgueroso and Juanes [@Cueto-Felgueroso08; @Cueto-Felgueroso09] have also advanced our understanding of the topic. The analysis of Cueto-Felgueroso and Juanes [@Cueto-Felgueroso08; @Cueto-Felgueroso09] proposes a macroscopic phase-field model during unsaturated flow. They also employed linear stability analysis to stress the importance of the role of the water saturation in the growth of the instabilities. In essence, their model introduces a non-linear term to the classical Richards equation to account for the appearance of perturbations. This term is formally related to the surface tension at the interface of the wetting front. From mathematical calculations, they predict that finger width and finger velocity both increase with infiltration rate.
Nevertheless, despite many proposed modifications to existing models, linear stability analysis is enough to describe well the morphology and certain aspects of kinetics of the formation of water channels as will be discussed in the next section.
Model Application and Comparison with Experiments
=================================================
Owing to our quasi-$2$D rainfall set-up built with equally spaced capillary tubes, it takes a certain time $t_h$ for the front to establish homogeneously. We take into account the previously calculated parameter, $t_h$, as a delay during which the front becomes fully homogeneous. The effective time of appearance of the first channel is then given by:
$$t_{C_1} = t_h + \tau,
\label{Eq18}$$
where $\tau$ is calculated from Eq. 14 and $t_h$ still remains to be estimated.
Time scale for the establishment of the homogeneous zone
--------------------------------------------------------
We can compute for $t_h$ since it is mainly a function of the distance between the capillaries, $d_{cap}$, in the rain set-up and flow rate, $Q$, as schematically shown in Fig. 6a.
We let $Q$ be the total flow rate impacting the system. The raindrops will impact on the surface of certain initial volume. Successive impacts will increase the volume of the drop, which will eventually result to coalescence of neighboring droplet impact sites. The total volume then on one impact site underneath a capillary tube is a function of time, $V(t)$ = ${A}{Q}{t}$, where $A$ is the cross-sectional area of the cell (m$^2$). Note that $Q$ has units in velocity (m/s). When the front makes a depth in the medium equal to $d_{cap}$, the time it takes for the front to become sufficiently homogeneous can be calculated from the parameters of the sample cell, where:
$$t_h = \frac{\epsilon d_{cap}}{Q} ,
\label{Eq19}$$
where the spacing between capillary tubes in the rain set-up is set at $d_{cap}$ = $1$ cm, $Q$ is the total flow rate in units of velocity, and $\epsilon$ is the porosity of the medium. The size of the sample cell is also a factor but $l$ x $e$, where $l$ and $e$ are the length and thickness of the sample cell respectively, will just cancel out with the cross-section $A$ in the denominator. Thus, in this equation, the spacing of the capillary tubes in the rain source is an important criterion that influences the front homogenization. This means that if the spacing had been larger, e.g. $d_{cap} > 1$ cm, we can expect the front homogenization to occur much later due to the fact that coalescence of droplet impact sites is less favored when the distance between them increases.
Once the front becomes homogeneous, it propagates in a uniform manner downward until instabilities develop.
{width="6in"}
Time scale for the instability to develop
-----------------------------------------
At constant $U_T$, $Q$, and $\mu$ but at varying $D$, Eq. 18 agrees well with experiments as shown in Fig. 6b. In this figure, $t_{C_1}$ decreases with increasing $D$. In addition, as a function of flow rate $Q$ and for three different viscosity values, $\mu$, Eq. 18 also shows decent agreement with experiments as depicted in Fig. 6c. This figure shows interesting behaviors. At low viscosities (water), $t_{C_1}$ decreases with $Q$, meaning water channels will form faster at higher flow rates. However, as the viscosity increases the time of formation of water channels initially decreases at low $Q$ but then slowly increases as $Q$ further increases. This becomes even more prominent at higher viscosity values (8 mPa$\cdot$s), where larger flow rates increases the time it takes for water channels to form. Based on experimental observations at high viscosity and high flow rate, where $Q$ is close to the value of $\kappa_s$, the fluid initially infiltrates as one massive front, so water channels form at a later time.
The decent agreement between our experimental data and the theoretical results suggest that taking into account an additional time delay for the formation of the homogeneous front is necessary to obtain a more accurate description of the process, especially at constant $U_T$, $Q$, and $\mu$ as shown in Fig. 6b. In Fig. 6c however, the model, which is based on the linear stability analysis developed by Parlange and Hill, seems only to capture the time scale for the destabilization of the homogeneous front for low flow rates. The model appears to be far less accurate when $Q$ approaches $\kappa_s$, which happens at conditions of higher $\mu$ and higher $Q$ values. It would then be worth testing these experimental data using other models in literature, in particular the recent model developed by Cueto-Felgueroso and Juanes [@Cueto09] to check whether the predictions are better. This is the subject of future investigations.
Channel finger velocity
-----------------------
We can derive the velocities of the channel from Eq. 16. Similarly, we apply this equation to two different cases, first at constant $Q$, $\mu$, and $U_T$ but at varying $D$ and second at constant $D$ and $U_T$ but at varying $Q$ and $\mu$. The results of the first case are shown in Fig. 7a. This figure shows that the velocity of the water channel is also dependent on the particle size. Larger particles have larger pores and thus have greater water flow velocity, allowing water to easily flow down.
The results of the second case, on the other hand, are shown in Fig. 7b. In both cases in applying Eq. 16, we use $C = 0.2$.
In Fig. 7a, higher channel finger velocities are predicted in larger bead diameters due to the accompanying larger pore size. In Fig. 7b, the dependence of the channel finger velocity, $v$, on $\sqrt{Q}$ is demonstrated for all fluid viscosities as predicted in Eq. 17.
From these results, the model derived from linear stability analysis fits reasonably well with the experimental data.
{width="6in"}
Number of channels
------------------
Now that we have shown the model fits for channel width, characteristic time of channel formation, and channel finger velocity, in this section we apply the same model to predict the number of channels, $N$. At constant $U_T$, $Q$ and $\mu$ but varying $D$, we observe two channels plotted in Fig. 7c and the separation distance between them roughly remains constant as shown in our companion paper [@Wei13].
At constant $D$ and $U_T$ but at varying $Q$ and $\mu$, shown in Fig. 7d, the number of channels observed for less viscous fluids ($1$ mPa$\cdot$s) such as water generally increases with $Q$. The value of $N$ in more viscous fluids, however, decreases with increasing $Q$. At low $Q$, there are less number of channels in less viscous fluids but already more channels in more viscous fluids. This result already provides a clue that the fluid viscosity modifies the temporal dynamics of the instability formation.
Similar to aforementioned approaches, we rescale the experimental data by using the equation for channel width obtained from linear stability analysis as a starting point.
By mass conservation, the total volumetric flow rate, $QA$, is equal to the volumetric flow rate in each finger multiplied by the number of fingers, $N$. This relationship can be written as:
$$Q A = N q_f,
\label{Eq20}$$
where $Q$ is the total flux into the granular system (m/s), $A$ is the total cross-sectional area, $q_f$ is the flux through each finger, which is a function of channel velocity, $v$ (Eq. 15), and channel finger width, $d$ (Eq. 11). Further expansion of Eq. 20 approximately results to:
$$N \sim \frac{A}{D^2}\left( \frac{\mu Q}{\rho g K_0 D^2} \right)^{\frac{1}{2}}\left(1-\frac{\mu Q}{\rho g K_0 D^2} \right)^2
\label{Eq21}$$
In Eq. 21, we can see that $N$ exhibits two behaviors as a function of fluid viscosity and flow rate at constant bead diameter. It increases with $\sqrt{\mu Q}$ but decreases with $(1 -\frac{ \mu Q}{\rho g K_0 D^2})^2$ . The curve is presented in Fig. 7d.
In this figure, the number of channels initially increases at low viscosity fluids ($1$ mPa$\cdot$s) and low flow rates. However, at higher fluid viscosities ($8$ mPa$\cdot$s) and higher flow rates, $N$ reduces in value not because there are no individual channels that form but because the fluid eventually infiltrates as one massive stable front. Additional points are indeed necessary to ascertain the decrease of the curve especially within the range of $3$x$10^{-6}$ $<$ $\mu Q$ $<$ $6$x$10^{-6}$. It is however experimentally difficult due to the limitations of the size of the rain source and the available capillary tubes.
At constant bead size, increasing the viscosity reduces the hydraulic conductivity of that particular fluid; thus, the flow rate, $Q$, slowly approaches the value of $\kappa_s$ ($Q \rightarrow \kappa_s$). When this happens, the number of channels decreases as predicted by the equation. Physically this implies that for a certain total cross-sectional area of the cell, we can predict the number of channels that can appear during the infiltration of a fluid within that area. And that there is a maximum number of channels that can form within the limits of the cross-sectional area as a function of rainfall flow rate and fluid viscosity. For example, if rain impacts a cross-sectional area of $20$ cm$^2$, then we will obtain the maximum number of channels when $\mu Q \approx 1.5$x$10^{-6}$ $\mu$Pa$\cdot$m. This means the maximum is achieved either using low viscosity fluids but infiltrating at high flow rates or using higher viscosity fluids infiltrating at lower flow rates.
Conclusion
==========
Preferential water paths are drainage outlets. Once they form, they effectively reduce the water content around the root zone. Studies have shown how this is affected by properties of the granular material, such as pore size. But through extensive experimental results, we have also explored the influence of the fluid properties as well on the formation of water channels. These properties include the viscosity of the fluid source and its flow rate, both of which have not been widely investigated.
The results on kinetics presented here are well described by a model developed by Parlange and Hill [@Parlange76], which is an extension of the model developed by Saffman and Taylor [@Saffman58]. But in the application of this model, we nevertheless also take into account an additional parameter that represents the characteristic time of formation of the wetting front. The wetting front becomes homogeneous when droplet volumes at neighboring impact sites coalesce.
While perhaps this does not fully represent actual rainfall since raindrops impact randomly, it still offers an understanding of how the manner in which a fluid is injected uniformly unto the surface affects water distribution in the soil. Fluid properties and spacing between droplets influence the aggregation of droplet impact sites thus providing information that the introduction of the fluid to the porous medium is also crucial to the establishment of a wetting front.
Moreover, results show that instabilities at the wetting front and thus formation of water channels initially decreases with flow rate, particularly for low flow rates. However, depending on the viscosity, the behavior may change at larger flow rates. At low viscosities, water channels form quicker at larger flow rates. But at higher viscosities, the time of formation of channels gradually increases at higher flow rates. While our results do not yet fully explain the exact dynamics of the instability, however it does demonstrate that, within a given cross-sectional area, the number of channels that form is a function of the fluid$^\prime$s viscosity and flow rate. In terms of velocity, water channels expectedly propagate faster in larger pore sizes at constant viscosity and flow rate. In addition, water channel velocities increase with flow rate at constant particle diameter and constant viscosity. But at constant particle diameter and constant flow rate, low viscosity fluids propagate faster than larger viscosity fluids.
We believe continuous investigations primarily focusing on the finger instability dynamics at the wetting front will further help bring to light certain aspects that remain unclear such as how the instability develops and in which particular part of the front does it develop. These experimental results presented in this paper could also be used to test existing models particularly to confirm data at conditions when the flow rate value approaches the value of the hydraulic conductivity. This typically happens when using high flow rates and high viscosity fluids. It is therefore our interest to apply these experimental results to other models of unstable multiphase flow proposed in literature with more recent ones such as the model of Cueto-Felgueroso and Juanes [@Cueto09] for example. Nevertheless, apart from contributing to the advancement of our understanding of the subject, these fundamental results could also provide insights into developing techniques and applications to better control the drainage of water via preferential water channels.
We thank Jean-Christophe Castaing, Zhiyun Chen, and Larry Hough for helpful discussions. This work is financially supported by Solvay, CNRS, and by the National Science Foundation through grants MRSEC / DMR-1120901 (Y.W. and D.J.D.) and DMR-1305199 (D.J.D).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Hirotaka Hayashi,'
- 'Sung-Soo Kim,'
- 'Kimyeong Lee,'
- and Futoshi Yagi
title: '6d SCFTs, 5d Dualities and Tao Web Diagrams'
---
We thank Kang-Sin Choi, Hee-Cheol Kim, Seok Kim, Xiao Liu, Soo-Jong Rey, Ashoke Sen and Piljin Yi for useful discussions. We especially thank Masato Taki for early collaboration and discussions. We are thankful to the Workshop on Geometric Correspondences of Gauge Theories at SISSA. S.K. is grateful to “APCTP focus program on holography and its application” (2015) where a part of work is done. The work of H.H. is supported by the grant FPA2012-32828 from the MINECO, the REA grant agreement PCIG10-GA-2011-304023 from the People Programme of FP7 (Marie Curie Action), the ERC Advanced Grant SPLE under contract ERC-2012-ADG-20120216-320421 and the grant SEV-2012-0249 of the “Centro de Excelencia Severo Ochoa” Programme. The work of HH is also supported in part by Perimeter Institute for Theoretical Physics where a part of the work was done. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through Ministry of Economic Development & Innovation. The work of K.L. is supported in part by the National Research Foundation of Korea (NRF) Grants No. 2006-0093850.
[10]{}
J. J. Heckman, D. R. Morrison, and C. Vafa, [*[On the Classification of 6D SCFTs and Generalized ADE Orbifolds]{}*]{}, [*JHEP*]{} [**1405**]{} (2014) 028, \[[[arXiv:1312.5746]{}](http://arxiv.org/abs/1312.5746)\].
D. Gaiotto and A. Tomasiello, [*[Holography for (1,0) theories in six dimensions]{}*]{}, [*JHEP*]{} [**1412**]{} (2014) 003, \[[[arXiv:1404.0711]{}](http://arxiv.org/abs/1404.0711)\].
M. Del Zotto, J. J. Heckman, A. Tomasiello, and C. Vafa, [*[6d Conformal Matter]{}*]{}, [*JHEP*]{} [**1502**]{} (2015) 054, \[[[arXiv:1407.6359]{}](http://arxiv.org/abs/1407.6359)\].
J. J. Heckman, [*[More on the Matter of 6D SCFTs]{}*]{}, [[arXiv:1408.0006]{}](http://arxiv.org/abs/1408.0006).
J. J. Heckman, D. R. Morrison, T. Rudelius, and C. Vafa, [*[Atomic Classification of 6D SCFTs]{}*]{}, [[ arXiv:1502.05405]{}](http://arxiv.org/abs/1502.05405).
L. Bhardwaj, [*[Classification of 6d N=(1,0) gauge theories]{}*]{}, [[arXiv:1502.06594]{}](http://arxiv.org/abs/1502.06594).
B. Haghighat, A. Iqbal, C. Kozcaz, G. Lockhart, and C. Vafa, [*[M-Strings]{}*]{}, [*Commun. Math. Phys.*]{} [**334**]{} (2015), no. 2 779–842, \[[[arXiv:1305.6322]{}](http://arxiv.org/abs/1305.6322)\].
B. Haghighat, C. Kozcaz, G. Lockhart, and C. Vafa, [*[Orbifolds of M-strings]{}*]{}, [*Phys. Rev.*]{} [**D89**]{} (2014), no. 4 046003, \[[[arXiv:1310.1185]{}](http://arxiv.org/abs/1310.1185)\].
B. Haghighat, G. Lockhart, and C. Vafa, [*[Fusing E-strings to heterotic strings: E+E $\rightarrow$ H]{}*]{}, [*Phys. Rev.*]{} [**D90**]{} (2014), no. 12 126012, \[[[arXiv:1406.0850]{}](http://arxiv.org/abs/1406.0850)\].
J. Kim, S. Kim, K. Lee, J. Park, and C. Vafa, [*[Elliptic Genus of E-strings]{}*]{}, [[arXiv:1411.2324]{}](http://arxiv.org/abs/1411.2324).
B. Haghighat, A. Klemm, G. Lockhart, and C. Vafa, [*[Strings of Minimal 6d SCFTs]{}*]{}, [*Fortsch.Phys.*]{} [**63**]{} (2015) 294–322, \[[[arXiv:1412.3152]{}](http://arxiv.org/abs/1412.3152)\].
A. Gadde, B. Haghighat, J. Kim, S. Kim, G. Lockhart, and C. Vafa, [*[6d String Chains]{}*]{}, [[ arXiv:1504.04614]{}](http://arxiv.org/abs/1504.04614).
B. Haghighat, S. Murthy, C. Vafa, and S. Vandoren, [*[F-Theory, Spinning Black Holes and Multi-string Branches]{}*]{}, [[arXiv:1509.00455]{}](http://arxiv.org/abs/1509.00455).
K. Ohmori, H. Shimizu, and Y. Tachikawa, [*[Anomaly polynomial of E-string theories]{}*]{}, [*JHEP*]{} [**1408**]{} (2014) 002, \[[[arXiv:1404.3887]{}](http://arxiv.org/abs/1404.3887)\].
K. Ohmori, H. Shimizu, Y. Tachikawa, and K. Yonekura, [*[Anomaly polynomial of general 6d SCFTs]{}*]{}, [*PTEP*]{} [**2014**]{} (2014), no. 10 103B07, \[[[arXiv:1408.5572]{}](http://arxiv.org/abs/1408.5572)\].
K. Intriligator, [*[6d, $ \mathcal{N}=\left(1,\;0\right) $ Coulomb branch anomaly matching]{}*]{}, [*JHEP*]{} [**1410**]{} (2014) 162, \[[[arXiv:1408.6745]{}](http://arxiv.org/abs/1408.6745)\].
J. J. Heckman, D. R. Morrison, T. Rudelius, and C. Vafa, [*[Geometry of 6D RG Flows]{}*]{}, [[arXiv:1505.00009]{}](http://arxiv.org/abs/1505.00009).
J. J. Heckman and T. Rudelius, [*[Evidence for C-theorems in 6D SCFTs]{}*]{}, [[arXiv:1506.06753]{}](http://arxiv.org/abs/1506.06753).
C. Cordova, T. T. Dumitrescu, and K. Intriligator, [*[Anomalies, Renormalization Group Flows, and the a-Theorem in Six-Dimensional (1,0) Theories]{}*]{}, [[arXiv:1506.03807]{}](http://arxiv.org/abs/1506.03807).
N. Bobev, M. Bullimore, and H.-C. Kim, [*[Supersymmetric Casimir Energy and the Anomaly Polynomial]{}*]{}, [[ arXiv:1507.08553]{}](http://arxiv.org/abs/1507.08553).
K. Ohmori, H. Shimizu, Y. Tachikawa, and K. Yonekura, [*[6d $\mathcal{N}=(1,0)$ theories on $T^2$ and class S theories: part I]{}*]{}, [[arXiv:1503.06217]{}](http://arxiv.org/abs/1503.06217).
M. Del Zotto, C. Vafa, and D. Xie, [*[Geometric Engineering, Mirror Symmetry and 6d (1,0) $\rightarrow$ 4d, N=2]{}*]{}, [[arXiv:1504.08348]{}](http://arxiv.org/abs/1504.08348).
K. Ohmori, H. Shimizu, Y. Tachikawa, and K. Yonekura, [*[6d $\mathcal{N}{=}(1,0)$ theories on $S^1/T^2$ and class S theories: part II]{}*]{}, [[arXiv:1508.00915]{}](http://arxiv.org/abs/1508.00915).
D. Gaiotto and S. S. Razamat, [*[$ \mathcal{N}=1 $ theories of class $
{\mathcal{S}}_k $]{}*]{}, [*JHEP*]{} [**07**]{} (2015) 073, \[[[arXiv:1503.05159]{}](http://arxiv.org/abs/1503.05159)\].
S. Franco, H. Hayashi, and A. Uranga, [*[Charting Class $\mathcal S_k$ Territory]{}*]{}, [*Phys. Rev.*]{} [**D92**]{} (2015), no. 4 045004, \[[[arXiv:1504.05988]{}](http://arxiv.org/abs/1504.05988)\].
A. Hanany and K. Maruyoshi, [*[Chiral theories of class S]{}*]{}, [[arXiv:1505.05053]{}](http://arxiv.org/abs/1505.05053).
H. Hayashi, S.-S. Kim, K. Lee, M. Taki, and F. Yagi, [*[A new 5d description of 6d D-type minimal conformal matter]{}*]{}, [*JHEP*]{} [**08**]{} (2015) 097, \[[[arXiv:1505.04439]{}](http://arxiv.org/abs/1505.04439)\].
I. Brunner and A. Karch, [*[Branes at orbifolds versus Hanany Witten in six-dimensions]{}*]{}, [*JHEP*]{} [**9803**]{} (1998) 003, \[[[hep-th/9712143]{}](http://arxiv.org/abs/hep-th/9712143)\].
A. Hanany and A. Zaffaroni, [*[Branes and six-dimensional supersymmetric theories]{}*]{}, [*Nucl.Phys.*]{} [**B529**]{} (1998) 180–206, \[[[hep-th/9712145]{}](http://arxiv.org/abs/hep-th/9712145)\].
S.-S. Kim, M. Taki, and F. Yagi, [*[Tao Probing the End of the World]{}*]{}, [*PTEP*]{} [**2015**]{} (2015), no. 8 083B02, \[[[arXiv:1504.03672]{}](http://arxiv.org/abs/1504.03672)\].
K. Yonekura, [*[Instanton operators and symmetry enhancement in 5d supersymmetric quiver gauge theories]{}*]{}, [*JHEP*]{} [**07**]{} (2015) 167, \[[[arXiv:1505.04743]{}](http://arxiv.org/abs/1505.04743)\].
Y. Tachikawa, [*[Instanton operators and symmetry enhancement in 5d supersymmetric gauge theories]{}*]{}, [*PTEP*]{} [**2015**]{} (2015), no. 4 043B06, \[[[arXiv:1501.01031]{}](http://arxiv.org/abs/1501.01031)\].
G. Zafrir, [*[Instanton operators and symmetry enhancement in 5d supersymmetric USp, SO and exceptional gauge theories]{}*]{}, [[arXiv:1503.08136]{}](http://arxiv.org/abs/1503.08136).
D. Gaiotto and H.-C. Kim, [*[Duality walls and defects in 5d N=1 theories]{}*]{}, [[arXiv:1506.03871]{}](http://arxiv.org/abs/1506.03871).
A. Sen, [*[F theory and orientifolds]{}*]{}, [*Nucl.Phys.*]{} [**B475**]{} (1996) 562–578, \[[[hep-th/9605150]{}](http://arxiv.org/abs/hep-th/9605150)\].
G. Zafrir, [*[Brane webs, $5d$ gauge theories and $6d$ $\mathcal{N}$$=(1,0)$ SCFT’s]{}*]{}, [[arXiv:1509.02016]{}](http://arxiv.org/abs/1509.02016).
O. DeWolfe, A. Hanany, A. Iqbal, and E. Katz, [*[Five-branes, seven-branes and five-dimensional E(n) field theories]{}*]{}, [*JHEP*]{} [**9903**]{} (1999) 006, \[[[hep-th/9902179]{}](http://arxiv.org/abs/hep-th/9902179)\].
K. A. Intriligator, D. R. Morrison, and N. Seiberg, [*[Five-dimensional supersymmetric gauge theories and degenerations of Calabi-Yau spaces]{}*]{}, [ *Nucl.Phys.*]{} [**B497**]{} (1997) 56–100, \[[[hep-th/9702198]{}](http://arxiv.org/abs/hep-th/9702198)\].
O. Bergman and G. Zafrir, [*[5d fixed points from brane webs and O7-planes]{}*]{}, [[arXiv:1507.03860]{}](http://arxiv.org/abs/1507.03860).
A. Hanany and E. Witten, [*[Type IIB superstrings, BPS monopoles, and three-dimensional gauge dynamics]{}*]{}, [*Nucl.Phys.*]{} [**B492**]{} (1997) 152–190, \[[[hep-th/9611230]{}](http://arxiv.org/abs/hep-th/9611230)\].
S.-S. Kim and F. Yagi, [*[5d E$_{n}$ Seiberg-Witten curve via toric-like diagram]{}*]{}, [*JHEP*]{} [**06**]{} (2015) 082, \[[[arXiv:1411.7903]{}](http://arxiv.org/abs/1411.7903)\].
F. Benini, S. Benvenuti, and Y. Tachikawa, [*[Webs of five-branes and N=2 superconformal field theories]{}*]{}, [*JHEP*]{} [**0909**]{} (2009) 052, \[[[arXiv:0906.0359]{}](http://arxiv.org/abs/0906.0359)\].
H. Hayashi, H.-C. Kim, and T. Nishinaka, [*[Topological strings and 5d $T_N$ partition functions]{}*]{}, [*JHEP*]{} [**1406**]{} (2014) 014, \[[[arXiv:1310.3854]{}](http://arxiv.org/abs/1310.3854)\].
M. Aganagic, N. Haouzi, and S. Shakirov, [*[$A_n$-Triality]{}*]{}, [[arXiv:1403.3657]{}](http://arxiv.org/abs/1403.3657).
O. Bergman and G. Zafrir, [*[Lifting 4d dualities to 5d]{}*]{}, [*JHEP*]{} [ **1504**]{} (2015) 141, \[[[ arXiv:1410.2806]{}](http://arxiv.org/abs/1410.2806)\].
H. Hayashi, Y. Tachikawa, and K. Yonekura, [*[Mass-deformed T$_{N}$ as a linear quiver]{}*]{}, [*JHEP*]{} [**1502**]{} (2015) 089, \[[[arXiv:1410.6868]{}](http://arxiv.org/abs/1410.6868)\].
N. Seiberg, [*[Five-dimensional SUSY field theories, nontrivial fixed points and string dynamics]{}*]{}, [*Phys.Lett.*]{} [**B388**]{} (1996) 753–760, \[[[hep-th/9608111]{}](http://arxiv.org/abs/hep-th/9608111)\].
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We present a DFT study utilizing the Hubbard U correction to probe structural and magnetic disorder in $\mathrm{NaO_{2}}$, primary discharge product of Na-O$_2$ batteries. We show that $\mathrm{NaO_{2}}$ exhibits a large degree of rotational and magnetic disorder; a 3-body Ising Model is necessary to capture the subtle interplay of this disorder. MC simulations demonstrate that energetically favorable, FM phases near room temperature consist of alternating bands of orthogonally-oriented $\mathrm{O_{2}}$ dimers. We find that bulk structures are insulating, with a subset of FM structures showing a moderate gap ($<2$ eV) in one spin channel.'
author:
- Oleg Sapunkov
- Vikram Pande
- Abhishek Khetan
- Venkatasubramanian Viswanathan
bibliography:
- 'PD\_PB\_Bibl.bib'
title: 'The role of disorder in NaO$_2$ and its implications for Na-O$_2$ batteries'
---
Electrification of road transport recently led to increased interest in high-energy-density rechargeable batteries, in particular, metal-air batteries. [@luntz2014nonaqueous; @sapunkov2015quantifying] Li-O$_2$ battery, one of the most promising, owing to its high energy density, suffers from electrolyte[@burke2015enhancing; @abraham1996polymer; @abraham1990li+; @bryantsev2011predicting; @bryantsev2013identification; @bryantsev2012predicting; @walker2013rechargeable; @elia2014advanced; @khetan2014solvent; @burke2015enhancing] and electrode[@ottakam2012carbon; @thotiyl2013stable; @mccloskey2012twin; @mccloskey2012limitations] instability, as well as limited discharge capacity.[@viswanathan2011electrical; @aetukuri2015solvating; @chen2013charging; @bruce2012li; @johnson2014role] Hartmann *et al.* recently demonstrated a rechargeable Na-O$_2$ battery, with sodium superoxide (NaO$_2$) as the primary observed discharge product, which showed superior cycle life to state-of-the-art Li-O$_2$ batteries.[@hartmann2013comprehensive; @hartmann2013rechargeable] Subsequently, there have been numerous theoretical and experimental efforts to understand the fundamental mechanisms that govern electrochemistry in Na-O$_2$ batteries.[@kang2014nanoscale; @yang2015intrinsic; @hartmann2013rechargeable; @das2014sodium; @hartmann2013comprehensive; @ellis2012sodium; @kundu2015emerging; @kwak2015nanoconfinement; @ortiz2015rate]
Accurate description of the Na-O$_2$ battery reaction mechanisms[@kim2013sodium; @kang2014nanoscale] requires detailed understanding of electronic structure throughout the phase space of sodium oxides. The roles of nucleation,[@krishnamurthy2016universality; @ortiz2015rate; @hartmann2015discharge] nanoscale stabilization,[@kang2014nanoscale] and surface energetics of various sodium-oxygen compounds have been examined through a combination of density functional theory (DFT) calculations and electrochemical measurements. This has led to an improved understanding of Na-O$_2$ battery reactions, which, as proposed, constitute a combination of surface and solution mechanisms of the NaO$_2$ discharge product.[@adelhelm2015lithium; @sapunkov2015quantifying] In certain cases, it should be noted, Na$_2$O$_2$ has been reported as the discharge product,[@jian2014high; @araujo2015unveiling] and selectivity between NaO$_2$ and Na$_2$O$_2$ is not yet fully characterized.
Alkali superoxides are known to be highly disordered materials, both in geometric structure and magnetic ordering.[@smith1966antiferromagnetism; @ren2013low] Presently, there is very limited understanding of geometric and magnetic disorder in NaO$_2$, due to challenges in experimentally synthesizing NaO$_2$. It is crucial to map out disorder in room-temperature NaO$_2$, considering its importance in determining electronic structure, surface energetics, and growth properties, relevant for Na-O$_2$ batteries. In this work, we aim to fill the gaps in the current understanding of the electronic structure of NaO$_2$. We perform DFT calculations incorporating the Hubbard U correction and use that to understand the geometric and magnetic disorder in NaO$_2$. We employ an Ising model accounting for magnetic and geometric degrees of freedom and find that both 2-body and 3-body nearest-neighbor interactions, which take into account both the geometric and magnetic disorder of the structure, are critical to accurately describe bulk NaO$_2$. The Ising model is used within a metropolis monte carlo framework to characterize the complete phase space as a function of temperature. In order to determine the effect of disorder on the electronic properties of NaO$_2$, we employ hybrid density functional theory calculations and find that the calculated bandgap is strongly affected by disorder. All investigated structures exhibit a direct bandgap around 4 eV, with the exception of a small subset of ferromagnetic structures, whose direct bandgap is below 2 eV in one spin channel.
The discharge process of the Na-O$_2$ battery involves the $\mathrm{Na^{+}}$ ion coupled electron transfer reaction with O$_2$ at the cathode, which produces NaO$_2$, the primary reported discharge product.[@hartmann2013comprehensive; @hartmann2013rechargeable] $$\mathrm{Na^{+} + O_{2} + e^{-} \rightleftharpoons {NaO_{2}}_{(s)}}$$ A few studies do report the formation of Na$_2$O$_2$,[@jian2014high] which is the thermodynamically stable structure according to the Na-O phase diagram at standard temperature and pressure.[@kang2014nanoscale] The preferential formation of NaO$_2$ as a discharge product in room-temperature Na-O$_2$ batteries, over the thermodynamically stabler Na$_2$O$_2$, still remains a puzzle. The stable phase of bulk NaO$_2$ itself is also temperature-dependent.[@wriedt1987sodium; @carter1953polymorphism; @templeton1950crystal; @kang2014nanoscale] Below 196 K, NaO$_2$ takes the $\mathrm{Pnnm}$ space group. Between 196 and 223 K, NaO$_2$ takes the Pa3 space group. Above 223 K, NaO$_2$ takes the Fm3m space group, the primary relevant structure for room-temperature Na-O$_2$ batteries.
To explore the reaction mechanism of the Na-O$_2$ battery, we need to understand the bulk composition of NaO$_2$ and accurately describe its electronic structure and dependent properties. The oxygen dimer in NaO$_2$ behaves similar to the superoxide anion, O$_{2}^{-}$.[@partridge1992theoretical] Within the molecular orbital picture of ground-state (triplet) O$_2$, the 2 highest occupied orbitals are the $\mathrm{\pi_{2p}^{*}}$ antibonding orbitals, with one electron in each orbital, following Hund’s Rule of Maximum Multiplicity.[@boyd1984quantum] The highest occupied molecular orbital of O$_{2}^{-}$ remains $\mathrm{\pi_{2p}^{*}}$, now occupied by 3 electrons, as illustrated in the appendix. This means that the O$_{2}^{-}$ dimer has $1$ unpaired electron, making O$_{2}^{-}$ magentic. The magnetic ordering of these spins is one of the key contributions to the configurational disorder of bulk NaO$_2$.
We considered bulk NaO$_2$ in the Fm3m space group structure. In this structure, sodium occupies the body centers and corners of the cubic cell, and oxygen dimers occupy face centers and edge centers of the cell, as illustrated in the appendix. We introduce the following naming scheme for these structures to describe both their geometric and magnetic arrangement. The first 4 letters, which can be A, B, C, or D, refer to the 4 possible geometric orientations of oxygen dimers. The next 4 letters, which can be P or N, refer to the net positive or negative magnetic moment of the corresponding dimers. Thus, a structure designated AAAA-PPPP has all dimers mutually parallel to each other (AAAA), and all dimers of identical, positive magnetization (PPPP), providing a ferromagnetic structure. Alternatively, a structure designated ABCD has all dimers mutually orthogonal to each other, while a structure designated PNPN or PPNN is antiferromagnetic. Formation enthalpies were calculated for 19 possible structures to explore the phase space of bulk NaO$_2$.
Self-consistent DFT calculations were performed using the Projector Augmented Wave Method as implemented in GPAW,[@enkovaara2010electronic] with the Revised PBE (RPBE) exchange correlation functional.[@hammer1999improved] To correct electron localization in NaO$_2$, we incorporated the Hubbard U, applied on oxygen $2p$ states in NaO$_2$.[@hubbard1963electron; @anisimov1991band; @himmetoglu2014hubbard; @garcia2012importance; @meredig2010method] Calculations were run with a real-space grid of 0.18 [Å]{}, and 6$\times$6$\times$6 k-point sampling, following the Monkhorst-Pack scheme.[@monkhorst1976special] Fermi-Dirac smearing of 0.01 eV was used to facilitate convergence and broyden-type mixing of electron densities was used in the calculation.[@johnson1988modified; @srivastava1984broyden]
Two schemes were used to calculate formation enthalpy of bulk NaO$_2$. The formation enthalpy of NaO$_2$ is given by: $$\mathrm{{\Delta}H_{F_{NaO_{2}}} = E_{NaO_{2}}^{DFT} - E_{Na}^{Ref} - E_{{O_2}}^{Ref} + \Delta pV}$$ The pressure-volume work term, $\mathrm{\Delta pV}$, can be disregarded, as it is typically about 5 orders of magnitude smaller than internal energy contributions in formation enthalpy calculations.[@aydinol1997ab; @obrovac2014alloy] Internal energies of Na, O$_2$, and NaO$_2$ were calculated using DFT simulations.
It is well known that molecular oxygen is poorly described in DFT.[@droghetti2008predicting] The energy of O$_2$ can be more accurately calculated using water as the reference molecule.[@rossmeisl2005electrolysis] In the first scheme used, reference energy of O$_2$ was computed using the DFT-calculated internal energies of H$_2$O and gaseous H$_2$, as well as the experimental formation enthalpy of H$_2$O: $$\mathrm{E_{O_2}^{Ref} = 2E_{H_{2}O}^{DFT} - 2E_{H_{2}}^{DFT} - {\Delta}H_{H_{2}O}^{Exp}}$$ The Hubbard U correction was applied on the oxygen atoms in both H$_2$O and NaO$_2$.
The second scheme was used to calculate reference energy of bulk Na. It was demonstrated in prior work that formation enthalpies of alkali oxides, peroxides and superoxides are best described when the oxidation state of the metal in the reference compound is matched to the oxidation state of the metal in the compound under investigation.[@christensen2015reducing] Following this scheme, the Na reference energy was calculated using the calculated internal energies of NaCl and gaseous Cl$_2$, as well as the experimental formation enthalpy of NaCl: $$\mathrm{E_Na^{Ref} = E_NaCl^{DFT} - \frac{1}{2}E_{{Cl}_2}^{DFT} - {\Delta}H_NaCl^{Exp}}$$
Currently, there is no universally accepted method for selecting the optimal Hubbard U to fit calculations to experimental data. In this study, we follow the scheme that compares the calculated formation enthalpy of a substance under investigation to the experimentally measured formation enthalpy.[@wang2006oxidation] The experimentally measured formation enthalpy of NaO$_2$ is $\mathrm{{\Delta}_{f}H^{0}_{solid} = -2.7 eV}$.[@chase1998nist] We find that the formation enthalpies of all considered bulk configurations of NaO$_2$ lie within 0.3-0.6 eV/NaO$_2$, at all investigated values of the Hubbard U, as shown in Fig. 1(a). The fully-organized, ferromagnetic (AAAA-PPPP) and fully disorganized, antiferromagnetic (ABCD-PPNN) structures were found to be least energetically stable, while the moderately disorganized structures with intermediate magnetic order, such as AABB-PPPN, were most stable. Many of these moderately disorganized structures were energetically degenerate at all examined values of the Hubbard U. We observe that U = 3 eV is matches the experimental formation energy and will be used for further analysis. It is also worth highlighting that the conclusions regarding disorder in bulk NaO$_2$ remain consistent for all examined values of the Hubbard U.
Given the high degeneracy observed among examined structures of NaO$_2$, it becomes crucial to examine the role of thermal disorder and its role in determining electronic properties of NaO$_2$. The computation of disorder in larger supercells directly through DFT is computationally intractable. We employ the approach of determining a lattice Hamiltonian[@hamer1981finite] that can be used to describe the energetics of bulk NaO$_2$.
In order to map out energetic interactions between the magnetic and geometric degrees of freedom, we utilize a modified Ising Model for the lattice Hamiltonian. As implemented, the model consists of a lattice of $N$ sites $i$, whose filling is described by occupation terms, $\sigma_{i}$. System energy contributions due to particle interactions in neighboring sites are captured by the 2-body and 3-body interaction terms, $j_{2_{i,k}}$ and $j_{3_{i,k,l}}$. The total energy of the $N$-site lattice is calculated as: $$\mathrm{E = -\sum_{\braket{ik}} j_{2_{i,k}}\sigma_{i}\sigma_{k} -\sum_{\braket{ikl}} j_{3_{i,k,l}}\sigma_{i}\sigma_{k}\sigma_{l}}$$
Formation enthalpies of $\mathrm{NaO_{2}}$ calculated by DFT simulations were used to fit Ising Model coefficients for our system. Formation enthalpy data at each value of the Hubbard U was used to derive a set of corresponding interaction coefficients, though a least-squares regression fit. The sole differences among the structures lay in the relative geometric and magnetic arrangement of the oxygen dimers, and these differences are reflected in the nearest-neighbor interaction terms $j$. A total of 4 different types of $j_{2_{i,k}}$ terms are identified, along with 6 different types of $j_{3_{i,k,l}}$ terms. Our naming scheme relies on comparison of dimers, e.g. $j_{2AAPP}$, corresponds to the interaction between 2 dimers of parallel geometry and identical magnetization, while $j_{2ABPN}$, corresponds to the interaction between 2 dimers of orthogonal geometry and opposite magnetization. Likewise, $j_{3AAAPPP}$, corresponds to the interaction among 3 dimers of mutually parallel orientation and mutually identical magnetic moment, while $j_{3ABCPPN}$ corresponds to the interaction among 3 oxygen dimers of mutually orthogonal orientation, where 2 dimers share an identical magnetization that differs from that of the third dimer. The plot of fitted interaction coefficients is shown in Fig. 1(b). The plot demonstrates that among the 2-body interaction terms, $j_{2AAPN}$ was most energetically stabilizing, while among the 3-body interaction terms, $j_{3AABPPP}$ was most energetically stabilizing. Our DFT calculations suggested that the most thermodynamically stable configurations correspond to bulk structures of intermediate disorder, both in geometric and magnetic arrangement (e.g. AABB-PPPN). Subsequent Ising Model calculations supported this observation, with $j_{2AAPP}$ the most-stabilizing 2-body interaction term, and $j_{3AABPPP}$ the most-stabilizing 3-body interaction term. It should be noted that using a purely 2-body Ising Model of NaO$_2$ finds that the most-stabilizing nearest-neighbor interaction term was $j_{2ABPN}$, implying that the fully-disorganized antiferromagnetic structure (e.g. ABCD-PNPN) would be the predominant energetically stable structure. This is not consistent with our DFT calculations, and therefore shows the need to incorporate 3-body interaction terms into the model.
Interaction coefficients $j$ derived from the Ising Model were used in Metropolis-Hastings Monte Carlo (MHMC) simulations, to characterize larger bulk cells with higher degree of disorder, at finite temperature. Supercell structures were constructed as N$\times$N$\times$N arrays of cubic unit cells of $\mathrm{Na_{4}O_{8}}$, the structures studied in DFT simulations. Simulations were run for supercells of size 4 to 7, to study convergence of bulk properties as the overall simulated system size increases. Periodic boundary conditions in all directions were implemented and the systems were studied using the Markov Chain Monte Carlo Method,[@gilks2005markov; @hastings1970monte] modified with the Metropolis-Hastings Algorithm.[@chib1995understanding] Interaction coefficients corresponding to a Hubbard U correction of 3 eV were used with the MHMC simulations, to most closely match the experimental formation enthalpy of NaO$_2$. The results would remain qualitatively consistent for other values of U, and there would be a minor change in the observed phase transition temperatures.
At the initialization of each MHMC simulation, a fully-organized supercell bulk structure was set up, with all dimers identical (e.g. AAAA-PPPP). Each individual trial step consisted of switching a randomly selected $\mathrm{O_{2}}$ dimer in the bulk structure to one of the other available configurations; orientational or magnetic change. The effect of this switch on system energy was calculated. The standard Metropolis Algorithm was used to decide whether to accept the trial step or not. At the end of every step, system formation enthalpy and entropy were recorded for further analysis.
Configurational entropy was calculated as the logarithm of the number of configurations with the same proportions of oxygen dimers of different types. If there are $M$ distinct dimer types available, the total number of possible configurations, $\Omega$, of the full structure can be calculated using the complete multinomial coefficient: $$\Omega = \prod_{j=1}^{M} {{\sum_{i=j}^{M} N_{i}} \choose N_{j} }$$ where $i$, $j$ refer to the available configurations of dimers and $N_{i}$, $N_{j}$ refer to the number of dimers of a particular configuration present in the supercell. The configurational entropy is then simply calculated by Boltzmann’s Entropy Formula: $S_{Conf} = k_{B} \log{\Omega}$.
To explore high-temperature, high-energy phases of bulk $\mathrm{NaO_{2}}$, each system simulation was initially raised to a temperature of 1252 K. Once the energy of the structure was stabilized at the initial high temperature, the system was annealed[@vanderbilt1984monte] in temperature steps of 0.25 K, down to 2 K. At each 50 K increment, formation enthalpy and configurational entropy data was collected for analysis. This scheme worked robustly and the thermodynamic quantities converged for all supercell sizes investigated.
MHMC simulations were used to determine effect of temperature on the structural disorder and in turn the free energy of bulk $\mathrm{NaO_{2}}$, as shown in Fig. 1(c) and Fig. 1(d). When the temperature is high, most of the structures could have become accessible, as the entropic term dominates. Thus, the structure at high temperature was predominantly fully-disordered, both in geometry and in magnetic moment as shown in Fig. 1(e). Around 650 K, the structure underwent a phase change, wherein the bulk $\mathrm{NaO_{2}}$ formed large, parallel planes of alternating geometry of $\mathrm{O_{2}}$ dimers, as depicted in Fig. 1(f). In this phase, the magnetic moment was still largely disordered, within and across the monolithic-geometry planes. Around 350 K, the bulk structure underwent a second phase change, where the magnetic moment of all dimers was locked in phase together and the geometry maintained the alternating planar structure. The degree of disorder in the system could strongly affect the electronic properties of NaO$_2$, so we explored this further using hybrid density functional theory calculations.
{width="\textwidth"}
Non-self-consistent hybrid density functional theory calculating employing the HSE06[@heyd2003hybrid] functional was used to determine direct bandgap for all phases of NaO$_2$, as an indicator of electrical conductivity. HSE06 has been shown to accurately capture bandgaps for various semiconducting and insulating materials with mean absolute error (MAE) varying from 0.26$-$0.4 eV for different classes.[@heyd2005energy; @tran2009accurate; @henderson2011accurate; @chan2010efficient] Bandgap values were calculated across all k-points and the minimum was reported for each phase of NaO$_2$ for Hubbard U = 3-7 eV. We find that there is an increase in bandgap with increasing U as shown in Table II in S.I. The widening of the bandgap with increasing U is due to enhanced localization of electrons in the $\pi^*_{2p}$ oxygen bands in NaO$_2$. We also observe that the bandgap in both spin channels is identical for antiferromagnetic structures, and markedly different for ferromagnetic structures. This behavior is observed as the additional electrons in O$_{2}^{-}$ ions only occupy one spin channel. In the AFM structures, electrons are equally distributed in the spin channels, and the gap is between pairs of $\pi^*_{2p}$ states. In the FM structures, the bandgap in the occupied spin channel is between $\pi^*_{2p}$ and $\sigma^*_{2p}$ state and is thus higher than AFM; while for the other channel, it is between deeper $\pi^*_{2p}$ states and is thus lower than that in AFM. For mixed cases (e.g. PPPN), we expect the bandgap to be bounded by the AFM and FM cases, due to electron occupation of the $\pi^*_{2p}$ bands. Finally, we observe that the most symmetric structures, AAAA-PPPP and ABCD-PPPP.
Structure Spin 1(+) Spin 2(-)
------------ ----------- ----------- -- -- -- --
AAAA(PPPP) 6.65 1.26
AAAB(PPPP) 7.43 3.69
AABB(PPPP) 7.03 3.96
AABC(PPPP) 7.3 3.92
ABCD(PPPP) 5.91 1.32
AAAA(PPNN) 3.82 3.82
AAAB(PPNN) 3.91 3.91
AABB(PPNN) 4.13 4.13
AABB(PNPN) 4.21 4.21
AABC(PPNN) 4.39 4.13
ABCD(PPNN) 3.94 3.94
: Calculaed bandgap values (eV) for various ferromagnetic and antiferromagnetic configurations of NaO$_2$, for U=3 eV using HSE06 functional.[]{data-label="bg"}
In conclusion, we demonstrate the need and effectiveness of a 2-body and 3-body nearest-neighbor Ising Model in describing energetics of bulk NaO$_2$ across a large temperature range. Our study demonstrates that at low temperatures, close to room temperature and below, the predominant phase of NaO$_2$ is ferromagnetic, with alternating planes of oxygen dimers in consistent geometric orientations. The system is expected to exhibit some degree of magnetic disorder even at room temperature. Our study on the electronic properties shows that NaO$_2$ is a wide-bandgap insulator, with a bandgap around 4 eV and we expect it to have poor electrical conductivity at room-temperature. In the context of Na-O$_2$ batteries, this implies that growth of the discharge product is most likely occurring due to the solution mechanism pathway involving a chemical dissolution of NaO$_2$ into Na$^+$ and O$_2^-$, similar to what is seen in Li-O$_2$ batteries. However, our analysis shows that preferential nucleation of certain magnetic phases could be possible through appropriate electrode choice. Further, our approach is expected to be important for other metal-oxygen batteries, such as K-O$_2$ batteries.
Acknowledgment is made to the Donors of the American Chemical Society Petroleum Research Fund and National Science Foundation CAREER award CBET-1554273 for partial support of this research.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Real networks often form interacting parts of larger and more complex systems. Examples can be found in different domains, ranging from the Internet to structural and functional brain networks. Here, we show that these multiplex systems are not random combinations of single network layers. Instead, they are organized in specific ways dictated by hidden geometric correlations interweaving the layers. We find that these correlations are significant in different real multiplexes, and form a key framework for answering many important questions. Specifically, we show that these geometric correlations facilitate: (i) the definition and detection of multidimensional communities, which are sets of nodes that are simultaneously similar in multiple layers; (ii) accurate trans-layer link prediction, where connections in one layer can be predicted by observing the hidden geometric space of another layer; and (iii) efficient targeted navigation in the multilayer system using only local knowledge, which outperforms navigation in the single layers only if the geometric correlations are sufficiently strong. Our findings uncover fundamental organizing principles behind real multiplexes and can have important applications in diverse domains.'
author:
- 'Kaj-Kolja Kleineberg'
- Marián Boguñá
- 'M. Ángeles Serrano'
- Fragkiskos Papadopoulos
title: Hidden geometric correlations in real multiplex networks
---
Real networks are often not isolated entities but instead can be considered constituents of larger systems, called multiplexes or multilayer networks [@multilayer:kivel; @arenas:radicchi:2013; @ginestra:natphys; @radicci:percolation; @arenas:multiplex; @Thurner:multi; @Menichetti; @Halu]. Examples can be found everywhere. One is the multiplex consisting of the different social networks that a person may belong to. Other examples include the Internet’s IPv4 and IPv6 topologies, or the structural and functional networks in the brain. Understanding the relations among the networks comprising a multiplex is crucial for understanding the behavior of a wide range of real world systems [@our:model; @ecology20; @worldmodel; @DeDomenico2014; @Simas2015]. However, despite the burst of recent research in studying the properties of multiplex networks, e.g., [@arenas:multiplex; @multilayer:kivel; @Boccaletti20141], a universal framework describing the relations among the single networks comprising a multiplex and what implications these relations may have when it comes to applications remains elusive.
Here, we show that real multiplexes are not random combinations of single network layers. Instead, we find that their constituent networks exhibit significant *hidden geometric correlations*. These correlations are called “hidden" as they are not directly observable by looking at each individual network topology. Specifically, each single network can be mapped (or embedded) into a separate hyperbolic space [@Krioukov2010; @Boguna2010; @frag:hypermap; @frag:hypermap_cn], where node coordinates abstract the *popularity* and *similarity* of nodes [@Serrano2008; @boguna:popularity]. We find that node coordinates are significantly correlated across layers of real multiplexes, meaning that distances between nodes in the underlying hyperbolic spaces of the constituent networks are also significantly correlated.
The discovered geometric correlations yield a very powerful framework for answering important questions related to real multiplexes. Specifically, we first show that these correlations suggest the existence of multidimensional communities, which are sets of nodes that are similar (close in the underlying space) in multiple layers. Further, we show that strong geometric correlations imply accurate trans-layer link prediction, where connections in one layer can be predicted by knowing the hyperbolic distances among nodes in another layer. This is important for applications where we only know the connections among nodes in one context, e.g., structural connections between brain regions, and we want to predict connections between the same nodes in some other context, e.g., likelihood of functional connections between the same brain regions.
Finally, to study the effects of geometric correlations on dynamical processes, we consider targeted navigation. Targeted navigation is a key function of many real networks, where either goods, people, or information is transferred from a source to a destination using the connections of the network. It has been shown that single complex networks, like the IPv4 Internet or the network of airport connections, can be navigated efficiently by performing *greedy routing* in their underlying geometric spaces [@Boguna2008; @Boguna2010; @Papadopoulos2010; @Krioukov2010]. In greedy routing, nodes forward incoming messages to their neighbors that are closest to the destination in the geometric space so that they only need local knowledge about the coordinates of their neighbors. In multilayer systems, messages can switch between layers and so a node forwards a message to its neighbor that is closest to the destination in any of the layers comprising the system. We call this process *mutual greedy routing* (MGR).
Mutual greedy routing follows the same line of reasoning as greedy routing in Milgram’s experiment [@milgram1969], which was indeed performed using multiple domains. For example, in the case of a single network, to reach a lawyer in Boston one might want to forward a message to a judge in Los Angeles (greedy routing). However, in the case of two network layers, it might be known that the lawyer in Boston is also a passionate vintage model train collector. An individual who knows a judge in Los Angeles and the owner of a vintage model train shop in New York would probably choose to forward the message to the latter (MGR). Similarly, air travel networks can be supported by train networks to enhance the possibilities to navigate the physical world, individuals can use different online social networks to increase their outreach, and so on. In this work, we consider the real Internet, which is used to navigate the digital world, and show that mutual greedy routing in the multiplex consisting of the IPv4 and IPv6 Internet topologies [@as_topo] outperforms greedy routing in the single IPv4 and IPv6 networks. We also use synthetic model networks to show that geometric correlations improve the navigation of multilayer systems, which outperforms navigation in the single layers if these correlations are sufficiently strong. As we will show, if optimal correlations are present, the fraction of failed deliveries is mitigated superlinerarly with the number of layers, suggesting that more layers with the right correlations quickly make multiplex systems almost perfectly navigable.
Geometric organization of real multiplexes {#geometric-organization-of-real-multiplexes .unnumbered}
==========================================
It has been shown that many real (single layer) complex networks have an effective or hidden geometry underneath their observed topologies, which is hyperbolic rather than Euclidean [@Krioukov2009; @Krioukov2010; @Boguna2010; @boguna:popularity; @Serrano2011]. In this work, we extend the hidden geometry paradigm to real multiplexes and prove that the coordinates of nodes in the different underlying spaces of layers are correlated.
Geometry of single layer networks {#geometry-of-single-layer-networks .unnumbered}
---------------------------------
Nodes of real single-layered networks can be mapped to points in the hyperbolic plane, such that each node $i$ has the polar coordinates, or hidden variables, $r_i, \theta_i$. The radial coordinate $r_i$ abstracts the node popularity. The smaller the radial coordinate of a node, the more popular the node is, and the more likely it attracts connections. The angular distance between two nodes, $\Delta \theta_{ij}=\pi-|\pi-|\theta_i-\theta_j||$, abstracts their similarity. The smaller this distance, the more similar two nodes are, and the more likely they are connected. The hyperbolic distance between two nodes, very well approximated by $x_{ij} \approx r_i+r_j+2\ln{\sin{(\Delta\theta_{ij}/2)}}$ [@Krioukov2010], is then a single-metric representation of a combination of the two attractiveness attributes, radial popularity and angular similarity. The smaller the hyperbolic distance between two nodes, the higher is the probability that the nodes are connected, meaning that connections take place by optimizing trade-offs between popularity and similarity [@boguna:popularity].
Techniques based on Maximum Likelihood Estimation for inferring the popularity and similarity node coordinates in a real network have been derived in [@Boguna2010] and recently optimized in [@frag:hypermap; @frag:hypermap_cn]. It has been shown that through the constructed hyperbolic maps one can identify soft communities of nodes, which are groups of nodes located close to each other in the angular similarity space [@Boguna2010; @Serrano2011; @boguna:popularity; @soft:comm]; predict missing links with high precision [@Serrano2011; @frag:hypermap; @frag:hypermap_cn]; and facilitate efficient greedy routing in the Internet, which can reach destinations with more than $90$% success rate, following almost shortest network paths [@Boguna2010; @frag:hypermap; @frag:hypermap_cn].
Geometry of real multiplexes {#geometry-of-real-multiplexes .unnumbered}
----------------------------
![**Hyperbolic mapping of the IPv4/IPv6 Internet.** **A:** IPv4 topology—for clarity only nodes with degrees greater than $3$ are shown. **B:** IPv6 topology. \[fig\_internet\_embedding\]](figure1.pdf){width="1\linewidth"}
**Name** **Type** **Nodes** **Layer 1, Layer 2, ...** **Source**
------------ --------------- ------------------------- -------------------------------------------------------------------- ----------------------------
Internet Technological Autonomous Systems IPv4 AS topology, IPv6 AS topology [@as_topo]
Air/Train Technological Airports/Train stations Indian airport network, Indian train network [@Halu]
Drosophila Biological Proteins Suppressive genetic interaction, additive genetic interaction [@biogrid; @arenas:reduce]
C. Elegans Biological Neurons Electric, chemical monadic, chemical polyadic synaptic junctions [@pnas:celegans; @muxviz]
Brain Biological Brain regions Structural network, functional network [@Simas2015]
SacchPomb Biological Proteins 5 different interactions (direct interaction, colocalization, ...) [@biogrid; @arenas:reduce]
Rattus Biological Proteins Physical association, direct interaction [@biogrid; @arenas:reduce]
arXiv Collaboration Authors 8 different categories (physics.bio-ph, cond-mat.dis-nn, ...) [@prx:modular]
Physicians Social Physicians Advice, discussion, friendship relations [@physicians:data]
We now consider nine different real-world multiplex networks from diverse domains. An overview of the considered datasets is given in Table \[tab\_datasets\]—see Supplementary Information section I for a detailed description. For each multiplex, we map each network layer independently to an underlying hyperbolic space using the *HyperMap* method [@frag:hypermap; @frag:hypermap_cn] (see Supplementary Information section II), thus inferring the popularity and similarity coordinates $r,\theta$ of all of its nodes. A visualization of the mapped IPv4 and IPv6 Internet layers is shown in Fig. \[fig\_internet\_embedding\]. For each of our real multiplexes, we find that both the radial and the angular coordinates of nodes that exist in different layers are correlated.
The radial popularity coordinate of a node $i$ depends on its observed degree in the network $k_i$ via $r_i \sim \ln{N}-\ln{k_i}$, where $N$ is the total number of nodes [@Boguna2010; @frag:hypermap; @frag:hypermap_cn]. Therefore, radial correlations are equivalent to correlations among node degrees, which have been recently found and studied [@pre:multiplex:correlations; @pre:degree:correlations:2; @prl:degree:correlations; @scirep:degree:corr; @self:similar:multiplex]. Consistent with these findings, radial correlations are present in our real multiplexes and are encoded in the conditional probability $P(r_2|r_1)$ that a node has radial coordinate $r_2$ in layer 2 given its radial coordinate $r_1$ in layer 1. $P(r_2|r_1)$ for different real multiplexes is shown in Supplementary Fig. 1, where we observe significant radial correlations.
{width="1\linewidth"}
The correlation among the angular similarity coordinates, on the other hand, is a fundamentally new result that has important practical implications. Fig. \[fig\_histogram\] shows the distribution of nodes that have angular coordinates $(\theta_1,\theta_2)$ in layers $1$ and $2$ of the real Internet, Drosophila, and arXiv multiplexes. The figure also shows the corresponding distributions in the reshuffled counterparts of the real systems, where we have destroyed the trans-layer coordinate correlations by randomly reshuffling node ids (see Supplementary Information section III). We observe an overabundance of two-dimensional similarity clusters in the real multiplexes. These clusters consist of nodes that are similar, i.e., are located at small angular distances, in both layers of the multiplex. These similarity clusters do not exist in the reshuffled counterparts of the real systems, and are evidence of significant angular correlations. Similar results hold for the rest of the real multiplexes that we consider (see Supplementary Information section IV). In Supplementary Information section X, we quantify both the radial and the angular correlations present in all the considered real multiplexes.
The generalization of community definition and detection techniques from single layer networks to multiplex systems has recently gained attention [@Mucha876; @multiplex:community:cs; @multiplex:communities; @prx:modular]. The two-dimensional similarity histograms in Fig. \[fig\_histogram\] allow one to naturally observe two-dimensional soft communities, defined as sets of nodes that are similar—close in the angular similarity space—in both layers of the multiplex simultaneously. They also provide a measure of distance between the different communities. We note that these communities are called “soft" as they are defined by the geometric proximity of nodes [@Boguna2010; @Serrano2011; @boguna:popularity; @soft:comm] rather than by the network topology as is traditionally the case [@Fortunato201075]. We also note that we do not develop an automated multidimensional community detection algorithm here. However, our findings could serve as a starting point for the development of such an algorithm (see discussion in Supplementary Information section IV). In Supplementary Information section V, we show how certain geographic regions are associated to two-dimensional communities in the IPv4/IPv6 Internet.
The problem of link prediction has been extensively studied in the context of predicting missing and future links in single layer networks [@link:prediction:1; @Clauset:2008:Hierarchicalstructure] and its generalization to real-world multilayer systems is recently gaining attention [@multiplex:link:prediction:1]. In our case, the radial and angular correlations across different layers suggest that by knowing the hyperbolic distance between a pair of nodes in one layer, we can predict the likelihood that the same pair is connected in another layer. Fig. \[fig\_link\_prediction\] (top) validates that this is indeed the case. The figure shows the empirical trans-layer connection probability, $P(1|2)$ ($P(2|1)$), that two nodes are connected in one of the layers of the multiplex given their hyperbolic distance in the other layer. In Fig. \[fig\_link\_prediction\] (top), we observe that this probability decreases with the hyperbolic distance between nodes in the considered real multiplexes (see Supplementary Information section IV for the results for the rest of the multiplexes). By contrast, in their reshuffled counterparts, which do not exhibit geometric correlations, this probability is almost a straight line. Fig. \[fig\_link\_prediction\] (bottom) shows that the trans-layer connection probability decreases with the angular distance between nodes, which provides an alternative empirical validation of the existence of significant similarity correlations across the layers. In Supplementary Information section VI, we quantify the quality of trans-layer link prediction in all the real multiplexes we consider, and show that our approach outperforms a binary link predictor that is based on edge overlaps.
{width="01\linewidth"}
Modeling geometric correlations and implications to mutual greedy routing {#modeling-geometric-correlations-and-implications-to-mutual-greedy-routing .unnumbered}
=========================================================================
The IPv4 Internet has been found to be navigable [@Boguna2010; @frag:hypermap; @frag:hypermap_cn]. Specifically, it has been shown that greedy routing (GR) could reach destinations with more than 90% success rate in the constructed hyperbolic maps of the IPv4 topology in 2009. We find a similar efficiency of GR in both the IPv4 and IPv6 topologies considered here, which correspond to January 2015. Specifically, we perform GR in the hyperbolic map of each topology among $10^5$ randomly selected source-destination pairs that exist in both topologies. We find that GR reaches destinations with $90\%$ and $92\%$ success rates in IPv4 and IPv6, respectively. Furthermore, we also perform angular GR, which is the same as GR but using only angular distances. We find that the success rate in this case is almost $60\%$ in both the IPv4 and IPv6 topologies. However, hyperbolic mutual greedy routing (MGR) between the same source-destination pairs, which travels to any neighbor in any layer closer to destination, increases the success rate to $95\%$, while the angular MGR that uses only the angular distances, increases the success rate along the angular direction to $66\%$. More details about the MGR process are found in Supplementary Information section XI.
The observations above raise the following fundamental questions. (i) How do the radial and angular correlations affect the performance of MGR? (ii) Under which conditions does MGR perform better than single-layer GR? (iii) How does the performance of MGR depend on the number of layers in a multilayer system? And (iv), how close to the optimal—in terms of MGR’s performance—are the geometric correlations in the IPv4/IPv6 Internet? Answering these questions requires a framework to construct realistic synthetic topologies (layers) where correlations—both radial and angular—can be tuned without altering the topological characteristics of each individual layer.
{width="1\linewidth"}
To this end, we develop the geometric multiplex model (GMM) (see Methods section B). The model generates multiplexes with layer topologies embedded in the hyperbolic plane and where correlations can be tuned independently by varying the model parameters $\nu \in [0,1]$ (radial correlations) and $g \in [0,1]$ (angular correlations). We consider two-, three- and four-layer multiplexes constructed using our framework with different values of the correlation strength parameters $\nu$ and $g$.
From Figs. \[fig\_layers\_panel\]A,B and Supplementary Figs. 16–18, we observe that in general both MGR and angular MGR perform better as we increase the correlation strengths $\nu$ and $g$. When both radial and angular correlations are weak, we do not observe any significant benefits from mutual navigation. Indeed, in Figs. \[fig\_layers\_panel\]C-E we observe that in the uncorrelated case ($\nu \to 0$, $g\to 0$) MGR performs almost identical to the single-layer GR, irrespectively of the number of layers. This is because when a message reaches a node in one layer after the first iteration of the MGR process, the probability that this node will have a neighbor in another layer closer to the destination is small. That is, even though a node may have more options (neighbors in other layers) for forwarding a message, these options are basically useless and messages navigate mainly single layers.
Increasing the strength of correlations makes MGR more efficient, as the probability to have a neighbor closer to the destination in another layer increases. However, increasing the strength of correlations also increases the edge overlap between the layers (see Supplementary Information section VIII), which reduces the options that a node has for forwarding a message. We observe that very strong radial and angular correlations may not be optimal at low temperatures (cf. Supplementary Figs. 16–18 for $T=0.1$). This is because when the layers have the same nodes and the same average degree and power law exponent, $\bar{k}, \gamma$, then as $\nu \to 1$, $g \to 1$, the coordinates of the nodes in the layers become identical (see Supplementary Information section VII). Further, if the temperature of the layers is low, the connection probability $p(x_{ij})$ in each layer becomes the step function, where two nodes $i,j$ are deterministically connected if their hyperbolic distance is $x_{ij} \leq R~\sim \ln{N}$ (see Methods section B). That is, as $T \to 0$, $\nu \to 1$, $g \to 1$, all layers become identical, and MGR degenerates to single-layer GR. We observe (Figs. \[fig\_layers\_panel\]A,B and Supplementary Figs. 16–18) that the best MGR performance is always achieved at high angular correlations, and either high radial correlations if the temperature of the individual layers is high, or low radial correlations if the temperature of the layers is low. The best angular MGR performance is always achieved at high angular and low radial correlations. In other words, MGR performs more efficient when the layers comprising the multiplex are similar but not “too” similar. High temperatures induce more randomness so that even with maximal geometric correlations the layers are not too similar. However, at low temperatures, very strong geometric correlations make the layers too similar and the best performance occurs for intermediate correlations.
From Figs. \[fig\_layers\_panel\]C-E we observe that, for a fixed number of layers and for optimal correlations, the failure rate ($1-$success rate) is reduced by a constant factor, which is independent of the navigation type (MGR or angular MGR) and the layer temperature. This factor, which we call *failure mitigation factor*, is the inverse of the slope of the best-fit lines in Figs. \[fig\_layers\_panel\]C-E. Fig. \[fig\_layers\_panel\]F shows the failure mitigation factor for our two-, three- and four-layer multiplexes for both uncorrelated and optimally correlated coordinates. Remarkably, if optimal correlations are present, the failure mitigation factor grows superlinerarly with the number of layers, suggesting that more layers with the right correlations can quickly make multiplex systems almost perfectly navigable. On the contrary, more layers without correlations do not have a significant effect on mutual navigation, which performs virtually identical to single-layer navigation.
![**Performance of mutual navigation as a function of radial and angular correlation strengths ($\nu, g$) in a two-layer synthetic multiplex that best mimics the real IPv4/IPv6 Internet.** **A:** Hyperbolic mapping of the real IPv4 topology where nodes marked by red also exist in the IPv6 topology. **B:** Hyperbolic mapping of layer 1 of our Internet-like synthetic multiplex, where nodes marked by red also exist in layer 2. **C**: Performance of angular MGR. **D**: Performance of MGR. In **C, D** the black star indicates the achieved performance with the estimated correlation strengths in the real IPv4/IPv6 Internet, $\nu_E \approx 0.4, g_E \approx 0.4$. \[fig\_psi\_fitted\_graphs\]](figure5.pdf){width="1\linewidth"}
Finally, we investigate how close to the optimal—in terms of mutual navigation performance—are the radial and angular correlations in the IPv4/IPv6 Internet. To this end, we use an extension of the GMM framework that can account for different layer sizes (see Methods section C) to construct an Internet-like synthetic multiplex. For those nodes that exist in both layers of our multiplex (see Figs. \[fig\_psi\_fitted\_graphs\]A,B), we tune the correlations among their coordinates as before, by varying the parameters $\nu$ and $g$, and perform mutual navigation. Figs. \[fig\_psi\_fitted\_graphs\]C,D show the performance of angular MGR and of MGR, respectively. We observe again that increasing the correlation strengths improves performance. In angular MGR, the success rate is $63\%$ with uncorrelated coordinates, while with optimal correlations it becomes $75\%$. In MGR, the success rate with uncorrelated coordinates is $85\%$ and with optimal correlations is $91\%$. The star in Figs. \[fig\_psi\_fitted\_graphs\]C,D indicates the achieved performance with the empirical correlation strengths in the IPv4/IPv6 Internet, $\nu_E \approx 0.4, g_E \approx 0.4$, which are estimated using the inferred radial and angular coordinates of nodes (see Supplementary Information section IX). At $\nu=\nu_E, g=g_E$, the success rate of angular MGR is $71\%$, which is closer to the rate obtained with optimal correlations than to the uncorrelated case. For MGR, the success rate is $88\%$, which lies in the middle of the uncorrelated and optimally correlated cases.
Outlook {#outlook .unnumbered}
=======
Hidden metric spaces underlying real complex networks provide a fundamental explanation of their observed topologies and, at the same time, offer practical solutions to many of the challenges that the new science of networks is facing nowadays. On the other hand, the multiplex nature of numerous real-world systems has proven to induce emerging behaviors that cannot be observed in single layer networks. The discovery of hidden metric spaces underlying real multiplexes is, thus, a natural step forward towards a fundamental understanding of such systems. The task is, however, complex due to the inherent complexity of each single layer and that, in principle, the metric spaces governing the topologies of the different single networks could be unrelated. Our results here clearly indicate that this is not the case.
Our findings open the door for multiplex embedding, in which all the layers of a real multiplex are simultaneously and not independently embedded into hyperbolic spaces. This would require a reverse engineering of the geometric multiplex model proposed here, which generates multiplexes where radial and angular correlations can be tuned independently by varying its parameters. Multiplex embeddings can have important applications in diverse domains, ranging from improving information transport and navigation or search in multilayer communication systems and decentralized data architectures [@Contreras11122015; @helbing:digital_democracy], to understanding functional and structural brain networks and deciphering their precise relationship(s) [@Simas2015], to predicting links among nodes (e.g., terrorists) in a specific network by knowing their connectivity in some other network.
Methods {#methods .unnumbered}
=======
Computation of trans-layer connection probability
-------------------------------------------------
To compute the trans-layer connection probability, we consider all nodes that exist in both layers. In each of the layers, we bin the range of hyperbolic distances between these nodes from zero to the maximum distance into small bins. For each bin we then find all the node pairs located at the hyperbolic distances falling within the bin. The percentage of pairs in this set of pairs that are connected by a link in the other layer, is the value of the empirical trans-layer connection probability at the bin.
Geometric multiplex model (GMM)
-------------------------------
Our framework builds on the (single-layer) network construction procedure prescribed by the newtonian $\mathbb{S}^{1}$ [@Serrano2008] and hyperbolic $\mathbb{H}^{2}$ [@Krioukov2010] models. The two models are isomorphic and here we present the results for the $\mathbb{H}^{2}$ version even if for calculations it is more convenient to make use of the $\mathbb{S}^{1}$. We recall that to construct a network of size $N$, the $\mathbb{H}^{2}$ model firsts assigns to each node $i=1,\ldots, N$ its popularity and similarity coordinates $r_i, \theta_i$. Subsequently, it connects each pair of nodes $i, j$ with probability $p(x_{ij})=1/(1+e^{\frac{1}{2T}(x_{ij}-R)})$, where $x_{ij}$ is the hyperbolic distance between the nodes and $R \sim \ln{N}$ (see Supplementary Information section VII A). The connection probability $p(x_{ij})$ is nothing but the Fermi-Dirac distribution. Parameter $T$ is the *temperature* and controls clustering in the network [@Dorogovtsev10-book], which is the probability that two neighbors of a node are connected. The average clustering $\bar{c}$ is maximized at $T=0$, nearly linearly decreases to zero with $T \in [0,1)$, and is asymptotically zero if $T>1$. As $T \to 0$ the connection probability becomes the step function $p(x_{ij}) \to 1$ if $x_{ij} \leq R$, and $p(x_{ij}) \to 0$ if $x_{ij} > R$. It has been shown that the $\mathbb{S}^{1}$ and $\mathbb{H}^{2}$ models can construct synthetic networks that resemble real networks across a wide range of structural characteristics, including power law degree distributions and strong clustering [@Serrano2008; @Krioukov2010]. Our framework constructs single-layer topologies using these models, and allows for radial and angular coordinate correlations across the different layers. The strength of these correlations can be tuned via model parameters $\nu \in [0,1]$ and $g \in [0,1]$, without affecting the topological characteristics of the individual layers, which can have different properties and different sizes (see Supplementary Information section VII). The radial correlations increase with parameter $\nu$—at $\nu=0$ there are no radial correlations, while at $\nu=1$ radial correlations are maximized. Similarly, the angular correlations increase with parameter $g$—at $g=0$ there are no angular correlations, while at $g=1$ angular correlations are maximized.
Extended GMM and the IPv4/IPv6 Internet
---------------------------------------
To construct the Internet-like synthetic multiplex of Fig. \[fig\_psi\_fitted\_graphs\] we use an extention of the GMM that can construct multiplexes with different layer sizes and where correlations among the coordinates of the common nodes between the layers can be tuned as before via parameters $\nu$ and $g$ (see Supplementary Information section VII D).
Layer 1 in the Internet-like synthetic multiplex has approximately the same number of nodes as in the IPv4 topology, $N_1=37563$ nodes, as well as the same power law degree distribution exponent $\gamma_1=2.1$, average node degree $\bar{k}_1\approx 5$, and average clustering $\bar{c}_1 \approx 0.63$ ($T_1=0.5$). Layer 2 has approximately the same number of nodes as in the IPv6 topology, $N_2=5163$ nodes, and the same power law exponent $\gamma_2=2.1$, average node degree $\bar{k}_2\approx 5.2$, and average clustering $\bar{c}_2 \approx 0.55$ ($T_2=0.5$).
The IPv4 topology is significantly larger than the IPv6 topology, and there are $4819$ common nodes (Autonomous Systems) in the two topologies. We find that nodes with a higher degree in IPv4 are more likely to also exist in IPv6. Specifically, we find that the empirical probability $\psi(k)$ that a node of degree $k$ in IPv4 also exists in IPv6 can be approximated by $\psi(k)=1/(1+15.4k^{-1.05})$ (see Supplementary Fig. 10). We capture this effect in our synthetic multiplex by first constructing layer 1, and then sampling with the empirical probability $\psi(k)$ nodes from layer 1 that will also be present in layer 2 (see Supplementary Information section VII D). A visualization illustrating the common nodes in the real Internet and in our synthetic multiplex is given in Figs. \[fig\_psi\_fitted\_graphs\]A,B. We note that the fact that nodes with higher degrees in the larger layer have higher probability to also exist in the smaller layer has also been observed in several other real multiplexes [@pre:multiplex:correlations]. However, our model for constructing synthetic multiplexes with different layer sizes is quite general, and allows for any sampling function $\psi(k)$ to be applied (see Supplementary Information section VII D).
[48]{}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 [****, ()]{} [****, ()](\doibase 10.1103/PhysRevX.4.031046) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevE.82.036106) [****, ()](\doibase 10.1038/ncomms1063) [****, ()](\doibase 10.1109/TNET.2013.2294052) [****, ()](\doibase
10.1103/PhysRevE.92.022807) [****, ()](\doibase 10.1103/PhysRevLett.100.078701) @noop [****, ()]{} [****, ()](\doibase 10.1038/nphys1130) [ , ()](http://dl.acm.org/citation.cfm?id=1833515.1833893) [****, ()](http://www.jstor.org/stable/2786545) @noop [“,” ]{} (), [****, ()](\doibase
10.1103/PhysRevE.80.035101) [****, ()](\doibase 10.1039/C2MB05306C) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1073/pnas.0506806103) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevX.5.011027) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevE.92.032805) [****, ()](\doibase
10.1103/PhysRevE.89.042811) [****, ()](\doibase 10.1103/PhysRevLett.111.058702) [****, ()](\doibase 10.1038/srep09120) [****, ()](http://stacks.iop.org/1367-2630/17/i=5/a=053033) @noop [****, ()]{} @noop [ , ****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, , )
[**Acknowledgements**]{}\
This work was supported by: the European Commission through the Marie Curie ITN “iSocial” grant no. PITN-GA-2012-316808; a James S. McDonnell Foundation Scholar Award in Complex Systems; the ICREA Academia prize, funded by the [*Generalitat de Catalunya*]{}; the MINECO project no. FIS2013-47282-C2-1-P; and the [*Generalitat de Catalunya*]{} grant no. 2014SGR608. Furthermore, M. B. and M. A. S. acknowledge support from the European Commission FET-Proactive Project MULTIPLEX no. 317532.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'On a Riemannian 2-torus $(T^2,g)$ we study the geodesic flow in the case of low complexity described by zero topological entropy. We show that this assumption implies a nearly integrable behavior. In our previous paper [@GK] we already obtained that the asymptotic direction and therefore also the rotation number exists for all geodesics. In this paper we show that for all $r \in \mathbb{R} \cup \{\infty\}$ the universal cover ${\mathbb{R}}^2$ is foliated by minimal geodesics of rotation number $r$. For irrational $r \in \mathbb{R}$ all geodesics are minimal, for rational $r \in \mathbb{R} \cup \{\infty\}$ all geodesics stay in strips between neighboring minimal axes. In such a strip the minimal geodesics are asymptotic to the neighboring minimal axes and generate two foliations.'
address: 'Faculty of Mathematics, Ruhr University Bochum, 44780 Bochum, Germany'
author:
- Eva Glasmachers and Gerhard Knieper
title: Minimal Geodesic Foliation on $T^2$ in case of vanishing topological entropy
---
Introduction
============
This paper continues our work [@GK] on geodesic flows on the unit tangent bundle of a two-dimensional Riemannian torus $(T^2,g)$. We also like to mention that earlier versions of the results obtained in [@GK] are already contained in the thesis of the first author [@G].
The goal of our work is to study dynamical and geometrical implications of vanishing topological entropy. Recall that the topological entropy of a continuous dynamical system represents the exponential growth rate of orbit segments distinguishable with arbitrarily fine but finite precision. It therefore describes the total exponential orbit complexity with a single number. Due to a theorem of A. Katok [@K80] positive topological entropy and the existence of a horseshoe are equivalent provided the phase space of the flow is 3-dimensional.
It turns out that zero topological entropy yields strong restrictions on the behavior of geodesics. Important results in this direction are due to J. Denvir and R. S. MacKay [@DM]. Their work implies that contractible closed geodesics on $T^2$ do not exist in case of vanishing topological entropy. An independent proof was also given in [@G]. Using variational methods S. V. Bolotin and P. H. Rabinowitz [@BR] studied the complexity of the geodesic flow on $T^2$ and obtained positive topological entropy under certain conditions.
In [@GK] we showed that absence of positive topological entropy implies nearly integrable behavior. In particular, the lifts of all geodesics on ${\mathbb{R}}^2$ (not just the minimal ones) stay in tubular neighborhoods of Euclidean lines. Hence, all geodesics have an asymptotic direction and define a nontrivial continuous constant of motion. The main tools used in our approach are curve-shortening techniques that allow to globally control geodesics in the presence of certain intersection patterns of geodesic segments.
In this paper we strengthen our previous results. We show that each asymptotic direction yields a geodesic foliation on the Riemannian universal covering $(\mathbb{R}^2, \tilde{g})$ of $(T^2
= {\mathbb{R}}^2 / {\mathbb{Z}}^2 ,g)$ consisting of globally minimizing geodesics. For irrational directions the foliation is unique and all geodesics with irrational directions are minimal. In particular, each geodesic $c$ on $\mathbb{R}^2$ with irrational rotation number does not intersect its translates, i.e. $\tau (c) \cap c =
\emptyset$ for all $\tau \in {\mathbb{Z}}^2\setminus \{(0,0)\}$. However, for all rational directions the foliations are unique if and only if the metric is flat. We remark that for monotone twist maps analogous result were obtained by S. Angenent [@A-1992]. Hence, our results extend well known relations between minimal orbits of monotone twist-maps (Aubry-Mather theory [@Aubry],[@Mather]) and minimal geodesics on $T^2$ which were studied by V. Bangert [@B1988] as well as M. L. Bialy and L. V. Polterovich [@BP]. However, our approach does not use monotone twist maps but again relies on the curve shortening flow.
Our main results can be summarized in the following theorem:\
[**Theorem.**]{} [*Let $(T^2,g)$ be a Riemannian torus with $h_{top}(g) =0$. Then for all $r \in {\mathbb{R}}\cup \{\infty\}$ the torus is foliated by minimal geodesics with rotation number $r$. If $r$ is irrational the foliation is unique. For all $r \in {\mathbb{Q}}\cup \{\infty\}$ the foliations are unique iff $T^2$ is flat. Moreover, each lift of a geodesic on $\mathbb{R}^2$ with irrational rotation number does not intersect their translates.*]{}
1. We remark that each such minimal foliation on $T^2$ corresponds to the graph of a Lagrangian torus on the unit tangent bundle $ST^2$ invariant under the geodesic flow. Therefore, a result of I. V. Polterovich [@Polt] implies that each irrational minimal geodesic is dense in $T^2$. In particular, he shows that the metric is flat provided one irrational minimal geodesic has no focal points.
2. Further analogies to properties of orbits of monotone twist maps in the case of vanishing topological entropy presented in [@A-int] were derived in [@GK] and will be summarized in this paper in Theorem \[summarize\].
Topological entropy and properties of minimal geodesics
=======================================================
A fundamental concept in our investigation is the topological entropy of a continuous dynamical system. It is invariant under topological conjugations and measures, as described in the introduction, the exponential orbit complexity with a single non-negative number. The precise meaning becomes apparent in the following definition introduced by R. E. Bowen [@BO].
\[topent\] Let $(Y,d)$ be a compact metric space, $\phi^t: Y \to Y$ a continuous flow and $d_{T}(\cdot,\cdot)$ the dynamical metric defined by $d_{T}(v,w): = \max_{0 \leq t \leq T}d(\phi^tv, \phi^tw)$ for all $v,w
\in Y$. For a given $\varepsilon>0$ a subset $F\subset Y$ is called $(\phi,\varepsilon,d_{T} )$-[*separated*]{} set of $Y$, if for $x_1 \not =x_2 \in F$ we have $d_{T}(x_1,x_2)> \varepsilon.$ The topological entropy of $\phi^t$ is defined as $$h_{top}(g) = h_{top}(\phi) =\lim_{\varepsilon \to 0} \limsup_{T \to
\infty}
\left(\frac{1}{T} \log r_T(\phi, \varepsilon) \right),$$ where $r_T(\phi, \varepsilon)$ is the maximum of the cardinalities of any $(\phi,\varepsilon, d_T)$-separated set of $Y$.
For more details and properties of the topological entropy see for example [@KH] or [@Wo]. In the investigation of the dynamics of the geodesic flow on $T^2 = {\mathbb{R}}^2 / {\mathbb{Z}}^2$ the notion of rotation number is of central importance.
We say that a geodesic $c:{\mathbb{R}}\to {\mathbb{R}}^2$ has an asymptotic direction if the limit $$\delta(c) :=\lim\limits_{t\to \infty} \frac{c(t)}{\|c(t)\|} \in S^1$$ exists. If $\delta(c) =(x,y)$ we call the projection onto $\mathbb{P}_1({\mathbb{R}})$ given by $$\rho(c) = \left\{ \begin{array}{cc}
\frac{y}{x}, & \text{if} \; x \not=0\\
\infty, & \text{otherwise}
\end{array} \right.$$ the rotation number of $c$.
A geodesic $c: {\mathbb{R}}\to {\mathbb{R}}^2$ on the Riemannian universal covering $({\mathbb{R}}^2, \tilde g)$ of $(T^2, g)$ is called an axis if there exists a nontrivial translation element $\tau \in {\mathbb{Z}}^2$ such that $\tau c(t) = c(t + l)$ for some $l \in {\mathbb{R}}$ and all $t \in {\mathbb{R}}$.
The projection of an axis $c: {\mathbb{R}}\to {\mathbb{R}}^2$ of a nontrivial translation element $\tau \in {\mathbb{Z}}^2$ onto $T^2$ corresponds to a closed geodesics in the homology class given by $\tau$.
For surfaces of genus strictly larger than one Morse [@Morse] began in 1924 a systematic investigation of minimal geodesics, i.e., geodesics which lift to minimal geodesics on the universal covering. Somewhat later Hedlund [@Hedlund] obtained similar results in the case of the 2-torus.
Minimal geodesics on $({\mathbb{R}}^2, \tilde g ) $ and their projections on $(T^2, g)$ will play an important role in this paper. Central properties are that two different minimal geodesics on $({\mathbb{R}}^2,\tilde g)$ cross at most once and minimal geodesics have no self-intersections.
There are well known connections between minimal geodesics on $({\mathbb{R}}^2,\tilde g)$ and rotation numbers: A fundamental result of Hedlund [@Hedlund] yields that for each $r \in {\mathbb{R}}\cup
\{\infty\}$ there exists a minimal geodesic $c:{\mathbb{R}}\to {\mathbb{R}}^2$ with rotation number $r$. Furthermore, there exists a constant $D>0$, such that to each minimal geodesic $c: {\mathbb{R}}\to {\mathbb{R}}^2$ corresponds a Euclidean line $l$, and to each Euclidean line $l$ corresponds a minimal geodesic $c$ such that $$d (l, c(t)) \leq D, \quad \text{for all $t \in {\mathbb{R}}.$}$$ In particular, this implies the existence of the rotation number for each minimal geodesic.
The set of minimal geodesics with a fixed irrational rotation number is totally ordered, i.e., in this set no pair of geodesics intersects. In the set of minimal geodesics with a fixed rational rotation number the subset of minimal axes is totally ordered as well. Two minimal axes $\alpha_1, \alpha_2$ with the same asymptotic directions bounding a strip containing no further minimal axes are called neighboring minimals. There exists a minimal geodesic $c: {\mathbb{R}}\to {\mathbb{R}}^2$ of asymptotic type $A(\alpha_1, \alpha_2)$, i.e., $d(c(t), \alpha_1({\mathbb{R}})) \to 0$ for $t \to \infty$ and $d(c(t), \alpha_2({\mathbb{R}})) \to 0$ for $t \to
-\infty$. If $\alpha_1 \not= \alpha_2$, then each pair $c_1$ and $c_2$ of minimal geodesics of asymptotic type $A(\alpha_1,
\alpha_2)$ and $ A(\alpha_2, \alpha_1)$, respectively, has a unique intersection. Moreover, all minimal geodesics with the same rational rotation number and the same asymptotic type are totally ordered on ${\mathbb{R}}^2$. For more details see [@B1988].\
In a recent paper [@GK] we derived under the assumption of vanishing topological entropy strong properties for all geodesics on the universal covering. Early versions of the results where already obtained in [@G]. The main properties are summarized in the following theorem.
\[see [@GK]\] \[summarize\] Let $g$ be a Riemannian metric on $T^2$ with vanishing topological entropy. Then, on the Riemannian universal covering $( {\mathbb{R}}^2, \tilde{g})$, every geodesic $c$ is escaping, i.e., $\lim\limits_{t \to \pm \infty}\|c(t)\| =\infty$, has no self-intersections and for every geodesic the asymptotic direction $\delta(c)$ exists with the additional property that $\delta(c)=-\delta(c^-)$ if $c^-(t) =c(-t)$ is the geodesic traversed in the opposite direction. Furthermore, each geodesic $c$ lies in a strip in ${\mathbb{R}}^2$ bounded by two parallel Euclidean lines.\
Moreover, the asymptotic direction defines a continuous function $\delta: S\mathbb{R}^2 \to S^1$ given by $\delta(v) =\delta(c_v)$ such that for each $x \in \mathbb{R}^2$ the restriction $\delta_x: S_x\mathbb{R}^2\to S^1$ to the fibers of $ \mathbb{R}^2$ is surjective.
Theorem \[summarize\] implies $\delta(v) = - \delta(-v)$ and $\rho(v)=\rho(-v)$ for all $v \in ST^2$. Furthermore, $\delta$ and hence $\rho$ induces a flow invariant continuous function on $ST^2$.
The main technical ingredient used in the proof of this Theorem is the following Fundamental Lemma derived in [@GK]. It will be crucial in this paper as well.
Let $g$ be a Riemannian metric on $T^2$ and $\alpha: {\mathbb{R}}\to {\mathbb{R}}^2$ a minimal axis of the translation element $\tau$. Let $c_1:[0, a] \to {\mathbb{R}}^2$ and $c_2:[0, b] \to {\mathbb{R}}^2$ be two geodesic segments with endpoints on $\alpha$ and $$c_1((0, a)) \cap \alpha({\mathbb{R}})=\emptyset, \; c_2((0, b)) \cap \alpha({\mathbb{R}})=\emptyset.$$ Assume that there exists a translation element $\eta$, with $\eta\alpha({\mathbb{R}}) \cap \alpha({\mathbb{R}})= \emptyset$ such that $$\eta \alpha({\mathbb{R}}) {\;\;\makebox[0pt]{$|$}\makebox[0pt]{$\cap$}\;\;}c_1([0, a]) \not= \emptyset \; \; \text{and}
\; \; \eta^{-1} \alpha({\mathbb{R}}) {\;\;\makebox[0pt]{$|$}\makebox[0pt]{$\cap$}\;\;}c_2([0, b]) \not= \emptyset.$$ Then the metric $g$ has positive topological entropy.
The proof of this lemma heavily relies on the curve shortening flow, see [@GO]. As shown in the proof of Theorem II in [@GK] the Fundamental Lemma generalizes also to broken geodesic segments $c_1,c_2$ such that the exterior angles in the singularities of $c_1,c_2$ are smaller than $\pi$. The exterior angles of $c_1$ are the interior angles in the unbounded connected component of $H_1 \setminus c_1([0,a])$ where $H_1$ is the halfplane in ${\mathbb{R}}^2\setminus \alpha({\mathbb{R}})$ containing $c_1(0,a)$.
Structure of the geodesics in case of zero topological entropy
==============================================================
\[1\] Let $(T^2,g)$ be a Riemannian torus with vanishing topological entropy and $({\mathbb{R}}^2, \tilde{g})$ the Riemannian universal cover with the lifted metric $\tilde{g}$. Then any pair of geodesics $c_1, c_2 : {\mathbb{R}}\to {\mathbb{R}}^2$ with $\rho(c_1)
\not = \rho(c_2)$ has at most one intersection.
According to Theorem \[summarize\] no geodesic on the universal covering has self-intersections provided the topological entropy is zero. Assume that there exists a pair $c_1, c_2: {\mathbb{R}}\to {\mathbb{R}}^2$ of geodesics with two intersections and such that $\rho(c_1) \not = \rho(c_2)$. After reparameterization and change of orientation of the geodesics we may assume: $c_1(0)=c_2(0)$ and $c_1(t_1)=c_2(t_2)$ for $t_1,t_2 >0$. Since $\rho(c_2) \not = \rho(c_1)$ and $\delta(c_2) = - \delta(c_2^{-})$, there also exist times $t_3,t_4 <0$ such that $c_1(t_3) =c_2(t_4)$. Consider the piecewise geodesic curves we get by gluing $c_1([0, \infty))$ with $c_2((-\infty, 0])$ and analogously gluing $c_2([0, \infty))$ with $c_1((-\infty, 0])$.
\[c\][$c_1(- \infty,0]$]{} \[c\][$c_2(- \infty, 0]$]{} ![$V$-shaped broken geodesic segments in the proof of Lemma \[1\].](v-geod-neu.eps "fig:")
There exists a minimal axis $\alpha$ and a translation element $\eta$ such that subsegments of these $V$-shaped broken geodesic curves by construction fulfill the assumption of the Fundamental Lemma in [@GK]. Hence, we conclude positive topological entropy in contradiction to the assumption.
For $x \in {\mathbb{R}}^2$ we define a lift $\tilde \delta_x:{\mathbb{R}}\to {\mathbb{R}}$ of the asymptotic direction $\delta_x: S_x\mathbb{R}^2\to S^1$. Let $e_1 = (1,0)$ and $e_2 = (0,1)$ be the standard orthonormal basis in $T_x{\mathbb{R}}^2 \cong {\mathbb{R}}^2$. Choose an orthonormal basis $v_1, v_2 \in T_x{\mathbb{R}}^2$ with respect to the metric $g_x$ and the same orientation as $e_1,e_2$. Consider the coverings $p_1 : {\mathbb{R}}\to S_x\mathbb{R}^2$ and $p_2 :{\mathbb{R}}\to S^1$ given by $p_1(t) = \cos t v_1 + \sin t v_2$ and $p_2(t) = \cos t e_1 + \sin t e_2$. Since $\delta_x: S_x\mathbb{R}^2\to S^1$ is continuous there exists a lift $\tilde \delta_x:{\mathbb{R}}\to {\mathbb{R}}$, unique up to a multiple of $2 \pi$, defined by $$p_2(\tilde \delta_x(t)) = \delta_x(p_1(t)) \enspace.$$
\[2\] Let $(T^2,g)$ be a Riemannian torus with zero topological entropy. Then, for all $x \in T^2$ the lift $\tilde \delta_x:{\mathbb{R}}\to {\mathbb{R}}$ of the asymptotic direction $\delta_x: S_xT^2 \to S^1$ is a monotone function. Moreover, $ \tilde \delta_x(t + 2 \pi) -\tilde \delta_x(t) = 2 \pi$ and therefore the degree of $\delta_x$ is one.
Choose $0 \le t_1 < t_2 < \pi$. Then $w_1 = p_1(t_1)$, $w_2 = p_1(t_2)$ are positively oriented and $c_{w_1}({\mathbb{R}})$ and $c_{w_2}({\mathbb{R}})$ are contained in Euclidean strips corresponding to the directions $\delta_x(w_1)$ and $\delta_x(w_2)$, respectively. If $\delta_x(w_1) \not= \delta_x(w_2)$ the ray $c_{w_2}((0, \infty)$ is by Lemma \[1\] contained in a single connected component of ${\mathbb{R}}^2 \setminus c_{w_1}({\mathbb{R}})$. This implies that $\delta_x(w_1)= p_2(\tilde \delta_x(t_1))$, $\delta_x(w_2)=p_2(\tilde \delta_x(t_2))$ have the same orientation as $w_1, w_2$ and therefore are positively oriented as well. Hence, $ 0 <\tilde \delta_x(t_2)-\tilde \delta_x(t_1)< \pi$. Furthermore, $\tilde \delta_x(t +\pi)-\tilde \delta_x(t) = \pi$ for all $t$ which yields the second assertion.
Lemma \[2\] implies the following Corollary:
\[c1\] Let $(T^2,g)$ be a Riemannian torus with zero topological entropy and $({\mathbb{R}}^2, \tilde{g})$ the Riemannian universal cover. Then for all $x \in
{\mathbb{R}}^2$, $S_x {\mathbb{R}}^2$ is a disjoint union of the closed sets $$S_x^{r} = \{v \in S_x {\mathbb{R}}^2 \mid \rho(c_v) =r\} \enspace,$$ where $c_v: {\mathbb{R}}\to {\mathbb{R}}^2$ is a geodesic with $\dot c_v(0)=v$ and $r \in {\mathbb{R}}\cup
\{\infty\}$. For each $r$ the set $S_x^{r}$ consists of two connected closed antipodal components.
In the sequel we will need the following theorem which summarizes important results on minimal rays obtained by Bangert [@B1994].
\[B\] Let $(T^2,g)$ be a Riemannian torus and $({\mathbb{R}}^2, \tilde{g})$ the Riemannian universal cover.
1. Let $c_1: [0, \infty) \to \mathbb{R}^2 $ and $c_2: [0, \infty) \to \mathbb{R}^2 $ two minimal rays with $c_1(0) = c_2(0)$ and $\dot{c}_1(0) \not= \pm \dot{c}_2(0) $. Then, for each $\epsilon >0$ the rays $c_1: [-\epsilon, \infty) \to \mathbb{R}^2 $ and $c_2: [-\epsilon, \infty)\to \mathbb{R}^2 $ are not minimal, provided $c_1,c_2$ are asymptotic or $c_1,c_2$ have the same irrational asymptotic direction.
2. For any minimal ray $c:[0, \infty) \to \mathbb{R}^2$ with a rational rotation number $r$ there exists a unique pair of neighboring minimal axes $ \alpha_1, \alpha_2$ with rotation number $r$ and the following properties:
1. $c[0, \infty)$ is contained in the strip bounded by $ \alpha_1$ and $ \alpha_2$.
2. $c:[0, \infty) \to \mathbb{R}^2$ is asymptotic to $ \alpha_1$ or $ \alpha_2$.
3. For each $x$ between a pair of neighboring minimal axes $\alpha_1, \alpha_2$ with the same asymptotic direction there exist four different asymptotic minimal geodesic rays $c_1^+, c_1^-, c_2^+$ and $c_2^-$ with $c_i^{\pm}(0) = x$ for $i\in \{1,2\}$ and $$\lim\limits_{t \to \infty} \|c_i^{\pm}(t) - \alpha_i(\pm t+ s_k)\|=0$$ for some constants $s_k$ with $k \in \{1,2,3,4\}$.
Under the assumption that $c_1$ and $c_2$ are asymptotic, assertion (1) is an easy consequence of the triangle inequality. If $c_1$ and $c_2$ have the same irrational asymptotic direction assertion (1) follows from Theorem 3.6 in [@B1994].\
The assertions (2) and (3) follow from Theorem 3.7 in [@B1994].
\[3\] Let $(T^2,g)$ be a Riemannian torus with zero topological entropy and $({\mathbb{R}}^2, \tilde{g})$ the Riemannian universal cover. Let $x \in {\mathbb{R}}^2$ be fixed.\
Then for all $r \in {\mathbb{R}}\cup \{\infty\}$ and each $v
\in \partial S_x^r$ any geodesic $c_w:{\mathbb{R}}\to {\mathbb{R}}^2$ with $w \in S_x {\mathbb{R}}^2 \setminus \{v \}$ intersects $c_v$ only in $x$. In particular, $c_v:[0, \infty) \to {\mathbb{R}}^2$ and $c_{-v}:[0, \infty) \to {\mathbb{R}}^2$ are minimal rays.\
If $r$ is rational there exists a pair of minimal neighboring axes $\alpha_1$, $\alpha_2$ with the following properties: Each of the minimal rays $c_v$ corresponding to $v \in \partial S_x^r$ is asymptotic to $\alpha_1$ or $\alpha_2$, or it coincides with one of the axes.
For $r \in {\mathbb{R}}\cup \{\infty\}$ and $x \in {\mathbb{R}}^2$ consider $v \in \partial S_x^{r}$ and $w \in S_x {\mathbb{R}}^2$ with $w \not= v$. Assume that $c_w$ intersects $c_{v}( {\mathbb{R}}) \setminus \{x\}$. According to Lemma \[1\] this implies $\rho(w) = \rho(v)$. Choose a sequence $v_n \in S_x {\mathbb{R}}^2 \setminus S_x^{r}$ converging to $v$. By the continuous dependence of geodesics on initial conditions $c_{v_n}( {\mathbb{R}}) \setminus \{x \}$ intersects $c_w$ for sufficiently large $n$ as well. Since $\rho(v_n) \not= \rho(w)$ this contradicts Lemma \[1\]. In particular, both rays $c_v: [0, \infty) \to {\mathbb{R}}^2$ and $c_{-v}: [0, \infty) \to {\mathbb{R}}^2$ are minimal. If $r$ is rational and $v \in \partial S_x^{r}$, then the minimality of $c_v:[0, \infty) \to {\mathbb{R}}^2$ and Theorem \[B\] imply that $c_v$ is asymptotic to one of the two neighboring minimal axes with rotation number $r$.
\[4\] Let $(T^2,g)$ be a Riemannian torus with vanishing topological entropy. Then the set $\{\partial S_x^r \mid x \in \mathbb{R}^2\}$ is invariant under the geodesic flow. If $r$ is irrational, then each $S_x^{r}$ consists of a pair of antipodal vectors $\{ \pm v \}$.
Consider $v \in \partial S_x^r$ and choose $t > 0$. Since $\rho$ is flow invariant we have $\phi^t(v) =:v_t \in S_{c_v(t)}^r$. We claim that $v_t \in \partial S_{c_v(t)}^r$. Assume first that $r$ is irrational. If $v_t \not \in \partial S_{c_v(t)}^r$, choose $w \in \partial S_{c_v(t)}^r$. But this contradicts Theorem \[B\] since $c_{v_t}: [-t, \infty) \to \mathbb{R}^2$ is minimal. Assume that $r$ is rational. Then $c_v$ is forward asymptotic to a minimal axis $\alpha$. If $v_t \not \in \partial S_{c_v(t)}^r$, there exists $w \in \partial S_{c_v(t)}^r$ such that $c_w$ is also asymptotic to $\alpha$. But then $c_{v_t}$ is asymptotic to $ c_w$ which contradicts Theorem \[B\]. Hence, we obtain $v_t=w$.\
Since $-v \in \partial S_x^r$ we have $\phi^t(-v) \in \partial S_{c_{-v}(t)}^r = \partial S_{c_{v}(-t)}^r$ for all $t >0$. But this implies $\phi^{-t}(v) = -\phi^t(-v) \in \partial S_{c_{v}(-t)}^r$ and therefore the flow invariance of $\{\partial S_x^r \mid x \in \mathbb{R}^2\}$ for negative times.
The next Theorem gives a complete characterization of minimal geodesics in the case of vanishing topological entropy:
\[5\] Let $(T^2,g)$ be a Riemannian torus with zero topological entropy and $({\mathbb{R}}^2, \tilde{g})$ the Riemannian universal cover. Then, for all $x \in {\mathbb{R}}^2$ it follows:
1. For all $r \in {\mathbb{R}}\cup \{\infty\}$ the geodesics $c_v$ with $v \in \partial S_x^r$ are minimal.
2. For all $r \in {\mathbb{R}}\setminus {\mathbb{Q}}$ we have $S_x^r =\partial
S_x^r =\{\pm v\}$. Hence, for each $r \in {\mathbb{R}}\setminus {\mathbb{Q}}$ there exists an up to orientation and parametrization unique geodesic with rotation number $r$ passing through $x$. Furthermore, this geodesic is minimal.
3. For $r \in {\mathbb{Q}}\cup \{\infty\}$ consider the pair of neighboring minimal axes $\alpha_1, \alpha_2$ with rotation number $r$ bounding a strip $S$ containing $x$.
1. If $\alpha_1= \alpha_2$ we have $S_x^r =\partial S_x^r =\{\pm v\}$. Hence, there exists an up to orientation and parametrization unique geodesic with rotation number $r$ passing through $x$. Furthermore, this geodesic is a minimal axis.
2. If $\alpha_1 \not= \alpha_2$ the set $S_x^r$ is a disjoint union of two antipodal connected sets bounded by two minimal geodesics, where each geodesic is forward and backward asymptotic to the pair of neighboring axes $\alpha_1, \alpha_2$. The interior of $S_x^r$ contains no minimal rays.\
Moreover, for all points $y$ in the interior of the strip $S$, the interior of the set $S_y^r$ is non-empty.
<!-- -->
1. Consider $v \in \partial S_x^{r}$ and $t>0$. Proposition \[4\] implies $\phi^{-t}(v) \in \partial S_{c_{v}(-t)}^r$ and hence by Lemma \[3\] the geodesic $c_v :[-t, \infty) \to {\mathbb{R}}^2$ is minimal. Since $t$ is arbitrary $c_v: {\mathbb{R}}\to {\mathbb{R}}^2$ is minimal.
2. For $r \in {\mathbb{R}}\setminus {\mathbb{Q}}$ consider $v, w \in \partial S_x^r$ and $v \not= -w$. Using (1) the geodesics $c_v$ and $c_w$ are minimal which by Theorem \[B\] implies $v=w$.
3. Consider $r \in {\mathbb{Q}}\cup \{\infty \}$ and $v \in \partial S_x^r$. Then $c_v$ is asymptotic to one in the pair of neighboring minimal axes $\alpha_1, \alpha_2$ with rotation number $r$.
1. If $\alpha_1= \alpha_2$ , the geodesic $c_v$ coincides with the axis up to orientation and $\partial S_x^r = \{ \pm v \}$.
2. If $\alpha_1 \not= \alpha_2$ and $c_v $ is asymptotic to say $\alpha_1$ there is $w \in \partial S_x^r$ with $w \not= \{\pm v\}$ such that $c_w $ is asymptotic to $\alpha_2$.\
Let $c_u$ be another minimal geodesic ray with $u \in S_x^r\setminus \partial S_x^r$. Then by Theorem \[B\] (2) $c_u$ is asymptotic to one of the minimal neighboring axes and hence to a geodesic $c_v $ with $v \in \partial S_x^r$. Since $c_v $ is minimal this contradicts Theorem \[B\] (1). The last assertion follows from Corollary \[c1\] and Theorem \[B\] (3).
Now we are able to prove our main theorem stated in the introduction:
By Theorem \[5\] for $r \in {\mathbb{R}}\setminus {\mathbb{Q}}$ the universal covering is foliated by minimal geodesics with rotation number $r$. Since the foliation is unique it is preserved by translation elements. Hence, geodesics with irrational rotation numbers do not intersect their translates. For $r \in {\mathbb{Q}}\cup \{\infty\}$ we distinguish the following two cases:
1. either the universal covering is foliated by minimal axes with rotation number $r$,
2. or between a pair of neighboring axes $\alpha_1, \alpha_2$ with rotation number $r$ there exist two different foliations by minimal geodesics asymptotic to $\alpha_1$ and $\alpha_2$ with rotation number $r$.
If for all rational rotation numbers the foliation is unique it consists of minimal axes. Then all geodesics are minimal and there exist no conjugate points. Then, by Hopf’s Theorem [@H] the Riemannian two-torus is flat.
Surprisingly the only known metrics on $T^2 = \mathbb{R}^2/ \mathbb{Z}^2$ with zero topological entropy are the so called Liouville metrics which are of the form $$ds^2 = (f(x) + g(y)) (dx^2 + dy^2)$$ where $f,g: \mathbb{R} \to \mathbb{R}$ are strictly positive smooth $1$-periodic functions. The geodesic flow of such a metric is integrable (see e.g. [@BF]). If $f,g$ are not constant the only non unique minimal geodesic foliations correspond to the directions $(1,0)$ and $(0,1)$. If only $f$ or $g$ is constant there is exactly one such direction. If both are constant no such direction exists since the metric is flat (see [@Sch]).
We like to close with some important and intriguing open problems.
$\;$
1. Is the converse of our main theorem true, i.e. does the following hold? Given a Riemannian metric on $T^2$ such that for each rotation number $r \in \mathbb{R} \cup \{\infty \}$ there exists a foliation of lifted minimal geodesics on $\mathbb{R}^2$ with rotation number $r$. Does this imply that the topological entropy is zero?
2. Has each metric with zero topological entropy on $T^2$ only finitely many rational directions with non unique minimal geodesic foliations? Is there at least one rational direction where the foliation is unique?
3. Is in case of zero topological entropy the geodesic flow of a Riemannian metric on $T^2$ integrable in the sense of Liouville and Arnold?
4. Finally we like to add the following longstanding open question which was raised by Kozlov, Fomenko, Sinai and others (see e.g. [@BF]). Do there exist besides the Liouville metric other metrics on $T^2$ with integrable geodesic flows? It has been conjectured by Fomenko and Kozlov (see [@BF]) that the answer is no.
[99]{} S. B. Angenent, *Monotone recurrence relations, their Birkhoff orbits and topological entropy*, Ergod. Th. and Dynam. Sys., **10** (1990), 15-41. S. B. Angenent, *A remark on the topological entropy and invariant circles of an area preserving twistmap*, In ’Twist mappings and their Applications’ IMA Volumes in Mathematics, **44**, Eds R.Mc Gehee K. Meyer, Springer (1992), 1-5. S. Aubry, P. Y. Le Daeron, *The discrete Frenkel-Kontorova model and its generalizations*, Physica, **8D** (1983), 381-422. V. Bangert, *Mather Sets for Twist Maps and Geodesics on Tori*, Dynamics Reported **1** (1988), 1-56. V. Bangert, *Geodesic rays, Busemann functions and monotone twist maps*, Calc. Var. Partial Differential Equations 2, **1** (1994), 49–63. M. L. Bialy and L. V. Polterovich, *Geodesic Flows on the two-dimensional Torus and Phase Transitions ’Commensurability-Noncommensurability’*, English translation: Functional Anal. Appl. **20** (1986) no. 4, 260-266. S. V. Bolotin and P. H. Rabinowitz, *Some geometrical conditions for the existence of chaotic geodesics on a torus*, Ergod. Th. and Dynam. Sys. **22** (2002), 1407-1428. A. V. Bolsinov, A. T. Fomenko *Integrable Geodesic Flows on two-dimensional surfaces*, Monographs in Contemporary mathematics, (1999). R. E. Bowen, *Entropy for Group Endomorphisms and Homogeneous spaces*, Trans. of Am. Math. Soc. **153** (1971), 401-414. J. Denvir and R. S. MacKay, *Consequences of contractible geodesics on surfaces*, Trans. Amer. Math. Soc. **350** (1998), no. 11, 4553-4568. E. Glasmachers, *Characterization of Riemannian metrics on $T^2$ with and without positive topological entropy*, Thesis 2007.\
`http://www-brs.ub.rub.de/netahtml/HSS/Diss/GlasmachersEva/` E. Glasmachers, G. Knieper, *Characterization of Geodesic Flows on $T^2$ with and without Positive Topological Entropy*, GAFA, Geom. Funct. Anal. **20** (2010) no. 5, 1259-1277. M. A. Grayson, *Shortening embedded curves*, Ann. of Math. (2) **129** (1989), no. 1, 71-111. G. A. Hedlund, *Geodesics on a two-dimensional Riemannian manifold with periodic coefficients*, Ann. of Math. **33** (1932), 719-739. E. Hopf, *Closed surfaces without conjugate points*, Proc. Natl. Acad. Sci. **34** (1948), 47-51. A. Katok, B. Hasselblatt, *Introduction to the Modern Theory of Dynamical Systems*, Cambridge University Press (1995). A. Katok, *Lyapunov Exponents, Entropy and Periodic Orbits for Diffeomorphisms*, Inst. Hautes Études Sci. Publ. Math. **51** (1980), 137-173. J. N. Mather, *Existence of quasi-periodic orbits for twist homeomorphisms of the annulus*, Topology, **21** (1982), 457-67. H. M. Morse, *A fundamental class of geodesics on any closed surface of genus greater than one*, Trans. Amer. Math. Soc. **26** (1924), 25-60. I. V. Polterovich, *On a characterization of flat metrics on 2-torus*, J. of Dynamical and Control Systems **2** (1996), 89-101. J. P. Schröder, *Beschreibung der Geodätischen auf einem Torus mit Liouvillemetrik*, Bachelorarbeit, Ruhr-University Bochum, 2011. P. Walters, *An Introduction to Ergodic Theory*, Graduate Texts in Mathematics, Springer-Verlag, New York, Berlin, Heidelberg (1982).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Recent discoveries of supposedly pure $\alpha$-tetragonal boron require to revisit its structure. The system is also interesting with respect to a new type of geometrical frustration in elemental crystals, which was found in $\beta$-rhombohedral boron. Based on density functional theory calculations, the present study has resolved the structural and thermodynamic characteristics of pure $\alpha$-tetragonal boron. Different from $\beta$-rhombohedral boron, the conditions for stable covalent bonding (a band gap and completely filled valence bands) are almost fulfilled at a composition B$_{52}$ with two $4c$ interstitial sites occupied. This indicates that the ground state of pure $\alpha$-tetragonal boron is stoichiometric. However, the covalent condition is not perfectly fulfilled because non-bonding in-gap states exist that cannot be eliminated. The half occupation of the $4c$ sites yields a macroscopic amount of residual entropy, which is as large as that of $\beta$-rhombohedral boron. Therefore, $\alpha$-tetragonal boron can be classified as an elemental crystal with geometrical frustration. Deviations from stoichiometry can occur only at finite temperatures. Thermodynamic considerations show that deviations $\delta$ from the stoichiometric composition (B$_{52+\delta}$) are small and positive. For reported high-pressure syntheses conditions $\delta$ is predicted to be about 0.1 to 0.2. An important difference between pure and C- or N-containing $\alpha$-tetragonal boron is found in the occupation of interstitial sites: the pure form prefers to occupy the $4c$ sites, whereas in C- or N-containing forms a mixture of $2a$, $8h$, and $8i$ sites are occupied. The present article provides relations of site occupation, $\delta$ values, and lattice parameters, which enable us to identify pure $\alpha$-tetragonal and distinguish the pure form from other ones.'
author:
- Naoki Uemura
- Koun Shirai
- Hagen Eckert
- Jens Kunstmann
title: 'Structure, non-stoichiometry, and geometrical frustration of $\alpha$-tetragonal boron'
---
Introduction {#sec:intro}
============
The existence of $\alpha$-tetragonal ($\alpha$-T) boron has long been a controversial issue. Historically, $\alpha$-T boron was first isolated in 1943 [@Laubengayer43], and tentatively identified as B$_{50}$, which is composed of four tetrahedrally coordinated B$_{12}$ icosahedra and two interstitial B atoms (see Fig. \[fig:struct\]).[@Hoard51] But soon this structural model was doubted by theorists. Longuet-Higgins and Roberts showed that the B$_{50}$ structure is unstable due to electron deficiency.[@LH55] Later, experimental groups showed that the actual crystals contained C or N atoms [@Amberger71; @Ploog72; @Will76] and the chemical compositions were approximately B$_{50}$N$_{2}$ or B$_{50}$C$_{2}$. This point of view was further supported by DFT calculations.[@Morrison92; @Lee92a] Since then, it is almost generally accepted that pure $\alpha$-T form does not exist and that the structure is stabilized only by inclusion of foreign atoms.
In the 2000’s, nanostructures were repeatedly reported to have the structure of pure $\alpha$-T boron.[@Zhang02; @Wang03; @Xu04; @Yang04; @Kirihara05] Hayami and Otani explained the existence of pure $\alpha$-T boron in nanostructures by the presence of low surface energies.[@Hayami08] Subsequently, they predicted the possibility of pure $\alpha$-T boron bulk structures with the composition B$_{52}$.[@Hayami10] After this theoretical work, the high-pressure syntheses of bulk $\alpha$-T boron were reported by several groups [@Ekimov11; @Ekimov11a; @Qin12; @Kurakevych12; @Kurakevych13; @Solozhenko13]. Notwithstanding the prediction by Hayami and Otani, many different forms of tetragonal boron were discovered, including a known form of $\beta$-tetragonal boron [@Ma03]. The current situation occurs to be rather complicated and many questions are raised: Does pure $\alpha$-T boron truly exists? And if it does, what stabilizes its structure and what was wrong with the previous theory? The purpose of this paper is to answer these questions.
To say it shortly, the qualitative argument of Longuet-Higgins and Roberts is not wrong. The structure of B$_{50}$ is indeed not stable. However, there are ingenious ways to circumvent this instability.[@Shirai12] One way is a deviation from stoichiometry. A suitable deviation from stoichiometry can lead to complete valence band filling and thus stabilize the structure of a covalent crystal. This explanation however raises a new question: Why was pure $\alpha$-T boron not found in bulk phases until recently? A simple answer is that there are other phases which are more stable at ambient conditions, [*i.e.*]{}, $\alpha$- and $\beta$-rhombohedral ($\alpha$-R and $\beta$-R) boron. However, the relative stability of these polymorphs can be changed at high pressures, which is a different way for stabilizing $\alpha$-T boron. To clarify the relative stability of boron phases, a detailed comparison is required, which will be addressed in another study. Here, we concentrate on $\alpha$-T boron, only and identify the best structure that pure $\alpha$-T boron can have, if it exists.
There are several theoretical studies on the structure of $\alpha$-T boron.[@Lee92a; @Morrison92; @Hayami10; @Aydin11; @Aydin12] However, these studies assumed stoichiometry. From our point of view [@Shirai12], this assumption leads to an incorrect description of the metal/insulator behavior. The important role of non-stoichiometry has been recognized only recently and it will therefore be described in Sec. \[sec:background\]. Some words are also given there concerning the terminologies related to defect states, because they are used differently in different fields of science. This section also provides an overview and analysis of recent experiments, which is necessary because the current situation is very complicated. Through this analysis, we are able to delineate a clear approach for identifying pure $\alpha$-T boron. After describing the calculation method (Sec. \[sec:method\]), the article studies the electronic structures of individual interstitial sites (IS) in Sec. \[sec:bondnature\]. By considering all the involved IS, the driving force for the deviation from stoichiometry is clarified in Sec. \[sec:supercell\]. Based on this analysis, a deviation is predicted for various synthesis conditions (Sec. \[sec:high-pressure\]). An assessment on the characterization of recently discovered $\alpha$-T boron is attempted in Sec. \[sec:lattice-para\]. Behind the deviation from stoichiometry, we can say more about why $\alpha$-T boron is not satisfied with a simple structure B$_{52}$. Section \[sec:frustration\] is devoted to discuss this topics from a view point of geometrical frustration. In the last section, we summarize our results.
Theoretical background and experimental facts {#sec:background}
=============================================
For a long time solid-state theory of boron crystals had a fundamental problem. For many boron-rich crystals, band structure calculations showed metallic behavior, whereas in experiments all the crystals were found to be semiconductors. We call this problem the [*metal/insulator*]{} problem. Table \[tbl:stoichiometry\] summarizes this discrepancy. This table is adapted from a paper of R. Schmechel and H. Werheit [@Schmechel99], who first pointed out the importance of deviation from stoichiometry for the metal/insulator problem. The last entry of the table corresponds to $\alpha$-T boron, where for convenience the conclusion of this study is already given. We can notice in this table that, whenever the metal/insulator problem occurs, deviation from stoichiometry is observed. In boron-rich solids, the crystals are composed of a regular arrangement of icosahedra and various kinds of atoms at interstitial sites. Those interstitial sites are either fully occupied sites (FOS) or partially occupied sites (POS). The deviation from stoichiometry is a result of the presence of POS.
--------------------- -------------- -------------- -------------- ------------ -------------- --------------------------- ----------
Crystal
Sym. $N_{\rm at}$ $N_{\rm el}$ Prediction $N_{\rm at}$ Ground state Elect.
$\alpha$-rhomb. $D_{\rm 3d}$ 12 36 semicon. 12 H-sym. semicon.
St.
$\beta$-rhomb. $D_{\rm 3d}$ 105 [*315*]{} metal 106.5 L-sym. semicon.
NSt.
${\rm B_{13}C_{2}}$ $D_{\rm 3d}$ 15 [*47*]{} metal 15 L-sym. semicon.
(on average) NSt.
$\alpha$-tetra. $D_{\rm 4h}$ 50 150 metal 52+$\delta$ $\delta=0$ for $T=0$ semicon.
52 156 semicon. $\delta \neq 0$ for $T>0$
--------------------- -------------- -------------- -------------- ------------ -------------- --------------------------- ----------
: A comparison of hypothetical (theoretically considered) and real structures of boron-rich solids. For most systems theory was unable to correctly predict their semiconducting properties (second column). The experimentally determined electronic properties are given in the last column. For the real structures, some ground-state properties are indicated, such as high- or low-symmetry (H-Sym./L-Sym.) and stoichiometric or non-stoichiometric compositions (St./NSt.). The experimentally determined electronic properties are given in the last column. $N_{\rm at}$ is the number of atoms per unit cell. Odd numbers of electrons $N_{\rm el}$ are indicated by italicized numbers. The last entry of this table contains our conclusion about the structure of $\alpha$-T boron, that is presented in this article.[]{data-label="tbl:stoichiometry"}
It is well known that boron-based materials are characterized as [*electron deficient*]{} systems in the chemical literature.[@Pauling] Electron deficiency and the metal/insulator problem occur unrelated and therefore we had not be aware until recently that these two different matters are actually intimately related. In the last decade, the effect of POS was extensively studied by DFT calculations on $\beta$-R boron.[@Masago06; @Setten07; @Widom08; @Ogitsu09] It was found that the occupation of IS in boron-rich solids cannot be considered as defects, contrary to the usual notion in the literature about defect physics. The interstitial atoms are a part of the host crystals, in a sense that the energy is lower than that of the perfect crystals. The deviation from stoichiometry in $\beta$-R boron led Ogitsu and Widom to conceive a new concept of [*geometrical frustration*]{}.[@Ogitsu09; @Ogitsu10; @Widom08] Shirai and Uemura have finally established a relationship between the deviation from stoichiometry and the metal/insulator problem.
The mechanism behind the insulating property of boron crystals, which are otherwise predicted as metals, can be elucidated by the following way.[@Shirai12] (i) The system has an odd number of electrons ($N_{\rm el}$), which is a consequence of electron deficiency. Then band theory unequivocally predicts it to be a metal. (ii) The system has strong covalency. This requires an even number of electrons. These two conditions (i) and (ii) are mutually incompatible. This competition yields a strong driving force to modify chemical bonds. Reconstruction of chemical bonds is usually difficult because of high energy barriers which exist between different configurations. By combining with the third condition (iii) of a large unit cell size (large number of atoms $N_{\rm at}$ per cell), the energy barrier can largely be reduced. The deviation from stoichiometry is an efficient solution for reconciling the electron deficiency and the covalency. Using this mechanism, difficult problems with the electronic properties of B$_{13}$C$_{2}$ were resolved.[@Shirai14; @Ektrarawong15]
The above-mentioned mechanism should naturally work for $\alpha$-T boron, too. The initially assumed structure of $\alpha$-T boron is B$_{50}$, which is composed of four icosahedra and two $2b$-site B atoms (Fig. \[fig:struct\]). The $2b$ site is almost perfectly occupied (see Table \[tbl:alpha-tetra\]). Therefore B$_{50}$ was believed to be stoichiometric. Later we will show, however, that this model should be revised. Although the hypothetical structure B$_{50}$ has an even number of electrons and condition (i) does not hold, the valence-electron counting indicates that the covalent condition (ii) is not satisfied, because the system is short of ten electrons to completely fill up the valence band. This is the reason why Longuet-Higgins and Roberts cast doubt on the existence of B$_{50}$.[@LH55] Their conclusion is essentially correct even in modern DFT calculations. See, for example, Fig. 3 of Ref. 23 where the top five bands of the valence band are unoccupied.
Aside from the $2b$ site, many IS, which are partially occupied, have been reported (see Fig. \[fig:struct\]). The occupancies are listed in Table \[tbl:alpha-tetra\]. As described in the Introduction, the $\alpha$-T boron crystals reported in the last century are now believed to contain C or N impurities and the crystals reported by Hoard [@Hoard58] may belong to this class, too. Hereafter we will call these crystals [*traditional*]{} $\alpha$-T boron. Such traditional $\alpha$-T boron always contains impurities such as C or N, and exhibits a deviation from stoichiometry.
Author Hoard [@Hoard58] Hyodo [@HyodoTh] Ekimov [@Ekimov11] Qin [@Qin12] Kurakevych [@Kurakevych12]
------------- ------------------ ----------------------- ----------------------- -------------------- ------------------ ---------------------------- --------------------
Formula ${\rm B_{50}}$ ${\rm B_{50}C_{1.9}}$ ${\rm B_{50}N_{1.8}}$ ${\rm B_{52.2}}$ ${\rm B_{51.5}}$ ${\rm B_{50}}$ various C conc.
including 0%
$a_{0}$ (Å) 8.750 8.753 8.634 8.808 9.0508 8.71 $ 8.775 \sim 8.93$
0 $+0.03$ $-1.3$ $+0.63$ $+2.8$ $-0.46$ $+0.28 \sim +2.1$
$c_{0}$ (Å) 5.060 5.093 5.128 5.047 5.1341 5.00 $5.064 \sim 5.08$
0 $+0.65$ $+1.3$ $-0.27$ $+1.5$ $-0.01$ $+0.08 \sim +0.40$
Site
$2a$ - 12.8 1.4 11 -
$2b$ 100 \(C) 90.8 \(N) 92.9 93 100
$4c$ - - - 0 31
$4g$ - - - - 6
$4d$ - - - 0 -
$8h$ - 11.2 2.6 2 -
$8i$ - 9.8 23 24 -
: A comparison of different $\alpha$-tetragonal boron structures that were reported in the literature. The lattice parameters and occupancies of partially occupied sites vary greatly among the different samples. Hoard’s and Will’s forms are traditional $\alpha$-T boron, and the others are recently discovered forms. The lattice parameters $a_0$ and $c_0$ are compared by evaluating the difference from Hoard’s crystal in %. For Qin’s structure, the name ${\rm B_{50}}$ is only nominal and does not indicate the accurate composition.[]{data-label="tbl:alpha-tetra"}
Recently, motivated by a quest for pure $\alpha$-T boron, Hayami and Otani showed that B$_{52}$, with two occupied $4c$ sites, is the lowest-energy structure, and they suggested that pure $\alpha$-T boron exists in bulk form.[@Hayami10] Hereafter, when speaking of non-stoichiometry, we refer to ${\rm B_{52}}$ as the stoichiometric composition, for a reason which will be clear below. The deviation from stoichiometry is expressed as ${\rm B_{52+\delta}}$, with a small fractional number $\delta$. Incidentally, several articles reporting the synthesis of pure $\alpha$-T boron have been published.[@Ekimov11; @Ekimov11a; @Qin12; @Kurakevych12; @Kurakevych13; @Solozhenko13] These newly discovered $\alpha$-T forms should be seen as distinct from the traditional boron, because in most cases they were obtained by solid-state phase transformation from $\beta$-R boron. Thus they are very likely to be truly pure boron crystals. Unfortunately, the chemical compositions in the new $\alpha$-T borons are not well analyzed. For example, Qin’s $\alpha$-T boron was designated as ${\rm B_{50}}$.[@Qin12] However, this is only because they could not measure the chemical composition accurately. Owing to the lack of reliable characterization, it is not even certain if the reported $\alpha$-T borons are all the same. In this paper, we shall refer to the reported $\alpha$-T borons by the name of the first author, for example, Ekimov’s $\alpha$-T boron. Kurakevych’s $\alpha$-T borons are actually a series of crystals containing a C coagent ranging from 0 to 5 at.%.[@Kurakevych12]
Hyodo analyzed Kirihara’s samples, reporting a slight non-stoichiometry ($\delta$=+0.2).[@HyodoTh] However, Kirihara’s $\alpha$-T boron comes from nanostructures[@Kirihara05] and therefore his analysis will not be discussed on the same ground as the results for bulk crystals. Only Ekimov [*et al*]{} clarified the chemical composition of their bulk $\alpha$-T boron, the value being 51.5 ($\delta$=-0.5).[@Ekimov11; @Ekimov11a] As shown in Table \[tbl:alpha-tetra\], Ekimov’s $\alpha$-T boron has the largest lattice parameter $a_{0}$ ever obtained. Our calculations show that the lattice parameters shrink when C or N impurities occupy the $2b$ site. This was also reported experimentally.[@Ekimov13] From this, Ekimov [*et al*]{} claimed that their crystals are pure $\alpha$-T boron. Because of their synthesis method (pyrolysis of decaborane ${\rm B_{10}H_{14}}$) the possibility of hydrogen incorporation is not excluded. Their chemical analysis, however, did not indicate H content as high as it influences the lattice parameter.
On the basis of the above analysis, we can summarize the current state of $\alpha$-T boron research as follows: If the recently synthesized $\alpha$-T borons are pure ones, there is little doubt that they are non-stoichiometric; in contrast theory predicts stoichiometric B$_{52}$ to be the lowest-energy structure. The question is thus, if the deviation from stoichiometry in real crystals is an intrinsic property (lowest-energy state) or not. For $\beta$-R boron[@Setten07; @Widom08; @Ogitsu09] and ${\rm B_{13}C_{2}}$ [@Shirai14], the deviation is an intrinsic property. If this is not the case for $\alpha$-T boron, the deviation must be an extrinsic property, caused by entropic effects at high temperatures. Then the degree of deviation would depend on the preparation conditions. Accordingly, one should characterize the real crystals with respect to the preparation conditions.
Computational Methods {#sec:method}
=====================
The electronic structures of $\alpha$-T boron were studied by density functional theory using a pseudopotential method and the Osaka2k code.[@Osaka2k] It uses the LDA parameterization by Perdew and Zunger [@PZ81], the Perdew-Burke-Ernzerhof form of the generalized gradient approximation (GGA) [@PBE96] and Troullier-Martins pseudopotentials [@TM91] with the fully separable Kleinman-Bylander form [@KB82]. In all the cases, the kinetic cutoff energy was 40 Ry. Various $k$-point sampling methods were used. For calculations using the primitive unit cell, a $4 \times 4 \times 4$ grid was used and for supercell calculations $\Gamma$-point only sampling. The convergence was well tested in our previous studies.[@Masago06; @Dekura11; @Shirai14] The formation energy $\Delta E_{f}$ is defined as the difference of the total energy with respect to a reference state, which was one form of $\alpha$-T boron B$_{m}$ (the stoichiometric composition, $m$=52, in most cases). Then, $\Delta E_{f}$ of a composition B$_{m+n}$ is obtained by, $$\Delta E_{f}[{\text B}_{m+n}] = E[{\text B}_{m+n}] - \frac{m+n}{m}E[{\text B}_{m}],
\label{eq:formE}$$ where $E[{\text B}_{m}]$ is the total energy of B$_{m}$. With this definition of $\Delta E_{f}$, a negative value implies a more stable structure. A similar definition is used to indicate the change of the specific volume $\Delta V$, where $E$ is substituted by $V$ in Eq. (\[eq:formE\]) and $\Delta V$ is evaluated in units of eV/GPa. Structural optimizations were performed with respect to atomic positions and cell parameters and no constraints on the crystal symmetry were imposed, except the tetragonal symmetry for the lattice parameters. Optimizing the cell parameters is important in view of the study of high-pressure phases. LDA was used only for supercell calculations, because of its computational efficiency, otherwise GGA was used. Of course, in this case, comparisons were made only between cells with the same size.
Comparisons between impurity-containing $\alpha$-T boron (impurity $X$=C or N) were also performed when necessary. In this case, compositions ${\rm B_{50}X_{2}}$ were assumed, where $X$-atoms were place on the $2b$ sites and the interstitial B atoms were placed either at the $8h$ or the $8i$ sites.
Results and Discussion
======================
The structure of $\alpha$-tetragonal boron {#sec:bondnature}
------------------------------------------
### Formation energy of interstitial sites {#sec:formaton-energy}
The formation energies of various IS were extensively studied by Hayami and Otani.[@Hayami10] In this work we want to study the roles of IS in more detail. We constructed a B$_{50}$ structure, where the $2b$ sites are fully occupied, and then placed B atoms at various IS, one by one. The results for $\Delta E_{f}$ are shown in Table \[tbl:HfofB50\]. However, care is needed for the interpretation of these values because usually an energy barrier is to be expected between two local energy minimum sites. Between the $4c$ and $8i$ sites, Hayami and Otani reported a small energy barrier, less than 10 meV. [@Hayami10] The values for $\Delta E_{f}$ in Table \[tbl:HfofB50\] were obtained by performing steepest-descent minimization with typically seven or eight iterations, which usually is enough to locate local minimum configurations. However, started from an $8i$ site, we found that continuing the iteration by more than 30 times finally brought the $8i$-site atom to the $4c$ site (see Supplemental Material). In this sense, $8i$ or $8h$ are metastable sites. However, to facilitate comparison with other calculations and experiments, we will still consider these sites as distinct, below.
The configuration B$_{52}$ with two $4c$ sites in the out of plane configuration (two $4c$-site atoms located in different $4c$ planes and not being neighbors) is the lowest-energy configuration. Hereafter, we designate this configuration as B$_{50}$ + 2B$_{4c}$. As mentioned above, all formation energies are defined with respect to this configuration.
Let us now consider only the blocks with $N_{\rm at}=51$ or less in Table \[tbl:HfofB50\]. We see that $\Delta E_{f}$ of many IS are lower than that of B$_{50}$ (3.14 eV). This fact already indicates that these interstitial B atoms are not defects but are a part of the host structure. The order of the stability of these IS is $4c$ (-1.45), $8i$ (-1.35), and $8h$ (-1.28) (the values in parenthesis are differences in $\Delta E_{f}$ with respect to B$_{50}$). The $2b$ site is always a stable site, as seen from its perfect occupancy, i.e., removing a $2b$-site atom from B$_{50}$ requires an energy of +1.45 eV. The sites $2a$ (0.66) and $4g$ (0.98) are not stable.
-------------- ----------------------------- ------ ------------ --------
$N_{\rm at}$ Sites $\Delta V$
(eV/GPa)
49 $-2b$ 4.59 (4.76) 0.068
50 3.14 (3.40) 0.052
51 $4c$ 1.69 (1.66) 0.034
$8h$ 1.72 (1.69) -0.047
$8i$ 1.80 (1.81) 0.035
$4g$ 4.12 (4.13) 0.036
$2a$ 3.80 (3.52) 0.049
52 $2\times 4c$ (out-of-plane) 0.0 (0.0) 0.0
$2\times 4c$ (in-plane) 0.19 (0.15) -0.006
$2\times 4c$ (nearest) 1.80 (1.82) -0.006
$4c+8i$ 0.19 (0.13) -0.005
$4c+8h$ 1.46 (1.49) 0.021
53 $3\times 4c$ 1.31 (1.11) -0.028
$2\times 4c+8i$ 2.68 (2.65) -0.036
$2\times 4c+8i'$ 2.67 (2.65) -0.037
$2\times 4c+8h$ 1.33 (1.10) -0.024
54 $4\times 4c$ 2.40 (2.00) -0.069
-------------- ----------------------------- ------ ------------ --------
: The formation energy $\Delta E_{f}$ of interstitial sites in $\alpha$-T boron within the primitive unit cell. The results indicate that B$_{52}$, with two 4c sites occupied, is the lowest-energy structure. The reference for $\Delta E_{f}$ and the change in volume $\Delta V$ is B$_{52}$ with two $4c$-site atoms in the out-of-plane configuration. In the second column ’-2b’ indicates a removal of a 2b atom. $2\times 4c$ should be read as two atoms at $4c$ sites. All the data are obtained by GGA, while LDA results are added for $\Delta E_{f}$ in parentheses. []{data-label="tbl:HfofB50"}
The energy ordering in $\Delta E_{f}$ can be understood by inspecting the bonding environment of these IS as shown in Fig. \[fig:POS\]. As seen, the $2b$ site is most perfectly tetrahedrally coordinated, which is the best configuration for a $sp^{3}$-type covalent bond. The short bond length of 1.63 Å also supports its strong covalent character. Hence, it is reasonable to obtain the perfect occupancy for this site. The $2a$ site also has perfect tetrahedral coordination, too. But, the nearest neighbor distance of 2.15 Å is too long. Furthermore, none of the atoms of the neighboring B$_{12}$ icosahedron are oriented in the direction linking the $2a$ site to B$_{12}$. Therefore, the bonding of $2a$-site atoms is weak. The site $4c$ is also close to tetrahedral coordination with a bond length of 1.79 Å, however the bond angles are largely distorted in the $ab$ plane. Although the $8i$ and $8h$ sites have short bond lengths (1.62 and 1.73 Å, respectively), there are only two of these short bonds, and the overall bonding is weaker than of $4c$-site atoms.
To summarize, the IS of $\alpha$-T boron are not defect states. Their role is not just to fill space between the icosahedra, but it is a more active one, i.e., the enhancement of the host structure by forming tetrahedral bonds. The closer the sites are to the perfect tetrahedral coordination, the stronger is the bonding.
\
### Comparison with C- or N-containing $\alpha$-tetragonal boron {#sec:impurity-tetra}
Although the positions of $8h$ sites $(0,1/2,u)$ and $8i$ sites $(u,0,1/2)$ are distinct from $4c$ sites $(0,1/2,0)$ by symmetry, they can be continuously connected to the $4c$ site by varying $u$. However, the bonding environments of $8h$ and $8i$ are very different from that of $4c$. The coordination of the former is two-fold, while the latter is four-fold. For pure $\alpha$-T boron, $\Delta E_{f}$ of the $8i$ and $8h$ sites is bigger than that of the $4c$ sites by 0.03 and 0.11 eV, respectively. On the other hand, for C- or N-containing $\alpha$-T boron, the energy difference between these sites almost vanishes (less than 0.001 eV or even negative). Only the energy of the $8h$ site of the N-containing form is higher by 0.008 eV. The present calculations imply that the $4c$ site is the major interstitial for pure $\alpha$-T boron, whereas for C/N-containing $\alpha$-T boron comparable amounts of $8i$-site atoms are present and the occupation of the $8h$ site is the smallest.
This different order of $\Delta E_{f}$ for $4c$ and $8i$ sites between pure and C/N-containing $\alpha$-T boron is consistent with the experiment in Table \[tbl:alpha-tetra\]. If the recently discovered $\alpha$-T forms are regarded as pure: $8h$ and $8i$ sites are observed in impurity-containing forms and $4c$ sites are observed in the pure form. In our calculations, the bond lengths of C or N at the $2b$-site to their neighbors are short, [*i.e.*]{}, 1.57 and 1.59 Å, respectively. The corresponding value for pure B$_{50}+ 2$B$_{4c}$ is 1.70 Å. In general, in icosahedron-based boron crystals, [*inter*]{}icosahedral bonds are the shortest ones but the value is at most 1.69 Å. The experimental value of 1.63 Å for traditional $\alpha$-T boron (see Fig. \[fig:POS\]) is too short for B-B bonds, so that the measured value may be evidence for C or N at the $2b$ site.
A clearer contrast between pure and C/N-containing forms is found in the occupation of the $2a$ site. $\Delta E_{f}$ of the pure form is large (3.80 eV), while that of the latter one is small (0.3 eV), so that for the pure form there are practically no $2a$-site atoms. Again, if recently discovered $\alpha$-T forms are pure ones, the present results for the site occupancies are consistent with the experiment, as seen in Table \[tbl:alpha-tetra\].
The different site occupations, [*i.e.*]{}, $4c$ in the pure form and $8h$, $8i$, and $2a$ in the impurity-containing forms, provide a useful guideline for identifying impurities. From the site occupations, we can say that at least Ekimov’s $\alpha$-T boron is not a traditional $\alpha$-T form. To identify it as pure $\alpha$-T boron, however, we may need further evidence, because here the term ’impurity-containing form’ is used in a restricted sense,, referring to C- or N-impurities, only. We cannot exclude the existence of other impurities. Further assessments of Ekimov’s $\alpha$-T boron are made below.
### Valence filling {#sec:valenc-filling}
Let us now consider to vary the composition $n$ in B$_{50+n}$ (hereafter, B$_{50+n}$ is also designated as B$_{50}+ n$B$_{s}$ to indicate $n$ atoms at site $s$). This is seen by inspecting the whole range of $N_{\rm at}$ in Table \[tbl:HfofB50\]. Hayami and Otani studied IS of $\alpha$-T boron over a compositional range $n=0-4$, concluding that B$_{50}+2$B$_{4c}$ is the most stable structure.[@Hayami10] The present calculations confirm their conclusion. Among various combinations of two $4c$-site atoms, the out-of-plane configuration has the lowest energy (For details of the structures, see Supplemental Material). A configuration in which two $4c$-site atoms occupy neighboring sites is highest in energy. There is a general tendency in boron-rich crystals that atoms of the same IS avoid each another. The present result for $4c$ sites is in accordance with this tendency.
The fact that the formation energy is minimized at $n=2$ can be understood by examining the valence filling for different values of $n$. Figure \[fig:POSfilling\] shows the evolution of the density of states (DOS) by successively placing B atoms at $4c$-sites. For B$_{50}$, there are 160 valence states (80 bands), and the top 10 states are unoccupied. Among these 10 unoccupied states, the lower 6 states are the tail part of the valence band and they are mainly contribution derived from the [*inter*]{}icosahedral $t-t$ bonds (in the notations of Ref. ). The upper 4 states are gap states, which are mainly derived from $p_{z}$ orbitals of two $4c$-site atoms.
As $n$ increases by one, the unoccupied states are filled by the additional 3 valence electrons, almost like a rigid-band shift. Although it is by no means a rigid-band shift, the valence-electron counting explains this rigid-band-like behavior very well.[@Shirai12] In $\alpha$-T boron, each equatorial ($e$ site) B atom of the B$_{12}$ icosahedron has one [*inter*]{}icosahedral bond. On placing a B atom at a $4c$ site, four [*inter*]{}icosahedral $e-e$ bonds are replaced by four tetrahedral bonds of $4c$-site atoms. In total, the number of orbitals is not changed, and consequently three electrons of a B atom are used for filling the unoccupied valence states. This way proceeds until $n$=2, where the valence states are completely filled. Further adding of interstitial atoms creates additional gap states, and it is therefore undesirable. B$_{52}$ with two $4c$-site B atoms is the best configuration to meet the valence requirement in a B$_{50+n}$ series.
![(Upper panel) Valence filling for B$_{50+n}$ as the number $n$ of $4c$-site B atoms increases. Only B$_{52}$ can completely fill up the valence bands, but 12 in-gap states remain. This indicates that the conditions for stable covalent bonding are only partially fulfilled. The numbers of occupied states are indicated by the \# symbol. The Fermi levels are indicated by red vertical lines. Light-blue areas indicate unoccupied states, and the numbers of unoccupied states is given in brackets. (Lower panel) Similar plot for impurity-containing ${\rm B_{50}C_{2}}$ and ${\rm B_{50}N_{2}}$. ${\rm B_{50}N_{2}}$ can properly fulfill the the conditions for stable covalent bonding, i.e. all valence bands are occupied and no in-gap states are created.[]{data-label="fig:POSfilling"}](Figs/dos_B50_4c "fig:"){width="8"}\
![(Upper panel) Valence filling for B$_{50+n}$ as the number $n$ of $4c$-site B atoms increases. Only B$_{52}$ can completely fill up the valence bands, but 12 in-gap states remain. This indicates that the conditions for stable covalent bonding are only partially fulfilled. The numbers of occupied states are indicated by the \# symbol. The Fermi levels are indicated by red vertical lines. Light-blue areas indicate unoccupied states, and the numbers of unoccupied states is given in brackets. (Lower panel) Similar plot for impurity-containing ${\rm B_{50}C_{2}}$ and ${\rm B_{50}N_{2}}$. ${\rm B_{50}N_{2}}$ can properly fulfill the the conditions for stable covalent bonding, i.e. all valence bands are occupied and no in-gap states are created.[]{data-label="fig:POSfilling"}](Figs/dos_B50C2_B50N2 "fig:"){width="8"}
This valence filling scheme becomes clearer, when examining the DOS of ${\rm B_{50}C_{2}}$ and ${\rm B_{50}N_{2}}$ (see the lower panel of Fig. \[fig:POSfilling\]). For ${\rm B_{50}N_{2}}$, 10 valence electrons from ${\rm N_{2}}$ completely fill the unoccupied states of ${\rm B_{50}}$. Therefore, ${\rm B_{50}N_{2}}$ is the most desirable compound for $\alpha$-T type structures. ${\rm B_{50}C_{2}}$ may be the next desirable form. In contrast, in pure $\alpha$-T boron the unoccupied states are successively filled with increasing $n$, but this has the side effect of also creating gap states. For the interstitial configurations examined here, we found no configuration that completely separates the valence and conduction bands by using interstitial B atoms only. Probably, there is really no pure configuration without gap states. The significance of this behavior is discussed in Sec. \[sec:frustration\].
Let us now discuss the energy gap of the B$_{52}$ structure. Hayami and Otani claimed that B$_{52}$ is a narrow energy-gap semiconductor with an energy gap of 0.07 eV. (according to a private communication with the authors Fig. 8 of their paper [@Hayami10], which is expected to indicate this gap, is incorrect.). The corresponding “gap" is seen in Fig. \[fig:POSfilling\]. However, in our opinion the empty bands, indicated by the light-blue area, should rather be regarded as gap states. This means that the LDA band gap is actually 2.7 eV. In experiment, the energy gap was reported to be 1.55 eV by electrical resistivity measurement[@Shaw57] and similar values even at high pressures [@Ekimov13a]. Unfortunately, even the most comprehensible data handbook[@LB-III41c] contains only very little data of the optical properties of $\alpha$-T boron. An optical measurement of a material with a similar structure ${\rm B_{48}Al_{3}C_{2}}$ has a fundamental gap of about 2.0 eV with a tail extending down to 0.5 eV.[@Werheit00a] According to our experience with boron carbide B$_{13}$C$_{2}$, the fundamental gap itself is large but there are many gap states, which renders the tail part of the conduction band extending deeply into the gap.[@Shirai14] Therefore, it is reasonable to expect a similar gap structure for pure $\alpha$-T boron.
Deviation from stoichiometry {#sec:deviation}
----------------------------
The conclusion that pure $\alpha$-T boron has the stoichiometric composition B$_{52}$, is only provisional. It was obtained by calculations using primitive unit cells and will change if we extend the cell size and use supercells. In this section, we examine this possibility.
### Supercell calculations {#sec:supercell}
For studying non-stoichiometry, Ekimov’s $\alpha$-T boron is appealing, because only in this case the site occupancy and the composition were determined. An interesting point of Ekimov’s $\alpha$-T boron is the occupation of IS, as indicated in Table \[tbl:alpha-tetra\]. The occupancy 0.31 for the $4c$ site amounts to approximately $1 \frac{1}{4}$ atoms per unit cell and 0.06 for the $4g$ site to approximately 1/4 atoms per unit cell. These values suggest an approximate structural model for Ekimov’s $\alpha$-T boron as $$4 \times {\rm B}_{50} + 5 \times {\rm B}_{4c} + {\rm B}_{4g} = {\rm B}_{206},
\label{eq:EkimovModel}$$ which can be realized in a $2 \times 2 \times 1$ supercell. A similar model structure for ${\rm B_{13}C_{2}}$ was studied before.[@Shirai14]
![The formation energies of different $2 \times 2 \times 1$ supercell model structures (B$_{206}$) for Ekimov’s $\alpha$-tetragonal boron. Overall one finds small positive energy differences, indicating that Ekimov’s $\alpha$-tetragonal boron is not the lowest-energy structure. The abscissa is the number of nearest neighbor $4c$-site atoms surrounding a $4g$-site atom and the energy reference state is $ 4 \times {\rm B}_{52} = {\rm B}_{208}$.[]{data-label="fig:ekimov-confg"}](Figs/Ekimov-plot){width="10"}
In this supercell model (\[eq:EkimovModel\]), one $4g$-site B atom is introduced, and four $4c$-site B atoms are placed, one in each primitive cell. The fifth B atom is randomly placed in one of the four cells, completing the B$_{206}$ model, which corresponds to a deviation from stoichiometry of $\delta=-0.5$. The corresponding formation energy $\Delta E_{f}$, plotted as a function of the number of nearest neighbor $4c$-site atoms surrounding the $4g$-site atom, is shown in Fig. \[fig:ekimov-confg\]. In the figure, the energy reference state is B$_{208}$, which corresponds to the stoichiometric B$_{50} + $2B$_{4c}$.
The formation energy $\Delta E_{f}$ has slight dependence on how many $4c$-site atoms gather around the $4g$ site, but it is not significant. The energy differences between the B$_{206}$ models and the stoichiometric B$_{208}$ are at least 1.3 eV/atom. The positive $\Delta E_{f}$ shows that the composition of Ekimov’s $\alpha$-T is not the lowest-energy composition. Entropic effect cannot change this situation. There are four $4c$ sites in a primitive unit cell. The configurational entropy is maximized at half occupation, $n$=2. And the entropic contribution results only in an improved stability of B$_{52}$, which is already the minimum of the total energy. Our survey for the lowest-energy structure is by no means exhaustive. But, it is unlikely to find lowest-energy configurations than the present ones by further extending the size of the supercell, when retaining the present structural model.
$N_{\rm at}$ ($\delta$) Config. $g$ $\Delta E_{f}$ (eV) $\Delta V$(eV/GPa)
------------------------- -- ------------------ --------------------- --------------------- ---------------------
207 $4c$ 4 0.84 0.010
$8h$ 8 0.68 0.005
$8i$ 8 0.85 0.009
208 $2\times 4c$ 2 0
$2\times 4c'$ 2 0.20 -0.022
$4c+8i$ 4$\times$2 0.07 -0.031
$4c+8h$ 4$\times$2 -0.06 -0.008
$8i+8h$ 8$\times$2 0.22 -0.043
209 $3\times 4c$ 4 1.10 -0.046
$2\times 4c+8h$ 4$\times$4 0.50 -0.027
$2\times 4c+8i$ 4$\times$4 1.04 -0.046
$4c+2\times 8i$ 4$\times$4 0.76 0.002
$2\times 4c+8i'$ 4$\times$4 1.15 -0.052
$4c+8h+8i$ 4$\times$6$\times$2 0.77 -0.051
: The formation energy of interstitial sites and the change of the specific volume $\Delta V$ in a $2 \times 2 \times 1$ supercell model of $\alpha$-T boron. Also in supercell calculations B$_{52}$ is the lowest energy structure. The reference configuration is $ 4 \times ({\rm B}_{50}+ 2 {\rm B}_{4c}) = {\rm B}_{208}$. The multiplicity $g$ of the interstitial sites refers to a primitive unit cell. []{data-label="tbl:HfofB200"}
Next, let us remove some restrictions of the composition and take other IS into account. A slightly different way of constructing the $2 \times 2 \times 1$ supercell will be employed, now. It is constructed by gathering four unit cells of the lowest-energy configuration B$_{50} + $ 2B$_{4c}$. Then, various combinations of IS are studied in one of the four cells only, and the remaining cells are left to be in the B$_{50} + $2B$_{4c}$ configuration. The formation energies of those configurations are listed in Table \[tbl:HfofB200\]. The energy zero is taken to be that of the stoichiometric B$_{208} = {\rm 4 \times (B_{50} + 2B_{4c})}$. The multiplicity $g_{i}$ of atom configurations for a specific type $i$ of IS is the number of configurations which are energetically degenerate. For convenience, $g_{i}$ is counted within a primitive unit cell. High-energy configurations were omitted beforehand.
Although a low-energy state different from ${\rm B_{50} + 2B_{4c}}$ was found for a combination of $4c$ and $8h$, we still continue to use B$_{50} + 2B_{4c}$ as energy zero. The energy decrease is very small, 0.06 eV. Rather, we should regard $2 \times 4c$, together with $4c+8i$ and $4c+8h$ to form a degenerate ground state of B$_{52}$. In this sense, pure $\alpha$-T boron has non-zero residual entropy even in the stoichiometric composition.
The minimum $\Delta E_{f}$ is still found to be B$_{52}$. From this, we have convinced ourselves that the lowest-energy structure of pure $\alpha$-T boron is indeed the stoichiometric B$_{52}$. However, the statistical distribution in $\Delta E_{f}$ brings a new feature into the structure B$_{52}$. Furthermore, an imbalance of $\Delta E_{f}$ is found between $\delta=-0.25$ and $+0.25$ (see Table \[tbl:HfofB200\]). On the average the case $\delta=+0.25$ has lower energy states than $\delta=-0.25$. In particular, the configuration $2 \time 4c+8h$ has the lowest formation energy among the non-stoichiometric compositions and the value 0.5 eV is not large for high-temperature synthesis. More importantly, the case $\delta=+0.25$ contains larger multiplicities than $\delta=-0.25$. This can yield a positive deviation ($\delta>0$) from B$_{52}$ at high temperatures. In the next section, we will discuss the deviation from stoichiometry $\delta$ at finite temperatures and finite pressure, based on the calculated values in Table \[tbl:HfofB200\].
### High-pressure high-temperature properties {#sec:high-pressure}
Considering that most of the recent syntheses of pure $\alpha$-T boron were performed at high pressures $p$, it is important to take the pressure dependence of the formation enthalpy $H$ into account. This can be done by calculating $\Delta H = \Delta E + p \Delta V$. As shown in Table \[tbl:HfofB200\], the magnitude of $\Delta V$ is of the order of 0.02 eV/GPa; so it has a sizable effect on $\Delta H$ at around $p=$ 10 GPa. For the evaluation of $\Delta H$, $\Delta E$ values from supercell calculations were used, because they are more accurate. Each atomic configuration $j$ in Table \[tbl:HfofB200\] is characterized by the formation enthalpy $\Delta H_{j}$, the multiplicity $g_{j}$, and the deviation from stoichiometry $\delta_{j}$. The thermal average of the deviation form stoichiometry $\left< \delta \right> $ is calculated by $$\left< \delta \right> = \frac{1}{Z} \sum_{j} \delta_{j} g_{j} \exp \left( -\frac{\Delta H_{j}}{kT} \right),
\label{eq:AverDelta}$$ where $Z$ is the partition function given by, [^1] $$Z = \sum_{j} g_{j} \exp \left( -\frac{\Delta H_{j}}{kT} \right).
\label{eq:PartitionF}$$ The quantity $\left< \delta \right> $ has pressure dependence through the enthalply. It is plotted as function of $p$ and $T$ in Fig. \[fig:delta\_pt\].
![Contour plot showing the deviation $\delta$ from the stoichiometric composition B$_{52+\delta}$ as a function of pressure $p$ and temperature $T$. The increment of the contours is 0.025. The results indicate that for high $p$-$T$ synthesis a positive deviation from stoichiometry is to be expected.[]{data-label="fig:delta_pt"}](Figs/delta-pT){width="10"}
The obtained values of $\left< \delta \right> $ are of the order of 0.1. With increasing $T$, $\left< \delta \right>$ becomes more positive. At high pressure, negative changes in the specific volume $\Delta V$ are favorable because of LeChatelier’s principle.[@Yamanaka00; @Dekura11] From Table \[tbl:HfofB200\], we see that configurations of $N_{\rm at}=209$ are more favorable at high $p$ than configurations of $N_{\rm at}=207$. Accordingly, $\left< \delta \right>$ increases with $p$. We see that $\left< \delta \right>=+0.14$ is expected for synthesis conditions of $p=10$ GPa and $T=2000$ K. The corresponding occupancies are 0.20, 0.08, 0.09 for $4c$, $8h$ and $8i$ sites, respectively. Therefore, even at high $T$, the $4c$ site predominates over other IS in pure $\alpha$-T boron.
The predicted value of $\left< \delta \right> \sim 0.1$ is not as large as $\delta=+1.5$ for $\beta$-R boron, where the deviation is an intrinsic property. This is a consequence of the fact that $\alpha$-T boron has an even number of valence electrons, and thus the driving force for the deviation is weak. Values of $\left< \delta \right> \sim 0.1$ are also found in ${\rm B_{50}C_{2}}$ and ${\rm B_{50}N_{2}}$ (see Table \[tbl:alpha-tetra\]). According to the discussion in Sec. \[sec:valenc-filling\], the structures of ${\rm B_{50}C_{2}}$ and ${\rm B_{50}N_{2}}$ meet the valence requirement most properly, and hence they are almost likely to be stoichiometric. Therefore, it is reasonable to find similar values of $\left< \delta \right> $ for pure and C/N-containing $\alpha$-T boron.
At present, there is little experimental information available about $\delta$. Among them, Ekimov’s samples are best characterized. They show a negative value $\delta=-0.5$, contrary to the present prediction. By considering their preparation conditions ($p=10$ GPa, and $T=1600^{\circ}$C) this discrepancy is well beyond the error of our calculations. As noted before, their samples were synthesized by pyrolysis of decaborane ${\rm B_{10}H_{14}}$. So there is the possibility of hydrogen inclusion, but the authors reported not to detect hydrogen. Hyodo’s samples with a positive value $\delta=+0.2$ are indeed within the present prediction, however no occupation of the $4c$ site was reported. In this case, the question is whether bulk samples and nanostructures can be compared on the same ground. More reliable measurements on the chemical compositions of the recently discovered $\alpha$-T borons are indispensable for further developments.
### Lattice parameters {#sec:lattice-para}
As indicated in Table \[tbl:alpha-tetra\], some trends in the lattice parameters exist between traditional and recently discovered $\alpha$-T borons. When compared with the early-days crystals by Hoard, it is seen that the lattice parameters $a_{0}$ are small for the traditional forms and large for the recently discovered ones. In general, the error of GGA for lattice parameters is less than 1 %, and hence GGA calculations can resolve subtle differences in the lattice parameters which are expected to exist among different $\alpha$-T forms.
![The lattice parameters of $\alpha$-T boron as a function of the composition $N$. Red circles indicate the present calculations, symbols with authors’s name are experimental data. For Kurakevych’s and Qin’s $\alpha$-T boron, the chemical compositions are not known and therefore the data are plotted at the rightmost side. Clear differences are discernible between the C/N containing forms and Ekimovs’s $\alpha$-T boron on the one hand and the recently discovered and the calculated forms on the other hand.[]{data-label="fig:lattice"}](Figs/avsn "fig:"){width="7.6"} ![The lattice parameters of $\alpha$-T boron as a function of the composition $N$. Red circles indicate the present calculations, symbols with authors’s name are experimental data. For Kurakevych’s and Qin’s $\alpha$-T boron, the chemical compositions are not known and therefore the data are plotted at the rightmost side. Clear differences are discernible between the C/N containing forms and Ekimovs’s $\alpha$-T boron on the one hand and the recently discovered and the calculated forms on the other hand.[]{data-label="fig:lattice"}](Figs/cvsn "fig:"){width="7.6"}
Our calculations of the lattice parameters are shown in Fig. \[fig:lattice\] as a function of the composition B$_{50+n}$. The calculated values of $a_{0}$ of pure $\alpha$-T boron are in most cases about 1% larger than that of the traditional $\alpha$-T boron over a range $n$ = 0 to 4. When compared with the calculated structures B$_{50}$C$_{2}$ and B$_{50}$N$_{2}$ (not shown in the figure), $a_{0}$ of pure B$_{52}$ is larger than that of B$_{50}$C$_{2}$ by 1.9% and of B$_{50}$N$_{2}$ by 2.8%. This result is to be expected, because, as discussed in Sec. \[sec:bondnature\], these impurities enhance the covalency of the crystal, resulting in contraction of the cell.
An opposite trend is found for $c_{0}$; both the calculated values for pure $\alpha$-T boron and the experimental values for the recently discovered ones are smaller than that of the traditional ones. This opposite behavior between $a_{0}$ and $c_{0}$ is a consequence of the distortion of the tetrahedral bonds of the IS. Remember that there is a strong anisotropy between $a_{0}$ and $c_{0}$, despite the tetrahedral arrangement of four icosahedra B$_{12}$. As mentioned repeatedly, the role of interstitial atoms is not small. The bonding of interstitial atoms indeed has to be strong in order to be able to support the local arrangement of icosahedra in the unit cell. An interstitial-site atom, in particular the $4c$ site, has a tetrahedral bonding environment (see Fig. \[fig:POS\]). However, these tetrahedral geometries are strongly flatten in the $ab$ plane. Accordingly, the four icosahedra which are connected to the interstitial atom are also deformed in the $ab$ plane. By replacing the interstitial B with C or N atoms, the covalency of the related tetrahedral bondings becomes more ideal, [*i.e.*]{}, isotropic, and the anisotropy between $a_{0}$ and $c_{0}$ is reduced.
Let us check the experimental data. Kurakevych’s $\alpha$-T boron contain various concentrations of C including 0 at.%. They showed an increase in $a_{0}$ as the C content is decreased [@Kurakevych12], which is consistent with our results. Total agreement in $a_{0}$ and $c_{0}$ with our calculation suggests that Kurakevych’s $\alpha$-T boron with no C content is pure $\alpha$-T boron. For Qin’s case a judgment is more difficult. Although they claimed in their paper [@Qin12] that there is a good agreement in the lattice parameters with previous experiments, TEM measurements, which they employed, are generally less accurate.
The lattice parameter $a_{0}$ of Ekimov’s $\alpha$-T boron is evidently larger than the traditional ones. However, his value is too large even when compared with the calculated one for pure $\alpha$-T boron. Judging from the lattice parameters and the negative deviation from stoichiometry $\delta=-0.5$, it is safe to identify Ekimov’s $\alpha$-T boron as non-pure boron. Further studies on Ekimov’s $\alpha$-T boron are in progress.
Geometrical frustration {#sec:frustration}
-----------------------
As discussed in Sec. \[sec:valenc-filling\], and different from impurity-containing forms, a profound feature of pure $\alpha$-T boron is the presence of gap states. Let us consider the electronic nature of the gap states. For B$_{50}$N$_{2}$, where two $2b$ sites are occupied by two N atoms and half of the $4c$ sites are occupied by two B atoms, the 10 unoccupied states of B$_{50}$ are completely filled by ten valence electrons from two N atoms. On the other hand, for pure $\alpha$-T boron, B$_{50}\ + 2$B$_{4c}$, the lower 6 valence states are filled, leaving 4 topmost states unoccupied. These topmost unoccupied states are filled by the extra 4 electrons in the case of B$_{50}$N$_{2}$, but are left as unoccupied gap states in the pure form.
![Partial density of states (DOS) projected on two $4c$-site B atoms (red area) for B$_{50} + 2$B$_{4c}$ indicates that the in-gap states mainly come from interstitial 4c-site atoms. The black line indicates the total DOS. The scales of partial and total DOS are different.[]{data-label="fig:pDOS2c"}](Figs/pDOS2c){width="10"}
The electronic nature of these topmost valence states for the pure form B$_{50}+ 2$B$_{4c}$ can be seen by calculating the partial DOS. Figure \[fig:pDOS2c\] shows the partial DOS with respect to two $4c$-site B atoms. The gap states are mainly coming from the $4c$-site atoms. As mentioned before, the bonding arrangement of the $4c$ site is a strongly flattened tetrahedron; roughly speaking, almost similar to $sp^{2}$ in-plane bonding. The wave functions of the gap states have non-bonding $p_{z}$-orbital character along the direction of the $c$-axis (see Supplemental Material). For the B$_{50}$N$_{2}$ case, the four extra electrons, coming from the two N atoms, can fill these gap states in a similar way as a rigid-band shift.
However, for pure $\alpha$-T boron, the valence filling does not work in this way. When successively adding B atoms to B$_{50}$, each time three unoccupied states are eliminated, until $n=2$. However, this valence filling by B atoms creates new gap states (see Fig. \[fig:POSfilling\]). Further increase of $n$ is even worse as it does not eliminate gap states. In this way, we see that satisfying one bond is connected to the creation of new non-bonding states elsewhere. There is no way to completely remove unoccupied states. This is just the principle of geometrical frustration, which was also found in a similar boron material, [*i.e.*]{} $\beta$-R boron.[@Ogitsu09]
In a general sense, the icosahedral geometry is intimately related to geometrical frustration.[@Sadoc06] But, the special feature of boron is that the geometrical frustration is caused by POS. The driving force is strong, because of the covalent nature of the bonding. A pictorial interpretation of the above electronic situation can be obtained by using an idea given by Ogitsu [*et al*]{} for $\beta$-R boron.[@Ogitsu09; @Ogitsu10; @Ogitsu13] In analogy with antiferromagnetism (AF), it is useful to understand what happens in B$_{50}+ 2$B$_{4c}$. As noted previously, in boron crystals there is a tendency of interstitial atoms to spatially avoid each other. This situation can be modeled by spin systems. Let us denote an occupied site with spin up, and an empty site with spin down. A perfectly ordered AF phase in a cubic lattice is shown in the left-hand side of Fig. \[fig:frustration\]. Let us apply this spin model to B$_{50}+ 2$B$_{4c}$. The right-hand side of the figure corresponds to the adapted spin model. It shows a primitive unit cell as in Fig. \[fig:struct\] and the vertices of the eight cuboids are the locations of the $2a$, $2b$ and $4c$ sites. The spin arrangement enclosed in green indicates that the involved spin directions are fixed, that is to say, the $2b$ site is always occupied and the $2a$ site is always unoccupied. Then, we are free to choose the spin directions at the $4c$ sites. Let us consider a spin arrangement in the first layer (denoted as I). The arrangement, as shown in the figure, meets the AF configuration. Every pair of neighboring spins is antiparallel. In the next layer (II), the $2a$ and $2b$ sites interchange their positions. Then, an AF spin arrangement within the second layer can be realized as shown in the figure. However, this spin arrangement causes parallel spins between layers I and II. In this way, we will find that there is no perfect ordered state for the occupation of the two $4c$-sites.
From the geometrical point of view, this frustration can be regarded as arising from the incompatibility of the symmetry, that is, a conflict between the presence of inversion symmetry for icosahedra and the absence of it for the tetrahedral arrangement of four icosahedra in a unit cell. This conflict yields an asymmetry in the occupation between fully occupied $2b$ sites and vacant $2a$ sites, which surround each icosahedron. Two symmetry equivalent $2b$ sites are connected by inversion with respect to the center of an icosahedron, and the same holds for $2a$ sites. This symmetric occupation of $2b$ and $2a$ sites is incompatible with the tetrahedral arrangement of the four icosahedra.
In the stoichiometric B$_{50}+ 2$B$_{4c}$ structure only half of the equivalent $4c$ sites are occupied. Therefore it has non-zero entropy $S_{0}$ at $T=0$. This residual entropy is (1/52) $\ln 6$ = 0.034 per atom (in units of $k_{\rm B}$, Boltzmann’s constant), provided that there is no interaction between $4c$-site atoms. This value is small compared with a value (1/3)$\ln (3/2)$ = 0.135 of water [@Ben-Naim09], which is a prototypical for geometrical frustration, but it is yet a macroscopic amount. In fact, this simple estimation shows that the residual entropy of $\alpha$-T boron is as large as that of $\beta$-R boron, which was recognized as a geometrically frustrated system. For $\beta$-R boron, the primary IS is a replacement of one atom at the B(13) site with one interstitial atom at the B(16) site. Both of those sites have six equivalent sites. This simple estimate gives the residual entropy (1/105) $\ln (6\times 6)$ = 0.034 per atom. The numerical agreement with the above estimate for $\alpha$-T boron is of course accidental. A more accurate estimate was done by Ogitsu [*et al*]{}.[@Ogitsu10] They determined a value of $S_{0}$ = 0.04 for $\beta$-R boron. The value is very close to the present estimation. Thus the residual entropy of $\alpha$-T boron is comparable to that of $\beta$-R boron.
---------------------------- -------------- ---------------------------------- ------------------ --------------------------------
Crystal $N_{\rm at}$ $S$ Notes
$k_{\rm B}$/atom
$\beta$-R boron [@Slack88] 105+1.7 74.5 (B13), 27.2(B16), 8.5 (B17) 0.137 Equivalent site numbers
6.6 (B18), 6.8 (B19), 3.7 (B20) are all 6, except 12 for B(20)
$\alpha$-T boron 52 +0.14 20 ($4c$), 8 ($8h$), 9 ($8i$) 0.128
---------------------------- -------------- ---------------------------------- ------------------ --------------------------------
: A comparison of the site occupancies and the configurational entropy between $\alpha$-T boron synthesized at $p = 10$ GPa and $T=2000$ K (modeled in this work) and $\beta$-R boron (from literature).[@Ogitsu10] Both crystals have a similar amount of configurational entropy, caused by geometrical frustration. $N_{\rm at}$ is presented by separating the stoichiometric composition and $\delta$.
\[tab:entropy\]
Let us now proceed with a more elaborate estimate of the residual entropy. At high temperatures at which the $\alpha$-T boron crystals are synthesized, many POS appear, non-stoichiometric systems are formed (see Sec. \[sec:high-pressure\]), and the configurational entropy increases. We estimate the configurational entropy $S$ by using the following formula, $$S = k_{\rm B} \sum_{j} \left[ x_{j} \ln x_{j} + (1-x_{j}) \ln (1-x_{j}) \right],
\label{eq:entropy}$$ where $x_{j}$ is the occupancy of the IS $j$. By using this formula, it is implied that all the IS are independent. Experimental values of occupancies for $\beta$-R boron by Slack [*et al*]{} [@Slack88] and our calculated values for $\alpha$-T boron are compared in Table \[tab:entropy\]. From these values, we obtain $S$ = 0.137 for $\beta$-R boron and $S$ = 0.128 for $\alpha$-T boron. So the configurational entropy of $\alpha$-T boron is comparable to that of $\beta$-R boron. On one hand this is an unexpected result, because the deviation from stoichiometry in $\alpha$-T boron is only small. On the other hand it is actually reasonable, because the numbers of equivalent sites (the multiplicity $g$) is bigger in $\alpha$-T boron (mostly 8) than in $\beta$-R boron (mostly 6). Moreover, the number of atoms per unit cell in $\beta$-R boron is twice as large as that in $\alpha$-T boron.
Based on these arguments, $\alpha$-T boron can be considered as a new member of the group geometrical frustrated boron systems, because (i) an unavoidable conflict between the elimination of the existing unoccupied states and the creation of new unoccupied states, and (ii) the existing of a macroscopic amount of residual entropy. An interesting point is that the geometrical frustration of $\alpha$-T boron does not require non-stoichiometry, differently from the situation in $\beta$-R boron. This makes the theoretical analysis simpler. Recently, a novel view, the so-called *correlated disorder* has appeared, which is a universal approach connecting different classes of disorders, that is, frustrated over-constraint and configurational under-constraint.[@Keen15] The structure of $\alpha$-T boron is a particularly a good system for developing this view, because the structure seems to have both characters of disorder in it, because there is no way to perfectly satisfy covalent conditions (over- constraint) and there are many ways of arranging POS (under-constraint).
Conclusions
===========
We have examined the possibility of the existence of non-stoichiometric, pure $\alpha$-T boron. Supercell calculations have shown that the lowest-energy state of pure $\alpha$-T boron, if it exists, is the stoichiometric B$_{52}$ with two occupied $4c$ sites. Therefore, a deviation from stoichiometry occurs only by entropic effects at high temperatures. This is different from $\beta$-R boron, where non-stoichiometry is a property of the lowest-energy state. This difference essentially comes from the fact that B$_{52}$ has an even number of electrons and therefore the driving force for a deviation from stoichiometry is relatively weak. Finite temperatures cause a deviation from stoichiometry B$_{52+\delta}$ by a small positive amount, with $4c$ site atoms as the main interstitial species. For high-pressure and high-temperature synthesis under reported conditions, $\delta$ is estimated to be around $\delta$=0.1 to 0.2.
The present results on the site occupancies, the $\delta$ values as a function of pressure and temperature, and the lattice parameters provide a good test for the experimental identification of pure $\alpha$-T boron. The dominant interstitial site of the pure form is $4c$, while for C- or N-containing forms the $2a$, $8h$ and $8i$ sites are favored. Judging from the limited experimental data, we conclude that Ekimov’s $\alpha$-T boron is probably not a pure form, because of its negative $\delta$ together with other inconsistencies. Judging from the lattice parameters, other recently discovered $\alpha$-T boron forms could indeed be pure ones, though not all. For a final conclusion, however, more detailed information about the chemical composition and the site occupancies are required.
Despite of B$_{50}$ + 2B$_{4c}$ having a band gap, the valence bond requirement is not ideally fulfilled. Unoccupied $p_{z}$ orbitals at $4c$ sites form gap states that cannot be eliminated by adding further interstitial atoms. There is no way to fully satisfy the conditions for covalent bonding. Despite of being stoichiometric, B$_{52}$ has a macroscopic amount of residual entropy that is as large as that of $\beta$-R boron. Because of having such a lowest-energy state, macroscopic residual entropy and no atomic arrangement that can satisfy the conditions for covalent bonding, pure $\alpha$-T boron is identified as a geometrically frustrated elemental crystal.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank K. Kimura and W. Hayami for initiating our interests in this subject. We also thank E. Ekimov for discussing the characterization of his samples. JK acknowledges financial support from the German Research Foundation (DFG) (project KU 2347/2-2). HE and JK acknowledge financial support from the German Academic Exchange Service (DAAD) (project 54368630)
[53]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , , ****, ().
, , , , , , ****, ().
, , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , , , ****, ().
, ****, ().
, ****, ().
, ****, (),
, , , ****, ().
, , , , , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ** (, , ).
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , , , , ****, ().
, , , , , , , ****, ().
, , , ****, ().
, , , , , ****, ().
, , , ****, ().
, ** ().
, , , , ****, ().
.
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, ** (, , ).
, , , , ****, ().
, , , , ****, ().
, ** (, , ).
, , , ****, ().
, ** (, , ).
, , , ****, ().
, ****, ().
[^1]: Caution is needed for the values of $\delta_{j}$ when evaluating Eq. ([\[eq:AverDelta\]]{}). For example, $\delta_{j}=+1$ should be used for $N_{\rm at}=209$ systems (not 1/4). This is because we are considering the change of the total energy of the crystal that is induced by one interstitial atom; so the size of a supercell is irrelevant.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Using typical experimental techniques it is difficult to separate the effects of carrier density from disorder on a two-dimensional superconducting transition. To address this problem, we have produced graphene sheets decorated with a non-percolating network of nanoscale tin clusters. These metal clusters both efficiently dope the graphene substrate and induce long-range superconducting correlations. This allows us to study the superconducting transition at fixed disorder and variable carrier concentration via the field effect. We find that despite structural inhomogeneity on mesoscopic length scales (10-100 nm), this material behaves electronically as a homogenous dirty superconductor. Our simple self-assembly method establishes graphene as an ideal tunable substrate for studying induced two-dimensional electronic systems at fixed disorder and our technique can readily be extended to other order parameters such as magnetism.'
author:
- 'B. M. Kessler'
- 'Ç. Ö. Girit'
- 'A. Zettl'
- 'V. Bouchiat'
title: 'Tunable Superconducting Phase Transition in Metal-Decorated Graphene Sheets'
---
The superconducting transition in two-dimensions is of interest for both the fundamental understanding of electronic order in reduced dimensions and applications involving superconducting thin films. An open question is how the transition behaves as a the density of carriers mediating the superconductivity is varied. In particular, the strength of disorder appears to play a fundamental role separating qualitatively different behavior[@Kapitulnik+Phase+Diagram]. In two-dimensions, the electric field-effect provides the most versatile method for tuning the carrier density of a system at fixed disorder. However, the field effect places more stringent limits on the dimensionality of the system since the film must be thinner than the Debye length, which is typically much smaller than the penetration depth that places the limit on two-dimensional superconductors. Despite this limitation, the field effect has been used to tune the superconducting transition in specific materials such as thin films with anomalously low carrier density[@Goldman_FET_Bi] and interfacial states between complex oxides[@Triscone_FET]. An alternative approach is to couple superconducting correlations directly into a truly two-dimensional electronic system, graphene.
The bipolar two-dimensional electron gas (2DEG) present in graphene[@Geim_Rev] is markedly different from the buried 2DEGs found at oxide interfaces or in GaAs heterostructures in that it is ‘open’ to the environment with a stable and inert surface. Using standard 2DEGs it is only feasible to capacitively couple the electron gas to materials deposited on their surface[@Clarke_DPT; @Wagenblast+Dissipation+Theory], whereas graphene is expected to allow direct coupling and thus offer access to different regions of phase space[@Kim_QPT_graphene]. We sought to determine whether bulk materials deposited directly onto the graphene surface, such as metal clusters, can act as dopants and efficiently couple through the electron gas, whose carrier density and type can be tuned by an applied gate voltage. The low carrier density in graphene, relative to bulk values, and weak intrinsic interactions such as spin-orbit coupling, should limit backaction of the electron gas on dopant properties. Thus, exposed graphene sheets could provide a near-ideal substrate for the manipulation and general study of proximity-induced electronic phases.
Graphene has been shown to effectively carry proximity-induced Josephson currents injected from contacting electrodes[@Pablo_JJ; @Andrei_JJ]. However, a finite coherence length limits the length of such junctions to approximately one micron, reducing the physics to one dimension. To maintain coherence over longer distances in two dimensions while retaining the unique properties of the graphene sheet we employ a geometry (Fig. 1a) where a large array of nanoscale dopant islands is placed in a non-percolating network on top of the graphene sheet[@Feigelman_graphene].
We avoid complicated lithographic patterning and exploit the poor wettability of graphite to simply and reliably produce an array of submicron islands. Low melting point metals such as the elemental superconductor Sn readily form self-assembled islands when deposited on pristine graphene at room temperature (Fig. 1b) similar to previous results on graphite[@Metois_Pb]. Analysis of scanning electron micrographs and atomic force micrographs indicates that 10 nm of nominal deposition thickness typically results in islands with $80 \pm 5$ nm diameter and $25 \pm 10$ nm gaps between them (Fig 1b). In general, many different materials with different electronic order parameters can be deposited via this process by controlling the graphene substrate temperature during deposition[@Zayed_recrystal+In], and other deposition methods such as chemical functionalization and wet self-assembly could be used as well.
Samples were prepared by exfoliating Kish graphite[@Geim_Rev] onto degenerately doped ($\rho < 0.005$ m$\Omega$-cm) silicon wafers coated with 285 nm of thermal oxide. Single-layer graphene flakes were identified by optical contrast and confirmed via analysis with a micro-Raman spectrometer[@Ferrari_Raman]. Four-probe contacts were defined via electron beam lithography and a Pd/Au 10/50 nm bilayer was evaporated as metal electrodes (Fig. 1, c). To produce the island network, Sn (99.999% purity) was evaporated using an electron gun in high vacuum ($10^{-7}$ torr) onto graphene substrates at room temperature. The samples were thermally anchored to the cold stage of He-3 or Dilution cryostats and connected to highly filtered lines. Linear response and differential conductance were measured with standard low frequency lock-in techniques using low excitation currents in the range 10-100 nA. Figure 1c displays the room-temperature field-effect characteristics of a device before and after Sn is deposited. Although 40% of the graphene surface is coated by Sn islands after the deposition, many of the original electronic properties of graphene remain intact, including bipolar transport and field-effect mobilities $\mu > 1000$ $\mbox{cm}^2/(\mbox{V} \cdot \mbox{s})$. The three main effects of Sn deposition are a rigid shift in the charge neutrality point (Dirac point $V_D$) to more negative voltages, a factor of five decrease in mobility, and a pronounced asymmetry between electron and hole transport. Note that the maximum resistance at charge neutrality remains unchanged, indicating that the high coverage of low resistance Sn islands does not directly shunt the current in the graphene sheet.
![\[Fig. 1\] (Color online) a) Schematic of device configuration and measurement setup. Blue islands correspond to Sn clusters. b) Scanning electron micrograph of Sn island morphology on the graphene sheet (Scale bar = 100 nm. Inset: optical image of a typical device showing the four probe configuration (Scale bar 10 microns) c) Four-terminal sheet resistance as a function of gate voltage for Sample A before (red online) and after (blue online) Sn deposition. The dotted lines indicate the charge neutral point and the arrow indicates the shift after Sn deposition.[]{data-label="setup"}](Figure1.pdf){width="45.00000%"}
All three of these effects are well described by inhomogeneous doping due to charge transfer from the metal islands to the graphene sheet[@Avouris_assymetry]. From the shift in the charge neutrality point ($V_D$) on three separate samples and the known gate capacitance ($C_g$ = 115 aF / $\mu \mbox{m}^2$) we can calculate the charge induced in the graphene sheet by the Sn, $n_{ind} = C_g \Delta V_D$. Normalizing by the observed Sn coverage, we infer that Sn transfers $9 \pm 2 \times 10^{12}$ $\mbox{cm}^{-2}$ electrons to the graphene underneath it. This is expected from the difference in work functions between the two materials ($\Phi_G = 4.5$ eV, $\Phi_{Sn} = 4.42$ eV)[@Kelly_doping+contacts] and in agreement with recent experiments performed using other metals[@Kern_Contact_Effects]. This induced charge reduces the mobility of both types of carriers via charged impurity scattering[@Fuhrer_Impurities] while the asymmetry in transport occurs because holes experience the pinned Fermi level under the Sn islands as a potential barrier, while electrons experience a potential well[@Avouris_assymetry; @Goldhaber-Gordon_contacts].
More interesting than the influence of the Sn islands on the normal state properties of graphene is the effect the superconducting correlations in the Sn have on transport via the proximity effect. Figure 2 shows the sheet resistance versus temperature for gate voltages on both the hole (Fig. 2a) and electron (Fig. 2b) sides of the charge neutrality point. Each curve exhibits two distinct features, a high temperature partial drop in resistance that occurs at $\sim 3.5$ K independent of gate voltage, and a broad transition between 3 K and 1 K to a state of zero resistance that is strongly dependent on the gate voltage.
![\[Fig. 2\] Sheet resistance versus temperature for various gate voltages, $V_g$, referenced to the charge neutrality point $V_D = +40$ V for this device. In a) $\Delta V_g = V_g-V_D < 0$ corresponds to hole transport, whereas, $\Delta V_g > 0$ in b) corresponds to electron transport through the graphene sheet. The arrow labeled $T_{c0}$ indicates the first partial resistance drop corresponding to the mean-field pairing transition of the Sn islands. Inset: Same data on a log scale.[]{data-label="transition"}](Figure2.pdf){width="40.00000%"}
The first partial resistance drop (arrows in Fig. 2a,b) is due to condensation of Cooper pairs in the Sn islands ($T^{Bulk}_c=3.72$ K). Analysis of the drop shows that it can be fit by $\Delta \sigma(T) \propto \ln (T/T_{c0})^{-1}$ (Fig. 3a) typical of Aslamazov-Larkin fluctuation-enhanced conductivity[@Aslamasoz+Larkin] in two dimensions. Fitting each curve from $3.8-4.5$ K we extract a mean-field pairing temperature ($T_{c0}$) of $3.54\pm0.02$ K independent of gate voltage (solid squares in Fig. 3c). Note that the amplitude of this drop is not directly proportional to the Sn coverage, indicating that the islands do not act as simple superconducting shunts. In two-dimensional superconducting systems, it is well known that although the amplitude of the superconducting wave function is well defined below the pairing temperature $T_{c0}$, thermally induced phase fluctuations (vortices) destroy global phase coherence and produce dissipation due to a finite flux flow resistance[@Berezinskii2; @Kosterlitz+Thouless]. However, below the critical Berezinskii-Kosterlitz-Thouless unbinding temperature, $T_{BKT}$, the attractive interaction between vortices with opposite orientation causes them to form bound pairs allowing a finite supercurrent to flow.
The vortex-unbinding temperature can be identified from the universal form of the flux flow resistance[@Minnhagen_Rev] above the transition $R_{\Box} (T) \propto \exp[ b (T- T_{BKT})^{-1/2} ]$, where $b$ is a constant of order unity governing the vortex-antivortex interaction strength and $T_{BKT}$ is the vortex unbinding temperature. To extract this form we plot $\left( d \ln(R_{\Box})/dT \right)^{-2/3}$ vs. T, which produces a straight line with $T_{BKT}$ given by the x-intercept for curves following the universal form. In Fig. 3b the resulting fits are given showing $T_{BKT}$ extracted from the x-intercepts for three different gate voltages. Through a large intermediate range the curves follow the universal form (straight line). However, at low temperatures the curves level off due to finite-size effects, which cut off the attractive vortex-antivortex interaction[@Minnhagen_Rev]. This departure from the universal form is particularly evident near the charge neutrality point (Fig. 2a and b), possibly indicating proximity to a superconductor-normal quantum critical point at the lowest charge densities[@Kivelson+SNT]. Figure 3c summarizes the resulting $T_{BKT}$ (open circles) extracted from the resistance versus temperature curves at each gate voltage.
To analyze the gate voltage dependence of the vortex unbinding transition, we consider the sheet as a dirty two-dimensional superconductor where the gate voltage allows us to tune the normal state resistance. This is justified since the length scales of disorder are much smaller than the superconducting coherence length, ie. $r \sim d \sim \ell_{mfp} < \xi_0$, where $r$ is the size of the islands, $d$ is the distance between islands, $ \ell_{mfp} \sim 20-30$ nm is the mean free path extracted from field effect measurements at 6 K and the superconducting coherence length Sn $( \xi^{Sn}_0 \sim 300$ nm). For a dirty 2D superconductor, one can use the jump in superfluid stiffness at the vortex unbinding transition to relate $T_{BKT}$ to the normal state resistance of the film[@Beasley_BKT; @Goldman_RG1],
$$\frac{ T_{c0} }{ T_{BKT} } \left\{ \frac{ \Delta(T_{BKT}) }{\Delta(0)} \tanh \left[ \frac{ \Delta(T_{BKT}) }{2k_b T_{BKT} } \right] \right\} = \frac{ \epsilon_v R_N }{ R_0 }$$
where $\Delta(T)$ is the superconducting energy gap, $R_N$ is the normal state sheet resistance, $R_0=\frac{2.18 \hbar}{e^2}\approx 8.96 k\Omega$, and $\epsilon_v$ is an effective dielectric constant that describes the material dependent screening of the attractive vortex-antivortex interaction[@Goldman_RG1]. Using the weak-coupling BCS limit for the superconducting gap and the sheet resistance measured at 6 K, we fit the $T_{BKT}$ extracted above using $\epsilon_v$ as the only adjustable parameter and find $\epsilon_v=2.40\pm0.05$ (solid curve in Fig.3), which is twice the value seen in related systems[@Goldman_RG1]. This implies that the vortices and antivortices are relatively weakly bound in our system.
![\[Fig. 3\] (Color online) a) Fits of the sheet resistance versus temperature to fluctuation-enhanced conductivity of the Aslamazov-Larkin form. b) Rescaling of the sheet resistance versus temperature to the BKT form to extract the vortex unbinding temperature $T_{BKT}$ c) The mean-field pairing temperature, $T_{c0}$ (black squares) and vortex-unbinding temperature, $T_{BKT}$ (open circles) as a function of gate voltage. The solid line is a fit of $T_{BKT}$ using equation 1 and the measured normal state properties of the device. (see text).[]{data-label="fitsummary"}](Figure3.pdf){width="49.00000%"}
At temperatures below $T_{BKT}$, vortices and antivortices form bound pairs and a finite critical current develops which saturates to a gate-voltage dependent value for $T \ll T_{BKT}$. Current-voltage characteristics at 100 mK for different applied gate voltages are shown in Figure 4. The gate-tunable critical current is qualitatively similar to isolated graphene Josephson junctions[@Pablo_JJ; @Andrei_JJ] with the exception that, in our devices, critical current densities ($I_c/width \sim 1$ A/m) comparable to submicron graphene Josephson junctions are maintained over distances of tens of microns, demonstrating the fully two-dimensional phase coherence in this system.
We have demonstrated a simple method to produce a two-dimensional superconductor on a graphene substrate and tune the transition via an electrostatic gate. This allowed us to systematically tune the carrier density at fixed disorder. Although structurally inhomogenous, this material behaved electronically as a weakly disordered two-dimensional superconductor. While we have probed the properties of this system using electron transport, the readily accessible interface allows application of a myriad of local characterization techniques such as scanning probe microscopy, optical spectroscopy, etc. We expect arranging the islands into regular arrays or superlattices should lead to interesting frustration effects as a function of applied magnetic and electric fields[@Jayaprakash_JJ_in_B; @Louie_superlattices]. This tunable superconducting material may find applications in bolometers for sensing applications or as an element in circuits for quantum information processing. The ease of fabrication and considerable versatility of deposition materials make graphene an attractive platform for investigating other electronic orders such as magnetism in two dimensions.
![\[Fig. 4\] (Color online) Current-voltage curves taken at zero magnetic field and 100 mK, corresponding to hole transport in the graphene sheet for gate voltages, $\Delta V_g$ relative to the Dirac point $V_D=+12$ V for this sample. Inset: similiar curves for electron transport. []{data-label="groundstate"}](Figure4.pdf){width="45.00000%"}
B.K., Ç.G., and A.Z. were supported by the Director, Office of Energy Research, Office of Basic Energy Sciences, Materials Sciences, and Engineering Division, of the U.S. Department of Energy under contract DE-AC02-05CH11231, through the sp2-bonded nanostructures program. V.B. acknowledges support from the Miller Institute for Basic Research in Science, CNRS/MPPU and ANR-JC/NEMESIS. We thank M. Feigelman, M. Skvortsov, J. Moore, G. Deutscher and P. Ghaemi for helpful discussions.
[27]{} 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: 'Based on the $SO(2)$-invariant slave-boson scheme, the static charge, orbital, and excitonic susceptibilities in the extended Falicov-Kimball model are calculated. Analyzing the phase without long-range order we find instabilities towards charge order, orbital order, and the excitonic insulator (EI) phase. The instability towards the EI is in agreement with the saddle-point phase diagram. We also evaluate the dynamic excitonic susceptibility, which allows the investigation of uncondensed excitons. We find qualitatively different features of the exciton dispersion at the semimetal-EI and at the semiconductor-EI transition supporting a crossover scenario between a BCS-type electron-hole condensation and a Bose-Einstein condensation of preformed bound electron-hole pairs.'
author:
- 'B. Zenker$^1$, D. Ihle$^2$, F. X. Bronold$^1$, and H. Fehske$^{1,3}$'
title: |
Slave-boson field fluctuation approach to the extended Falicov-Kimball model:\
charge, orbital, and excitonic susceptibilities
---
Introduction
============
At low temperatures electronic correlations can cause anomalies at the semimetal-semiconductor (SM-SC) transition. [@HR68] Half a century ago, Mott [@M61] argued that in a SM with a very low carrier density the Coulomb attraction between electrons and holes should lead to the spontaneous formation of electron-hole bound states (excitons), and the system would become insulating. Shortly afterwards, Knox [@Kno63] noticed that a SC is unstable against the spontaneous formation of excitons if the exciton binding energy overcomes the gap energy separating valence and conduction band. Both arguments suggest a new distorted phase, an exciton condensate known as the excitonic insulator (EI), to be the crystal ground state. The SM-EI transition is mathematically similar to the BCS theory of superconductivity, while the SC-EI transition can be treated as a Bose-Einstein condensation (BEC) of preformed excitons. Hence, the EI is discussed in view of a BCS-BEC crossover scenario in a solid. [@CN82a; @BF06; @IPBBF08; @PBF10] Whilst theoretically predicted a long time ago, [@C65] (for recent reviews see Ref. ) no conclusive experimental proof of the existence of the EI has been achieved yet. However, there are a few promising candidates. In the mixed valence compound TmSe$_{0.45}$Te$_{0.55}$ detailed studies of the pressure-induced SC-SM transition suggest that excitons are created in a large number and condense below 20 K. [@NW90] More recently, several transition-metal dichalcogenides were reported to exhibit an EI phase. Angle-resolved photo emission spectra (ARPES) measurements of Ta$_2$NiSe$_5$ traced back the extreme valence band top flattening at low temperature to an EI ground state. [@WaEtAl09] ARPES data of $1T$-TiSe$_2$ indicate that the EI is the driving force for the charge-density-wave (CDW) transition in this material. [@CeEtAl07]
![(Color online) Hartree-Fock ground-state phase diagram of the EFKM in two dimensions for Coulomb strength $U=2$. The difference between CDW and SOO is explained in the text. The black solid line represents the second-order transition from an EI to a BI (band insulator), the dashed line represents the first-order CDW/SOO-EI transition. The label $0^+$ emphasizes that the EI phase is present only for finite $f$-bandwidths.[]{data-label="fig1"}](fig1.eps){width="0.95\linewidth"}
From a theoretical point of view, the description of the EI with a Falicov-Kimball-type model seems promising. The original Falicov-Kimball (FKM) model [@FK69] contains itinerant $c$-electrons (with bandcenter $E_c$ and hopping amplitude $t_c$) that interact via a local Coulomb repulsion $U$ with localized $f$-electrons (with energy level $E_f$), where the spin is neglected. Since the local $f$-electron number is strictly conserved in the FKM, $f$-$c$-coherence cannot be established. [@SB88] One way to overcome this shortcoming is to include an $f$-$c$ hybridization. [@KMM76] As shown in Refs. , the extension by a finite $f$-bandwidth also induces $f$-$c$ coherence. The model with a direct $f$-$f$ hopping (with hopping amplitude $t_f$) is called the extended Falicov-Kimball model (EFKM) and has previously been used to describe different properties of the EI phase. [@Br08; @IPBBF08; @ZIBF10; @PBF10] The ground-state phase diagram of the EFKM was determined with a constraint path Monte Carlo (CPMC) technique for one and two dimensions (1D and 2D) in the strong [@Ba02b] and intermediate coupling regime [@BGBL04] as well as in the Hartree-Fock (HF) approximation for 2D, [@Fa08] 3D, [@Fa08] and infinite dimensions. [@SC08] All approaches yield a qualitatively similar phase diagram. Figure \[fig1\] displays the HF ground-state phase diagram for $U=2$ in 2D, depicting the generic situation. It was shown previously that Fig. \[fig1\] agrees with the CPMC data even quantitatively. [@Fa08] Besides full $c$- and full $f$-band insulator (BI) regions, the EFKM ground-state phase diagram exhibits three symmetry broken phases: the EI, a CDW, and a staggered orbital order (SOO). The EI is characterized by a nonvanishing average $\langle c^\dagger f^{} \rangle$. The CDW is described by a periodic modulation in the total electron density comprising both $f$- and $c$-electrons. The SOO is characterized by a periodic modulation in the difference between the $f$-electron and the $c$-electron density, which may be accompanied by a CDW. The SOO (CDW) establishes the ground state for the symmetric case ($E_f=E_c$) for all ratios of $-t_f/t_c$ (for the CDW the point $|t_f|=|t_c|$ has to be excluded, see below). These phases are rapidly suppressed in favor of the EI if $E_f\neq E_c$. Between the uniform EI phase and the CDW or SOO phase there is a first-order phase transition. The EI-BI transition is of second order. For $t_f=0$ the FKM is recovered, and the EI phase cannot be realized.
For the investigation of electron correlation effects the Gutzwiller approximation [@Gu63] is an established technique. Kotliar and Ruckenstein introduced a scalar slave-boson (SB) scheme which reproduces the Gutzwiller solution of the Hubbard model as a saddle-point. [@KR86] A manifestly spin-rotation invariant form of the SB representation has been worked out for the Hubbard model [@LWH89] and for multiband Hubbard models. [@LGKP07] We have developed an $SO(2)$-invariant SB approach for the EFKM (Ref. ) that reproduces the HF result for the EI phase boundary at $T=0$, but leads to a substantial reduction of the critical temperature. It is the aim of this work to include Gaussian fluctuations around the saddle-point [@La90; @LSW91; @DFKTI94] in the $SO(2)$-invariant SB scheme at zero and finite temperature. This offers an opportunity to calculate susceptibilities for investigating instabilities against long-range ordered phases and the formation of excitons.
The paper is organized as follows. In Sec. \[Theory\] the model Hamiltonian and the SB scheme are introduced. Moreover the saddle-point approximation is given and the calculation of response functions within the SB scheme is explained. In Sec. \[numerics\] we present numerical results for the instabilities toward the CDW, the SOO, and the EI phase. Finally we investigate the formation of excitons in the phase without long-range order. Section \[summary\] summarizes our results.
Theory {#Theory}
======
Model Hamiltonian {#Hamiltonian}
-----------------
Expressing the orbital flavor by a pseudospin variable $\sigma=\uparrow,\downarrow$, where $c_{i\uparrow}^{(\dagger)}\equiv f_i^{(\dagger)}$ and $c_{i\downarrow}^{(\dagger)}\equiv c_i^{(\dagger)}$, the EFKM can be written as an asymmetric Hubbard model, $$H=\sum_{i,\sigma} (E_\sigma-\mu) c_{i\sigma}^\dagger c_{i\sigma}^{ } - \sum_{\langle i,j\rangle,\sigma} t_\sigma c_{i\sigma}^\dagger c_{j\sigma}^{} + U\sum_i n_{i\uparrow} n_{i\downarrow}\;, \label{EFKM}$$ where $c_{i\sigma}^{(\dagger)}$ annihilates (creates) a $\sigma$-band electron at the Wannier site $i$ and $n_{i\sigma}^{ }=c_{i\sigma}^\dagger c_{i\sigma}^{ }$ is the corresponding number operator. $E_\sigma$ denotes the bandcenter of the $\sigma$-electron band, $\mu$ gives the chemical potential, $t_\sigma$ is the hopping amplitude, and $U$ measures the Coulomb interaction strength. In what follows we consider $E_\downarrow=0$, $E_\uparrow<0$, $t_\downarrow=1$, and $t_\uparrow<0$. All energies are measured in units of $t_\downarrow$. We restrict ourselves to $t_\downarrow t_\uparrow<0$, i.e., the valence band top and the conduction band minimum are located at the Brillouin zone center. Moreover we exclusively investigate the half-filled band case, i.e., $\tfrac{1}{N} \sum_{i,\sigma} \langle n_{i\sigma} \rangle =1$, where $N$ is the number of lattice sites.
Slave-boson functional integral representation {#SB-representation}
----------------------------------------------
Following Refs. the Hilbert space is enlarged by introducing auxiliary bosons: $e_i^{(\dagger)}$, related to an empty site, $d_i^{(\dagger)}$, related to a doubly occupied site, and $\underline{p}_i^{(\dagger)}$, related to a singly occupied site, $$\begin{aligned}
|0_i\rangle\;\;\rightarrow &\;\;e_i^\dagger |\mathrm{vac}\rangle\,,\\
|2_i\rangle\;\;\rightarrow &\;\;\tilde{c}_{i\uparrow}^\dagger
\tilde{c}_{i\downarrow}^\dagger d_i^\dagger|\mathrm{vac}\rangle\,,\\
|\sigma_i \rangle\;\;\rightarrow &\;\; \sum_{\rho}
\tilde{c}_{i\rho}^\dagger p_{i\rho\sigma}^\dagger |\mathrm{vac}\rangle\,,\end{aligned}$$ where $ |\mathrm{vac}\rangle$ means the vacuum state. The matrix operator $\underline{p}_i^{(\dagger)}$ is given as $$\underline{p}_i^{(\dagger)}=\frac{1}{2}\sum_{\mu}
\underline{\tau}_\mu p_{i\mu}^{(\dagger)}
=\frac{1}{2}\begin{pmatrix}
p_{i0}^{(\dagger)} +p_{iz}^{(\dagger)} & p_{ix}^{(\dagger)} -ip_{iy}^{(\dagger)}
\\[0.1cm] p_{ix}^{(\dagger)} +ip_{iy}^{(\dagger)} & p_{i0}^{(\dagger)}-p_{iz}^{(\dagger)}
\end{pmatrix}\,,$$ where $\mu=0,x,y,z$. $\underline{\tau}_0$ denotes the unit matrix and $\vec{\underline{\tau}}=(\underline{\tau}_x, \underline{\tau}_y, \underline{\tau}_z)^{\rm T}$ is the vector of the Pauli matrices. The fermionic degrees of freedom are captured by the pseudofermions $ \tilde{\mathbf{c}}^\dagger_i=(\tilde{c}_{i\uparrow}^\dagger , \tilde{c}_{i\downarrow}^\dagger)$ and $ \tilde{\mathbf{c}}^{}_i=(\tilde{c}_{i\uparrow}^{} , \tilde{c}_{i\downarrow}^{})^{\rm T}$.
Unphysical states of the extended fermion-boson Fock space are excluded by two sets of local constraints, $$\begin{aligned}
C^{(1)}_i &=& e^\dagger_i e_i^{}\;
+ 2\, \mbox{Tr}\,
{\underline{p}}^\dagger_i{\underline{p}}^{}_i \,+ d^\dagger_id_i^{}\,- 1\, =0\,,\label{c1}\\
\underline{C}^{(2)}_i &=& \tilde{\mathbf{c}}^{}_i \tilde{\mathbf{c}}^\dagger_i +
2 \, {\underline{p}}^\dagger_i{\underline{p}}^{}_i
+ d^\dagger_id_i^{}\,\underline{\tau}_0 -\underline{\tau}^{}_0 =0\,.
\label{c2}\end{aligned}$$
Since the bosonic occupation number of one site is coupled to the fermionic occupation, the bosons have to change simultaneously when an electron is created or annihilated. This is achieved by introducing the bosonic hopping operator $\underline{z}_i$, $$c_{i\sigma} = \sum_\rho z_{i\sigma \rho} \tilde{c}_{i\rho} \label{bosonHop1} \;.$$ The choice of $\underline{z}_i$ is not unique. We choose [@LWH89; @ZIBF10] $${\underline{z}_i} = {\underline{L}_i}e^\dagger_i
{{M}_i}{\underline{p}^{}_i} {\underline{N}_i}
+ {\underline{L}_i} {\tilde{\underline{p}}^{\dagger}_i}
{{M}_i} d^{}_i {\underline{N}_i}\;
\label{E_zsri}$$ with $$\begin{aligned}
{\underline{L}}^{}_i &=& [( 1 - d_i^\dagger d^{}_i )\underline{\tau}_0 -
2{\underline{p}}^\dagger_i{\underline{p}}^{}_i ]^{-1/2}\;,\\
{\underline{N}_i} &=& [( 1 - e_i^\dagger e^{}_i )\underline{\tau}_0 -
2{\underline{\tilde{p}}}^\dagger_i{\underline{\tilde{p}}}^{}_i
]^{-1/2}\;,\\
M^{}_i &=& [1 + e^\dagger_i e^{}_i + d^\dagger_i d^{}_i
+ 2\, \mathrm{Tr}\, {\underline{p}}^\dagger_i{\underline{p}}^{}_i]^{1/2}\:,
\label{LMR_d}\end{aligned}$$ and , which guarantees the correct free-fermion result on the mean-field level. The Hubbard interaction term is bosonized via $n_{i\uparrow} n_{i\downarrow} \rightarrow d_i^\dagger d_i^{}$.
The resulting coupled fermion-boson system is evaluated within a functional integral representation. Then, the bosons become complex fields and the fermions are given by complex Grassmann fields. The Lagrange multipliers $\lambda_i^{(1)}$, $\lambda_{i0}^{(2)}$, $\lambda_{ix}^{(2)}$, $\lambda_{iy}^{(2)}$, and $\lambda_{iz}^{(2)}$ are introduced to enforce the constraints and . Exploiting the gauge symmetry of the action and transforming the Lagrange multipliers into real time-dependent Bose fields we can remove the phases of $p_{i0}$, $p_{iz},$ and $e_i$. Using the Grassmann integration formula, we obtain the grand canonical partition function given by a functional integral over Bose fields only, $$\begin{aligned}
Z=\int &D[e] D [p^{}_0] D [p^{\ast}_x,p^{}_x] D [p^{\ast}_y,p^{}_y] D [p^{}_z] D[d^\ast,d] \nonumber \\
&D[\lambda^{(1)}] D[\lambda^{(2)}_0] D[\vec{\lambda}^{(2)}] \,\mathrm{e}^{-S}
\label{E_zusta2}\end{aligned}$$ with the effective bosonic action $$\begin{aligned}
S =& \int\limits_0^\beta d\tau \bigg\{\sum_i
\Big[- \lambda^{(1)}_i+
\lambda^{(1)}_i e^2_i + \sum_\mu ( \lambda^{(1)}_i -
\lambda^{(2)}_{i0} ) |p_{i\mu}|^2 \nonumber \\
&\qquad\quad\qquad- p^{ }_{i0}(\vec{p}^{\,\ast }_i+\vec{p}^{ }_i)
\vec{\lambda}^{(2)}_{i}
-i\vec{\lambda}^{(2)}_i (\vec{p}^{\,\ast}_i\times \vec{p}^{ }_i)
\nonumber \\[0.1cm]
&\qquad\quad\qquad+(\lambda_i^{(1)} + U-2\lambda_{i0}^{(2)}) |d_{i}|^2
\nonumber \\
&\qquad\quad\qquad+ p^\ast_{ix}\partial_\tau p_{ix}+p^\ast_{iy}\partial_\tau p_{iy}+d_i^\ast \partial_\tau d_i^{}
\Big]\bigg\}\nonumber \\
&\;\;- \mbox{Tr}\,
\ln\Big\{-G^{-1}_{\langle ij\rangle,\rho \rho '}(\tau, \tau ')\Big\}\,,
\label{E_seff}\end{aligned}$$ where $\vec{p}_i=(p_{ix},p_{iy},p_{iz})$ and $\vec{\lambda}_i^{(2)}=(\lambda_{ix}^{(2)}, \lambda_{iy}^{(2)},\lambda_{iz}^{(2)})$. The inverse Green propagator is given by $$\begin{aligned}
G^{-1}_{\langle ij\rangle,\rho \rho '}(\tau, \tau ') =&
\Big[
\big( -\partial^{ }_\tau + \mu - \lambda^{(2)}_{i0}\big)\delta^{ }_{\rho \rho '}
\nonumber\\
&- \tfrac{E_\uparrow}{2} (\underline{\tau}_0
+\underline{\tau}_z)_{\rho \rho'}-\vec{\lambda}^{(2)}_{i}
\vec{\tau}_{\rho \rho '}
\Big]\delta^{ }_{i j}\,\delta (\tau - \tau ')\nonumber\\
&+
(\underline{z}^\ast_i\,\underline{t}\,\underline{z}^{}_j)^{}_{\rho \rho ',
\tau \tau '}(1-\delta^{ }_{i j})\,,
\label{E_Gij}\end{aligned}$$ where $\underline{t}=\begin{pmatrix} t_\uparrow & 0 \\ 0 & t_\downarrow\end{pmatrix}$. The trace in Eq. (\[E\_seff\]) extends over time, space, and spin variables. For the half-filled band case Eqs. and are an exact representation of the partition function of the EFKM. One obtains $\underline{z}_i=z_i \underline{\tau}_0$.
Saddle-point approximation {#saddle-point}
--------------------------
To proceed we approximate all bosonic fields by their time-averaged values (static approximation), i.e., the bosonic fields are taken to be real. Moreover, we look for uniform solutions, that is, the Bose fields are taken to be independent of the lattice site.
We restrict ourselves to the phase without long-range order, which we denote as paraphase. The saddle-point equations for the paraphase ($p_{x}=p_{y}=\lambda_{x}^{(2)}=\lambda_{y}^{(2)}=0$) are $$\begin{aligned}
p_0 p_z &=& \frac{1}{2}\left( n_\uparrow-n_\downarrow \right) \;, \label{p0pz} \\
p_0^2 &=& \frac{1}{2} +\sqrt{n_\uparrow n_\downarrow (1-z^2)} \;, \label{p02} \\
d^2 &=& \frac{1}{2 z^2} \biggl[ z^2 \left( 2-p_0^2-p_z^2\right) + 2p_0^2\nonumber \\
&& -2p_0 \sqrt{z^2\left(2-p_0^2-p_z^2\right)+z^4p_z^2+p_0^2}\biggr]\;,\label{d} \\
\lambda_z^{(2)} &=& -\frac{p_z}{p_0}\left(\frac{z^2}{2d^2}-\frac{1}{p_0^2-p_z^2}\right) z^2 \epsilon(0) \;,\label{lambda_z} \\
U&=&-\frac{2d^2-p_0^2+z^2p_z^2}{p_0^2d^2} z^2 \epsilon(0)-2\lambda_z^{(2)} \frac{p_z}{p_0}\;,\label{corrEqn} \\
n_\sigma &=& \frac{1}{N}\sum_{\bf k} n_{{\bf k}\sigma} \label{occupation_number} \;,\\
\epsilon(0) &=& \frac{1}{N} \sum_{\bf k} \left( t_\uparrow \gamma_{\bf k} n_{{\bf k}\uparrow} + t_\downarrow \gamma_{\bf k} n_{{\bf k}\downarrow}\right) \;,\label{epsilon}\end{aligned}$$ where $$\begin{aligned}
n_{{\bf k}\sigma} &=& [\exp( \beta E_{{\bf k}\sigma})+1]^{-1} \;, \label{n_k} \\
E_{{\bf k}\sigma}& =& E_\sigma +\sigma \lambda_z^{(2)}-\tilde\mu-z^2 t_\sigma \gamma_{\bf k} \;, \label{Ek_updown} \\
\tilde \mu &=&\mu-\lambda_0^{(2)} \;. \label{mu_tilde}\end{aligned}$$ On a D-dimensional hypercubic lattice, $ \gamma_{\bf k}=2\sum_{l=1}^D \cos k_l$. The chemical potential is determined by the condition $$\frac{1}{N}\sum_{{\bf k},\sigma} n_{{\bf k}\sigma}=1\,.
\label{mu2}$$
The quasiparticle gap $E_g$ indicates the splitting of the $\uparrow$- and $\downarrow$-band (in the paraphase), which is caused by the correlation-induced quasiparticle bandshift $\lambda_z^{(2)}$. For a D-dimensional hypercubic lattice, $E_g$ is given by $$E_g=|E_\uparrow|+|2\lambda_z^{(2)}|-2Dz^2(|t_\uparrow|+|t_\downarrow|) \;. \label{Hartree}$$ For a SM, $E_g\leq0$ and for a SC, $E_g>0$.
We obtain the EI phase boundary by solving the SB gap equation, $$1= \frac{1}{p_0 p_z} \lambda_z^{(2)} \frac{1}{N} \sum_{\bf k} \frac{n_{{\bf k}\uparrow} -n_{{\bf k}\downarrow}}{E_{{\bf k}\uparrow}-E_{{\bf k}\downarrow}} \;, \label{SBGapEq}$$ resulting from Eqs. (63) and (65) of Ref. .
The gap equation captures both the BCS and the BEC situation, but it cannot discriminate between them. To this end, we follow an idea from Ihle et al. (Ref. ) and investigate the excitonic susceptibility in the paraphase.
Gaussian fluctuations {#Gaussian-fluct}
---------------------
In order to study response functions, we take into account Gaussian fluctuations around the saddle point for the paraphase, that is, $\Phi_{ia} = \bar\Phi_a + \delta\Phi_{ia}$, where $\Phi_{ia}=\{e_i,p_{i0},p_{ix},p_{iy},p_{iz},d_i,\lambda_i^{(1)},\lambda_{i0}^{(2)}, \lambda_{ix}^{(2)}, \lambda_{iy}^{(2)}, \lambda_{iz}^{(2)}\}$. Then, the action is given by $$S = \bar{S} + \sum_{q,a,b} \delta\Phi_a(-q) {\cal S}_{ab}(q) \delta\Phi_b(q) \;,
\label{fluct_expansion}$$ where the bar denotes the saddle-point value.
In order to achieve comparability with the saddle-point results, we start the fluctuation calculation from the same level of approximation as for the saddle-point calculation, i.e., we first perform the static approximation and consider only the fluctuations of the 11 real-valued fields $\Phi_{ia}$.
The fluctuation matrix can be calculated according to $$\begin{aligned}
{\cal S}&_{ab}(q,q') \nonumber \\
=&\; \frac{1}{2N\beta} \sum_{R_i,R_j} e^{-iqR_i} \frac{\partial^2 S}{\partial \Phi_{ia} \partial \Phi_{jb}}\Bigg|_{\tiny\begin{matrix}\Phi_i=\Phi_j=\bar{\Phi}\\\tau=\tau'\end{matrix}} e^{-iq'R_j}\nonumber\\=&\;{\cal S}_{ab}(q)\delta_{q,-q'}\;.
\label{fluct_matrix1}\end{aligned}$$ Here, we use the shorthand notation ${R}_i=({\bf R}_i, \tau)$ and ${q}=({\bf q}, \omega_m)$, where $\tau$ is the imaginary time, $\omega_m=2\pi m/\beta$ denote the bosonic Matsubara frequencies, ${\bf R}_i$ is the position vector, and ${\bf q}$ is the wave vector.
The response functions can be expressed in terms of the SB field fluctuations using the local constraints and . The charge susceptibility reads $$\begin{aligned}
\chi_c(q) &=& \langle \delta\left[n_{\uparrow}(-q) + n_\downarrow(-q)\right] \delta\left[n_{\uparrow}(q) + n_\downarrow(q)\right] \rangle \nonumber \\
&=& 4 \left( e^2 \langle \delta e(-q) \delta e(q) \rangle - 2ed \langle \delta e(-q) \delta d(q)\rangle \right. \nonumber \\
&&\left.+ d^2 \langle \delta d(-q) \delta d(q) \rangle \right) \label{charge_sus}\;.\end{aligned}$$ The orbital susceptibility is given by $$\begin{aligned}
\chi_o(q) &=& \langle \delta\left[n_{\uparrow}(-q) - n_\downarrow(-q)\right] \delta\left[n_{\uparrow}(q) - n_\downarrow(q)\right] \rangle \nonumber \\
&=&4 \left( p_z^2 \langle \delta p_0(-q) \delta p_0(q) \rangle + 2p_z p_0\langle \delta p_0(-q) \delta p_z(q) \rangle \right.\nonumber \\
&&\left.+ p_0^2 \langle \delta p_z(-q) \delta p_z(q) \rangle \right)\label{spin_sus}\;.\end{aligned}$$ Considering the creation operator of an onsite electron-hole pair [@IPBBF08] $$b_i^\dagger = c_{i\downarrow}^\dagger c_{i\uparrow}^{\,} \;, \hspace{0.5cm}
b_{\bf q}^\dagger = \frac{1}{\sqrt{N}}\sum_{\bf k} c_{{\bf k}+{\bf q}\downarrow}^\dagger c_{{\bf k}\uparrow}^{} \;,\label{X-op2}$$ the electron-hole susceptibility, hereafter denoted as excitonic susceptibility, is given by $$\begin{aligned}
\chi_X(q) &=& \langle \delta b_{\bf q}^{} \,\delta b_{\bf q}^\dagger \rangle \nonumber \\
&=& p_0^2\left[ \langle \delta p_x(-q)\,\delta p_x(q)\rangle + \langle \delta p_y(-q)\,\delta p_y(q)\rangle \right.\nonumber\\
&&\left. -i\langle \delta p_y(-q)\,\delta p_x(q)\rangle+i\langle \delta p_x(-q)\,\delta p_y(q)\rangle \right] \;. \nonumber \\\label{X-prop2}\end{aligned}$$
The correlation functions may be expressed as functional integrals over Bose fields: $$\langle \delta\Phi_a(-q) \delta\Phi_b(q)\rangle = \frac{1}{Z} \int D[\Phi] \; \delta\Phi_a(-q) \delta\Phi_b(q)\; e^{-{\cal S}(q)}\;.\label{CorrFuncGen}$$ Hence, the correlation functions are related to the inverse fluctuation matrix by $$\langle \delta\Phi_a(-q) \delta\Phi_b(q)\rangle = \frac{1}{2} {\cal S}_{ab}^{-1}(q)\;. \label{CorFluct}$$
It turns out that for the paraphase the $11\times 11$ fluctuation matrix decomposes into a $7\times 7$ matrix containing the charge fluctuations ($\delta e$, $\delta p_0$, $\delta d$, $\delta \lambda^{(1)}$, $\delta \lambda_0^{(2)}$) and the orbital fluctuations ($\delta p_z$, $\delta \lambda_z^{(2)}$) and into a $4\times 4$ matrix containing the electron-hole pair fluctuations ($\delta p_x$, $\delta \lambda_x^{(2)}$, $\delta p_y$, $\delta \lambda_y^{(2)}$). The SB fields are obtained by solving the saddle-point equations – self-consistently.
The description of the CDW and SOO requires the inclusion of inhomogeneous solutions with a periodic modulation in the densities, $\langle n_{i\sigma} \rangle = n_\sigma + \delta_\sigma \cos({\bf QR}_i)$, where the order vector in 3D is given by ${\bf Q}=(\pi,\pi,\pi)$. The CDW and SOO order parameters are $\delta_{\rm CDW} =\tfrac{1}{2}(\delta_\uparrow+\delta_\downarrow)$ and $\delta_{\rm SOO} = \tfrac{1}{2}(\delta_\uparrow-\delta_\downarrow)$, respectively. [@ZFB10] If $|\delta_\uparrow|\neq |\delta_\downarrow|$, the CDW and SOO describe the same symmetry broken state. We can investigate the formation of both phases without generalizing the SB formalism to a bipartite lattice by calculating the static ($\omega=0$) charge and orbital susceptibility with order vector ${\bf q}={\bf Q}$, given by $$\begin{aligned}
\chi_c &=&\chi_c({\bf Q},0)\nonumber \\
&=& 2\left[e^2({\cal S}^{-1})_{ee}+d^2({\cal S}^{-1})_{dd}-2ed({\cal S}^{-1})_{ed}\right]\;,\label{charge_sus2} \\
\chi_o &=& \chi_o({\bf Q},0) \nonumber \\
&=& 2\left[p_z^2({\cal S}^{-1})_{p_0p_0}+p_0^2({\cal S}^{-1})_{p_zp_z}-2p_0p_z({\cal S}^{-1})_{p_0p_z}\right]\;.\nonumber \\\label{spin_sus2}\end{aligned}$$ The inversion of the $7\times7$ matrix is performed numerically.
After analytic continuation ($i\omega_m\rightarrow \omega +i0^+$) the excitonic susceptibility yields $$\chi_X({\bf q},\omega) = \frac{\chi_X^{(0)}({\bf q},\omega)}{-\tfrac{{\cal S}_{p_x p_x}}{p_0^2}\, \chi_X^{(0)}({\bf q},\omega) +1 } \;, \label{X-prop3}$$ with $$\chi_X^{(0)}({\bf q},\omega)=\frac{1}{N} \sum_{\bf k} \frac{n_{{\bf k}\uparrow}-n_{{\bf k}+{\bf q}\downarrow} }{\omega + E_{{\bf k}\uparrow}-E_{{\bf k}+{\bf q} \downarrow} } \; \label{G0}$$ and $${\cal S}_{p_x p_x}=\left( \frac{1}{p_0^2} - \frac{1}{2} \frac{p_0^2-p_z^2}{p_0^2d^2} z^2 \right) z^2\,\epsilon(0) + \frac{p_z}{p_0} \lambda_z^{(2)} \;. \label{S_pxpx}$$ For the BI at $T=0$ the random phase approximation result [@IPBBF08] is recovered, $-\frac{S_{p_x p_x}}{p_0^2}= U$.
To determine the EI phase we compute the static excitonic susceptibility $\chi_X({\bf q},0)$. The direct band gap situation gives the order vector of the EI phase as ${\bf q}=0$. Using Eq. the fluctuation matrix element ${\cal S}_{p_x p_x}$ \[Eq. \] reduces to $${\cal S}_{p_x p_x} = \frac{p_0}{p_z} \lambda_z^{(2)} \;. \label{S_pxpx_gap}$$ It is easy to see that the condition for the divergence of $\chi_X(0,0)$ equates to the gap equation .
The poles of $\operatorname{Re}\chi_X^{(0)}({\bf q},\omega)$ \[Eq. \] give the continuum of electron-hole excitations, i.e., $\omega_k({\bf q})=E_{{\bf k}+{\bf q} \downarrow}-E_{{\bf k}\uparrow}$. Excitonic pairing of electrons and holes is described by the pole of $\operatorname{Re}\chi_X({\bf q},\omega)$ \[Eq. \] outside the electron-hole continuum, [@IPBBF08] i.e., by $$\operatorname{Re}\chi_X^{(0)}({\bf q},\omega_X) = \frac{p_0^2}{S_{p_x p_x} } \;, \label{X-cond}$$ with $$0 < \omega_X({\bf q}) < \omega_C({\bf q}) \;, \label{X-cond_energy}$$ where $\omega_C({\bf q})=\omega_k({\bf q})|_{\rm min}$. The exciton binding energy is given by $$E_X^B ({\bf q}) = \omega_C({\bf q}) - \omega_X({\bf q}) \;. \label{X-bind}$$ We want to emphasize that $\omega_X$, $\omega_C$ and $E_X^B$ are explicitly ${\bf q}$-dependent in contrast to Ref. , where only excitons with ${\bf q}=0$ are considered, and Ref. , where the exciton binding energy is assumed to be independent of ${\bf q}$.
Numerical results {#numerics}
=================
Instabilities against CDW and SOO {#CDW-SOO}
---------------------------------
To obtain results for the 3D EFKM we transform the ${\bf k}$-summation into an energy integral using the tight-binding density of states (DOS) for a simple cubic lattice.
From the charge and orbital susceptibility we derive information about the CDW and SOO formation, respectively.
![(Color online) Static charge and orbital susceptibility of the 3D EFKM for $T=0$ and $U=4$ as a function of $E_\uparrow$.[]{data-label="fig2"}](fig2.eps){width="0.95\linewidth"}
For asymmetric bands ($|t_\uparrow|\neq |t_\downarrow|$) the charge and orbital susceptibility diverge at the same critical $E_\uparrow$, as shown in Fig. \[fig2\], implying $|\delta_\uparrow|\neq |\delta_\downarrow|$. The analogy between CDW and SOO vanishes if the bandwidths are equal, as can be seen for $t_\uparrow=-1.0$ in Fig. \[fig2\]. In this case, the orbital susceptibility diverges contrary to the charge susceptibility, thus, a CDW will not develop and $\delta_\uparrow=-\delta_\downarrow$. We conclude that the density inhomogeneity $\delta_\sigma$ is largely affected by the bandwidth.
![(Color online) Static charge and orbital susceptibility of the 3D EFKM for $T=0$ and $t_\uparrow=-0.8$ as a function of $E_\uparrow$.[]{data-label="fig3"}](fig3.eps){width="0.95\linewidth"}
Figure \[fig3\] shows $\chi_o$ and $\chi_c$ for $t_\uparrow=-0.8$. The susceptibilities diverge at the same critical $E_\uparrow$. With increasing strength of the Coulomb interaction the critical $|E_\uparrow|$ for CDW (SOO) formation increases, because for a larger interaction the charge (orbital) order becomes more favorable. Figure \[fig3\] clearly shows that the CDW and SOO region is confined close to the symmetric case $E_\uparrow=0$.
For small band splitting either the CDW (SOO) or the EI, separated by a first-order phase transition, can be realized, and one has to compare the free energies to identify the true ground state. Hence, to determine the SB ground-state phase diagram (analogous to the HF case shown in Fig. \[fig1\]) the generalization of the saddle-point equations to a bipartite lattice is inevitable, which is beyond the scope of this work. To investigate the EI in the following, we choose the band-structure parameters $E_\uparrow=-2.4$ and $t_\uparrow=-0.8$, where a CDW (SOO) is not realized (see Fig \[fig3\]).
Instability against EI {#EI}
----------------------
Figure \[fig4\] shows that the EI phase boundary in the weak-coupling as well as in the strong-coupling regime is reproduced by poles of the uniform static excitonic susceptibility, as demonstrated analytically in Sec. \[Gaussian-fluct\].
![(Color online) EI phase boundary (black solid line) of the 3D EFKM in the $U$-$T$ plane. The red diamonds give the poles of the uniform static excitonic susceptibility for some fixed values of $U$. The blue circles give the critical $U_X$ for exciton formation with center-of-mass momentum ${\bf q}=0$. The black dashed line separates the SM and the SC phase. Inside the EI phase we suggest that the black dashed line gives a crude estimate for the BCS-BEC crossover region (see text). The band-structure parameters are $E_\uparrow=-2.4$ and $t_\uparrow=-0.8$.[]{data-label="fig4"}](fig4.eps){width="0.95\linewidth"}
To determine the region where free excitons can exist, we evaluate the condition for exciton formation subjected to the constraint . The exciton binding energy has to be positive. For numerical reasons we set the threshold to $\min(E_X^B)=10^{-6}$. For the 3D case we restrict ourselves to excitons with a center-of-mass momentum ${\bf q}=0$. In Fig. \[fig4\] the critical Coulomb interaction strength for the exciton formation $U_X(T)$ is shown by blue circles, where zero-momentum excitonic states exist for $U\geq U_X$. For low temperatures $U_X(T)$ coincides with the EI phase boundary in the strong-coupling regime. This suggests that in this regime the EI is developed by a BEC of preformed excitons. Within our analysis, for high temperatures $U_X(T)$ deviates slightly from the critical $U_g(T)$, determined from Eq. , which separates the SM ($U\leq U_g$) and the SC ($U>U_g$), except for the point where $U_g(T)$ coincides with the EI phase boundary. This deviation turns out to be a result of the required finite binding energy of the excitons. In a SM excitons with ${\bf q}=0$ cannot exist. Here, the EI develops due to a BCS-type pairing of electrons and holes. We suggest that the BCS-BEC crossover region is estimated by extrapolating $U_g(T)$ into the EI phase.
To consider excitons with an arbitrary center-of-mass momentum, the ${\bf q}$-resolved excitonic susceptibility is calculated for the 2D EFKM, in order to keep the numerical effort manageable. Again we choose the band-structure parameters $E_\uparrow=-2.4$ and $t_\uparrow=-0.8$, for which the EI phase is stable in 2D, [@BGBL04; @Fa08] see Fig. \[fig1\]. To evaluate the SB parameters, the ${\bf k}$-summation is transformed into an energy integral using the tight-binding DOS for a square lattice. For the computation of the excitonic susceptibility the ${\bf k}$-summation is explicitly performed.
![Phase diagram of the 2D EFKM for the band-structure parameters $E_\uparrow=-2.4$ and $t_\uparrow=-0.8$. The solid line shows the EI phase boundary and the dashed line separates the SM and the SC phase. The exciton dispersion at the marked points a,b,c,d is shown in Fig. \[fig8\], and the exciton dispersion for e and f is shown in Fig. \[fig9\].[]{data-label="fig5"}](fig5.eps){width="0.95\linewidth"}
The phase diagram is shown in Fig \[fig5\]. Qualitatively there is no difference between the phase diagram of the 2D and 3D EFKM. Hence, the following results for 2D should hold qualitatively for 3D, too.
Figure \[fig6\] shows the static excitonic susceptibility for zero temperature.
![(Color online) Static excitonic susceptibility for $T=0$ as a function of ${\bf q}$ (2D). For all $U$ we obtain $n_\uparrow=1$ and $n_\downarrow=0$.[]{data-label="fig6"}](fig6.eps){width="0.95\linewidth"}
The formation of the EI phase is indicated by the divergence of $\chi_X({\bf q},0)$ at ${\bf q}=0$ for the critical value $U_{\rm EI}=5.71$. For $U>U_{\rm EI}$ the static excitonic susceptibility remains finite for all ${\bf q}$.
Excitonic bound states {#excitons}
----------------------
Next we want to characterize the paraphase in the vicinity of the EI phase with a view to the formation of uncondensed excitons.
![(Color online) Real part of the dynamic excitonic susceptibility as a function of ${\bf q}$ (2D) for $U=5.5$ and $T=0.256$ (upper panel) and the corresponding $\operatorname{Re}\chi_X^{(0)}({\bf q},\omega)$ (lower panel). The densities are $n_\uparrow=0.994$ and $n_\downarrow=0.006$. The vertical solid lines show the lower boundary of the electron-hole excitation continuum. The black dotted line gives the critical value of $\operatorname{Re}\chi_X^{(0)}$ for the exciton formation.[]{data-label="fig7"}](fig7a.eps "fig:"){width="0.95\linewidth"} ![(Color online) Real part of the dynamic excitonic susceptibility as a function of ${\bf q}$ (2D) for $U=5.5$ and $T=0.256$ (upper panel) and the corresponding $\operatorname{Re}\chi_X^{(0)}({\bf q},\omega)$ (lower panel). The densities are $n_\uparrow=0.994$ and $n_\downarrow=0.006$. The vertical solid lines show the lower boundary of the electron-hole excitation continuum. The black dotted line gives the critical value of $\operatorname{Re}\chi_X^{(0)}$ for the exciton formation.[]{data-label="fig7"}](fig7b.eps "fig:"){width="0.95\linewidth"}
Figure \[fig7\] shows the real part of the dynamic excitonic susceptibility outside the continuum for several values of $\omega$. The divergences of $\operatorname{Re}\chi_X({\bf q},\omega)$ point out that the system is unstable against the formation of free excitons. With increasing excitation energy $\omega$ the exciton momentum ${\bf q}$ for the exciton formation increases due to the direct band gap situation. The divergence of $\operatorname{Re}\chi_X^{(0)}({\bf q},\omega)$ means $\omega=\omega_C({\bf q})$, shown as the vertical solid lines in Fig. \[fig7\]. For the ${\bf q}$ values where $\operatorname{Re}\chi_X({\bf q},\omega)$ (upper panel) and $\operatorname{Re}\chi_X^{(0)}({\bf q},\omega)$ (lower panel) is not plotted in Fig. \[fig7\], the given $\omega$ is larger than $\omega_C({\bf q})$. Hence, these ${\bf q}$ values are irrelevant for the exciton formation for the considered excitation energy $\omega$.
Having confirmed the existence of excitonic states, we now turn to the properties of these states. We denote the energy-momentum relation derived from Eq. as the dispersion of the exciton band. The smallest pole of $\operatorname{Re}\chi_X^{(0)}({\bf q},\omega)$ defines the ${\bf q}$-resolved lower boundary of the electron-hole excitation continuum, which we denote in the following as the continuum band. In Fig. \[fig8\] the exciton band $\omega_X({\bf q})$ and the continuum band $\omega_C({\bf q})$ are shown for characteristic points in the phase diagram (see Fig. \[fig5\]). In general, the continuum band is more sensitive to the temperature and Coulomb strength than the exciton band. Hence, the binding energy of the excitons is mainly affected by the continuum band when $T$ or $U$ is varied. Figure \[fig8\] suggests that the exciton dispersion has a cosine-like form, determined by the underlying lattice.
![Exciton band (solid line) in comparison with the continuum band (dashed line) at the points marked in Fig. \[fig5\]: (a) in the SC phase ($n_\uparrow=0.996$, $n_\downarrow=0.004$, $\tilde{\mu}=0.527$), (b) in the SC phase with a smaller band gap ($n_\uparrow=0.988$, $n_\downarrow=0.012$, $\tilde{\mu}=0.188$), (c) at the SC-SM transition ($n_\uparrow=0.973$, $n_\downarrow=0.027$, $\tilde{\mu}=0.056$), (d) in the SM phase ($n_\uparrow=0.960$, $n_\downarrow=0.040$, $\tilde{\mu}=-0.013$). The chemical potentials are measured relative to the valence band top. In all figures $q_y=0$ (2D).[]{data-label="fig8"}](fig8a.eps){width="\linewidth"}
![Exciton band (solid line) in comparison with the continuum band (dashed line) at the points marked in Fig. \[fig5\]: (a) in the SC phase ($n_\uparrow=0.996$, $n_\downarrow=0.004$, $\tilde{\mu}=0.527$), (b) in the SC phase with a smaller band gap ($n_\uparrow=0.988$, $n_\downarrow=0.012$, $\tilde{\mu}=0.188$), (c) at the SC-SM transition ($n_\uparrow=0.973$, $n_\downarrow=0.027$, $\tilde{\mu}=0.056$), (d) in the SM phase ($n_\uparrow=0.960$, $n_\downarrow=0.040$, $\tilde{\mu}=-0.013$). The chemical potentials are measured relative to the valence band top. In all figures $q_y=0$ (2D).[]{data-label="fig8"}](fig8b.eps){width="\linewidth"}
![Exciton band (solid line) in comparison with the continuum band (dashed line) at the points marked in Fig. \[fig5\]: (a) in the SC phase ($n_\uparrow=0.996$, $n_\downarrow=0.004$, $\tilde{\mu}=0.527$), (b) in the SC phase with a smaller band gap ($n_\uparrow=0.988$, $n_\downarrow=0.012$, $\tilde{\mu}=0.188$), (c) at the SC-SM transition ($n_\uparrow=0.973$, $n_\downarrow=0.027$, $\tilde{\mu}=0.056$), (d) in the SM phase ($n_\uparrow=0.960$, $n_\downarrow=0.040$, $\tilde{\mu}=-0.013$). The chemical potentials are measured relative to the valence band top. In all figures $q_y=0$ (2D).[]{data-label="fig8"}](fig8c.eps){width="\linewidth"}
![Exciton band (solid line) in comparison with the continuum band (dashed line) at the points marked in Fig. \[fig5\]: (a) in the SC phase ($n_\uparrow=0.996$, $n_\downarrow=0.004$, $\tilde{\mu}=0.527$), (b) in the SC phase with a smaller band gap ($n_\uparrow=0.988$, $n_\downarrow=0.012$, $\tilde{\mu}=0.188$), (c) at the SC-SM transition ($n_\uparrow=0.973$, $n_\downarrow=0.027$, $\tilde{\mu}=0.056$), (d) in the SM phase ($n_\uparrow=0.960$, $n_\downarrow=0.040$, $\tilde{\mu}=-0.013$). The chemical potentials are measured relative to the valence band top. In all figures $q_y=0$ (2D).[]{data-label="fig8"}](fig8d.eps){width="\linewidth"}
In Fig. \[fig8\](a), for any momentum a finite energy is needed to transfer one electron from the valence band to the conduction band, $\omega_C({\bf q})>0$, which is characteristic for the underlying SC band structure. Likewise the creation of an exciton requires energy, $\omega_X({\bf q})>0$ for all ${\bf q}$. The comparison of Fig. \[fig8\](b) with Fig. \[fig8\](a) unveils the influence of the Coulomb interaction strength. In Fig. \[fig8\](b), the exciton band is shifted to lower energies because the point (b) in the phase diagram (see Fig. \[fig5\]) is closer to the EI phase than (a). For the continuum bands $\omega_C^{(b)}({\bf q})<\omega_C ^{(a)}({\bf q})$ and, therefore, the binding energy of the excitons in Fig. \[fig8\](b) is smaller than in Fig. \[fig8\](a), since $U_{(b)}<U_{(a)}$, i.e., the Coulomb attraction between electrons and holes in Fig. \[fig8\](b) is smaller than in Fig. \[fig8\](a) as well, and the electrons and holes are more loosely bound.
The exciton and continuum dispersion at the SC-SM transition are shown in Fig. \[fig8\](c). The continuum band reaches $\omega=0$ for ${\bf q}=0$, due to the direct band gap situation. The excitonic band disappears for small, finite center-of-mass momenta, not only for ${\bf q}=0$. The reason is the required finite binding energy of the excitons, i.e., $\omega_X({\bf q})$ should be separated from $\omega_C({\bf q})$.
In Fig. \[fig8\](d) the SM band structure is characterized by the vanishing of positive $\omega_C({\bf q})$ at some finite value of ${\bf q}$. In a SM band structure excitonic states exist only with finite center-of-mass momenta, in contrast to Figs. \[fig8\](a) and \[fig8\](b). The existence of excitonic bound states in a SM is confirmed by a very recent EFKM study with the projector-based renormalization method. [@PBF11] The comparison of Figs. \[fig8\](b), \[fig8\](c), and \[fig8\](d) demonstrates that the energetic position of the continuum band decreases with increasing temperature and, as a result, the exciton binding energy is lowered.
The qualitatively different feature of the exciton and the continuum band in the SC and in the SM phase suggests that the condensation process at the SC-EI transition differs from the one at the SM-EI transition.
![Exciton band (solid line) and continuum band (dashed line) (a) at point marked in Fig. \[fig5\]: the SC-EI transition ($n_\uparrow=0.995$, $n_\downarrow=0.005$, $\tilde{\mu}=0.280$), and (b) at point marked in Fig. \[fig5\]: the SM-EI transition ($n_\uparrow=0.949$, $n_\downarrow=0.051$, $\tilde{\mu}=-0.375$). The chemical potentials are measured relative to the valence band top. In all figures $q_y=0$.[]{data-label="fig9"}](fig9a.eps){width="\linewidth"}
![Exciton band (solid line) and continuum band (dashed line) (a) at point marked in Fig. \[fig5\]: the SC-EI transition ($n_\uparrow=0.995$, $n_\downarrow=0.005$, $\tilde{\mu}=0.280$), and (b) at point marked in Fig. \[fig5\]: the SM-EI transition ($n_\uparrow=0.949$, $n_\downarrow=0.051$, $\tilde{\mu}=-0.375$). The chemical potentials are measured relative to the valence band top. In all figures $q_y=0$.[]{data-label="fig9"}](fig9b.eps){width="\linewidth"}
Figure \[fig9\](a) shows the exciton and the continuum dispersion at the SC-EI transition. As already stated, $\omega_C({\bf q})>0$ $\forall {\bf q}$ reflects the underlying SC band structure. The minimum of the exciton band (at ${\bf q}=0$) reaches $\omega=0$, so that the creation of a zero-momentum exciton does not require energy. Physically, comparing only the energies the exciton band touches the valence band top. In our analysis, this is the point where the BEC of excitons sets in.
On the contrary, the exciton dispersion at the SM-EI transition, shown in Fig. \[fig9\](b), gives no hint for a condensation process. In fact, it is a characteristic plot for a SM band structure: there are excitonic bound states, but only with a finite center-of-mass momentum, c.f. Fig. \[fig8\](d). In the SM regime there is a large and well-defined Fermi surface and the electron-hole condensation at the SM-EI transition can be imagined as a BCS-type process.
Summary
=======
In this paper we studied the extended Falicov-Kimball model within a $SO(2)$-invariant slave-boson treatment taking Gaussian fluctuations into account. We computed the static charge and orbital susceptibility in order to investigate the formation of a charge density wave and staggered orbital order. Analyzing the static excitonic susceptibility, the instability towards an excitonic insulator (EI) phase agrees with the saddle-point phase diagram. By calculating the dynamic excitonic susceptibility, we analyzed the formation of excitons in the phase without long-range order. We found that finite-momentum excitons form in the vicinity of the EI phase, not only at the semiconducting (SC) side but also at the semimetal (SM) side. At the transition to the EI phase we observed qualitatively different features at the SC and the SM side. At the SC-EI transition the excitation energy of the excitonic state with zero center-of-mass momentum vanishes, leading to a Bose-Einstein condensation of these preformed excitons. At the SM side there are no excitonic bound states with ${\bf q}=0$. Here, the EI forms due to a BCS-type pairing of electrons and holes, and the occurring excitonic states of finite momentum are not of importance for the phase transition.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank K. W. Becker, R. L. Heinisch, V.-N. Phan, and O. P. Shuskov for stimulating discussions. HF acknowledges a Gordon Godfrey professorial visiting fellowship at the UNSW, Australia. This work was supported by DFG through SFB 652.
[38]{}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 [****, ()]{} in @noop [**]{}, (, , ) p. @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{}; @noop [****, ()]{}; [, ]{}@noop [****, ()]{} @noop [**** ()]{}; [, ]{}@noop [****, ()]{} @noop [****, ()]{}; [, ]{}@noop [****, ()]{}; [, ]{}@noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{}; [, ]{}@noop [****, ()]{} @noop [****, ()]{}; [, ]{}@noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{}; [, ]{}@noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} V.-N. Phan, K. W. Becker, and H. Fehske, EPL [**95**]{}, 17006 (2011).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study optical conductivity of a two-dimensional electron gas with anisotropic $k$-cubic Rashba spin-orbit interaction formed at the LaAlO$_3$/SrTiO$_3$ interface. The anisotropic spin splitting energy gives rise to different features of the optical conductivity in comparison to the isotropic $k$-cubic Rashba spin-orbit interaction. For large carrier density and strong spin-orbit couplings, the density dependence of Drude weight deviates from the linear behavior. The charge and optical conductivities remain isotropic despite anisotropic nature of the Fermi contours. An infinitesimally small photon energy would suffice to initiate inter-band optical transitions due to degeneracy along certain directions in momentum space. The optical conductivity shows a single peak at a given photon energy depending on the system parameters and then falls off to zero at higher photon energy. These features are lacking for systems with isotropic $k$-cubic Rashba spin-orbit coupling. These striking features can be used to extract the information about nature of the spin-orbit interaction experimentally and illuminate some light on the orbital origin of the two-dimensional electron gas.'
author:
- |
Alestin Mawrie and Tarun Kanti Ghosh\
Department of Physics, Indian Institute of Technology-Kanpur, Kanpur-208 016, India
title: 'Optical conductivity of a 2DEG with anisotropic Rashba interaction at the interface of LaAlO$_3$/SrTiO$_3$'
---
Introduction
============
Spin-orbit interaction [@soc; @soc2] (SOI) plays an important role in understanding physical properties of different materials as it lifts the spin degeneracy due to the absence of either the structure inversion symmetry or the time reversal symmetry. In general, there are two different types of symmetry dependent SOI, Rashba [@Rashba1; @Rashba2] and Dresselhaus [@dres] SOIs, in various condensed matter systems. In two-dimensional electron gas (2DEG) formed at the III-V semiconductor heterostructures[@bdas] and in various topological insulating systems [@TI], the Rashba SOI (RSOI) is linear in momentum and of the form $H_R = i\alpha k_-\sigma_++\textrm{h.c.}$, where $\alpha$ is the strength of RSOI, $\sigma_\pm = \sigma_x \pm i \sigma_y$ with $\sigma_x$ and $\sigma_y$ are the Pauli’s spin matrices and $k_\pm = k_x \pm i k_y$ with $k_x$ and $k_y$ the components of the wave vector ${\bf k}$. Besides, the Rashba SOI in two-dimensional hole gas formed at the interface of p-type GaAs/AlGaAs heterostructures[@sherman; @winkler], 2DEG on the surface of SrTiO$_3$ single crystals [@nakamura] and in 2D hole gas formed in a strained Ge/SiGe quantum well [@strain] is cubic in momentum and is of the form $H_R^{\rm iso}=i\alpha k_-^3\sigma_++\textrm{h.c}$. The spin splitting energy due to this RSOI is always isotropic and hereafter we will mention this as isotropic cubic RSOI.
An extremely high mobility 2DEG was discovered at the interface of the complex oxides LaAlO$_3$ and SrTiO$_3$ [@nature1; @nature2; @high-mobility]. The mobility at oxide interfaces is relatively less than that of III-V semiconductor heterostructures[@hih-mobilityIIIV]. Unlike the conventional III-V semiconductor heterojunctions, the 2DEGs at LaAlO$_3$/SrTiO$_3$ interfaces are characterized by very strong spin-orbit interaction, high carrier densities, higher effective mass[@high-mobility2]. The LAO/STO interface structure now has a broken structure inversion symmetry as a result of the confinement along the axis normal to the interface, which leads to the lifting of the spin-degeneracy of the six $t_{2g}$ orbitals in STO [@lee]. Moreover, the $d_{xy}$ orbitals are confined in the $x$-$y$ plane and are localized at the interface due to impurities and electron-phonon coupling [@popovic], whereas the electrons associated with the $d_{xz}$ and $d_{yz}$ orbitals [@popovic] are itinerant and contribute to transport. One of the major concerns is to understand the nature of the SOI of the charge carriers at the oxide interface. In Refs. [@held; @lee], a $k$-linear Rashba SOI for $d_{xy}$ orbital and an isotropic $k$-cubic for $d_{xz}$ and $d_{yz}$ orbitals were proposed. The magneto-transport measurement of 2DEG at the oxide interface has indicated the existence of $k$-cubic RSOI and is modeled using the isotropic $k$-cubic SOI, $H_R^{\rm iso}$ [@nakamura; @nayak]. On the other hand, the first-principle calculations suggested anisotropic non-parabolic spin-split branches for the $d_{xz}$ and $d_{yz}$ orbitals [@fasolino]. Two recent polarization-dependent ARPES revealed non-isotropic Fermi contours of the 2DEG at the oxide interface [@arpes]. Very recent theoretical study [@xiao] predicted that these orbitals are characterized by $k$-cubic but anisotropic Rashba spin-orbit interaction whose form is given by $ H_{R}^{\rm ani} = \alpha (k_{x}^2 - k_{y}^2)({\bf k} \times {{\boldsymbol}\sigma}) \cdot \hat z $. The spin splitting energy and Fermi contours become highly anisotropic as a result of this anisotropic SOI. In this paper we will refer this as anisotropic RSOI. This form of the anisotropic RSOI [@xiao] enables to explain the experimental observations of the anisotropic spin susceptibility [@spin-sus; @spin-sus1] successfully. It is also shown [@xiao] that the anisotropic RSOI leads to different behavior of the spin Hall conductivity, in comparison to the isotropic $k$-cubic RSOI.
The spectroscopic measurement of the absorptive part of the optical conductivity can probe the spin-split energy levels. Theoretical studies of the optical conductivity of various charged systems with an isotropic $k$-cubic Rashba SOI have been carried out [@lin; @yang; @wong; @li; @mawrie]. It is shown that the optical transition takes place for a certain range of photon energy depending on the carrier density and spin-orbit coupling constant. At zero temperature, it takes a box-like function and its value is $\sigma_{xx}^{\rm iso} = 3e^2/(16 \hbar)$, independent of carrier density and spin-orbit coupling strength.
In this paper we study the Drude weight and optical conductivities of the 2DEG with anisotropic $k$-cubic RSOI formed at the oxide interface and compare our results with that of the isotropic $k$-cubic RSOI. Firstly, we present the characteristics of the zero-frequency Drude weight as a function of the charge density and strength of the anisotropic RSOI. We find that the Drude weight is strongly modified due to the presence of the anisotropic $k$-cubic SOI. It deviates from the linear density dependence for large carrier density and for strong spin-orbit coupling. The Drude weight decreases with the increase of the strength of RSOI. Secondly, we find that an infinitesimally small photon energy would initiate the inter-band optical transition. This is due to the vanishing spin-splitting energy along certain directions in the momentum space. There is a single peak in the optical conductivity and its value depends on the electron density and strength of the anisotropic RSOI. Moreover, the charge and optical conductivities are isotropic despite the fact that the RSOI is anisotropic. In conventional 2DEG the van Hove singularities largely affects the various physical properties like transport$^{32}$, character of plasmons$^{33}$, N-type kink in photoluminescence$^{34}$, dilute-magnetic semiconductor properties$^{35}$ etc. Here as well, the van Hove singularities drastically affects the optical conductivity, thereby responsible for the single peak observed in it. The van Hove singularities are of the same $M_1$ type. These features can be used to find out the nature of the RSOI experimentally.
This paper is organized as follows. In section II, we describe basic properties of the 2DEG with anisotropic $k$-cubic spin-orbit interaction. In section III, we present the analytical and numerical results of the Drude weight and the optical conductivity. The summary and conclusions of this paper are presented in section IV.
Description of the Physical System
==================================
The effective Hamiltonian of the electron in $d_{xz}$ and $d_{yz}$ orbitals at the interface of LAO/STO is given by[@xiao] $$\begin{aligned}
H = \frac{\hbar^2 {\bf k}^2}{2m^*} +
\alpha (k_x^2-k_y^2)({\bf k} \times {{\boldsymbol}\sigma}) \cdot \hat z,\end{aligned}$$ where $m^\ast$ is the effective mass of the electron, $\alpha$ is the strength of the anisotropic RSOI and ${{\boldsymbol}\sigma}=\sigma_x \hat{x}+\sigma_y\hat{y}$. The above Hamiltonian is valid within the narrow region around the $\Gamma$ point. The anisotropic dispersion relations and the corresponding eigenfunctions are given by $$\begin{aligned}
\label{spec}
E_\lambda({\bf k})=\frac{\hbar^2 k^2}{2m^\ast} + \lambda\alpha k^3 |\cos 2\theta|\end{aligned}$$ and $ \psi_{\bf k}^\lambda({\bf r}) = e^{i {\bf k} \cdot {\bf r}}
\phi_{\bf k}^\lambda({\bf r})/\sqrt{\Omega}$ with the spinor $$\begin{aligned}
\phi_{\bf k}^\lambda({\bf r}) = \frac{1}{\sqrt{2}}
\begin{pmatrix}
1\\
\lambda \eta_{\bf k} i e^{i\theta}
\end{pmatrix}.\end{aligned}$$ Here $\Omega$ is the surface area of the two-dimensional system, $\lambda=\pm$ denotes the spin-split branches and $\eta_{\bf k}=\cos 2\theta/|\cos 2\theta|$ with $\theta = \tan^{-1}(k_y/k_x)$ measures the anisotropy of the spectrum. The magnitude of the anisotropic spin-splitting energy is $E_g({\bf k}) = | E_+({\bf k}) - E_-({\bf k}) | = 2\alpha k^3|\cos 2\theta|$. The spin splitting energy vanishes at $\theta= (2p+1)\pi/4$ with $p=0,1,2,3$. On the other hand, the maximum spin-splitting ($E_{g}^{\rm max} = 2 \alpha k^3$) occurs at $\theta=p\pi/2$. To allow only the bound states, the wave-vector ${\bf k}$ should have an upper cut-off given by $k_c(\pi/4)=\hbar^2/(3m^*\alpha) $ which corresponds to the cut-off energy $ E_c= \alpha k_c^3/2$.
The spin texture on the $k_x$-$k_y$ plane can be obtained from the average values of spin vector (in units of $3\hbar/2$) $ {\bf P}_{\lambda}({\bf k}) = \la {{\boldsymbol}\sigma} \ra_{\lambda} =
\lambda \eta_{\bf k} \hat \theta $, where $\hat \theta = -\hat x \sin \theta + \hat y \cos \theta $ is the unit polar vector. The electron spin lies in the ${\bf k} $ plane and always locked at right angles to its momentum.
The Berry connection [@berry] is defined as ${\bf A_k} = i \langle \phi_{\bf k}^\lambda|\bigtriangledown_{\bf k}|\phi_{\bf k}^\lambda\rangle$, where $ \phi_{\bf k}^{\lambda} $ is the spinor part of the wave function $\psi_{\bf k}^{\lambda}({\bf r})$. The Berry connection for this system yields $ {\bf A}_{\bf k} = -\hat \theta/(2k)$. Using the expression of the Berry phase [@berry] $\gamma = \oint {\bf A_k} \cdot d{\bf k} $, we get $ \gamma_{\rm ani} = - \pi $ for anisotropic case, whereas $\gamma_{\rm iso} = 3 \pi $ for isotropic cubic RSOI [@mawrie].
![(color online) Plots showing the density of states in the units of $D_0$ for two different values of $\alpha $. Here $\alpha=0.004$ eV nm$^3$ (solid black) and $\alpha=0.006$ eV nm$^3$ (dotted red).[]{data-label="Fig1"}](fig3.eps){width="95mm" height="70mm"}
In order to obtain two anisotropic Fermi contours $k_{f}^{\lambda}(\theta)$, we need to calculate density of states (DOS) and Fermi energy $E_f$. The density of states of the spin-split energy branches are given by $$\begin{aligned}
D_\lambda(E) & = &
\int \frac{d^2k}{(2\pi)^2} \delta(E - E_{\lambda}({\bf k})) \nonumber \\
& = & \frac{D_0}{2\pi}\int_0^{2\pi}
\frac{ k_E^\lambda(\theta)\,d\theta}{|k_E^\lambda(\theta) +
\lambda 6 \pi \alpha D_0 {(k_E^\lambda(\theta))}^2 |\cos 2\theta||} \nonumber,\end{aligned}$$ where $D_0=2\pi m^\ast/h^2$ and $k_E^\lambda(\theta)$ being the solution of the equation $( \hbar k_E^\lambda)^2/2m^\ast +\alpha \lambda (k_E^\lambda)^3 |\cos 2\theta| - E = 0$. The density of states is obtained numerically and their characteristics for the two branches are shown in Fig. 1. The DOS of the anisotropic spin-split levels varies asymmetrically with respect to $D_0$. For fixed electron density $n_e$ and $\alpha$, the Fermi energy ($E_f$) is obtained from the conservation of electron number $n_e = \int_{0}^{E_f} \sum_\lambda D_\lambda(E) dE$. The variations of the Fermi energy with $n_e$ and $\alpha$ are shown in Fig. 2. The Fermi energy increases with the increase of the carrier density. On the other hand, the Fermi energy decreases with the increase of the spin-orbit coupling strength. The Fermi wave vectors $k_{f}^{\lambda}(\theta) $ can be obtained numerically from the solutions of the equation $\hbar^2 k^2/2m^\ast + \lambda \alpha
k^3 |\cos 2\theta| - E_f = 0$. The Fermi contours are depicted in Fig. 5 (color: black).
![(color online) Plots of the Fermi energy vs $n_e$ and $\alpha$. Left panel: plots of the Fermi energy vs density for different values of $\alpha$. Right panel: plots of the Fermi energy vs $\alpha $ for different values of carrier density $n_e$.[]{data-label="Fig1"}](Fermi.eps){width="90mm" height="65mm"}
Drude weight and Optical conductivity
=====================================
The complex charge conductivity for a two-level system of charge carriers in presence of a sinusoidal electric field (${\bf E}(\omega) \sim \hat {\bf x} E_0 e^{i\omega t} $) can be written as $
\Sigma_{xx}(\omega) = \sigma_D(\omega) + \sigma_{xx}(\omega),
$ where $ \sigma_D(\omega)$ is the intra-band induced dynamic Drude conductivity and $\sigma_{xx}(\omega) $ is the inter-band induced complex optical conductivity.
The absorptive part of the conductivity can be obtained by taking the real part of $\Sigma_{xx}(\omega)$ and is given by $${\mathrm{Re}\,}[\Sigma_{xx}(\omega) ] = D_w\delta(\omega) + {\mathrm{Re}\,}[\sigma_{xx}(\omega)].$$ Here, $D_w$ is known as the Drude weight measuring the Drude conductivity ($\sigma_d = \tau D_w/\pi$) for a DC electric field and $ {\mathrm{Re}\,}[\sigma_{xx}(\omega)] $ is the optical conductivity as a function of the frequency of the AC electric field with vanishing momentum $q\rightarrow 0$. The vanishing momentum of the electric field forces the charge carriers to make a transition from $\lambda=-1$ branch to $\lambda = +1$ branch such that the momentum is conserved.
[**Drude weight**]{}: The semi-classical expression for the Drude weight at low temperature is given by [@mahan] $$\begin{aligned}
D_w = \pi e^2 \sum_\lambda \int \frac{d^2{\bf k}}{(2\pi)^2} \,
\langle \hat{v}_x\rangle_{\lambda}^2 \delta(E_\lambda({\bf k})-E_f).\end{aligned}$$ Here $ \hat{v}_x $ is the $x$-component of the velocity operator. Using the Heisenberg’s equation of motion, $i \hbar \dot {\bf r} = [{\bf r}, H] $, the $x$- and $y$-components of the velocity operator are given by $$\begin{aligned}
\label{vx}
\hat{v}_x & = &
\frac{\hbar k_x}{m^\ast} \mathbb{I} +
\frac{\alpha}{\hbar}\big[(3k_x^2-k_y^2)\sigma_y-2k_xk_y\sigma_x\big]\end{aligned}$$ and $$\begin{aligned}
\hat{v}_y &=&\frac{\hbar k_y}{m^\ast} \mathbb{I} - \frac{\alpha}{\hbar}
\big[(3k_y^2-k_x^2)\sigma_x-2k_xk_y\sigma_y\big].\end{aligned}$$
For the system with anisotropic cubic RSOI, the calculation of the Drude conductivity yields $$\begin{aligned}
D_{w}^{\rm ani} & = & \Big(\frac{e}{2\pi \hbar }\Big)^2
\sum_\lambda \int_{0}^{2\pi}
m^*[v_f^\lambda(\theta)]^2 B_{\lambda}(\theta) d\theta,\end{aligned}$$ where $ v_f^{\lambda}(\theta) = \hbar k_f^{\lambda}(\theta)/m^* $ and $$\begin{aligned}
B_{\lambda}(\theta) =
\frac{[\cos\theta + \lambda \eta_{\bf k}\alpha V_{f}^{\lambda}(\theta)
(5\cos\theta + \cos 3\theta)/2]^2}{1 + \lambda 3 \alpha V_f^{\lambda}(\theta)}\end{aligned}$$ with $V_f^{\lambda}(\theta) = (m^*/\hbar^2) k_f^{\lambda}(\theta) $.
![(color online) Left panels: plots of the Drude weight $D_w$ (in units of $D_{\alpha} = \pi e^2/m^* l_{\alpha}^2 $) vs $n_e$ for $\alpha = 0.006$ eV nm${}^3$ (solid: blue), $\alpha=0.008$ eV nm${}^3$ (dotted-dashed: red) and $\alpha=0.012$ eV nm${}^3$ (dashed: black). Here $l_{\alpha} = m^* \alpha/\hbar^2$. Right panel: plots of $D_w$ (in units of $D_n = \pi e^2 n_e/m^*$) vs $\alpha $ for $n_e = 3\times 10^{16}$ m$^{-2}$ (dashed: black), $n_e=3.5\times 10^{16}$ m$^{-2}$ (doted-dashed: red) and $n_e=4 \times 10^{16}$ m$^{-2}$ (solid: blue).[]{data-label="Fig2"}](Drude.eps){width="102mm" height="75mm"}
For carrying out the numerical calculation, we adopt the following parameters used in Refs. [@held; @xiao]: $n_e=3.5 \times 10^{16}$ m$^{-2}$ and $m^\ast/m_0=1$, where $m_0$ is the bare mass of the electron. In Fig. 3, the variations of the Drude weight with the carrier density and with the strength of the Rashba spin-orbit interactions are shown. The plots of the Drude weight vs carrier density for three different values of $\alpha $ are shown in the left panels of Fig. 3. The analytical expression of $D_{w}^{\rm iso}$ obtained in Ref. [@mawrie] clearly shows the deviation from the linear density dependence. Because of the small value of $\alpha$ considered here, the deviation is not visible in this figure. On the other hand, the Drude weight vs $\alpha $ for three different values of carrier density are plotted in the right panels of Fig. 3. The Drude weight decreases with the increase of $\alpha$ but the decreasing nature of $D_w$ for the two different cases is quite different. This important feature would help to know the nature of the RSOI.
[**Optical Conductivity**]{}: The generalized Kubo formula of the optical conductivity in terms of the Matsubara Green’s function is given by[@mahan] $$\begin{aligned}
\label{NMac}
& &\sigma_{\mu \nu}(\omega)= -\frac{e^2 T}{i\omega}\frac{1}{(2\pi)^2}
\int d^2 {\bf k} \nonumber\\
& \times & \sum_l {\rm Tr}\langle \hat v_\mu \hat{G}({\bf k},\omega_l) \hat v_\nu
\hat{G}({\bf k},\omega_s+\omega_l) \rangle_{i\omega_s\rightarrow \omega + i\delta }.\end{aligned}$$ Here, $\mu, \nu =x,y$, $T$ being the temperature, $\omega_s=(2s+1)\pi T$ and $\omega_l=2l\pi T$ are the fermionic and bosonic Matsubara frequencies with $s$ and $l$ are integers, respectively.
The matrix Green’s function associated with the Hamiltonian given by Eq. (1) is $$\begin{aligned}
\label{gf}
G({\bf k},\omega_n) = \frac{1}{2}\sum_\lambda
\Big[ \mathbb{I} + {\bf P}_{\lambda}({\bf k})
\cdot {{\boldsymbol}\sigma}\Big]G_0^{\lambda}({\bf k},\omega_n).\end{aligned}$$ Here $\mathbb{I}$ is a $2 \times 2$ unit matrix and $G_0^{\lambda}({\bf k},\omega_n)=1/(i\hbar\omega_n + \mu_0 - E_{\lambda}({\bf k}) ) $ with $\mu_0$ being the chemical potential. It indicates that the optical spectral weight is directly related to the local spin texture ${\bf P}_{\lambda}({\bf k})$.
Substituting Eqs. (\[vx\]) and (\[gf\]) into Eq. (\[NMac\]), the $xx$-component of the longitudinal conductivity reduces to $$\begin{aligned}
\label{op-con}
\sigma_{xx}(\omega) & = -& \frac{e^2}{i(2\pi \hbar)^2\omega}
\int_{0}^{\infty} \int_{0}^{2\pi} \alpha^2 k^5 \, \cos^22\theta\sin^2\theta dk d\theta \nonumber\\
& \times &\Big[\frac{f(E_-)-f(E_+)}{\hbar\omega + i\delta - E_+ + E_-} +
(E_-\leftrightarrow E_+)\Big],\end{aligned}$$ where $f(E)=[e^{(E-\mu_0)\beta} + 1]^{-1}$ is the Fermi-Dirac distribution function with $\beta = 1/(k_BT) $.
We have carried out the same calculation for other components of the conductivity tensor $\sigma_{\mu\nu}(\omega)$. We find that $ \sigma_{yy}(\omega) = \sigma_{xx}(\omega) $ and $\sigma_{xy}(\omega) = \sigma_{yx}(\omega) = 0 $. Hence the optical conductivity remains isotropic despite the fact that the Fermi contours are anisotropic.
Using the fact that $\omega >0$ and after performing the $k$ integral, the expression for the absorptive part of the optical conductivity at $T=0$ is given by $$\begin{aligned}
\label{cond}
{\mathrm{Re}\,}[\sigma_{xx}(\omega) ] =\frac{e^2}{24 h}
\int_0^{2\pi}d\theta \sin^2\theta \Big[ \Theta(\mu_{+}) - \Theta(\mu_-) \Big],\end{aligned}$$ where $\Theta(x) $ is the unit step function and $ \mu_{\pm} = E_{\pm}(k_{\omega}) - \mu_0 $ with $k_\omega \equiv k_w(\theta) = (\hbar\omega/2\alpha|\cos 2\theta|)^{1/3}$. This integral cannot be solved analytically due to $\theta$ dependence of $k_{\omega}$.
![(color online): Top panel: Plots of $\epsilon_{\pm}(\theta) $ vs $\theta $. Middle panel: plots of the joint density of states vs $\hbar\omega$ with $k_0=\sqrt{2\pi n_e}$. Bottom panel: the real part of the optical conductivity as a function of photon energy $\hbar\omega $.[]{data-label="Fig2"}](cond.eps){width="150mm" height="125mm"}
On the other hand, in isotropic cubic Rashba SOI the closed form expression of the absorptive part of the optical conductivity at $T=0$ K is given by[@mawrie] $$\begin{aligned}
\label{isotropic}
{\mathrm{Re}\,}[\sigma_{xx}^{\rm iso}(\omega)]=\frac{3e^2}{16 \hbar}
\big[ \Theta(\tilde \mu_+) - \Theta(\tilde \mu_-) \big],\end{aligned}$$ where $ \tilde \mu_{\pm} = E_{\pm}(\tilde k_{\omega}) - \mu_0 $ with $\tilde k_\omega=(\hbar\omega/2\alpha)^{1/3}
= k_w(\theta = (2p+1)\pi/2)$. It leads to featureless optical conductivity which has box shape with the height $ \sigma_{xx}^{\rm iso} = 3e^2/(16 \hbar)$ which is independent of the carrier density and $\alpha$. Note that simultaneous presence of isotropic Rashba and Dresselhaus SOI leads to anisotropic Fermi contours, in turns produces interesting optical features. Whereas anisotropic RSOI alone gives rise to anisotropic Fermi contours and provides distinct optical features.
![(color online) Plots of the Fermi contours $k_f^+(\theta) $, $k_f^-(\theta) $, the constant-energy difference curves $C_1$: $ E_g({\bf k}) = \epsilon_1 $ and $C_2$: $ E_g({\bf k}) = \epsilon_2 $.[]{data-label="Fig1"}](contour.eps){width="80mm" height="70mm"}
Here we shall present how the anisotropic RSOI alone gives rise to some unique features of the optical conductivity. We first evaluate $ {\mathrm{Re}\,}[\sigma_{xx}(\omega)] $ numerically using the parameters $\alpha = 0.004$ eV nm$^3$, $n_e = 3.5 \times 10^{16}$ m$^{-2}$ and $m^\ast/m_0=1$ as used in Refs. [@held; @xiao] and shown in the lower panel of Fig. 4. For comparison with the isotropic case, we plot $ {\mathrm{Re}\,}[\sigma_{xx}^{\rm iso}(\omega)] $ which appears as the rectangular box on the right side of the lower panel of Fig. 4. We depict $ \epsilon_{\pm}(\theta) = 2\alpha [k_{f}^{\pm}(\theta)]^3 |\cos(2\theta)| $ in the top panel of Fig. 4. The contribution to optical conductivity arises from the shaded angular region. The optical transitions from $\lambda = - 1$ to $\lambda = + 1 $ occur when the photon energy satisfies the inequality $ 0 < \hbar \omega < \epsilon_{-}(\theta)$. One can see that an infinitesimally small photon energy can initiate the optical transition, in complete contrast to the isotropic SOI case. This is due to the presence of the degenerate lines $ \theta = (2 p+ 1)\pi/4 $. There is a single peak of the Re $[\sigma_{xx}(\omega)]$ at $\hbar\omega = \epsilon_+(p\pi/2) = 2\alpha {[k_{f}^{+}(p\pi/2)]}^3 $ and the optical conductivity becomes zero when $\hbar\omega \geq \epsilon_{-}(p\pi/2) = 2\alpha{[k_{f}^{-}(p\pi/2)]}^3 $. For better understanding of these features, we plot the constant energy-difference curves $E_g({\bf k}) = \epsilon_{\omega} $ for $ \epsilon_{\omega} = \epsilon_{+}(p \pi/2)=\epsilon_1 $ ($C_1$: dashed) and $\epsilon_{\omega} = \epsilon_{-}(p\pi/2)=\epsilon_2$ ($C_2$: solid) in Fig. 5. The area intercept by the curves $C_i$ with $i=1,2$ and the Fermi contours ($k_{f}^{\lambda}$) are responsible for the ${\bf k}$-selective optical transitions as shown in Fig. 5. It should be noted that the anisotropic $k$-cubic band is well separated from the other two bands with $k$-linear SOI for the parameters used in Refs.[@held; @xiao]. As a result, the contribution to the optical conductivity from other two bands having $k$- linear SOI is ruled out since they occur at $\hbar\omega$ much larger than $\epsilon_-(p\pi/2)\approx 0.9$ meV.
The overall behavior of the optical spectra can be understood from the joint density of states which is given as $$\begin{aligned}
D(\omega)= \int \frac{ d^2 {\bf k}}{(2\pi)^2}
[f(E_+({\bf k}))-f(E_-({\bf k}))] \delta(E_g({\bf k}) - \hbar\omega).\end{aligned}$$
![(color online) Left panel: Plots of $\sigma_{\rm peak} $ vs $ \alpha $ for fixed values of $n_e = 4.0 \times 10^{16} $ m$^{-2}$ (dashed) and $n_e = 3.0 \times 10^{16} $ m$^{-2}$ (solid). Right panel: Plots of $\sigma_{\rm peak} $ vs $n_e$ for different values of $\alpha = 0.004$ eV nm$^3 $ (solid) and $\alpha = 0.006$ eV nm$^3 $ (dashed). []{data-label="Fig2"}](peak.eps){width="95mm" height="70mm"}
![(color online) Left panel: Plots of $\Delta $ vs $ \alpha $ for fixed values of $n_e = 3.5 \times 10^{16} $ m$^{-2}$ (dashed) and $n_e = 3.0 \times 10^{16} $ m$^{-2}$ (solid). Right panel: Plots of $\Delta $ vs $n_e$ for different values of $\alpha = 0.004$ eV nm$^3 $ (solid) and $\alpha = 0.006$ eV nm$^3 $ (dashed). []{data-label="Fig2"}](Delta.eps){width="90mm" height="70mm"}
It can be reformulated as $$\begin{aligned}
\label{jdos}
D(\omega)=\frac{1}{(2\pi)^2}\int_C \frac{dC[f(E_+(k_\omega))-f(E_-(k_\omega))]}
{|\partial_k E_g({\bf k})|_{E_g=\hbar\omega}}.\end{aligned}$$ Here $C$ is the line element along the contour. The joint density of states vs $\hbar\omega$ is plotted in the middle panel of Fig. 4. The location of the single peak and the region of zero optical conductivity are nicely described by the joint density of states. It can be seen from Eq. (\[jdos\]) that any peak may arise whenever $|\partial_k E_g({\bf k})| $ attains a minimum value. For the present problem, the singular points are at ${\bf k}_s = (k, p\pi/4)$. The single peak appears at $\epsilon_{\omega} = \epsilon_{+}(p \pi/2)$ in the joint density of states corresponds to the well known van Hove singularity. The asymmetric spin-splitting at the Fermi contours along the $ k_y = k_x = 0 $ lines is the reason for the appearance of the peak at $ \epsilon_{+}(p\pi/2) $.
There are three different types of the singularity [@ssp-book] depending on the nature of change of the energy gap around the singular points ${\bf k}_s $. Using the Taylor series expansion of $ E_g({\bf k})$ around ${\bf k}_s$ as $ E_g({\bf k}) = E_g({\bf k}_s) + \sum_{\mu} a_\mu(p) (k_\mu - k_{s\mu})^2 $ with the expansion coefficients $ 2a_\mu(p) = \frac{\partial^2 E_g({\bf k})}{\partial k_{\mu}^2}\Big|_{{\bf k}_s} $. The co-efficients $a_\mu$ are as follows $ a_x(p) = \alpha k [5 + 7(-1)^p] $ and $ a_y(p) = \alpha k [5 - 7 (-1)^p]$. The sign of the coefficients will determine the type of classification of the various singular points. One can easily find that the signs of $a_{x}$ and $a_y$ at different singular points are $(-1)^p $ and $(-1)^{p+1}$, respectively. Therefore, every singularities are all of the same class i.e. $M_1$ type.
The variations of the peak height $(\sigma_{\rm peak})$ with $n_e$ and $\alpha $ are shown in Fig. 6. It strongly depends on the Fermi energy. We also define a width $\Delta = \epsilon_-(p\pi/2) - \epsilon_+(p\pi/2)$, the difference between peak position and the position beyond which $\sigma_{xx}(\omega) $ vanishes. Its variation with $\alpha$ as well as $n_e$ are shown in Fig. 7. It shows that $\Delta $ increases with the increase of $n_e$ as well as $\alpha $.
Summary and Conclusion
======================
We have studied the Drude weight and optical conductivity for 2DEG with $k$-cubic anisotropic RSOI at the oxide interface. We have presented the variation of the zero-frequency Drude weight with the carrier density as well as the strength of the anisotropic spin-orbit coupling. For anisotropic RSOI, the Drude weight deviates from the linear density dependence. It is indicated that the spectral weight is directly related to the local spin texture in momentum space. We found that the charge and optical conductivities remain isotropic although the Fermi contours are anisotropic. It is found that an infinitesimally small photon energy can trigger inter-band optical conductivity. This is due to the fact that the spin-splitting energy vanishes along the certain directions in ${\bf k}$ space. We found a single peak in the optical conductivity whose value depends on the Fermi energy. We have shown that the van Hove singularities responsible for the single peak in the optical conductivity are of the same $M_1$ type. The different features of the conductivity can determine the information of the nature of the spin-orbit interaction experimentally and would help in understanding the orbital origin of the two-dimensional electron gas at the oxide interface.
[55]{}
R. Winkler, Spin-orbit Coupling Effects in Two-Dimensional Electron and Hole systems (Springer, Berlin 2003).
I. Zutic, J. Fabian, and S. Das Sarma, Rev. Mod. Phys. [**76**]{}, 323 (2004).
E. I. Rashba, Sov. Phys. Solid State [**2**]{}, 1109 (1960).
Y. A. Bychkov and E. I. Rashba, J. Phys. C: Solid State Phys. [**17**]{}, 6039 (1984).
G. Dresselhaus, Phys. Rev. [**100**]{}, 580 (1955).
B. Das, D. C. Miller, S. Datta, R. Reifenberger, W. P. Hong, P. K. Bhattachariya, J. Sing, and M. Jaffe, Phys. Rev. B [**39**]{}, 1411 (1989); J. Nitta, T. Akazaki, H. Takayanagi, and T. Enoki, Phys. Rev. Lett. [**78**]{}, 1335 (1997).
M. Z. Hasan and C. L. Kane, Rev. Mod. Phys. [**82**]{}, 3045 (2010); X. L. Qi and S. C. Zhang, Rev. Mod. Phys. [**83**]{}, 1057 (2011); Y. Xu, I. Miotkowski, C. Liu, J. Tian, H. Nam, N. Alidoust, J. Hu, Chih-Kang Shih, M. Z. Hasan, and Y. P. Chen, Nature Phys. [**10**]{}, 956 (2014).
E. I. Rashba and E. Ya. Sherman, Phys. Lett. A [**129**]{}, 175 (1988).
R. Winkler, Phys. Rev. B [**62**]{}, 4245 (2000); G. M. Minkov, A. A. Sherstobitov, A. V. Germanenko, O. E. Rut, V. A. Larionova, and B. N. Zvonkov, Phys. Rev. B [**71**]{}, 165312 (2005). H. Nakamura, T. Koga, and T. Kimura, Phys. Rev. Lett. [**108**]{}, 206601 (2012).
R. Moriya, K. Sawano, Y. Hoshi, S. Masubuchi, Y. Shiraki, A. Wild, C. Neumann, G. Abstreiter, D. Bougeard, T. Koga, and T. Machida, Phys. Rev. Lett. [**113**]{}, 086601 (2014).
A. Ohtomo and H. Y. Hwang, Nature. [**427**]{}, 423 (2004).
A. D. Caviglia, S. Gariglio, N. Reyren, D. Jaccard, T. Schneider, M. Gabay, S. Thiel, G. Hammerl, J. Mannhart, and J.-M. Triscone, Nature. [**456**]{}, 624 (2008).
C. Bell, S. Harashima, Y. Kozuka, M. Kim, B. G. Kim, Y. Hikita, and H. Y. Hwang, Phys. Rev. Lett [**103**]{}, 226802 (2009).
A.D. Caviglia, S. Gariglio, C. Cancellieri, B. Sace, A.Fete, N. Reyren, M. Gabay, A. F. Morpurgo, and J.M. Triscone Phys. Rev. Lett [**105**]{}, 236802 (2010).
V. Umansky, R. de-Picciotto and M. Heiblum, Appl. Phys. Lett. [**71**]{}, 683 (1997)
G. K. B. Lee and A. H. MacDonald, Phys. Rev. B [**88**]{}, 041302 (2013).
Z. S. Popovic, S. Satpathy, and R. M. Martin, Phys. Rev. Lett. [**101**]{}, 256801 (2008).
Z. Zhong, A. Toth, and K. Held, Phys. Rev. B [**87**]{}, 161102(R) (2013).
Y. Kim, R. M. Lutchyn, and C. Nayak, Phys. Rev. B [**87**]{}, 245121 (2013).
L. W. van Heeringen, G. A. de Wijs, A. McCollam, J. C. Maan, and A. Fasolino, Phys. Rev. B [**88**]{}, 205140 (2013).
C. Cancellieri, M. L. Reinle-Schmitt, M. Kobayashi, V. N. Strocov, P. R. Willmott, D. Fontaine, Ph. Ghosez, A. Filippetti, P. Delugas, and V. Fiorentini, Phys. Rev. B [**89**]{}, 121412(R) (2014); P. D. C. King, S. M. Walker, A. Tamai, A. de la Torre, T. Eknapakul, P. Buaphet, S.-K. Mo, W. Meevasana, M. S. Bahramy, and F. Baumberger, Nat. Commun. [**5**]{}, 3414 (2014).
J. Zhou, W. Shan, and D. Xiao, Phys. Rev. B [**91**]{}, 241302(R) (2015).
L. Li, C. Richter, J. Mannhart, and R. C. Ashoori, Nat. Phys. [**7**]{}, 762 (2011).
J. A. Bert, B. Kalisky, C. Bell, M. Kim, Y. Hikita, H. Y. Hwang, and K. A. Moler, Nat. Phys. [**7**]{}, 767 (2011).
W. Xu and L. B. Lin, J. Phys.: Condens. Matter [**16**]{}, 1777 (2008).
C. H. Yang, W. Xu, Z. Zeng, F. Lu, and C. Zhang, Phys. Rev. B [**74**]{}, 075321 (2006).
A. Wong and F. Mireles, Phys. Rev. B [**81**]{}, 085304 (2010).
Z. Li, F. Marsiglio, and J. P. Carbotte, Sci. Rep. [**3**]{}, 2828 (2013).
M. V. Berry, Proc. R. Soc. London, Ser. A [**392**]{}, 45 (1984).
A. Mawrie and T. K. Ghosh, J. Appl. Phys. [**119**]{}, 044303 (2016)
J. A. Simmons, S. K. Lyo, N. E. Harff, and J. F. Klem Phys. Rev. Lett. [**73**]{}, 2256 (1994).
S. Cheng, and R. Gerhardts Phys. Rev. B [**65**]{}, 085307 (2002).
M. Orlita, R. Grill, P. Hlidek, M. Zvara, G. H. Dohler, S. Malzer, and M. Byszewski Phys. Rev. B [**72**]{}, 165314 (2005).
C. Simserides, Phys. Rev. B [**75**]{}, 195344 (2007). G. D. Mahan, Many Particle Physics, (Springer Science, New York-1981).
C. Hamaguchi, Basic Semiconductor Physics, (Springer-Verlag-2010).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'In several frustrating systems incommensurate behaviors are often observed. For the $S=1$ bilinear-biquadratic model, we show that the main oscillatory behavior, which is proportional to the free edge spins, is eliminated in the incommensurate subphase, considering the average of triplet and singlet energy spectra under open boundary conditions. In the same way, the $\pi$-mode oscillation is also removed in the commensurate subphase. Moreover, we find that higher order corrections are exponentially decaying from an analysis of small size data.'
author:
- Takahiro Murashima and Kiyohide Nomura
bibliography:
- 'hfm2006.bib'
title: |
Cancellation of oscillatory behaviors\
in incommensurate region
---
Introduction
============
Commensurate-incommensurate (C-IC) transitions are interesting phenomena in frustrating quantum spin systems. In the Haldane gap systems, incommensurabilities have often been regarded as troublesome problems and have rarely been discussed in detail until recently. In inelastic neutron scattering experiments, Xu [*et al*]{}[@Xu2000] have revealed that a quasi-one-dimensional oxide, Y$_{2-x}$Ca$_x$BaNiO$_5$, has an incommensurate double-peaked structure factor. However, an analytical interpretation for incommensurabilities has not been clear.
The spin-1 bilinear-biquadratic model, $${\cal H}=\sum_{i=1}^{N} h_i,\qquad
h_i= {{\bf S}}_i\cdot{{\bf S}}_{i+1}
+\alpha({{\bf S}}_i\cdot{{\bf S}}_{i+1})^2,
\label{eq_Ham}$$ which plays a role of prototype of Y$_{2}$BaNiO$_5$, has the C-IC change point which is corresponding to the Affleck-Kennedy-Lieb-Tasaki (AKLT) point $\alpha=\alpha_{\rm D}=1/3$ [@BXG1995; @SJG1996]. The AKLT point is solvable and has an energy gap above the valence-bond-solid (VBS) ground state[@AKLT1987; @AKLT1988]. The VBS state is recently getting more attention with reference to quantum entanglements[@VMDC2004; @FKR2004; @VC2004; @FKRHB2006].
Analyzing the structure factor among the commensurate and incommensurate subphases, and also the C-IC change point, we have deduced two candidates of the ${\it real}$ structure factor unifying these two regions and the AKLT point [@FS2000; @Nomura2003]. Following the S[ø]{}rensen-Affleck prescription[@SA1994], we have constructed Green functions from the candidates, and compared them with the energy gap numerically obtained under the open boundary conditions[@MN2006]. Thus, we have found that the Green function consists of two elements which have anomalies in upper- or lower-half plane.
In this paper we will show that incommensurate oscillatory behaviors in the incommensurate subphase can be canceled, using triplet and singlet energy spectra. In the same way, we can obtain comparable results in the commensurate case. Moreover we discuss higher order corrections considering small size systems.
Short review on singlet-triplet energy gap
==========================================
The gapped Haldane phase has nonvanishing nonlocal string order[@dNR1989] and effectively free $S=1/2$ spins at the ends of open chains [@Kennedy1990; @HKAHR1990]. These edge spins bring a low-lying excitation, triplet, which is degenerate with the ground state in the thermodynamic limit[@Kennedy1990]. While this excitation is clearly a boundary effect, it is linked to the bulk behavior because of the existence of nonlocal string order[@SJG1996].
According to S[ø]{}rensen and Affleck (SA)[@SA1994], the effective Hamiltonian can be connected with the Green function: $${\cal H}_{\rm eff} = (-1)^N {{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}
\lambda^2 \int \frac{dq d\kappa}{(2\pi)^2}G(q, \kappa)
\exp(i q N) \delta(\kappa), \label{eq_Heff}$$ where ${{{\bf S}'}}_{\rm L}$ and ${{{\bf S}'}}_{\rm R}$ are spin-1/2 operators at the ends of chain. Considering the Clebsch-Gordan coefficient, we obtain singlet and triplet expectation values for edge spins as $$\begin{split}
{\langle}{\rm S}| {{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}|{\rm S}{\rangle}=-3/4, \\
{\langle}{\rm T}| {{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}|{\rm T}{\rangle}=1/4,
\label{eq_CG}
\end{split}$$ where $|{\rm T}{\rangle}=|s^{\rm T}=1, s_z=\pm 1,0{\rangle}$ and $|{\rm S}{\rangle}= |s^{\rm T}=0, s_z=0{\rangle}$[@memo1]. Thus the energy difference between the singlet and triplet states can be described with the Green function: $$\begin{split}
\Delta E_{\rm ST} (N) &\equiv E_{\rm T} - E_{\rm S}\\
&=(-1)^N
\lambda^2 \int \frac{dq}{2\pi}G(q, 0)
\exp(i q N),
\end{split}$$ where $E_{\rm T}$ and $E_{\rm S}$ are the triplet and singlet energies, respectively.
![\[fig\_gap\] Oscillation of the energy gap of edge states. The energy gaps with different chain lengths ($N=9, \cdots, 13$) are plotted as a function of $\alpha$. ](fig1.eps){width="8cm"}
Figure \[fig\_gap\] shows the energy gap behaviors of the model with different chain lengths. Increasing the chain length on some fixed $\alpha$ in the commensurate subphase, we see that the energy gap oscillates between even and odd chains and decreases exponentially fast, while this behavior is not simple in the incommensurate subphase. The even-odd oscillation is modulated by the frustration in the incommensurate subphase.
In our previous study[@MN2006], which has been performed to explain these behaviors, we have found $$\Delta E_{\rm ST}
(N)=
\begin{cases}
(-1)^{N}\widetilde{A}\exp(-\widetilde{m}N)\sin(\sqrt{d}N),
&\quad (\alpha > 1/3)\\
0, & \quad (\alpha=1/3) \\
(-1)^{N}\widetilde{A}\exp(-\widetilde{m}N)\sinh(\sqrt{d}N).
&\quad (\alpha < 1/3)\\
\end{cases}
\label{eq_gap}$$ The parameters, $\widetilde{A}$, $\widetilde{m}$, and $d$, depend on $\alpha - \alpha_{\rm D}$ and have been determined with the nonlinear least-squares fitting method as shown in Ref. [@MN2006]. We have found that the C-IC change is characterized by the following Green function: $$G(q,0) \sim \frac{1}{(q - i\widetilde{m})^2 -d}
+ \frac{1}{(q + i\widetilde{m})^2 -d}.$$
Average of triplet-singlet energies
===================================
Now we define the average of triplet-singlet energies: $${\overline{E}_{\rm TS}}\equiv (E_{\rm S}+3E_{\rm T})/4.$$ From Eqs. and , we expect that we can eliminate the principal term caused by the anomalies of the Green function using this average.
![\[fig\_Ets\] Cancellation of triplet-singlet energies. Triplet and singlet energy spectra, and an average of triplet-singlet energies are plotted as a function of $1/N$ for $1\le N \le 15$ when (a) $\alpha=$ 0.3258 (commensurate case), and (b) 0.3425 (incommensurate case). ](fig2a.eps "fig:"){width="8cm"} ![\[fig\_Ets\] Cancellation of triplet-singlet energies. Triplet and singlet energy spectra, and an average of triplet-singlet energies are plotted as a function of $1/N$ for $1\le N \le 15$ when (a) $\alpha=$ 0.3258 (commensurate case), and (b) 0.3425 (incommensurate case). ](fig2b.eps "fig:"){width="8cm"}
Figure \[fig\_Ets\] shows bare triplet and singlet energies and averages of triplet-singlet energies plotted as a function of $1/N$ for (a) $\alpha=$ 0.3258 (commensurate case), and (b) 0.3425 (incommensurate case). We see that the average of triplet-singlet energies on a fixed $\alpha$ varies linearly with $1/N$ not only in the incommensurate subphase but also in the commensurate subphase.
![\[fig\_fitEts\] Finite size effect in an average of triplet-singlet energies and a least-squares fitting line ($c_0+ c_1/N$) are plotted when (a) $\alpha=0.3258$, (b) $0.3425$. ](fig3a.eps "fig:"){width="8cm"} ![\[fig\_fitEts\] Finite size effect in an average of triplet-singlet energies and a least-squares fitting line ($c_0+ c_1/N$) are plotted when (a) $\alpha=0.3258$, (b) $0.3425$. ](fig3b.eps "fig:"){width="8cm"}
We perform the least-squares fit to the averages of triplet-singlet energies using the following fitting function: $$f(N) = c_0 + c_1 /N. \label{eq_fit}$$ Figure \[fig\_fitEts\] shows the data of triplet-singlet averages and $f(N)$. When (a) $\alpha=0.3258$, each coefficients are obtained as $c_0=-0.68167581 \pm 2 \times 10^{-9}$ and $c_1=-1.1671 \times 10^{-4} \pm 2 \times 10^{-8}$ for $8 \leq N \leq 15$. In the same way, $c_0=-0.6483465831 \pm 2 \times 10^{-10}$ and $c_1=-1.77295 \times 10^{-4} \pm 2 \times 10^{-9}$ when (b) $\alpha=0.3425$. The average of triplet-singlet energies ${\overline{E}_{\rm TS}}/N$ seems to behave high linearly in small $\alpha$ region. In fact, we observe that coefficients of $O(1/N^2)$, $O(1/N^3)$ are very small. Hence, we can say that a higher power of $1/N$ does not appear.
![\[fig\_fit\_c\] Coefficients (a) $c_0$ and (b) $c_1$ in Eq. are plotted near the AKLT point. It appears that $c_0 \propto \alpha -\alpha_{\rm D}$ and $c_1 \propto
(\alpha -\alpha_{\rm D})^2$.](fig4a.eps "fig:"){width="8cm"} ![\[fig\_fit\_c\] Coefficients (a) $c_0$ and (b) $c_1$ in Eq. are plotted near the AKLT point. It appears that $c_0 \propto \alpha -\alpha_{\rm D}$ and $c_1 \propto
(\alpha -\alpha_{\rm D})^2$.](fig4b.eps "fig:"){width="8cm"}
Figure \[fig\_fit\_c\] shows coefficients $c_0$ and $c_1$ in Eq. obtained with the least-squares fitting. The second term in Eq. comes from surface effects. Since the surface effects are caused by the one-dimensionality, they are different from the effect of edge spins. Additionally, the surface effects are perfectly zero, namely $c_1=0$, at the AKLT point[@memo2]. Moreover, $c_1$ always shows negative value except the AKLT point. We see that the fitting parameters $c_0$ and $c_1$ behave approximately as $\alpha - \alpha_{\rm D}$ and $(\alpha-\alpha_{\rm D})^2$, respectively.
So far we have excluded data smaller than $N=8$ since these data differ from Eq. . Then we proceed to study small size corrections from Eq. . For a rough estimate, we consider the difference $$\Delta y (N)= {\overline{E}_{\rm TS}}(N)/N - (c'_0 + c'_1/N) \label{eq_dy}$$ for small $N$, although we use $c'_0$ and $c'_1$, which are determined from the data of $N=14,15$. Figure \[fig\_dy\] (a) shows $\log (|\Delta y|)$ for $1 \le N \le 8$ at $\alpha=0.3258$ in the commensurate region. Since the logarithm of $\Delta y$ decreases linearly with $N$, the difference $\Delta y$ results in $\Delta y \sim \exp(-N/\xi')$. We estimate $\xi'$ as shown in Fig. \[fig\_dy\] (b), and then we find $\xi'\sim \xi/2$ (The correlation length $\xi$ is obtained in the previous study[@MN2006])[@memo3].
 The logarithm of $|\Delta y|$ (Eq. ) are plotted against $N$. The dotted line is drawn for a guide to the eye. (b) Two different correlation lengths and the ratio $\xi'/\xi$ are plotted in the commensurate region.](fig5a.eps "fig:"){width="8cm"} The logarithm of $|\Delta y|$ (Eq. ) are plotted against $N$. The dotted line is drawn for a guide to the eye. (b) Two different correlation lengths and the ratio $\xi'/\xi$ are plotted in the commensurate region.](fig5b.eps "fig:"){width="8cm"}
In the incommensurate region, we see the oscillatory behavior again. We roughly estimate the wave number of this oscillation, and then we find $q'_{\rm IC} \sim 2(q_{\rm
IC}+\pi)$[@memo4]. We will show a detailed calculation in another paper.
Summary
=======
We have shown that the principal oscillatory behaviors, which are proportional to ${{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R}$, among the triplet and singlet energy spectra under open boundary conditions cancel out in the commensurate and incommensurate subphases.
We have found that the energy spectra of singlet and triplet states under open boundary conditions consist of the bulk, surface, and edge spin energies: $$E_{\rm C}/N = B + S/N +
{\langle}{\rm C}|{\cal H}_{\rm eff}|{\rm C}{\rangle},
\qquad ({\rm C=\{S, T\}})$$ where $B$ and $S$ are the bulk and surface energies, respectively. Considering small $N$, we have found an exponentially decaying correction term, the correlation length and the incommensurate wave number of which differ by factor 2 from those obtained by the energy gap of edge states. One possibility is that the correction will be $O(({{{\bf S}'}}_{\rm L}\cdot{{{\bf S}'}}_{\rm R})^2)$. Therefore, we will need to improve the SA theory so as to contain such higher terms.
We observe similar results for the $S=1$ next-nearest-neighbor model. Thus the cancellation of triplet-singlet energies can be found in general spin gap systems.
The numerical calculation in this work is based on the program packages TITPACK version 2, developed by Professor H. Nishimori.
This research is partially supported by a Grant-in-Aid for Scientific Research (C), 18540376 (2006), from the Ministry of Education, Science, Sports and Culture of Japan.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Let $I$ denote an $R_+$-primary homogeneous ideal in a normal standard-graded Cohen-Macaulay domain over a field of positive characteristic $p$. We give a linear degree bound for the Frobenius powers $I^{[q]}$ of $I$, $q=p^{e}$, in terms of the minimal slope of the top-dimensional syzygy bundle on the projective variety ${\operatorname{Proj}}R$. This provides an inclusion bound for tight closure. In the same manner we give a linear bound for the Castelnuovo-Mumford regularity of the Frobenius powers $I^{[q]}$.'
address: 'Department of Pure Mathematics, University of Sheffield, Hicks Building, Hounsfield Road, Sheffield S3 7RH, United Kingdom'
author:
- Holger Brenner
bibliography:
- 'bibliothek.bib'
title: A linear bound for Frobenius powers and an inclusion bound for tight closure
---
Mathematical Subject Classification (2000): 13A35; 13D02; 14J60
Introduction {#introduction .unnumbered}
============
Let $R$ denote a noetherian ring, let ${\mathfrak{m}}$ denote a maximal ideal in $R$ and let $I$ denote an ${\mathfrak{m}}$-primary ideal. This means by definition that ${\mathfrak{m}}$ is the radical of $I$. Then there exists a (minimal) number $k$ such that ${\mathfrak{m}}^k \subseteq I \subseteq {\mathfrak{m}}$ holds. If $R$ contains a field of positive characteristic $p$, then the Frobenius powers of the ideal $I$, that is $$I^{[q]} = \{ f^q: f \in I\} \, , \, \, \, q = p^{e} ,$$ are also ${\mathfrak{m}}$-primary and hence there exists a minimal number $k(q)$ such that ${\mathfrak{m}}^{k(q)} \subseteq I^{[q]}$ holds. In this paper we deal with the question how $k(q)$ behaves as a function of $q$, in particular we look for linear bounds for $k(q)$ from above. If ${\mathfrak{m}}^k \subseteq I$ and if $l$ denotes the number of generators for ${\mathfrak{m}}^k$, then we get the trivial linear inclusion $({\mathfrak{m}}^k)^{lq} \subseteq ({\mathfrak{m}}^k)^{[q]} \subseteq I^{[q]}$.
The main motivation for this question comes from the theory of tight closure. Recall that the tight closure of an ideal $I$ in a domain $R$ containing a field of positive characteristic $p$ is the ideal $$I^* = \{f \in R: \exists 0 \neq c \in R \mbox{ such that }
cf^q \in I^{[q]} \mbox{ for all } q=p^{e} \} \, .$$ A linear inclusion relation $ {\mathfrak{m}}^{\lambda q + \gamma} \subseteq I^{[q]}$ for all $q=p^{e}$ implies the inclusion ${\mathfrak{m}}^\lambda \subseteq
I^*$, since then we can take any element $0 \neq c \in {\mathfrak{m}}^\gamma$ to show for $f \in {\mathfrak{m}}^\lambda$ that $cf^q \in {\mathfrak{m}}^{\lambda q
+\gamma} \subseteq I^{[q]}$, hence $f \in I^*$. The trivial bound mentioned above yields ${\mathfrak{m}}^{kl} \subseteq I^*$, but in fact we have already $ {\mathfrak{m}}^{kl} \subseteq {\mathfrak{m}}^k\subseteq I$, so this does not yield anything interesting.
We restrict in this paper to the case of a normal standard-graded domain $R$ over an algebraically closed field $K=R_0$ of positive characteristic $p$ and a homogeneous $R_+$-primary ideal $I$. The question is then to find the minimal degree $k(q)$ such that $R_{\geq k(q)} \subseteq I^{[q]}$ or at least a good linear bound $k(q) \leq \lambda q+\gamma$. In this setting we work mainly over the normal projective variety $Y= {\operatorname{Proj}}R$, endowed with the very ample invertible sheaf ${\mathcal{O}}_Y(1)$. If $I=(f_1 { , \ldots , }f_n)$ is given by homogeneous ideal generators $f_i$ of degree $d_i = \deg (f_i)$, then we get on $Y$ the following short exact sequences of locally free sheaves, $$0 {\longrightarrow}{\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)(m) {\longrightarrow}\bigoplus_{i=1}^n {\mathcal{O}}_Y(m -qd_i)
\stackrel{f_1^q { , \ldots , }f_n^q}{{\longrightarrow}} {\mathcal{O}}_Y(m) {\longrightarrow}0 \, .$$ Another homogeneous element $h \in R$ of degree $m$ yields a cohomology class $\delta(h) \in H^1(Y,{\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)(m))$, and therefore the question whether $h \in (f_1^q { , \ldots , }f_n^q)=I^{[q]}$ is equivalent to the question whether $\delta(h)=0$. Since ${\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)(0) =F^{*e} ({\operatorname{Syz}}(f_1 { , \ldots , }f_n)(0))$ is the pull-back under the $e$-th absolute Frobenius morphism $F^{e}:Y {\rightarrow}Y$, our question is an instance of the following more general question: given a locally free sheaf ${\mathcal{S}}$ on a normal projective variety $(Y,{\mathcal{O}}_Y(1))$, find an (affine-linear) bound $\ell (q)$ such that for $m \geq \ell (q)$ we have $H^1(Y, {\mathcal{S}}^q (m))=0$, where we set ${\mathcal{S}}^{q}=F^{e*}({\mathcal{S}})$. Using a resolution ${\mathcal{G}}_\bullet {\rightarrow}{\mathcal{S}}{\rightarrow}0$ where ${\mathcal{G}}_j=
\bigoplus_{(k,j)} {\mathcal{O}}_Y( -\alpha_{k,j})$, we can shift the problem (at least if $Y= {\operatorname{Proj}}R$ with $R$ Cohen-Macaulay, so that $H^{i}(Y,{\mathcal{O}}_Y(m))=0$ for $0< i < \dim (Y)$) to the problem of finding a bound such that $H^t(Y, {\mathcal{S}}_t^q(m)) =0$, where ${\mathcal{S}}_t =
{\operatorname{kern}}({\mathcal{G}}_t {\rightarrow}{\mathcal{G}}_{t-1})$ and $t = \dim (Y)$. By Serre duality this translates to ${\operatorname{Hom}}( {\mathcal{S}}_t^q(m), \omega_Y)=0$. Now the existence of such mappings is controlled by the minimal slope of ${\mathcal{S}}_t^q(m)$. Let $\bar{\mu}_{\min} ({\mathcal{S}}_t)= \lim \inf_{q=p^{e}}
\mu_{\min}({\mathcal{S}}_t^q)/q$ and set $\nu =- \bar{\mu}_{\min} ({\mathcal{S}}_t)/
\deg (Y)$. With these notations applied to ${\mathcal{S}}={\operatorname{Syz}}(f_1 { , \ldots , }f_n)(0)$ our main results are the following theorems (Theorems \[theoreminclusion\] and \[tightinclusion\]).
\[theoreminclusionintro\] Let $R$ denote a standard-graded normal Cohen-Macaulay domain over an algebraically closed field $K$ of characteristic $p >0$. Suppose that the dualizing sheaf $\omega_Y$ of $Y= {\operatorname{Proj}}R$ is invertible. Let $I$ denote a homogeneous $R_+$-primary ideal. Then $R_{ > q \nu + \frac{\deg (\omega_Y)}{\deg(Y)} } \subseteq I^{[q]}$.
From this linear bound for the Frobenius powers we get the following inclusion bound for tight closure.
\[tightinclusionintro\] Under the assumptions of Theorem \[theoreminclusionintro\] we have the inclusion $R_{\geq \nu}
\subseteq I^*$, where $I^*$ denotes the tight closure of $I$.
This theorem generalizes [@brennerslope Theorem 6.4] from dimension two to higher dimensions. We also obtain an inclusion bound for the Frobenius closure (Corollary \[frobenius\]) and a linear bound for the Castelnuovo-Mumford regularity of the Frobenius powers $I^{[q]}$ (Theorem \[regularitybound\]), which improves a recent result of M. Chardin [@chardinregularitypowers].
I thank M. Blickle for useful remarks.
Some projective preliminaries
=============================
Let $K$ denote an algebraically closed field and let $Y$ denote a normal projective variety over $K$ of dimension $t$ together with a fixed ample Cartier divisor $H$ with corresponding ample invertible sheaf ${\mathcal{O}}_Y(1)$. The degree of a coherent torsion-free sheaf ${\mathcal{S}}$ (with respect to $H$) is defined by the intersection number $\deg({\mathcal{S}})= \deg (c_1({\mathcal{S}}) )= c_1({\mathcal{S}}) . H^{t-1}$, see [@maruyamagrauertmuelich Preliminaries] for background of this notion. The degree is additive on short exact sequences [@maruyamagrauertmuelich Lemma 1.5(2)].
The slope of ${\mathcal{S}}$ (with respect to $H$), written $\mu({\mathcal{S}})$, is defined by dividing the degree through the rank. The slope fulfills the property that $\mu ({\mathcal{S}}_1 \otimes {\mathcal{S}}_2)= \mu({\mathcal{S}}_1) +
\mu({\mathcal{S}}_2)$ [@maruyamagrauertmuelich Lemma 1.5(4)]. The minimal slope of ${\mathcal{S}}$, $\mu_{\min} ({\mathcal{S}})$, is given by $$\mu_{\min} ({\mathcal{S}}) = \inf \{ \mu({\mathcal{Q}}):\, {\mathcal{S}}{\rightarrow}{\mathcal{Q}}{\rightarrow}0
\mbox{ is a torsion-free quotient sheaf} \} \, .$$ If ${\mathcal{S}}_1
{ \subset \ldots \subset }{\mathcal{S}}_k= {\mathcal{S}}$ is the Harder-Narasimhan filtration of ${\mathcal{S}}$ [@maruyamagrauertmuelich Proposition 1.13], then $\mu_{\min} ({\mathcal{S}}) = \mu({\mathcal{S}}/{\mathcal{S}}_{k-1})$. If ${\mathcal{L}}$ is an invertible sheaf and $\mu_{\min}({\mathcal{S}}) > \deg({\mathcal{L}})$, then there does not exist any non-trivial sheaf homomorphism ${\mathcal{S}}{\rightarrow}{\mathcal{L}}$. The sheaf ${\mathcal{S}}$ is called semistable if $ \mu({\mathcal{S}})
=\mu_{\min}({\mathcal{S}})$.
Suppose now that the characteristic of $K$ is positive and let $F^{e}\!:Y {\rightarrow}Y$ denote the $e$-th absolute Frobenius morphism. We denote the pull-back of ${\mathcal{S}}$ under this morphism by ${\mathcal{S}}^q
=F^{e*}({\mathcal{S}})$, $q=p^{e}$. The slope behaves like $\mu({\mathcal{S}}^q) = q
\mu ({\mathcal{S}})$ (this follows from [@maruyamagrauertmuelich Lemma 1.6], for which it is enough to assume that the finite mapping is flat in codimension one; note that we compute the slope always with respect to ${\mathcal{O}}_Y(1)$, not with respect to $F^{*e}({\mathcal{O}}_Y(1))={\mathcal{O}}_Y(q)$). It may however happen that $\mu_{\min}
({\mathcal{S}}^q) < q \mu_{\min} ({\mathcal{S}})$. Therefore it is useful to consider the number (compare [@langersemistable]) $$\bar{\mu}_{\min} ({\mathcal{S}})=\liminf_{q=p^{e}} \mu_{\min} ({\mathcal{S}}^q)/q \, .$$ This limit exists, since there exists for some number $k$ a surjection $\oplus_j {\mathcal{O}}( \beta_j) {\rightarrow}{\mathcal{S}}(k) $ such that all $
\beta_j$ are positive. Then ${\mathcal{S}}(k)$ is a quotient of an ample bundle and so all its quotients have positive degree. This holds also for all its Frobenius pull-backs, hence $\mu_{\min}
(({\mathcal{S}}(k))^q) \geq 0$ and the limit is $\geq 0$. Thus $\mu_{\min}
({\mathcal{S}}^q) \geq -q k \deg ({\mathcal{O}}_Y(1))$ for all $q$. Moreover, a theorem of Langer implies that this limit is even a rational number, see [@langersemistable]. The sheaf ${\mathcal{S}}$ is called strongly semistable if $\mu({\mathcal{S}})=\bar{\mu}_{\min} ({\mathcal{S}})$; equivalently, if all Frobenius pull-backs ${\mathcal{S}}^q$ are semistable.
The degree of the variety $Y$ (with respect to $H$) is by definition the top self intersection number $\deg(Y) = \deg({\mathcal{O}}_Y(1))=H^t$. In the following we will impose on a polarized variety $(Y, {\mathcal{O}}_Y(1))$ of dimension $t$ the condition that $H^i(Y,{\mathcal{O}}_Y(m))=0 $ for $i=1
{ , \ldots , }t-1$ and all $m$. If $Y= {\operatorname{Proj}}R$, where $R$ is a standard-graded Cohen Macaulay ring, this property holds true due to [@brunsherzog Theorem 3.5.7].
\[sheafproposition\] Let $Y$ denote a normal projective variety of dimension $t \geq 1$ over an algebraically closed field $K$ of positive characteristic $p$. Let ${\mathcal{O}}_Y(1)$ denote a very ample invertible sheaf on $Y$ such that $H^i(Y,{\mathcal{O}}(m))=0 $ for $i=1
{ , \ldots , }t-1$. Suppose that the dualizing sheaf $\omega_Y$ on $Y$ is invertible. Let ${\mathcal{S}}$ denote a torsion-free coherent sheaf on $Y$. Suppose that the stalk ${\mathcal{S}}_y$ is free for every non-smooth point $y \in Y$. Let $$\cdots {\longrightarrow}{\mathcal{G}}_3 {\longrightarrow}{\mathcal{G}}_2 {\longrightarrow}{\mathcal{S}}{\longrightarrow}0$$ denote an exact complex of sheaves, where ${\mathcal{G}}_j$ has type ${\mathcal{G}}_j=
\bigoplus_{(k,j) } {\mathcal{O}}_Y(- \alpha_{k,j})$. Set ${\mathcal{S}}_j={\operatorname{im}}({\mathcal{G}}_{j+1}
{\rightarrow}{\mathcal{G}}_{j})= \ker( {\mathcal{G}}_{j} {\rightarrow}{\mathcal{G}}_{j-1})$, $j \geq 2$, and ${\mathcal{S}}_1={\mathcal{S}}$. Fix $i=1 { , \ldots , }t$. Then for $$m > - q \frac{ \bar{\mu}_{\min}( {\mathcal{S}}_{t-i+1}) }{\deg(Y)} + \frac{\deg(\omega_Y)}{\deg(Y)}$$ we have $H^i (Y, {\mathcal{S}}^q (m)) =0$.
Note first that the Frobenius acts flat on the exact complex and on the corresponding short exact sequences $0 {\rightarrow}{\mathcal{S}}_{j+1} {\rightarrow}{\mathcal{G}}_{j+1} {\rightarrow}{\mathcal{S}}_j {\rightarrow}0$. This can be checked locally and is true for the smooth points of $Y$. Over a singular point $y \in Y$ the sheaf ${\mathcal{S}}$ is free, so these short exact sequences split locally in a neighborhood of such a point and hence all the ${\mathcal{S}}_j$ are also free in $y$. So also in these points the Frobenius preserves the exactness of the complex.
Due to our assumption on ${\mathcal{O}}_Y(1)$ we have $H^i(Y,{\mathcal{G}}_j(m))=0 $ for $i=1 { , \ldots , }t-1$ and all $m$ and all $j \geq 2$. Hence from the short exact sequences $0 {\rightarrow}{\mathcal{S}}_{j+1}(m) {\rightarrow}{\mathcal{G}}_{j+1}(m) {\rightarrow}{\mathcal{S}}_j(m) {\rightarrow}0$ we can infer that $$\begin{aligned}
& & H^i (Y, {\mathcal{S}}_j(m)) \cong H^{i+1}(Y, {\mathcal{S}}_{j+1}(m)) \, \, \,
\mbox{ isomorphisms for } i=1 { , \ldots , }{t}-2, \cr & & H^{{t}-1} (Y, {\mathcal{S}}_j(m)) \subseteq H^{ {t}}(Y, {\mathcal{S}}_{j+1}(m)) \, \, \,
\mbox{ injection for } t \geq 2 , \cr & & H^{{t}} (Y, {\mathcal{G}}_{j+1}
(m)) {\rightarrow}H^{{t}} (Y, {\mathcal{S}}_j(m)) \, \, \, \mbox{ surjection}.\end{aligned}$$ The same is true if we replace $S_j$ and $G_j$ by their Frobenius pull-backs $S_j^q$ and $G_ j^q$. For $i=1 { , \ldots , }{t}$ we find $$H^{i}(Y, {\mathcal{S}}^q_1 (m)) \!\cong \! H^{i+1}(Y, {\mathcal{S}}^q_2 (m)) \! { \cong \ldots \cong }\!
H^{{t}-1} (Y, {\mathcal{S}}^q_{ {t}-i} (m)) \! \subseteq \! H^{{t}} (Y, {\mathcal{S}}^q _{{t}-i+1}(m)) .$$ So we only have to look at $H^{t}(Y, {\mathcal{S}}^q_{t-i+1}(m))$, which is by Serre duality dual to ${\operatorname{Hom}}({\mathcal{S}}^q_{t-i+1}(m), \omega_Y)$, see [@hartshornealgebraic Theorem III.7.6]. Suppose now that $m$ fulfills the numerical condition. Then $$\begin{aligned}
\mu_{\min}({\mathcal{S}}^q_{t-i+1}(m)) &=& \mu_{\min}({\mathcal{S}}^q_{t-i+1})+ m \deg(Y) \cr
&\geq & q \bar{\mu}_{\min} ({\mathcal{S}}_{t-i+1}) + m \deg (Y) \cr
&>& q \bar{\mu}_{\min}( {\mathcal{S}}_{t-i+1})
+ \big(- q \frac{ \bar{\mu}_{\min}( {\mathcal{S}}_{t-i+1}) }{\deg(Y)}
+ \frac{\deg(\omega_Y)}{\deg(Y)} \big) \deg(Y) \cr
&=& \deg(\omega_Y) \, .\end{aligned}$$ So for these $m$ there are no non-trivial mappings from ${\mathcal{S}}_{t-i+1}^q(m)$ to $\omega_Y$ and therefore $H^{t}(Y, {\mathcal{S}}_{t-i+1}^q(m))=0$.
The dualizing sheaf $\omega_Y$ on the projective variety $Y
\subseteq {\mathbb{P}}^N$ is invertible under the condition that $Y$ is locally a complete intersection in ${\mathbb{P}}^N$ and in particular if $Y$ is smooth (see [@hartshornealgebraic Theorem III.7.11 and Corollary III.7.12]. If $\omega_Y$ is not invertible, but torsion-free, then we may replace $\deg(\omega_Y)$ by $\mu_{\max}(\omega_Y)$ to get the same statement as in Proposition \[sheafproposition\].
An inclusion bound for tight closure {#inclusion}
====================================
We first fix the following situation, with which we will deal in this section.
\[situation\] Let $K$ denote an algebraically closed field of characteristic $p >0$. Let $R$ denote a standard-graded normal Cohen-Macaulay domain of dimension $t+1 \geq 2$ over $K$ with corresponding projective normal variety $Y={\operatorname{Proj}}R$. Suppose that the dualizing sheaf $\omega_Y$ of $Y$ is invertible. Let $I
\subseteq R$ denote a homogeneous $R_+$-primary ideal. Let $$\cdots {\longrightarrow}F_2= \bigoplus_{(k,2)} R(- \alpha_{k,2})
{\longrightarrow}F_1= \bigoplus_{(k,1)} R(- \alpha_{k,1}) {\longrightarrow}I {\longrightarrow}0 \, ,$$ denote a homogeneous complex of graded $R$-modules which is exact on $D(R_+)$. Let $$\cdots {\longrightarrow}{\mathcal{G}}_2=\bigoplus_{(k,2)} {\mathcal{O}}(- \alpha_{(k,2)}) {\longrightarrow}{\mathcal{G}}_1= \bigoplus_{(k,1)} {\mathcal{O}}(- \alpha_{k,1})
{\longrightarrow}{\mathcal{O}}_Y {\longrightarrow}0 \,$$ denote the corresponding exact complex of sheaves on $Y$. Denote by ${\operatorname{Syz}}_j = {\operatorname{kern}}({\mathcal{G}}_j {\rightarrow}{\mathcal{G}}_{j-1})$ the locally free kernel sheaves on $Y$, and set ${\operatorname{Syz}}_j(m)= {\operatorname{Syz}}_j \otimes {\mathcal{O}}_Y(m)$. Let $\nu =- \bar{\mu}_{\min} ({\operatorname{Syz}}_t) /\deg (Y)$, where $t$ is the dimension of $Y$.
\[theoreminclusion\] Suppose the situation and notation described in \[situation\]. Then for all prime powers $q=p^{e}$ we have the inclusion $R_{> q \nu + \frac{\deg(\omega_Y) }{\deg(Y)}}
\subseteq I^{[q]}$.
Since $I$ is primary all the syzygy sheaves occurring in the resolution on $Y$ are locally free and hence we may apply Proposition \[sheafproposition\]. Fix a prime power $q=p^{e}$. Let $h \in R$ denote a homogeneous element of degree $m > q \nu +
\frac{\deg(\omega_Y) }{\deg(Y)}$. This gives via the short exact sequence on $Y$, $$0 {\longrightarrow}{\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)(m) {\longrightarrow}\bigoplus_{i=1}^n {\mathcal{O}}_Y(m -qd_i)
\stackrel{f_1^q { , \ldots , }f_n^q}{{\longrightarrow}} {\mathcal{O}}_Y(m) {\longrightarrow}0$$ rise to a cohomology class $\delta(h) \in H^1(Y,{\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)(m))$, where $${\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)(m) =(F^{e*}({\operatorname{Syz}}(f_1
{ , \ldots , }f_n))) (m)= {\mathcal{S}}^q (m) \, ,$$ ${\mathcal{S}}= {\mathcal{S}}_1= {\operatorname{Syz}}(f_1
{ , \ldots , }f_n)$. It is enough to show that $\delta (h)=0$, for then $h \in I^{[q]} \Gamma(D(R_+), {\mathcal{O}}) =I^{[q]}$, since $R$ is normal. But this follows from Proposition \[sheafproposition\] applied to ${\mathcal{S}}= {\operatorname{Syz}}(f_1 { , \ldots , }f_n)$ and $i=1$.
\[resolutionremark\] We do not insist that the “resolution” of the ideal is exact on the whole ${\operatorname{Spec}}R$ nor that it is minimal, but it is likely that a minimal resolution will give us in general a better bound $\nu$. For example we can always use the Koszul complex given by ideal generators of the $R_+$-primary ideal $I$.
The next theorem gives an inclusion bound for tight closure. Recall that the tight closure of an ideal $I \subseteq R$ in a noetherian domain containing a field of positive characteristic $p$ is by definition the ideal $$I^* = \{f \in R: \exists 0 \neq c \in R \mbox{ such that } cf^q \in I^{[q]}
\mbox{ for all } q=p^{e} \} \, .$$ See [@hunekeapplication] for basic properties of this closure operation.
\[tightinclusion\] Suppose the situation described in \[situation\]. Then we have the inclusion $R_{\geq \nu} \subseteq
I^*$.
Let $f \in R$ be a homogeneous element of degree $\deg (f)=m
\geq \nu =- \bar{\mu}_{\min}({\operatorname{Syz}}_t) /\deg(Y)$. Due to the definition of tight closure we have to show that $cf^q \in I^{[q]}$ holds for some $c \neq 0$ and all prime powers $q$. Let $c \neq 0$ be any homogeneous element of degree $ > \deg (\omega_Y)/\deg (Y)$. Then $\deg (cf^q) = qm +\deg(c) > q \nu + \deg(\omega_Y)/\deg (Y)$ and therefore $cf^q \in I^{[q]}$ by Theorem \[theoreminclusion\].
Suppose that $R$ fulfills the condition of the situation described in \[situation\] and let $I=(f_1 { , \ldots , }f_n)$ denote an ideal generated by a full regular system of homogeneous parameters of degree $\deg(f_i)=d_i$ (so $n=t+1$). Then the Koszul resolution of these elements gives a resolution on $Y={\operatorname{Proj}}R$ such that the top-dimensional syzygy bundle is invertible, namely $${\operatorname{Syz}}_t (m)={\mathcal{G}}_{t+1}(m)= {\mathcal{O}}_Y(m-d_1 - \cdots -d_{t+1} ) \,$$ Then Theorem \[tightinclusion\] gives the known (even without the condition Cohen-Macaulay) inclusion bound $R_{\geq d_1 { + \ldots + }d_n} \subseteq (f_1 { , \ldots , }f_n)^*$, see [@hunekeparameter Theorem 2.9].
The next easiest case is then the $R_+$-primary homogeneous ideal $I$ has finite projective dimension (it is again enough to impose the exactness only on $D(R_+)$). In this case the resolution on $Y$ looks like $$0 {\longrightarrow}{\mathcal{G}}_{t+1} {\longrightarrow}{\mathcal{G}}_t { {\longrightarrow}\ldots {\longrightarrow}}{\mathcal{G}}_1 {\longrightarrow}{\mathcal{O}}_Y{\longrightarrow}0$$ and the top-dimensional syzygy bundle is ${\operatorname{Syz}}_t = {\mathcal{G}}_{t+1}
=\bigoplus_k {\mathcal{O}}_Y( - \alpha_{k,t+1})$, and therefore $$\mu_{\min}({\operatorname{Syz}}_t)= \deg(Y) \min_k \{ - \alpha_{k,t+1}\} = -
\deg(Y) \max_k \{ \alpha_{k,t+1} \} \, .$$ The corresponding inclusion bound was proved in [@hunekesmithkodaira Theorem 5.11]. Such a situation arises for example if $I$ is generated by a set of monomials in a system of homogeneous parameters.
The following easy corollary unifies two known inclusion bounds for tight closure given by K. Smith (see [@smithgraded Propositions 3.1 and 3.3]), namely that $R_{\geq \sum_{i=1}^n \deg (f_i)}
\subseteq I^*$ and that $R_{\geq \dim(R) \max _i \{ \deg(f_i)\}}
\subseteq I^*$.
\[tightcorollary\] Suppose the situation described in \[situation\] and suppose that the homogeneous $R_+$-primary ideal $I=(f_1 { , \ldots , }f_n)$ is generated by homogeneous elements of degree $d_i= \deg(f_i)$. Set $d= \max _{1 \leq i_1 < \ldots <
i_{\dim (R)} \leq n}( d_{i_1} { + \ldots + }d_{i_{\dim(R)}} )$. Then $R
_{\geq d} \subseteq I^*$.
We consider the Koszul resolution of $I=(f_1 { , \ldots , }f_n)$, which is exact outside the origin. This gives the surjection $$\bigoplus _{1 \leq i_1 < \ldots < i_{\dim (R)} \leq n}
{\mathcal{O}}(-d_{i_1} { - \ldots - }d_{i_{\dim(R)}} ) {\longrightarrow}{\operatorname{Syz}}_{\dim(R)-1 } {\longrightarrow}0$$ which shows that $$\begin{aligned}
\bar{\mu}_{\min} ({\operatorname{Syz}}_{\dim(R)-1}) &\geq& \bar{\mu}_{\min}
\big(\bigoplus _{1 \leq i_1 < \ldots < i_{\dim (R)} \leq n}
{\mathcal{O}}(-d_{i_1} { - \ldots - }d_{i_{\dim(R)}}) \big) \cr &=& - \max _{1
\leq i_1 < \ldots < i_{\dim (R)} \leq n} \{ d_{i_1} { + \ldots + }d_{i_{\dim(R)}} \} \deg (Y) \, .\end{aligned}$$ Hence $\nu =- \bar{\mu}_{\min} ( {\operatorname{Syz}}_{\dim(R)-1})/ \deg (Y) \leq
\max \{d_{i_1} { + \ldots + }d_{i_{\dim(R)}} \}$ and Theorem \[tightinclusion\] applies.
If the dimension of $R$ is two, then Theorem \[tightinclusion\] was proved in [@brennerslope Theorem 6.4] using somewhat more geometric methods. In this case $Y= {\operatorname{Proj}}R$ is a smooth projective curve and the top syzygy bundle is just the first syzygy bundle, and the result also holds in characteristic zero for solid closure. See [@brennerslope] and [@brennercomputationtight] for concrete computations of the number $\nu$ in this case. It is in general difficult to compute the number $\nu$ of the theorem, as it is difficult to compute the minimal slope of a locally free sheaf.
The following corollary gives an inclusion bound for tight closure under the condition that the top-dimensional syzygy bundle is strongly semistable. In the two-dimensional situation this bound is exact, in the sense that below this bound an element belongs to the tight closure only if it belongs to the ideal itself, see [@brennerslope Theorem 8.4].
\[topstable\] Suppose the situation described in \[situation\] and let $I=(f_1 { , \ldots , }f_n)$ be generated by homogeneous elements of degree $d_i=\deg(f_i)$. Let $F_\bullet {\rightarrow}I$ denote the Koszul complex and suppose that the top-dimensional syzygy bundle ${\operatorname{Syz}}_t$ is strongly semistable. Set $d= (\dim (R)-1) (d_1 { + \ldots + }d_n)/
(n-1)$. Then $R_{\geq d} \subseteq I^*$.
The condition strongly semistable means that $\mu({\operatorname{Syz}}_t)= \bar{\mu}_{\min}({\operatorname{Syz}}_t)$. So we only have to compute the degree and the rank of ${\operatorname{Syz}}_t$. It is easy to compute that $\det({\operatorname{Syz}}_t)= {\mathcal{O}}_Y(\binom{n-2}{t-1}(- \sum_{i=1}^n d_i))$, hence $$\deg( {\operatorname{Syz}}_t) = \binom{n-2}{t-1}(- \sum_{i=1}^n d_i) \deg (Y)$$ and ${\operatorname{rk}}({\operatorname{Syz}}_t)= \binom{n-1}{t}$. Therefore $$\mu( {\operatorname{Syz}}_t)= \binom{n-2}{t-1} (- \sum_{i=1}^n d_i) \deg (Y) / \binom{n-1}{t}
= \frac{t}{n-1} (- \sum_{i=1}^n d_i) \deg (Y)$$ and $\nu = \frac{t}{n-1} ( \sum_{i=1}^n d_i)$.
As the proofs of Theorem \[tightinclusion\] and Proposition \[sheafproposition\] show, Corollary \[topstable\] is also true under the weaker condition that there does not exist any non-trivial mapping ${\operatorname{Syz}}_t^q {\rightarrow}{\mathcal{L}}$ to any invertible sheaf ${\mathcal{L}}$ contradicting the semistability of ${\operatorname{Syz}}_t^q$ for all $q=p^{e}$.
Theorem \[tightinclusion\] applies in particular when $R$ is a normal complete intersection domain. Let $R=K[X_1 { , \ldots , }X_N]/(H_1
{ , \ldots , }H_r)$, where $H_j$ are homogeneous forms of degree $\delta_j$. Then $\omega_Y = {\mathcal{O}}(\sum_j \delta_j -N)$. Therefore the number $ \deg(\omega_Y)/ \deg(Y)= \sum_j \delta_j -N$ is just the $a$-invariant of $R$.
We want to apply Corollary \[topstable\] to the computation of the tight closure $(x^a,y^a,z^a,w^a)^*$ in $R=K[x,y,z,w]/(H)$, where $H$ is supposed to be a polynomial of degree $4$ defining a smooth projective (hyper-)surface $$Y= V_+(H)={\operatorname{Proj}}R \subset {\mathbb{P}}^3 ={\operatorname{Proj}}K[x,y,z,w]$$ of degree $4$; hence $Y$ is a $K 3$ surface. Our result will only hold true for generic choice of $H$. We look at the Koszul complex on ${\mathbb{P}}^3$ defined by $x^a,y^a,z^a,w^a$ and break it up to get $$0 {\longrightarrow}{\operatorname{Syz}}_2 \cong \bigwedge^2 {\operatorname{Syz}}{\longrightarrow}\bigoplus_6 {\mathcal{O}}_{{\mathbb{P}}^3}(-2a)
{\longrightarrow}\bigoplus_4 {\mathcal{O}}_{{\mathbb{P}}^3}(-a) {\longrightarrow}{\mathcal{O}}_{{\mathbb{P}}^3} {\longrightarrow}0 \, .$$ Suppose first that $K$ is an algebraically closed field of characteristic $0$. It is easy to see that the syzygy bundle ${\operatorname{Syz}}={\operatorname{Syz}}(x^a,y^a,z^a,w^a)$ is semistable on ${\mathbb{P}}^3$ [@brennerlookingstable Corollary 3.6 or Corollary 6.4]. Therefore also the exterior power ${\operatorname{Syz}}_2 \cong \bigwedge^2 {\operatorname{Syz}}$ is semistable on ${\mathbb{P}}^3$. By the restriction theorem of Flenner [@flennerrestriction Theorem 1.2] it follows that the restriction ${\operatorname{Syz}}_2 \!|_Y$ is also semistable on the generic hypersurface $Y=V_+(H)$.
On the other hand, due to the Theorem of Noether (see [@haramp §IV.4]), every curve on the generic surface of degree $4$ in ${\mathbb{P}}^3$ is a complete intersection and $R=K[x,y,z,w]/(H)$ is a factorial domain for generic $H$ of degree $4$. It follows that the cotangent bundle $\Omega_Y$ on $Y=V_+(H)$ is semistable. For the semistability of a rank two bundle we only have to look at mappings ${\mathcal{L}}{\rightarrow}\Omega_Y$, where ${\mathcal{L}}$ is invertible. But since ${\mathcal{L}}=
{\mathcal{O}}_Y(k)$, the semistability follows, since $Y$ is a $K3$ surface and so $\Omega_Y$ has degree $0$ but does not have any global non-trivial section (see [@griffithsharris IV. 5]).
So for $H$ generic the relevant second syzygy bundle ${\operatorname{Syz}}_2 \!|_Y$ and the cotangent bundle $\Omega_Y$ are both semistable in characteristic $0$. Since the ${\mathbb{Q}}$-rational points are dense in ${\mathbb{A}}^N_K$, there exist also such polynomials $H$ with rational coefficients and then also with integer coefficients. We consider such a polynomial $H$ with integer coefficients as defining a family of quartics over ${\operatorname{Spec}}{\mathbb{Z}}$. Since semistability is an open property, we infer that the second syzygy bundle and the cotangent bundle are also semistable on $Y_p=V_+(H_p)$ for $p \gg 0$.
By the semistability of $\Omega_{Y_p}$ ($p \gg 0$), the maximal slope of $\Omega_{Y_p}$ is $\leq 0$. A theorem of Langer [@langersemistable Corollary 2.4 and Corollary 6.3] shows then that every semistable bundle on $Y_p$ is already strongly semistable. Hence the second syzygy bundle is also strongly semistable. Therefore we are in the situation of Corollary \[topstable\] and we compute $d=8a/3$. Thus $$R_{8a/3}\subseteq (x^a,y^a,z^a,w^a)^*$$ holds in $R=K[x,y,z,w]/(H)$ for $H$ generic of degree $4$ and for $p \gg 0$. The first non-trivial instance is for $a=3$. In fact for the (non-generic) Fermat quartic $x^4+y^4+z^4+w^4=0$ it was proved by Singh in [@singhcomputation Theorem 4.1] directly that $x^2y^2z^2w^2 \in
(x^3,y^3,z^3,w^3)^*$.
For the next corollary we recall the definition of the Frobenius closure. Suppose that $R$ is a noetherian ring containing a field of positive characteristic $p >0$, and let $I$ denote an ideal. Then the Frobenius closure of $I$ is defined by $$I^F =\{ f \in R:\, \exists q=p^{e} \mbox{ such that } f^q \in I^{[q]} \} \, .$$ It is easy to see that the Frobenius closure of an ideal is contained in its tight closure.
\[frobenius\] Suppose the situation described in \[situation\]. Then $R_{> \nu} \subseteq I^F$, the Frobenius closure of $I$.
Let $f$ denote a homogeneous element of degree $m=\deg(f) >
\nu=- \bar{\mu}_{\min} ({\operatorname{Syz}}_t)/\deg(Y)$. Then we just have to take a prime power $q=p^{e}$ such that $\deg(f^q)=qm >q \nu+
\deg(\omega_Y)/ \deg(Y)$ holds. Then $f^q \in I^{[q]}$ holds due to Theorem \[theoreminclusion\].
Corollary \[frobenius\] is not true for $R_{\geq \nu}$ instead of $R_{> \nu }$. This is already clear for parameter ideals in dimension two, say for $(x,y)$ in $R=K[x,y,z]/(H)$, where $H$ defines a smooth projective curve $Y= {\operatorname{Proj}}R =V_+(H) \subset
{\mathbb{P}}^2$. Here we have the resolution $$0 {\longrightarrow}{\mathcal{O}}_Y(-2) \cong {\operatorname{Syz}}(x,y)(0) {\longrightarrow}{\mathcal{O}}_Y(-1) \oplus {\mathcal{O}}_Y(-1) \stackrel{x,y}{{\longrightarrow}}
{\mathcal{O}}_Y {\longrightarrow}0 \, .$$ Hence we get $\nu =2$, but an element of degree two (say $z^2$) does not in general belong to the Frobenius closure of $(x,y)$.
A problem of Katzman and Sharp (see [@katzmansharpfrobenius]) asks in its strongest form: does there exist a number $b$ such that whenever $f \in I^F$ holds, then already $f^{p^b} \in I^{[p^b]}$ holds. A positive answer (together with the knowledge of a bound for the number $b$) to this question would give a finite test to check whether a given element $f$ belongs to the Frobenius closure $I^F$ or not. For those elements which belong to $I^{F}$ because of Corollary \[frobenius\] (due to degree reasons, so to say), the answer is yes, at least in the sense that for $f$ fulfilling $\deg(f) \geq \nu + \epsilon$ ($\epsilon
>0$) we have $ \deg (f^q) = q \deg(f) \geq q \nu + q \epsilon$, so the condition $q \epsilon > \deg(\omega_Y)/\deg(Y)$ is sufficient to ensure that $f^q \in I^{[q]}$. It is however possible that elements of degree $\deg(f) \leq \nu$ belong to the Frobenius closure.
The Castelnuovo-Mumford regularity of Frobenius powers {#cmregularity}
======================================================
We recall briefly the notion of the Castelnuovo-Mumford regularity following [@brodmannsharp Definition 15.2.9]. Let $R$ denote a standard-graded ring and let $M$ denote a finitely generated graded $R$-module. Then the Castelnuovo-Mumford regularity of $M$ (or regularity of $M$ for short) is $${\operatorname{reg}}(M) = \sup \{ {\operatorname{end}}(H^i_{R_+} (M)) +i :\, 0 \leq i \leq \dim M \} \, ,$$ where ${\operatorname{end}}(N)$ of a graded $R$-module $N$ denotes the maximal degree $e$ such that $N_e \neq 0$. For a number $l$ we define the regularity ${\operatorname{reg}}^l (M)$ at and above level $l$ by $${\operatorname{reg}}^l(M)= \sup \{ {\operatorname{end}}(H^i_{R_+} (M))+i :\, l \leq i \leq \dim M \} \, ,$$
A question of M. Katzman raised in [@katzmanfrobenius Introduction] asks how the regularity of the Frobenius powers $I^{[q]}$ behaves, in particular whether there exists a linear bound ${\operatorname{reg}}(I^{[q]}) \leq C_1 q + C_0$. Such a linear bound for the regularity of the Frobenius powers of an ideal was recently given by M. Chardin in [@chardinregularitypowers Theorem 2.3]. The following theorem gives a better linear bound for the regularity of Frobenius powers of $I$ in terms of the slope of the syzygy bundles.
\[regularitybound\] Let $K$ denote an algebraically closed field of positive characteristic $p$. Let $R$ denote a standard-graded normal Cohen-Macaulay $K$-domain of dimension $t+1 \geq 2$. Let $I=(f_1 { , \ldots , }f_n) \subseteq R$ denote a homogeneous ideal generated by homogeneous elements of degree $d_i =\deg(f_i)$. Suppose that the dualizing sheaf $\omega_Y$ on $Y={\operatorname{Proj}}R$ is invertible. Suppose that the points $y \in \sup ({\mathcal{O}}_Y/{\mathcal{I}})$ are smooth points of $Y$. Let $F_\bullet {\rightarrow}I$ denote a graded free resolution with corresponding exact complex of sheaves on $Y$, $
{\mathcal{G}}_\bullet {\rightarrow}{\mathcal{I}}\subseteq {\mathcal{O}}_Y $. Set ${\operatorname{Syz}}_j = \ker({\mathcal{G}}_j
{\rightarrow}{\mathcal{G}}_{j-1})$. Then we have for the Castelnuovo-Mumford regularity of the Frobenius powers $I^{[q]}$ the linear bound ${\operatorname{reg}}(I^{[q]}) \leq C_1q + C_0 $, where $$C_1= \max \{ d_i, i=1 { , \ldots , }n, \,
- \frac{ \bar{\mu}_{\min}({\operatorname{Syz}}_j)}{\deg(Y)},\, j=1 { , \ldots , }t =\dim(Y) \} \mbox{ and }$$ $$C_0 = \max \{ {\operatorname{reg}}(R), \frac{\deg(\omega_Y)}{\deg(Y) } \}
\, .$$
The ideal generators define for $q=p^{e}$ the homogeneous short exact sequences $$0 {\longrightarrow}{\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q) {\longrightarrow}\bigoplus_{i=1}^n R(-qd_i)
\stackrel{f_1^q { , \ldots , }f_n^q}{ {\longrightarrow}} I^{[q]} {\longrightarrow}0$$ of graded $R$-modules. It is an easy exercise [@brodmannsharp Exc. 15.2.15] to show that for a short exact sequence $0
{\rightarrow}L {\rightarrow}M {\rightarrow}N {\rightarrow}0$ we have ${\operatorname{reg}}(N) \leq \max \{ {\operatorname{reg}}^1(L) -1,
{\operatorname{reg}}(M) \}$. We have ${\operatorname{reg}}(R(-qd)) = {\operatorname{reg}}(R) +qd$ and $${\operatorname{reg}}( \bigoplus_{i=1}^n R(-qd_i)) = \max_i \{{\operatorname{reg}}(R(-qd_i)) \}
={\operatorname{reg}}(R) + q \max_i \{ d_i \} \, ,$$ which gives the first terms in the definition of $C_1$ and $C_0$ respectively. Hence it is enough to give a linear bound for ${\operatorname{reg}}^1({\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q))$. Moreover, the long exact local cohomology sequence associated to the above short exact sequence gives $${\longrightarrow}H^0_{R_+}(I^{[q]}) {\longrightarrow}H^1_{R_+} ({\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)) {\longrightarrow}\bigoplus_{i=1}^n
H^1_{R_+} (R(-qd_i)) {\longrightarrow}\, .$$ The term on the right is $0$, since $R$ is Cohen-Macaulay, and the term on the left is $0$, since $R$ is a domain. Therefore $H^1_{R_+} ({\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q))
=0$ and we have to find a linear bound for $ {\operatorname{reg}}^2({\operatorname{Syz}}(f_1^q
{ , \ldots , }f_n^q)) ={\operatorname{reg}}^1({\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q))$. We have $H^i_{R_+}({\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q) )= H^{i-1} (D(R_+), {\operatorname{Syz}}(f_1^q
{ , \ldots , }f_n^q) \widetilde{\, } \, )$ for $i \geq 2$ due to the long exact sequence relating local cohomology with sheaf cohomology. Denote now by ${\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q) $ the corresponding torsion-free sheaf on $Y={\operatorname{Proj}}R$. On $Y$ we have the short exact sequences of sheaves $$0 {\longrightarrow}{\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)
{\longrightarrow}\bigoplus_{i=1}^n {\mathcal{O}}_Y(-qd_i) {\longrightarrow}{\mathcal{I}}^{[q]} {\longrightarrow}0 \, .$$ We may compute the cohomology as $$H^{i}(D_+(R), {\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q) \widetilde{\,} \, )_m
= H^i(Y, {\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q) (m)) \, .$$ Note that the syzygy bundle ${\operatorname{Syz}}(f_1 { , \ldots , }f_n)$ is free by assumption in the singular points of $Y$. Hence we are in the situation of Proposition \[sheafproposition\] with ${\mathcal{S}}={\operatorname{Syz}}(f_1 { , \ldots , }f_n)$; therefore $H^{i} (Y, {\operatorname{Syz}}(f_1^q { , \ldots , }f_n^q)(m))=0$ ($i=1 { , \ldots , }t$) holds for $m> \max_{j=1 { , \ldots , }t} \{- q \frac{ \bar{\mu}_{\min}
({\operatorname{Syz}}_j)}{ \deg(Y)} \} + \frac{\deg(\omega_Y)}{\deg(Y)}$, which proves the theorem.
The Castelnuovo-Mumford regularity of a standard-graded Cohen-Macaulay domain $R$ is just ${\operatorname{reg}}(R)= {\operatorname{end}}(H^{\dim ( R)}_{R_+} (R)) + \dim (R)$. The end of the top-dimensional local cohomology module of a graded ring is also called its $a$-invariant, see [@brodmannsharp 13.4.7], hence ${\operatorname{reg}}(R)=a + \dim (R)$. If $R$ is Gorenstein, then $R(a)$ is the canonical module of $R$ and $\omega_Y={\mathcal{O}}_Y(a)$ is the dualizing sheaf on $Y={\operatorname{Proj}}R$. So in this case the quotient $\deg (\omega_Y) / \deg (Y)= a \deg (Y)/ \deg (Y) =a$ equals also the $a$-invariant.
The surjection $\bigoplus_{(k,j+1)} {\mathcal{O}}_Y(- \alpha_{k,j+1}) {\rightarrow}{\operatorname{Syz}}_j {\rightarrow}0$ gives at once the bound $\bar{\mu}_{\min} ({\operatorname{Syz}}_j)
\geq \bar{\mu}_{\min}(\bigoplus_{(k,j+1)} {\mathcal{O}}_Y(- \alpha_{k,j+1}))
\!= \! - \max \{\alpha_{k,j+1} \} \deg(Y)$. Therefore we get for the constant $C_1$ coming from Theorem \[regularitybound\] the estimate $C_1 \leq \max \{ \alpha_{k,j}:\, j=1 { , \ldots , }t+1 = \dim
(R) \, \} =C_1'$. This number $C_1'$ is the coefficient for the linear bound which M. Chardin has obtained in [@chardinregularitypowers Theorem 2.3]. This bound corresponds to the inclusion bounds for tight closure of K. Smith which we obtained in Corollary \[tightcorollary\]. The following standard example of tight closure theory shows already the difference between the Chardin-Smith bound and the slope bound.
Consider the ideal $I=(x^2,y^2,z^2)$ in $R=K[z,y,z]/(x^3+y^3+z^3)$, ${\rm char} (K) \neq 3$. We compute the bound coming from Theorem \[regularitybound\] for the regularity of the Frobenius powers $I^{[q]}=(x^{2q}, y^{2q},z^{2q})$. We first observe that we may consider the curve equation $0=x^3+y^3+z^3=xx^2+yy^2+zz^2$ as a global section of the syzygy bundle of degree $3$. Since this section has no zero on $Y= {\operatorname{Proj}}R$, we get the short exact sequence $$0 {\longrightarrow}{\mathcal{O}}_Y {\longrightarrow}{\operatorname{Syz}}(x^2,y^2,z^2)(3) {\longrightarrow}{\mathcal{O}}_Y {\longrightarrow}0 \, .$$ This shows that the syzygy bundle is strongly semistable and therefore $\bar{\mu}_{\min} ( {\operatorname{Syz}}(x^2,y^2,z^2)(0))= -6 \deg(Y)/2=
-9 $. So $C_1=3$ and we get altogether the bound ${\operatorname{reg}}( I^{[q]})
\leq 3q+2$.
Since $\, {\operatorname{Syz}}(x^2,y^2,z^2)(3)\, $ is not generated by its global sections, because the section just mentioned is the only section. Hence a surjection $ \bigoplus_k {\mathcal{O}}(-\alpha_k) {\rightarrow}{\operatorname{Syz}}(x^2,y^2,z^2)(0)$ is only possible for $ \max _k \{ \alpha_k \}
\geq 4$. So the linear bound for the regularity which you get by considering only the degrees in a resolution is worse than the slope bound.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'The optimal binning is the optimal discretization of a variable into bins given a discrete or continuous numeric target. We present a rigorous and extensible mathematical programming formulation to solving the optimal binning problem for a binary, continuous and multi-class target type, incorporating constraints not previously addressed. For all three target types, we introduce a convex mixed-integer programming formulation. Several algorithmic enhancements such as automatic determination of the most suitable monotonic trend via a Machine-Learning-based classifier and implementation aspects are thoughtfully discussed. The new mathematical programming formulations are carefully implemented in the open-source python library OptBinning.'
author:
- 'Guillermo Navas-Palencia'
bibliography:
- 'optbinning\_bib.bib'
title: 'Optimal binning: mathematical programming formulation'
---
Introduction
============
Binning (grouping or bucketing) is a technique to discretize the values of a continuous variable into bins (groups or buckets). From a modeling perspective, the binning technique may address prevalent data issues such as the handling of missing values, the presence of outliers and statistical noise, and data scaling. Furthermore, the binning process is a valuable interpretable tool to enhance the understanding of the nonlinear dependence between a variable and a given target while reducing the model complexity. Ultimately, resulting bins can be used to perform data transformations.
Binning techniques are extensively used in machine learning applications, exploratory data analysis and as an algorithm to speed up learning tasks; recently, binning has been applied to accelerate learning in gradient boosting decision tree [@Ke2017]. In particular, binning is widely used in credit risk modeling, being an essential tool for credit scorecard modeling to maximize differentiation between high-risk and low-risk observations, and in expected credit loss modeling.
There are several unsupervised and supervised binning techniques. Common unsupervised techniques are equal-width and equal-size or equal-frequency interval binning. On the other hand, well-known supervised techniques based on merging are Monotone Adjacent Pooling Algorithm (MAPA), also known as Maximum Likelihood Monotone Coarse Classifier (MLMCC) [@Thomas2002] and ChiMerge [@Kerber1992], whereas other techniques based on decision trees are CART [@Breiman1984], Minimum Description Length Principle (MDLP) [@Fayyad1993] and more recently, condition inference trees (CTREE) [@Hothorn2006].
The binning process might require to satisfy certain constraints. These constraints might range from requiring a minimum number of records per bin to monotonicity constraints. This variant of the binning process is known as the optimal binning process. The optimal binning is generally solved by iteratively merging an initial granular discretization until imposed constraints are satisfied. Performing this fine-tuning manually is likely to be unsatisfactory as the number of constraints increases, leading to suboptimal or even infeasible solutions. However, we note that this manual adjustment has been encouraged by some authors [@Siddiqi2005], legitimating the existing interplay of “art and science” in the binning process.
There are various commercial software tools to solving the optimal binning problem[^1]. Software IBM SPSS and the MATLAB Financial Toolbox, use MDLP and MAPA as a default algorithm, respectively. The most advanced tool to solve the optimal binning problem is available in the SAS Enterprise Miner software. A limited description of the proprietary algorithm can be found in [@Oliveira2008], where two mixed-integer programming (MIP) formulations are sketched: a mixed-integer linear programming (MILP) formulation to obtain a fast probably suboptimal solution, and a mixed-integer nonlinear programming (MINLP) formulation to obtain an optimal solution. The suboptimal formulation is the default method due to computational time limitations (MILP techniques are considerably more mature). We note that the SAS implementation allows most of the constraints required in credit risk modelling, becoming an industry standard. Besides, there exist a few open-source solutions, but the existing gap comparing to the commercial options in terms of capabilities is still significant. Among the available alternatives, we mention the MATLAB implementation of the monotone optimal binning in [@Mironchyk2017], and the R specialized packages `smbinning` [@smbinning], relying on CTREE, and `MOB` [@mob], which merely include basic functionalities.
In this paper, we develop a rigorous and extensible mathematical programming formulation for solving the optimal binning problem. This general formulation can efficiently handle binary, continuous, and multi-class target type. The presented formulations incorporate the constraints generally required to produce a good binning [@Siddiqi2005], and new constraints not previously addressed. For all three target types, we introduce a convex mixed-integer programming formulation, ranging from a integer linear programming (ILP) formulation for the simplest cases to a mixed-integer quadratic programming (MIQP) formulation for those cases adding more involved constraints.
The remainder of the paper is organized as follows. Section \[section\_mip\_formulation\] introduces our general problem formulation and the corresponding mixed-integer programming formulation for each supported target. We focus on the formulation for binary target, investigating various formulation variants. Then, in Section \[section\_algorithms\] we discuss in detail several algorithmic aspects such as the automatic determination of the optimal monotonic trend and the development of presolving algorithms to efficiently solve large size instances. Section \[section\_experiments\] includes experiments with real-world datasets and compares the performance of supported solvers for large size instances. Finally, in Section \[section\_conclusions\], we present our conclusions and discuss possible research directions.
Mathematical programming formulation {#section_mip_formulation}
====================================
The optimal binning process comprises two steps: A pre-binning process that generates an initial granular discretization, and a subsequent refinement or optimization to satisfy imposed constraints. The pre-binning process uses, for example, a decision tree algorithm to calculate the initial split points. The resulting $m$ split points are sorted in ascending order, $s_1 \le s_2 \le \ldots \le s_m$ to create $n = m + 1$ pre-bins. These pre-bins are defined by the intervals $(-\infty, s_1), [s_1, s_2), \ldots, [s_m, \infty)$.
Given $n$ pre-bins, the decision variables consist of a binary lower triangular matrix (indicator variables) of size $n$, $X_{ij} \in \{0, 1\}, \forall (i, j) \in \{1, \ldots, n: i \ge j\}$. The starting point is a diagonal matrix, meaning that initially all pre-bins are selected. A basic feasible solution must satisfy the following constraints:
- All pre-bins are either isolated or merged to create a larger bin interval, but cannot be erased. Each column must contain exactly one $1$. $$\label{def_ct_sum_columns_one}
\sum_{i=1}^n X_{ij} = 1, \quad j = 1, \ldots, n.$$
- Only consecutive pre-bins can be merged. Continuity by rows, no $0-1$ gaps are allowed. $$\label{def_ct_flow_continuity}
X_{ij} \le X_{ij+1}, \quad i = 1, \ldots, n; \; j = 1, \ldots, i-1.$$
- A solution has a last bin interval of the form $[s_k, \infty)$, for $k \le n$. The binary decision variable $X_{nn} = 1$.
To clarify, Figure \[basic\_diagram\] shows an example of a feasible solution. In this example, pre-bins corresponding to split points $(s_2, s_3, s_4)$ and $(s_5, s_6)$ are merged, thus having an optimal binning with bin intervals $(-\infty, s_1), [s_1, s_4), [s_4, s_6), [s_6, \infty)$.
![Lower triangular matrix $X$. Initial solution after pre-binning (left). Optimal solution with 4 bins after merging pre-bins (right).[]{data-label="basic_diagram"}](images/basic_matrix_diagram.pdf)
The described problem can be seen as a generalized assignment problem. A direct formulation of metrics involving ratios such as the mean or most of the divergence measures on merged bins leads to a non-convex MINLP formulation, due to the ratio of sums of binary variables. Solving non-convex MINLP problems to optimality is a challenging task requiring the use of global MINLP solvers, especially for large size instances.
Investigating the binary lower triangular matrix in Figure \[basic\_diagram\], it can be observed, by analyzing the constraints in Equations (\[def\_ct\_sum\_columns\_one\]) and (\[def\_ct\_flow\_continuity\]) imposing continuity by rows, that a feasible solution is entirely characterized by the position of the first $1$ for each row. This observation permits the pre-computation of the set of possible solutions by rows, obtaining an aggregated matrix with the shape of $X$ for each involved metric. Consequently, the non-convex objective function and constraints are linearized, resulting in a convex formulation by exploiting problem information. Using this reformulation, we shall see that the definition of constraints for binary, continuous and multi-class target are almost analogous.
Mixed-integer programming formulation for binary target
-------------------------------------------------------
Given a binary target $y$ used to discretize a variable $x$ into $n$ bins, we define the normalized count of non-events (NE) $p_i$ and events (E) $q_i$ for each bin $i$ as $$p_i = \frac{r_i^{NE}}{r_T^{NE}}, \quad q_i = \frac{r_i^E}{r_T^E},$$ where $r_T^{NE}$ and $r_T^E$ are the total number of non-event records and event records, respectively. Next we define the Weight of Evidence (WoE) and event rate for each bin, $$\text{WoE}_i = \log\left(\frac{r_i^{NE}/ r_T^{NE}}{r_i^E / r_T^E}\right), \quad D_i = \frac{r_i^E}{r_i^E + r_i^{NE}}.$$ The Weight of Evidence $\text{WoE}_i$ and event rate $D_i$ for each bin are related by means of the functional equations $$\begin{aligned}
\text{WoE}_i &= \log\left(\frac{1 - D_i}{D_i}\right) + \log\left(\frac{r_T^E}{ r_T^{NE}}\right) = \log\left(\frac{r_T^E}{ r_T^{NE}}\right) - \text{logit}(D_i)\\
D_i &= \left(1 + \frac{r_T^{NE}}{r_T^{E}} e^{\text{WoE}_i}\right)^{-1} = \left(1 + e^{\text{WoE}_i - \log\left(\frac{r_T^{E}}{r_T^{NE}}\right)}\right)^{-1},\end{aligned}$$ where $D_i$ can be characterized as a logistic function of $\text{WoE}_i$, and $\text{WoE}_i$ can be expressed in terms of the logit function of $D_i$. This shows that WoE is inversely related to the event rate. The constant term $\log(r_T^{E} / r_T^{NE})$ is the log ratio of the total number of event and the total number of non-events.
Divergence measures serve to assess the discriminant power of a binning solution. The Jeffreys’ divergence [@Jeffreys1946], also known as Information Value (IV) within the credit risk industry, is a symmetric measure expressible in terms of the Kullback-Leibler divergence $D_{KL}(P || Q)$ [@kullback1951] defined by $$J(P|| Q) = IV = D_{KL}(P || Q) + D_{KL}(Q || P) = \sum_{i=1}^n (p_i - q_i) \log \left(\frac{p_i}{q_i}\right).$$ The IV statistic is unbounded, but some authors have proposed rules of thumb to settings quality thresholds [@Siddiqi2005]. Alternatively, the Jensen-Shannon divergence is a bounded symmetric measure also expressible in terms of the Kullback-Leibler divergence $$JSD(P || Q) = \frac{1}{2}\left(D(P || M) + D(Q || M)\right), \quad M = \frac{1}{2}(P + Q),$$ and bounded by $JSD(P||Q) \in [0, \log(2)]$. Note that these divergence measures cannot be computed when $r_i^{NE} = 0$ and/or $r_i^E = 0$. Other divergences measures without this limitation are described in [@Zeng2013].
A good binning algorithm for binary target should be characterized by the following properties [@Siddiqi2005]:
1. Missing values are binned separately.
2. Each bin should contain at least $5\%$ observations.
3. No bins should have $0$ accounts for non-events or events.
Property 1 is adequately addressed in many implementations, where missing and special values are incorporated as additional bins after the optimal binning terminates. Property 2 is a usual constraint to enforce representativeness. Property 3 is required to compute the above divergence measures.
Let us define the parameters of the mathematical programming formulation: $$\begin{aligned}
& n \in \mathbb{N} && \textrm{number of pre-bins.}\\
& r^{NE}_T \in \mathbb{N} && \textrm{total number of non-event records.}\\
& r^E_T \in \mathbb{N} && \textrm{total number of event records.}\\
& r^{NE}_i \in \mathbb{N} && \textrm{number of non-event records per pre-bin.}\\
& r^E_i \in \mathbb{N} && \textrm{number of event records per pre-bin.}\\
& r_i = r^{NE}_i + r^E_i && \textrm{number of records per pre-bin.}\\
& r^{NE}_{\min} \in \mathbb{N} && \textrm{minimum number of non-event records per bins.}\\
& r^{NE}_{\max} \in \mathbb{N} && \textrm{maximum number of non-event records per bins.}\\
& r^E_{\min} \in \mathbb{N} && \textrm{minimum number of event records per bins.}\\
& r^E_{\max} \in \mathbb{N} && \textrm{maximum number of event records per bins.}\\
& b_{\min} \in \mathbb{N} && \textrm{minimum number of bins.}\\
& b_{\max} \in \mathbb{N} && \textrm{maximum number of bins.}\end{aligned}$$
The objective function is to maximize the discriminant power among bins, therefore, maximize a divergence measure. The IV can be computed using the described parameters and the decision variables $X_{ij}$, yielding $$IV = \sum_{i=1}^n \left(\sum_{j=1}^i \left(\frac{r_z^{NE}}{r^{NE}_T} - \frac{r_z^E}{r^E_T}\right)X_{ij}\right) \log\left(\frac{\sum_{j=1}^i r_j^{NE} / r_T^{NE} X_{ij}}{\sum_{j=1}^i r_j^E / r_T^E X_{ij}}\right),$$ The IV is the sum of the IV contributions per bin, i.e., the sum by rows. As previously stated, given the constraints in Equations (\[def\_ct\_sum\_columns\_one\]) and (\[def\_ct\_flow\_continuity\]), an aggregated low triangular matrix $V_{ij} \in \mathbb{R}^+_0, \forall (i, j) \in \{1, \ldots, n: i \ge j\}$ with all possible IV values from bin merges can be pre-computed as follows $$\label{iv_obj_params}
V_{ij} = \left(\sum_{z=j}^i \frac{r_z^{NE}}{r^{NE}_T} - \frac{r_z^E}{r^E_T}\right) \log \left(\frac{\sum_{z=j}^i r_z^{NE} / r_T^{NE}}{\sum_{z=j}^i r_z^E / r_T^E}\right), \quad i=1, \ldots, n;\; j=1, \ldots, i.$$ The optimal IV for each bin is determined by using the remarked observation that a solution is characterized by the position of the first $1$ for each row, thus, using the continuity constraint in (\[def\_ct\_flow\_continuity\]), we obtain $$V_{i\cdot} = V_{i1} X_{i1} + \sum_{j=2}^i V_{ij} (X_{ij} - X_{ij-1}) \Longleftrightarrow V_{ii} X_{ii} + \sum_{j=1}^{i-1} (V_{ij} - V_{ij+1})X_{ij}.
\label{iv_obj_definition}$$ for $i=1, \ldots, n$. The latter formulation is preferred to reduce the fill-in of the matrix of constraints. Similarly, a lower triangular matrix of event rates $D_{ij} \in [0, 1], \forall (i, j) \in \{1, \ldots, n: i \ge j\}$ can be pre-computed as follows $$D_{ij} = \frac{\sum_{z=j}^i r_z^E}{\sum_{z=j}^i r_z}, \quad i=1, \ldots, n;\; j=1, \ldots, i.$$
The ILP formulation, with no additional constraints such as monotonicity constraints, can be stated as follows
$$\begin{aligned}
\underset{X}{\text{max}} \quad & \sum_{i=1}^n V_{ii} X_{ii} + \sum_{j=1}^{i-1} (V_{ij} - V_{ij+1}) X_{ij} \label{objective}\\
\text{s.t.} \quad & \sum_{i=j}^{n} X_{ij} = 1, & j = 1,\ldots, n\label{sum_columns_one}\\
\quad & X_{ij} - X_{ij+1} \le 0 , & i = 1, \ldots, n; \; j = 1, \ldots, i-1\label{flow_continuity}\\
\quad & b_{\min} \le \sum_{i=1}^n X_{ii} \le b_{\max} \label{min_max_bins}\\
\quad & r_{\min} X_{ii} \le \sum_{j=1}^i r_j X_{ij} \le r_{\max} X_{ii}, & i = 1, \ldots, n \label{min_max_bin_size}\\
\quad & r^{NE}_{\min} X_{ii} \le \sum_{j=1}^i r^{NE}_j X_{ij} \le r^{NE}_{\max} X_{ii}, & i = 1, \ldots, n \label{min_max_bin_n_nonevent}\\
\quad & r^E_{\min} X_{ii} \le \sum_{j=1}^i r^E_j X_{ij} \le r^E_{\max} X_{ii}, & i = 1, \ldots, n \label{min_max_bin_n_event}\\
\quad & X_{ij} \in \{0, 1\}, & \forall (i, j) \in \{1, \ldots, n: i \ge j\}\label{binary_indicator}\end{aligned}$$
Apart from the constraints (\[sum\_columns\_one\]) and (\[flow\_continuity\]) already described, constraint (\[min\_max\_bins\]) imposes a lower and upper bound on the number of bins. Other range constraints (\[min\_max\_bin\_size\]-\[min\_max\_bin\_n\_event\]) limit the number of total, non-event and event records per bin. Note that to increase sparsity, the range constraints are not implemented following the standard formulation to avoid having the data twice in the model. For example, constraint (\[min\_max\_bins\]) is replaced by $$d + \sum_{i=1}^n X_{ii} - b_{\max} =0, \quad 0 \le d \le b_{\max} - b_{\min}.$$
### Monotonicity constraints {#section_monotonicity}
Monotonicity constraints between the event rates of consecutive bins can be imposed to ensure legal compliance and business constraints. Three types of monotonic trends are considered: the usual ascending/descending, and two types of unimodal forms, concave/convex and peak/valley. This modeling flexibility can help to capture overlooked or unexpected patterns, providing new insights to enrich models. Note that work in [@Oliveira2008] uses the WoE approach instead.
Applying Equation (\[iv\_obj\_definition\]), the optimal event rate for each bin is given by $$D_{i\cdot} = D_{ii} X_{ii} + \sum_{j=1}^{i-1} (D_{ij} - D_{ij+1})X_{ij}, \quad i=1, \ldots, n.\label{event_rate_definition}$$
#### Monotonic trend: ascending and descending.
The formulation for monotonic ascending trend can be stated as follows, $$\begin{aligned}
&D_{zz} X_{zz} + \sum_{j=1}^{z-1}(D_{zj} - D_{z j+1}) X_{zj} + \beta(X_{ii} + X_{zz} - 1)\nonumber\\ &\le 1 + (D_{ii} - 1) X_{ii} + \sum_{j=1}^{i - 1} (D_{ij} - D_{ij+1})X_{ij}, \quad i=2, \ldots, n; \; z=1,\ldots i-1.\end{aligned}$$ The term $1 + (D_{ii} - 1) X_{ii}$ or simply $1 - X_{ii}$, is used to ensure that event rates are in $[0, 1]$, and the ascending constraint is satisfied even if bin $i$ is not selected. Note that this is a big-$M$ formulation $M + (D_{ii} - M) X_{ii}$ using $M=1$, which suffices given $D \in [0, 1]$, however, a tighter (non integer) $M = \max(\{D_{ij}: i=1, \ldots, n;\; i \ge j\})$, can be used instead. The parameter $\beta$ is the minimum event rate difference between consecutive bins. The term $\beta(X_{ii} + X_{zz} - 1)$ is required to ensure that the difference between two selected bins $i$ and $z$ is greater or equal than $\beta$. Similarly, for the descending constraint, $$\begin{aligned}
&D_{ii} X_{ii} + \sum_{j=1}^{i-1}(D_{ij} - D_{i j+1})X_{ij} + \beta(X_{ii} + X_{zz} - 1)\nonumber\\ &\le 1 + (D_{zz} - 1) X_{zz} + \sum_{j=1}^{z - 1} (D_{zj} - D_{zj+1})X_{zj}, \quad i=2, \ldots, n; \; z=1,\ldots i-1.\end{aligned}$$
#### Monotonic trend: concave and convex.
The concave and convex trend can be achieved by taking the definition of concavity/convexity on equally spaced points: $$\begin{aligned}
-x_{i+1} + 2 x_i - x_{i-1} &\ge 0 && \textrm{concave}\\
x_{i+1} - 2 x_i + x_{i-1} &\ge 0 && \textrm{convex}\end{aligned}$$ Thus, replacing Equation (\[event\_rate\_definition\]) in the previous definition of concavity we obtain the concave trend constraints, $$\begin{aligned}
&-\left(D_{ii}X_{ii} + \sum_{z=1}^{i-1}(D_{iz} - D_{iz+1}) X_{iz}\right) + 2 \left(D_{jj}X_{jj} + \sum_{z=1}^{j-1}(D_{jz} - D_{jz+1}) X_{jz}\right)\nonumber\\
&-\left(D_{kk}X_{kk} + \sum_{z=1}^{k-1}(D_{kz} - D_{kz+1}) X_{kz}\right) \ge X_{ii} + X_{jj} + X_{kk} - 3,\end{aligned}$$ for $i=3, \ldots n$; $j=2, \ldots, i-1$ and $k=1, \ldots, j-1$. Similarly, for convex trend we get $$\begin{aligned}
&\left(D_{ii}X_{ii} + \sum_{z=1}^{i-1}(D_{iz} - D_{iz+1}) X_{iz}\right) - 2 \left(D_{jj}X_{jj} + \sum_{z=1}^{j-1}(D_{jz} - D_{jz+1}) X_{jz}\right)\nonumber\\
&\left(D_{kk}X_{kk} + \sum_{z=1}^{k-1}(D_{kz} - D_{kz+1}) X_{kz}\right) \ge X_{ii} + X_{jj} + X_{kk} - 3,\end{aligned}$$ for $i=3, \ldots n$; $j=2, \ldots, i-1$ and $k=1, \ldots, j-1$. Note that term $X_{ii} + X_{jj} + X_{kk} - 3$ is used the preserve redundancy of constraints when not all bins $i$, $j$ and $k$ are selected, given that $D \in [0, 1]$.
#### Monotonic trend: peak and valley.
The peak and valley trend[^2] define an event rate function exhibiting a single trend change or reversal. The optimal trend change position is determined by using disjoint constraints, which can be linearized using auxiliary binary variables. The resulting additional constraints are as follows,
$$\begin{aligned}
i - n (1 - y_i) &\le t \le i + n y_i, & i = 1, \ldots, n\label{disjoint_constraint}\\
t &\in [0, n]\\
y_i &\in \{0, 1\}, & i = 1, \ldots, n\label{disjoint_aux_variables}\end{aligned}$$
where $t$ is the position of the optimal trend change bin, $y_i$ are auxiliary binary variables and $n$ in (\[disjoint\_constraint\]) is the smallest big-$M$ value for this formulation while preserving the redundancy of constraints. Furthermore, for the peak trend we incorporate the following constraints,
$$\begin{aligned}
&y_i + y_z + 1 + (D_{zz} - 1) X_{zz} + \sum_{j=1}^{z-1} (D_{zj} - D_{zj+1}) X_{zj}\nonumber\\
&\ge D_{ii} X_{ii} + \sum_{j=1}^{i-1}(D_{ij} - D_{ij+1}) X_{ij}, \quad i=2, \ldots, n; \; z=1, \ldots, i - 1,\\
& 2 - y_i - y_z + 1 + (D_{ii} - 1) X_{ii} + \sum_{j=1}^{i-1}(D_{ij} - D_{ij+1}) X_{ij}\nonumber\\
&\ge D_{zz} X_{zz} + \sum_{j=1}^{z-1} (D_{zj} - D_{zj+1}) X_{zj}, \quad i=2, \ldots, n; \; z=1, \ldots, i - 1.\end{aligned}$$
Similarly, for the valley trend we include,
$$\begin{aligned}
&y_i + y_z + 1 + (D_{ii} - 1) X_{ii} + \sum_{j=1}^{i-1} (D_{ij} - D_{ij+1}) X_{ij}\nonumber\\
&\ge D_{zz} X_{zz} + \sum_{j=1}^{z-1}(D_{zj} - D_{zj+1}) X_{zj}, \quad i=2, \ldots, n; \; z=1, \ldots, i - 1,\\
& 2 - y_i - y_z + 1 + (D_{zz} - 1) X_{zz} + \sum_{j=1}^{z-1}(D_{zj} - D_{zj+1}) X_{zj}\nonumber\\
&\ge D_{ii} X_{ii} + \sum_{j=1}^{i-1} (D_{ij} - D_{ij+1}) X_{ij}, \quad i=2, \ldots, n; \; z=1, \ldots, i - 1.\end{aligned}$$
Note that none of these constraints are necessary if the position of the change bin $t$ is fixed in advance. For example, given $t$, the valley trend constraints are replaced by two sets of constraints; one to guarantee a descending monotonic trend before $t$ and another to guarantee an ascending monotonic trend after $t$. Devising an effective heuristic to determine the optimal $t$ can yield probably optimal solutions while reducing the problem size significantly.
### Additional constraints {#section_additional_contraints}
#### Reduction of dominating bins.
To prevent any particular bin from dominating the results, it might also be required that bins have at most a certain number of (total/non-event/event) records using constraints (\[min\_max\_bin\_size\] - \[min\_max\_bin\_n\_event\]). Furthermore, we might produce more homogeneous solutions by reducing a concentration metric such as the standard deviation of the number of total/non-event/event records among bins. Three concentration metrics are considered: standard deviation, Herfindahl-Hirschman Index [@Mironchyk2017] and the difference between the largest and smallest bin.
The standard deviation among the number of records for each bin is given by $$std = \left(\frac{1}{m - 1} \sum_{i=1}^n \left(\sum_{j=1}^i r_j X_{ij} - \frac{X_{ii}}{m} \sum_{i=1}^n \sum_{j=1}^i r_j X_{ij}\right)^2\right)^{1/2},$$ where $m = \sum_{i=1}^n X_{ii}$ is the optimal number of bins. Let us define the following auxiliary variables $$\mu = \frac{1}{m} \sum_{i=1}^n \sum_{j=1}^i r_j X_{ij}, \quad w_i = \sum_{j=1}^i r_j X_{ij} - \mu X_{ii}, \quad i=1, \ldots n.$$ Taking $w = (w_1, \ldots, w_n)^T$, the standard deviation $t$ can be incorporated to the formulation with a different representation. Since $std = (w^T w / (m - 1))^{1/2}$ then $$\frac{||w||_2}{(m - 1)^{1/2}} \le t \Longleftrightarrow ||w||_2^2 \le (m - 1) t^2.$$ The non-convex MINLP formulation using the parameter $\gamma$ to control the importance of the term $t$,
$$\begin{aligned}
\underset{X, \mu, w}{\text{max}} \quad & \sum_{i=1}^n V_{ii} X_{ii} + \sum_{j=1}^{i-1} (V_{ij} - V_{ij+1}) X_{ij} - \gamma t \label{reduce_std_objective}\\
\text{s.t.} \quad & \text{(\ref{sum_columns_one} - \ref{binary_indicator})}\\
& \sum_{i=1}^n w_i^2 \le (m - 1) t^2\\
& \mu = \frac{1}{m} \sum_{i=1}^n \sum_{j=1}^i r_j X_{ij}\\
& w_i = \sum_{j=1}^i r_j X_{ij} - \mu X_{ii}, & i=1, \ldots n\\
& m = \sum_{i=1}^n X_{ii}\\
& m \ge 0\\
& \mu \ge 0\\
& w_i \in \mathbb{R}, & i=1, \ldots, n.\end{aligned}$$
A widely used metric to quantify concentration is the Herfindahl-Hirschman Index (HHI), which can be employed to asses the quality of a binning solution. Lower values of HHI correspond to more homogeneous bins. The HHI of the number of records for each bin is given by $$HHI = \frac{1}{r_T^2} \sum_{i=1}^n \left(\sum_{j=1}^i r_j X_{ij}\right)^2,$$ where $r_T = \sum_{i=1}^n r_i$ is the total number of records. The MIQP formulation using the parameter $\gamma$ to control the importance of HHI is stated as
$$\begin{aligned}
\underset{X}{\text{max}} \quad & \sum_{i=1}^n V_{ii} X_{ii} + \sum_{j=1}^{i-1} (V_{ij} - V_{ij+1}) X_{ij} - \frac{\gamma}{r_T^2} \sum_{i=1}^n \left(\sum_{j=1}^i r_j X_{ij}\right)^2 \label{reduce_hhi}\\
\text{s.t.} \quad & \text{(\ref{sum_columns_one} - \ref{binary_indicator})}\end{aligned}$$
An effective MILP formulation can be devised using a simplification of the standard deviation approach based on reducing the difference between the largest and smallest bin. The MILP formulation is given by
$$\begin{aligned}
\underset{X, p_{\min}, p_{\max}}{\text{max}} \quad & \sum_{i=1}^n V_{ii} X_{ii} + \sum_{j=1}^{i-1} (V_{ij} - V_{ij+1}) X_{ij} - \gamma (p_{\max} - p_{\min}) \label{reduce_diff_max_min_size}\\
\text{s.t.} \quad & \text{(\ref{sum_columns_one} - \ref{binary_indicator})}\\
& p_{\min} \le r_T (1 - X_{ii}) + \sum_{j=1}^i r_j X_{ij}, & i = 1, \ldots, n\\
& p_{\max} \ge \sum_{j=1}^i r_j X_{ij}, & i = 1, \ldots, n\\
& p_{\min} \le p_{\max}\\
& p_{\min} \ge 0.\\
& p_{\max} \ge 0.\end{aligned}$$
#### Maximum p-value constraint.
A necessary constraint to guarantee that event rates between consecutive bins are statistically different is to impose a maximum p-value constraint setting a significance level $\alpha$. Suitable statistical tests are the Z-test, Pearson’s Chi-square test or Fisher ’s exact test. To perform these statistical tests we require an aggregated matrix of non-event and event records per bin, $$R^{NE}_{ij} = \sum_{z=j}^i r_z^{NE}, \quad R^E_{ij} = \sum_{z=j}^i r_z^E, \quad i=1, \ldots, n;\; j=1, \ldots, i.$$
The preprocessing procedure to detect pairs of pre-bins that do not satisfy the p-value constraints using the Z-test is shown in Algorithm \[alg:max\_pvalue\].
$zscore = \Phi^{-1}(1 - \alpha / 2)$ $\mathcal{I} = \{\}$ $l = i + 1$ $x = R^E_{ij}$ $y = R^{NE}_{ij}$ $w = R^E_{kl}$ $z = R^{NE}_{kl}$ $\mathcal{I} = \mathcal{I} \cup (i, j, k, l)$
These constraints are added to the formulation by imposing that, at most, one of the bins violating the maximum p-value constraint can be selected, $$X_{ij} + X_{kl} \le 1, \quad \forall(i, j, k, l) \in \mathcal{I}.$$
### Mixed-integer programming reformulation for local and heuristic search {#section_localsolver}
The number of binary decision variables $X$ is $n (n + 1) / 2$. For large $n$ the $\mathcal{NP}$-hardness of the combinatorial optimization problem might limit the success of tree-search techniques. A first approach to tackle this limitation is reformulating the problem to reduce the number of decision variables. First, observe in Figure \[basic\_diagram\] that a solution is fully characterized by the diagonal of $X$. Thus, having the diagonal we can place the ones on the positions satisfying unique assignment (\[def\_ct\_sum\_columns\_one\]) and continuity (\[def\_ct\_flow\_continuity\]) constraints. On the other hand, to return indexed elements in any aggregated matrix in the original formulation, we require the position of the first one by row. We note that this information can be retrieved by counting the number of consecutive zeros between ones (selected bins) of the diagonal. To perform this operation we use two auxiliary decision variables: an accumulator of preceding zeros $a_i$ and the preceding run-length of zeros $z_i$. A similar approach to counting consecutive ones is introduced in [@Kalvelagen2018] The described approach is illustrated in Figure \[heuristic\_diagram\].
![New decision variables suitable to counting consecutive zeros.[]{data-label="heuristic_diagram"}](images/heuristic_matrix_diagram.pdf)
The positions in $z$ are zero-based indexes of the reversed rows of the aggregated matrices. For example, the aggregated lower triangular matrix $R^E$ is now computed backward: $R^E_{ij} = \sum_{z=i}^j r_z^E$ for $i=1, \ldots, n;\; j=1, \ldots, i$. Same for the aggregated matrices $V$, $D$, $R$ and $R^{NE}$. Let us define the parameters of the mathematical programming formulation: $$\begin{aligned}
& n \in \mathbb{N} && \textrm{number of pre-bins.}\\
& V_{[i, z_i]} \in \mathbb{R}^+_0 && \textrm{Information value.}\\
& D_{[i, z_i]} \in [0, 1] && \textrm{event rate.}\\
& R_{[i, z_i]} \in \mathbb{N} && \textrm{number of records.}\\
& R^{NE}_{[i, z_i]} \in \mathbb{N} && \textrm{number of non-event records.}\\
& R^E_{[i, z_i]} \in \mathbb{N} && \textrm{number of event records.}\\
& r^{NE}_{\min} \in \mathbb{N} && \textrm{minimum number of non-event records per bins.}\\
& r^{NE}_{\max} \in \mathbb{N} && \textrm{maximum number of non-event records per bins.}\\
& r^E_{\min} \in \mathbb{N} && \textrm{minimum number of event records per bins.}\\
& r^E_{\max} \in \mathbb{N} && \textrm{maximum number of event records per bins.}\\
& b_{\min} \in \mathbb{N} && \textrm{minimum number of bins.}\\
& b_{\max} \in \mathbb{N} && \textrm{maximum number of bins.}\end{aligned}$$ and the decision variables: $$\begin{aligned}
& x_i \in \{0, 1\} && \textrm{binary indicator variable.}\\
& a_i \in \mathbb{N}_0 && \textrm{accumulator of preceding zeros.}\\
& z_i \in \mathbb{N}_0 && \textrm{preceding run-length of zeros.}\end{aligned}$$
The new formulation with $3n$ decision variables is stated as follows
$$\begin{aligned}
\underset{X}{\text{max}} \quad & \sum_{i=1}^n V_{[i, z_i]} x_i\\\label{new_objective_cp}
\text{s.t.} \quad & x_n = 1\\
\quad & a_i = (a_{i-1} + 1) (1-x_i), & i= 1, \ldots, n\\\label{accum_zero_cp}
\quad & z_i = a_{i-1} (1-x_{i-1}) x_i, & i= 1, \ldots, n\\\label{count_zero_cp}
\quad & b_{\min} \le \sum_{i=1}^n x_i \le b_{\max}\\
\quad & r_{\min} \le \sum_{i=1}^n R_{[i, z_i]} x_i \le r_{\max}, & i = 1, \ldots, n\\
\quad & r^{NE}_{\min} \le \sum_{i=1}^n R^{NE}_{[i, z_i]} x_i \le r^{NE}_{\max}, & i = 1, \ldots, n\\
\quad & r^E_{\min} \le \sum_{i=1}^n R^E_{[i, z_i]} x_i \le r^E_{\max}, & i = 1, \ldots, n\\
\quad & x_i \in \{0, 1\}, & i = 1, \ldots, n\\
\quad & a_i \in \mathbb{N}_0, & i = 1, \ldots, n\\
\quad & z_i \in \mathbb{N}_0, & i = 1, \ldots, n\label{z_cp}\end{aligned}$$
This MINLP formulation is particularly suitable for Local Search (LS) and heuristic techniques, where decision variable $z_i$ can be used as an index, for example, $V_{[i, z_i]}$. The nonlinear constraints (\[accum\_zero\_cp\]) and (\[count\_zero\_cp\]) are needed to counting consecutive zeros. After the linearization of these constraints via big-$M$ inequalities or indicator constraints [@Kalvelagen2018], the formulation is adequate for Constraint Programming (CP). Additional constraints such as monotonicity constraints can be incorporated to (\[new\_objective\_cp\] - \[z\_cp\]) in a relatively simple manner:
Monotonic trend ascending $$D_{[i, z_i]} x_i + 1 - x_i \ge D_{[j, z_j]} x_j + \beta (x_i + x_j - 1), \quad i=2, \ldots, n;\; z=1, \ldots, i - 1.$$
Monotonic trend descending $$D_{[i, z_i]} x_i + \beta (x_i + x_j - 1) \le 1 - x_j + D_{[j, z_j]} x_j, \quad i=2, \ldots, n;\; z=1, \ldots, i - 1.$$
Monotonic trend concave $$-D_{[i, z_i]} x_i + 2 D_{[j, z_j]} x_j - D_{[k, z_k]} x_k \ge x_i + x_j + x_k - 3,$$ for $i=3, \ldots, n;\; j=2, \ldots, i - 1;\; k=1, \ldots, j - 1$.
Monotonic trend convex $$D_{[i, z_i]} x_i - 2 D_{[j, z_j]} x_j + D_{[k, z_k]} x_k \ge x_i + x_j + x_k - 3,$$ for $i=3, \ldots, n;\; j=2, \ldots, i - 1;\; k=1, \ldots, j - 1$.
Monotonic trend peak: constraints (\[disjoint\_constraint\] - \[disjoint\_aux\_variables\]) and
$$\begin{aligned}
y_i + y_j + 1 + (D_{[j, z_j]} - 1) x_j - D_{[i, z_i]} x_i &\ge 0\\
2 - y_i -y_j + 1 + (D_{[i, z_i]} - 1) x_i - D_{[j, z_j]} x_j &\ge 0,\end{aligned}$$
for $i=2, \ldots, n;\; z=1, \ldots, i - 1$.
Monotonic trend valley: constraints (\[disjoint\_constraint\] - \[disjoint\_aux\_variables\]) and
$$\begin{aligned}
y_i + y_j + 1 + (D_{[i, z_i]} - 1) x_i - D_{[j, z_j]} x_j & \ge 0\\
2 - y_i -y_j + 1 + (D_{[j, z_j]} - 1) x_j - D_{[i, z_i]} x_i &\ge 0\end{aligned}$$
for $i=2, \ldots, n;\; z=1, \ldots, i - 1$.
In Section \[section\_experiments\], we compare the initial CP/MIP formulation to the presented LS formulation for large size instances.
Mixed-integer programming formulation for continuous target
-----------------------------------------------------------
The presented optimal binning formulation given a binary target can be seamlessly extended to a continuous target. Following the methodology developed for Equations (\[iv\_obj\_params\]) and (\[iv\_obj\_definition\]), given a $p$-norm distance ($L_1$-norm or $L_2$-norm), an aggregated lower triangular matrix $L_{ij} \in \mathbb{R}^+_0, \forall (i, j) \in \{1, \ldots, n: i \ge j\}$ can be pre-computed as follows, $$L_{ij} = \left\|\mu_i - U_{ij}\right\|, \quad U_{ij} = \frac{\sum_{z=i}^j s_z}{\sum_{z=i}^j r_z},$$ where $\mu_i \in \mathbb{R}$ is the target mean for each pre-bin and $U_{ij} \in \mathbb{R}$ is the aggregated matrix of mean values, where $s_i$ is the sum of target values for each pre-bin. A more robust approach might replace the mean by an order statistic, but unfortunately, the aggregated matrix computed from the pre-binning data would require approximation methods. Replacing $L_{ij}$ in the objective function (\[objective\]) and changing the optimization sense to minimization, the resulting formulation is given by
$$\begin{aligned}
\underset{X}{\text{min}} \quad & \sum_{i=1}^n L_{ii} X_{ii} + \sum_{j=1}^{i-1} (L_{ij} - L_{ij+1}) X_{ij}\\
\text{s.t.} \quad & \text{(\ref{sum_columns_one} - \ref{binary_indicator})}\end{aligned}$$
### Monotonicity constraints {#monotonicity-constraints}
As for the binary target case, we can impose monotonicity constraints between the mean value of consecutive bins. Since establishing tight bounds for $U_{ij} \in \mathbb{R}$ is not trivial, we discard a big-$M$ formulation. Another traditional technique such as the use of SOS1 sets is also discarded to avoid extra variables and constraints. Instead, we state the ascending monotonic trend in double implication form as follows $$\begin{aligned}
&X_{ii} = 1 \textrm{ and } X_{zz} = 1 \Longrightarrow
U_{zz} X_{zz} + \sum_{j=1}^{z-1}(U_{zj} - U_{z j+1} X_{zj}) + \beta\nonumber\\ &\le U_{ii} X_{ii} + \sum_{j=1}^{i - 1} (U_{ij} - U_{ij+1})X_{ij}, \quad i=2, \ldots, n; \; z=1,\ldots i-1.\end{aligned}$$ The enforced constraint must be satisfied iff the two literals $X_{ii}$ and $X_{zz}$ are true, otherwise the constraint is ignored. This is half-reified linear constraint [@Feydy2011]. The parameter $\beta$ is the minimum mean difference between consecutive bins. Similarly, for the descending constraint, $$\begin{aligned}
&X_{ii} = 1 \textrm{ and } X_{zz} = 1 \Longrightarrow U_{ii} X_{ii} + \sum_{j=1}^{i-1}(U_{ij} - U_{i j+1} X_{ij}) + \beta\nonumber\\
&\le U_{zz} X_{zz} + \sum_{j=1}^{z - 1} (U_{zj} - U_{zj+1})X_{zj}, \quad i=2, \ldots, n; \; z=1,\ldots i-1.\end{aligned}$$
Furthermore, the concave and convex trend can be written in triple implication form using literals $X_{ii}$, $X_{jj}$ and $X_{kk}$. The concave trend constraints are $$\begin{aligned}
X_{ii} = 1, X_{zz} = 1 \textrm{ and } X_{kk} = 1 \Longrightarrow
&-\left(U_{ii}X_{ii} + \sum_{z=1}^{i-1}(U_{iz} - U_{iz+1}) X_{iz}\right)\nonumber\\
& + 2 \left(U_{jj}X_{jj} + \sum_{z=1}^{j-1}(U_{jz} - U_{jz+1}) X_{jz}\right)\nonumber\\
&-\left(U_{kk}X_{kk} + \sum_{z=1}^{k-1}(U_{kz} - U_{kz+1}) X_{kz}\right) \ge 0,\end{aligned}$$ for $i=3, \ldots n$; $j=2, \ldots, i-1$ and $k=1, \ldots, j-1$. Similarly, for convex trend we get $$\begin{aligned}
X_{ii} = 1, X_{zz} = 1 \textrm{ and } X_{kk} = 1 \Longrightarrow
&\left(U_{ii}X_{ii} + \sum_{z=1}^{i-1}(U_{iz} - U_{iz+1}) X_{iz}\right)\nonumber\\
& - 2 \left(U_{jj}X_{jj} + \sum_{z=1}^{j-1}(U_{jz} - U_{jz+1}) X_{jz}\right)\nonumber\\
&\left(U_{kk}X_{kk} + \sum_{z=1}^{k-1}(U_{kz} - U_{kz+1}) X_{kz}\right) \ge 0,\end{aligned}$$ for $i=3, \ldots n$; $j=2, \ldots, i-1$ and $k=1, \ldots, j-1$.
The formulation can be extended to support valley and peak trend. The peak trend requires constraints (\[disjoint\_constraint\] - \[disjoint\_aux\_variables\]) and
$$\begin{aligned}
X_{ii} = 1 \textrm{ and } X_{zz} = 1 \Longrightarrow
&M (y_i + y_z) + U_{zz} X_{zz} + \sum_{j=1}^{z-1} (U_{zj} - U_{zj+1}) X_{zj}\nonumber\\
&\ge U_{ii} X_{ii} + \sum_{j=1}^{i-1}(U_{ij} - U_{ij+1}) X_{ij}, \\
& M (2 - y_i - y_z) + U_{ii} X_{ii} + \sum_{j=1}^{i-1}(U_{ij} - U_{ij+1}) X_{ij}\nonumber\\
&\ge U_{zz} X_{zz} + \sum_{j=1}^{z-1} (U_{zj} - U_{zj+1}) X_{zj},\end{aligned}$$
for $i=2, \ldots, n; \; z=1, \ldots, i - 1$. The big-$M$ formulation to handle disjoint constraints in (\[disjoint\_constraint\] - \[disjoint\_aux\_variables\]) requires an effective bound, we suggest $M= \max(\{|U_{ij}|: i=1, \ldots, n: i \ge j\})$. Similarly, for the valley trend we include constraints and
$$\begin{aligned}
X_{ii} = 1 \textrm{ and } X_{zz} = 1 \Longrightarrow
&M (y_i + y_z) + U_{ii} X_{ii} + \sum_{j=1}^{i-1} (U_{ij} - U_{ij+1}) X_{ij}\nonumber\\
&\ge U_{zz} X_{zz} + \sum_{j=1}^{z-1}(U_{zj} - U_{zj+1}) X_{zj},\\
& M(2 - y_i - y_z) + U_{zz} X_{zz} + \sum_{j=1}^{z-1}(U_{zj} - U_{zj+1}) X_{zj},\nonumber\\
&\ge U_{ii} X_{ii} + \sum_{j=1}^{i-1} (U_{ij} - U_{ij+1}) X_{ij},\end{aligned}$$
for $i=2, \ldots, n; \; z=1, \ldots, i - 1$.
Finally, additional constraints described in Section \[section\_additional\_contraints\] can be naturally incorporated with minor changes.
Mixed-integer programming formulation for multi-class target
------------------------------------------------------------
A simple approach to support a multi-class target is to use the one-vs-rest scheme with $n_C$ distinct classes. This scheme consists of building a binary target for each class. The resulting mathematical formulation closely follows the formulation for binary target,
$$\begin{aligned}
\underset{X}{\text{max}} \quad & \sum_{c=1}^{n_C}\sum_{i=1}^n V^c_{ii} X_{ii} + \sum_{j=1}^{i-1} (V^c_{ij} - V^c_{ij+1}) X_{ij} \label{multiclass_objective}\\
\text{s.t.} \quad & \text{(\ref{sum_columns_one} - \ref{min_max_bin_size})}\end{aligned}$$
Note that for this formulation we need an aggregated matrix $V$ and $D$ for each class $c$. It is important to emphasize that the monotonicity constraints in Section \[section\_monotonicity\] act as linking constraints among classes, otherwise, $n_C$ optimal binning problems with binary target could be solved separately. Again, additional constraints described in Section \[section\_additional\_contraints\] can be naturally incorporated with minor changes.
Algorithmic details and implementation {#section_algorithms}
======================================
Automatic monotonic trend algorithm
-----------------------------------
Our approach to automate the monotonic trend decision employs a Machine Learning (ML) classifier that predicts, given the pre-binning data, the most suitable monotonic trend to maximize discriminatory power (binary and multi-class target) or minimize a prediction error (continuous target). In particular, we aim to integrate an off-line classifier, hence we are merely interested in ML classification algorithms easily embeddable. Recently, a similar approach was implemented in the commercial solver CPLEX to make automatic decisions over some algorithmic choices [@bonami2018].
For this study, a dataset is generated with 417 instances collected from public datasets. We design a set of 16 numerical features, describing the pre-binning instances in terms of number of pre-bins, distribution of records per pre-bin and trend features. The most relevant trend features are: number of trend change points, linear regression coefficient sense, area of the convex hull, and area comprised among extreme trend points.
The labeling procedure consists of solving all instances selecting the ascending (A), descending (D), peak (P) and valley (V) monotonic trend. Concave and convex trends are discarded due to being a special case of peak and valley, respectively. In what follows, without loss of generality, we state the procedure for the binary target case: if the relative difference between IV with ascending/descending monotonic trend and IV with peak/valley trend is less than 10%, the ascending/descending monotonic trend is selected, due to the lesser resolution times. Table \[table\_dataset\] summarizes the composition of the dataset with respect to assigned labels. We note that the dataset is slightly unbalanced, being predominant the descending label (D).
To perform experiments, the dataset is split into train and test subsets in a stratified manner to treat unbalanced data. The proportion of the test set is 30%. Three interpretable multi-class classification algorithms are tested, namely, logistic regression, decision trees (CART) and Support vector Machine (SVM) using the Python library Scikit-learn [@scikit-learn]. All three algorithms are trained using option `class_weight=balanced`. Throughout the learning process, we discard 8 features and perform hyperparameter optimization for all three algorithms. These experiments show that SVM and CART have similar classification measures, and we decide to choose CART (max depth 5) to ease implementation.
On the test set, the trained CART has a weighted average accuracy, precision and recall of 88%. See classification measures and the confusion matrix in Table \[table\_measures\_confusion\]. We observe that various instances of the minority class (V) are misclassified, indicating that more instances or new features might be required to improve classification measures. Improving this classifier is part of ongoing research.
Presolving algorithm
--------------------
The mathematical programming formulation is a hard combinatorial optimization problem that does not scale well as the number of pre-bins increases. To reduce solution times we need to reduce the search space to avoid deep tree searches during branching. The idea is to develop a presolving algorithm to fix bins not satisfying monotonicity constraints, after that the default presolver may be able to reduce the problem size significantly. The presolving algorithm applies to the binary target case and was developed after several observations about the aggregated matrix of event rates $D$. Algorithm \[alg:asc\_mono\] shows the implemented approach for the ascending monotonicity trend. Presolving algorithm for the descending monotonicity is analogous, only requiring inequalities change.
fix $X_{i,j} = 0$ fix $X_{i+j, i+j} = 0$
Binning quality score
---------------------
To assess the quality of binning for binary target, we develop a binning quality score considering the following aspects:
- Predictive power: IV rule of thumb [@Siddiqi2005] in Table \[IV\_rule\_of\_thumb\].
- Statistical significance: bin event rates must be statistically different, therefore large p-values penalize the quality score.
- Homogeneity: binning with homogeneous bin sizes or uniform representativeness, increases reliability.
To account for all these aspects, we propose a rigorous binning quality score function
Given a binning with Information Value $\nu$, p-values between consecutive bins $p_i$, $i=1, \ldots, n - 1$ and normalized bins size $s_i$, $i=1, \ldots, n$, the binning quality score function is defined as $$Q(\nu, p, s) = \frac{\nu}{c} \exp\left(-\nu^2/ (2c^2) + 1/2\right) \left(\prod_{i=1}^{n-1} (1 - p_i)\right) \left(\frac{1 - \sum_{i=1}^n s_i^2}{1 - 1/n}\right),$$ where $Q(\nu, p, s) \in [0, 1]$ and $c = \frac{1}{5}\sqrt{\frac{2}{\log(5/3)}}$ is the best a priori IV value in $[0.3, 0.5)$.
Given the rule of thumb in Table \[IV\_rule\_of\_thumb\], let us consider the set of statistical distributions with positive skewness, positive fat-tail, and support on the semi-infinite interval $[0, \infty)$. The function should penalize large values of Information Value $\nu$ and it is expected a fast decay after certain threshold indicating over-prediction. This fast decay is a required property that must be accompanied by the following statement: $\lim_{\nu\to 0} f(\nu) = \lim_{\nu \to \infty} f(\nu) = 0$ Among the available distributions satisfying aforementioned properties, we select the Rayleigh distribution, which probability density function is given by $$f(\nu;c) = \frac{\nu}{c^2} e^{-\nu^2 / (2c^2)}, \quad \nu\ge 0.$$
This is a statistical distribution, not a function, hence we need a scaling factor so that $\max_{\nu \in [0, \infty)}f(\nu; c) = 1$: the maximum value of the unimodal probability distribution is the mode $c$, thus $$\gamma = f(c,c) = \frac{1}{c \sqrt{e}} \Longrightarrow \frac{f(\nu,c)}{\gamma} = \frac{\nu \exp\left(-\nu^2/ (2c^2) + 1/2\right)}{c}.$$ The optimal $c$ such that $f(a) = f(b)$ for $b > a$ can be obtained by solving $f(b; c) - f(a; c) = 0$ for $c$, which yields $$c^* = \frac{\sqrt{b^2 - a^2}}{\sqrt{2\log(b/a)}}.$$ Term $\prod_{i=1}^{n-1} (1 - p_i)$ assesses the statistical significance of the bins. Furthermore, term $\frac{1 - \sum_{i=1}^n s_i^2}{1 - 1/n} = 1 - HHI^*$, where $HHI^*$ is the normalized Herfindahl Hirschman Index, assesses the homogeneity/uniformity of the bin sizes.
For example, if we consider that the boundaries of the interval with strong IV predictive power in Table \[IV\_rule\_of\_thumb\], $a = 0.3$ and $b = 0.5$, should produce the same quality score, $c^*(a,b) = \frac{1}{5}\sqrt{\frac{2}{\log(5/3)}}$. Table \[table\_iv\_metrics\] shows the value of $f(\nu, c^*)$ for various IV values $\nu$; note the fast decay of $f(\nu, c^*)$ when $\nu > 0.5$.
Implementation
--------------
The presented mathematical programming formulations are implemented using Google OR-Tools [@ortools] with the open-source MILP solver CBC [@cbc_2018], and Google’s BOP and CP-SAT solvers. Besides, the specialized formulation in Section \[section\_localsolver\] is implemented using the commercial solver LocalSolver [@localsolver2011]. The python library OptBinning[^3] has been developed throughout this work to ease usability and reproducibility.
Much of the implementation effort focuses on the careful implementation of constraints and the development of fast algorithms for preprocessing and generating the model data. A key preprocessing algorithm is a pre-binning refinement developed to guarantee that no bins have 0 accounts for non-events and events in the binary target case.
Categorical variables require special treatment: pre-bins are ordered in ascending order with respect to a given metric; the event rate for binary target and the target mean for a continuous target. The original data is replaced by the ordered indexes and is then used as a numerical (ordinal) variable. Furthermore, during preprocessing, the non-representative categories may be binned into an “others” bin. Similarly, missing values and special values are incorporated naturally as additional bins after the optimal binning is terminated.
Experiments {#section_experiments}
===========
The experiments were run on an Intel(R) Core(TM) i5-3317 CPU at 1.70GHz, using a single core, running Linux. Two binning examples are shown in Tables \[fico\_example\] and \[kaggle\_example\], using Fair Isaac (FICO) credit risk dataset [@FICO2018] ($N=10459$) and Home Credit Default Risk Kaggle competition dataset [@HomeCreditGroup2018] ($N=307511$), respectively.
Example in Table \[fico\_example\] uses the variable AverageMInFile (Average Months in File) as an risk driver. FICO dataset imposes monotonicity constraints to some variables, in particular, for this variable, the event rate must be monotonically decreasing. Moreover, the dataset includes three special values/codes defined as follows:
- -9: No Bureau Record or No Investigation
- -8: No Usable/Valid Trades or Inquiries
- -7: Condition not Met (e.g. No Inquiries, No Delinquencies)
For the sake of completeness, we also include a few random missing values on the dataset. As shown in Table \[fico\_example\], these values are separately treated by incorporating a Special and Missing bin. Regarding computation time, this optimal binning instance is solved in 0.08 seconds. The optimization time accounts for 91% of the total time, followed by the pre-binning time representing about 6%. The remaining 3% is spent in pre-processing and post-processing operations.
Example in Table \[kaggle\_example\] uses the categorical variable ORGANIZATION\_TYPE from Kaggle dataset. This variable has 58 categories, and we set the non-representative categories cut-off to $0.01$. Note that the bin just before the Special bin corresponds to the bin with non-representative categories, which is excluded from the optimization problem, hence monotonicity constraint does not apply. This optimal binning instance is solved in 0.25 seconds. For categorical variables, most of the time is spent on pre-processing, 71% in this particular case, whereas the optimization problem is solved generally faster.
Benchmark CP/MIP vs local search heuristic
------------------------------------------
For large instances, we compare the performance of Google OR-Tools’ solvers BOP (MIP) and CP-SAT against LocalSolver. For these tests we select two variables from Home Credit Default Risk Kaggle competition dataset [@HomeCreditGroup2018] ($N=307511$). We aim to perform a far finer binning than typical in many applications to stress the performance of classical solvers for large combinatorial optimization problems.
Tables \[benchmark\_preprocessing\_1\] and \[benchmark\_preprocessing\_2\] show results for varying number of pre-bins $n$ and monotonic trends. In test 1 from Table \[benchmark\_preprocessing\_1\] , LocalSolver does not improve after 10 seconds, not being able to reduce the optimality gap. In test 2, LocalSolver outperforms CP-SAT, finding the optimal solution after 5 seconds, 28x faster. In test 3, solution times are comparable. Results reported in Table \[benchmark\_preprocessing\_2\] are also interesting; in test 1, BOP and CP-SAT solvers cannot find an optimal solution after 1000 seconds. LocalSolver finds the best found feasible solution after 30 seconds. Nevertheless, we recall that the described heuristic for peak/valley trend introduced in Section \[section\_monotonicity\] could reduce resolution times substantially, obtaining times comparable to those when choosing ascending/descending monotonic trend.
Conclusions {#section_conclusions}
===========
We propose a rigorous and flexible mathematical programming formulation to compute the optimal binning. This is the first optimal binning algorithm to achieve solutions for nontrivial constraints, supporting binary, continuous and multi-class target, and handling several monotonic trends rigorously. Importantly, the size of the decision variables and constraints used in the presented formulations is independent of the size of the datasets; they are entirely controlled by the starting solution computed during the pre-binning process. In the future, we plan to extend our methodology to piecewise-linear binning and multivariate binning. Lastly, the code is available at <https://github.com/guillermo-navas-palencia/optbinning> to ease reproducibility.
[^1]: To the author’s knowledge, at the time of writing, these tools are restricted to the problem of discretizing a variable with respect to a binary target.
[^2]: In some commercial tools, peak and valley trend are called inverse U-shaped and U-shaped, respectively.
[^3]: <https://github.com/guillermo-navas-palencia/optbinning>
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We demonstrate the utility of the numerical Contractor Renormalization (CORE) method for quantum spin systems by studying one and two dimensional model cases. Our approach consists of two steps: (i) building an effective Hamiltonian with longer ranged interactions up to a certain cut-off using the CORE algorithm and (ii) solving this new model numerically on finite clusters by exact diagonalization and performing finite-size extrapolations to obtain results in the thermodynamic limit. This approach, giving complementary information to analytical treatments of the CORE Hamiltonian, can be used as a semi-quantitative numerical method. For ladder type geometries, we explicitely check the accuracy of the effective models by increasing the range of the effective interactions until reaching convergence. Our results in the perturbative regime and also away from it are in good agreement with previously established results. In two dimensions we consider the plaquette lattice and the lattice as non-trivial test cases for the numerical CORE method. As it becomes more difficult to extend the range of the effective interactions in two dimensions, we propose diagnostic tools (such as the density matrix of the local building block) to ascertain the validity of the basis truncation. On the plaquette lattice we have an excellent description of the system in both the disordered and the ordered phases, thereby showing that the CORE method is able to resolve quantum phase transitions. On the lattice we find that the previously proposed twofold degenerate $S=1/2$ basis can account for a large number of phenomena of the spin $1/2$ system. For spin $3/2$ however this basis does not seem to be sufficient anymore. In general we are able to simulate system sizes which correspond to an $8\times 8$ lattice for the plaquette lattice or a 48-site lattice, which are beyond the possibilities of a standard exact diagonalization approach.'
author:
- Sylvain Capponi
- Andreas Läuchli
- Matthieu Mambrini
title: Numerical Contractor Renormalization Method for Quantum Spin Models
---
Low-dimensional quantum magnets are at the heart of current interest in strongly correlated electron systems. These systems are driven by strong correlations and large quantum fluctuations - especially when frustration comes into play - and can exhibit various unconventional phases and quantum phase transitions.
One of the major difficulties in trying to understand these systems is that strong correlations often generate highly non trivial low-energy physics. Not only the groundstate of such models is generally not known but also the low-energy degrees of freedom can not be identified easily. Moreover, among the techniques available to investigate these systems, not many have the required level of generality to provide a systematic way to derive low-energy effective Hamiltonians.
Recently the Contractor Renormalization (CORE) method has been introduced by Morningstar and Weinstein [@weinstein96]. The key idea of the approach is to derive an effective Hamiltonian acting on a truncated local basis set, so as to exactly reproduce the low energy spectrum. In principle the method is exact in the low energy subspace, but only at the expense of having a priori long range interactions. The method becomes most useful when one can significantly truncate a local basis set and still restrict oneself to short range effective interactions. This however depends on the system under consideration and has to be checked systematically. Since its inception the CORE method has been mostly used as an analytical method to study strongly correlated systems [@weinstein01; @core-bosonfermion; @core-pyrochlore]. Some first steps in using the CORE approach and related ideas in a numerical framework have also been undertaken [@shepard97; @dynamicCORE; @core-tJladder; @malrieu].
The purpose of the present paper is to explore the numerical CORE method as a complementary approach to more analytical CORE procedures, and to systematically discuss its performance in a variety of low dimensional quantum magnets, both frustrated and unfrustrated. The approach consists basically of numerical exact diagonalizations of the effective Hamiltonians. In this way a large number of interesting quantities are accessible, which otherwise would be hard to obtain. Furthermore we discuss some criteria and tools useful to estimate the quality of the CORE approach.
The outline of the paper is as follows: In the first section we will review the CORE algorithm in general and discuss some particularities in a numerical CORE approach, both at the level of the calculation of the effective Hamiltonians and the subsequent simulations.
In section \[sec:ladder\] we move to the first applications on one-dimensional (1D) systems: the well known two-leg spin ladder and the 3-leg spin ladder with periodic boundary conditions in the transverse direction (3-leg torus). Both systems exhibit generically a finite spin gap and a finite magnetic correlation length. We will show that the numerical CORE method is able to get rather accurate estimates of the groundstate energy and the spin gap by successively increasing the range of the effective interactions.
In section \[sec:2D\] we discuss two-dimensional (2D) systems. As in 2D a long ranged cluster expansion of the interactions is difficult to achieve on small clusters, we will discuss some techniques to analyze the quality of the basis truncation. We illustrate these issues on two model systems, the plaquette lattice and the lattice. The plaquette lattice is of particular interest as it exhibits a quantum phase transition from a disordered plaquette state to a long range ordered Néel antiferromagnet, which cannot be reached by a perturbative approach. We show that a range-two effective model captures many aspects of the physics over the whole range of parameters. The lattice on the other hand is a highly frustrated lattice built of corner-sharing triangles. For spin $1/2$ it has been studied both numerically and analytically and it is one best-known candidate systems for a spin liquid groundstate. A very peculiar property is the exponentially large number of low-energy singlets in the magnetic gap. We show that already a basic range two CORE approach is able to devise an effective model which exhibits the same exotic low-energy physics. For higher half-integer spin, i.e. $S=3/2$, this simple effective Hamiltonian breaks down; we analyze how to detect this, and discuss some ways to improve the results.
In the last section we conclude and give some perspectives. Finally three appendices are devoted to (i) the density matrix of local building block, (ii) the calculation of observables by energy considerations and (iii) some general remarks on effective Hamiltonians coupling antiferromagnetic half-integer spin triangles.
CORE Algorithm
==============
The Contractor Renormalization (CORE) method has been proposed by Morningstar and Weinstein in the context of general Hamiltonian lattice models [@weinstein96]. Later Weinstein applied this method with success to various spin chain models [@weinstein01]. For a review of the method we refer the reader to these original papers [@weinstein96; @weinstein01] and also to a pedagogical article by Altman and Auerbach [@core-bosonfermion] which includes many details. Here, we summarize the basic steps before discussing some technical aspects which are relevant in our numerical approach.
[*CORE Algorithm :*]{}
- Choose a small cluster (e.g. rung, plaquette, triangle, etc) and diagonalize it. Keep $M$ suitably chosen low-energy states.
- Diagonalize the full Hamiltonian $H$ on a connected graph consisting of $N_c$ clusters and obtain its low-energy states $|n\rangle$ with energies $\varepsilon_n$.
- The eigenstates $|n\rangle$ are projected on the tensor product space of the states kept and Gram-Schmidt orthonormalized in order to get a basis $|\psi_n\rangle$ of dimension $M^{N_c}$. As it may happen that some of the eigenstates have zero or very small projection, or vanish after the orthogonalization it might be necessary to obtain more than just $M^{N_c}$ exact eigenstates.
- Next, the effective Hamiltonian for this graph is built as $$h_{N_c} = \sum_{n=1}^{M^{N_c}} \varepsilon_n
|\psi_n\rangle\langle \psi_n|.$$
- The connected range-$N_c$ interactions $h^{\rm conn}_{N_c}$ are determined by substracting the contributions of all connected subclusters.
- Finally, the effective Hamiltonian is given by a cluster expansion as $$H^{\mbox{\tiny CORE}}=\sum_i h_i +\sum_{\langle
ij\rangle}h_{ij}+\sum_{\langle ijk\rangle} h_{ijk} +\cdots$$
This effective Hamiltonian *exactly* reproduces the low-energy physics provided the expansion goes to infinity. However, if the interactions are short-range in the starting Hamiltonian, we can expect that these operators will become smaller and smaller, at least in certain situations. In the following, we will truncate at range $r$ and verify the convergence in several cases. This convergence naturally depends on the number $M$ of low-lying states that are kept on a basic block. In order to describe quantitatively how “good” these states are, we introduce the density matrix in section \[sec:2D\].
When the number of blocks increases, a full diagonalization is not always easy and one is tempted to use a Lanczos algorithm in order to compute the low-lying eigenstates. In that case, one has to be very careful to resolve the correct degeneracies, which is known to be a difficult task in the Lanczos framework. In practice such degeneracies arise when the cluster to be diagonalized is highly symmetric. If the degeneracies are ignored, often a wrong effective Hamiltonian with broken SU(2) symmetry is obtained. As a consequence we recommend to use specialized `LAPACK` routines whenever possible.
In the present work we investigate mainly SU(2) invariant Heisenberg models described by the usual Hamiltonian $$H=\sum_{\langle ij\rangle} J_{ij} \vec{S}_i \cdot \vec{S}_j$$ where the exchange constants $J_{ij}$ will be limited to short-range distances in the following. As a consequence of the SU(2) symmetry, the total spin of all states is a good quantum number. This also has some effects when calculating the effective Hamiltonian. It is possible to have situations where a low energy state has a non-zero overlap with the tensor product basis, but gets eliminated by the orthogonalization procedure because one has already exhausted all the states in one particular total spin sector by projecting states with lower energy.
Once an effective Hamiltonian has been obtained, it is still a formidable task to determine its properties. Within the CORE method different routes have been taken in the past. In their pioneering papers Morningstar and Weinstein have chosen to iteratively apply the CORE method on the preceding effective Hamiltonian in order to flow to a fixed point and then to analyze the fixed point. A different approach has been taken in Refs. \[\]: There the effective Hamiltonian after one or two iterations has been analyzed with mean-field like methods and interesting results have been obtained. Yet another approach - and the one we will pursue in this paper - consists of a single CORE step to obtain the effective Hamiltonian, followed by a numerical simulation thereof. This approach has been explored in a few previous studies [@shepard97; @dynamicCORE; @core-tJladder]. The numerical technique we employ is the Exact Diagonalization (ED) method based on the Lanczos algorithm. This technique has easily access to many observables and profits from the symmetries and conservation laws in the problem, i.e. total momentum and the total $S^z$ component. Using a parallelized program we can treat matrix problems of dimensions up to $\sim 50$ millions, however the matrices contain significantly more matrix elements than the ones of the microscopic Hamiltonian we start with.
Ladder geometries {#sec:ladder}
=================
In this section, we describe results obtained on ladder systems with 2 and 3 legs respectively.
![ (a) 2-leg ladder. Basic block is a $2\times 2$ plaquette. (b) 3-leg torus with rung coupling $J_\perp$ and inter-rung coupling $J_\parallel$. \[fig:Lattices1D\] ](Lattices1D){width="0.9\linewidth"}
We want to build an effective model that is valid from a perturbative regime to the isotropic case $J_{ij}=J=1$. We have chosen periodic boundary conditions (PBC) along the chains in order to improve the convergence to the thermodynamic limit.
Two-leg Heisenberg ladder {#sec:LadderGeometries}
-------------------------
The 2-leg Heisenberg ladder has been intensively studied and is known to exhibit a spin gap for all couplings [@2leg; @2legdata].
In order to apply our algorithm, we select a $2\times 2$ plaquette as the basic unit (see Fig. \[fig:Lattices1D\] (a)). The truncated subspace is formed by the singlet ground-state (GS) and the lowest triplet state.
Using the same CORE approach, Piekarewicz and Shepard have shown that quantitative results can be obtained within this restricted subspace [@shepard97]. Moreover, dynamical quantities can also be computed in this framework [@dynamicCORE].
![ Ground-state energy per site and spin gap of a $2\times L$ Heisenberg ladder using CORE method with various range $r$ using PBC. For comparison, we plot the best known extrapolations with arrows. \[fig:heisen\_2xL\] ](heisen_2xL){width="\linewidth"}
Since we are dealing with a simple system, we can compute the effective models including rather long-range interactions (typically, to obtain range-4 interactions, we need to compute the low-lying states on a $2\times 8$ lattice with Open Boundary Conditions which is feasible, although it requires a large numerical effort). It is desirable to compute long-range effective interactions since we wish to check how the truncation affect the physical results and how the convergence is reached.
In a second step, for each of these effective models, we perform a standard Exact Diagonalization (ED) using the Lanczos algorithm on finite clusters up to $N_c=12$ clusters ($N=48$ sites for the original model). The GS energy and the spin gap are shown in Fig. \[fig:heisen\_2xL\]. The use of PBC allows to reduce considerably finite-size effects since we have an exponential convergence as a function of inverse length. CORE results are in perfect agreement with known results and the successive approximations converge uniformly to the exact results. For instance, the relative errors of range-4 results are $10^{-4}$ for the GS energy and $10^{-2}$ for the spin gap. This fast convergence is probably due to the rather short correlation length in an isotropic ladder (typically 3 to 4 lattice spacings [@xi_2leg]).
3-leg Heisenberg torus
----------------------
As a second example of ladder geometry, we have studied a 3-leg Heisenberg ladder with PBC along the rungs. This property causes geometric frustration which leads to a finite spin-gap and finite dimerization for all interchain coupling $J_\perp$ [@kawano97; @cabra98], contrary to the open boundary condition case along the rungs, which is in the universality class of the Heisenberg chain.
[*Perturbation theory :*]{} The simple perturbation theory is valid when the coupling along the rung ($J_\perp$) is much larger than between adjacent rungs ($J_\parallel$). In the following, we fix $J_\perp=1$ as the energy unit and denote $\alpha=J_\parallel/J_\perp$.
On a single rung, the low-energy states are the following degenerate states, defined as $$\begin{aligned}
\mid\uparrow L\rangle&=&\frac{1}{\sqrt{3}}(\mid \uparrow \uparrow \downarrow
\rangle +\omega \mid\uparrow \downarrow \uparrow \rangle +\omega^2
\mid\downarrow \uparrow \uparrow \rangle), \label{eqn:TorusChirality} \\
\mid\downarrow L\rangle&=&\frac{1}{\sqrt{3}}(\mid\downarrow \downarrow
\uparrow \rangle +\omega \mid\downarrow \uparrow \downarrow \rangle
+\omega^2 \mid\uparrow \downarrow \downarrow \rangle), \nonumber \\
\mid\uparrow R\rangle&=&\frac{1}{\sqrt{3}}(\mid\uparrow \uparrow
\downarrow \rangle +\omega^2 \mid\uparrow \downarrow \uparrow \rangle
+\omega
\mid\downarrow \uparrow \uparrow \rangle), \nonumber\\
\mid\downarrow R\rangle&=&\frac{1}{\sqrt{3}}(\mid\downarrow \downarrow
\uparrow \rangle +\omega^2 \mid\downarrow \uparrow \downarrow \rangle
+\omega \mid\uparrow \downarrow \downarrow \rangle)
\nonumber\end{aligned}$$ where $\omega=\exp(i2\pi/3)$. The indices $L$ and $R$ represent the momentum of the 3-site ring $k_y=2\pi/3$ and $-2\pi/3$ respectively. They define two chiral states which can be viewed as pseudo-spin states with operators $\vec{\tau}$ on each rung defined by $$\begin{aligned}
\tau^+ \mid\cdot\, R\rangle = 0 \qquad \tau^+ \mid \cdot\, L\rangle = | \cdot\, R\rangle\\
\tau^- \mid \cdot\, R\rangle = | \cdot\, L\rangle \qquad \tau^- \mid \cdot\, L\rangle = 0 \\
\tau^z \mid \cdot\, R\rangle = \frac{1}{2} \mid \cdot\, R\rangle \quad \tau^z \mid \cdot\, L\rangle = -\frac{1}{2} \mid \cdot\, L\rangle \end{aligned}$$ These states have in addition a physical spin 1/2 described by $\vec{\sigma}$.
Applying the usual perturbation theory for the inter-rung coupling, one finds[@schulz; @kawano97] $$\label{H3leg_pert}
H_{pert}=-\frac{N}{4}+\frac{\alpha}{3}\sum_{\langle i j \rangle}
\vec{\sigma}_i\cdot \vec{\sigma}_j (1+4 (\tau_i^+ \tau_j^- + \tau_i^-
\tau_j^+))$$ where $N$ is the total number of sites.
This effective Hamiltonian has been studied with DMRG and ED techniques and it exhibits a finite spin gap $\Delta_S=0.28~J_\parallel$ and a dimerization of the ground state [@kawano97; @cabra98].
Here we want to use the CORE method to extend the perturbative Hamiltonian with an effective Hamiltonian in the same basis for *any* coupling.
[*CORE approach :*]{} As a basic unit, we choose a single 3-site rung. The subspace consists of the same low-energy states as for the perturbative result (Eq. (\[eqn:TorusChirality\])) which are 4-fold degenerate (2 degenerate $S=1/2$ states). We can apply our procedure to compute the effective interactions at various ranges, in order to be able to test the convergence of the method.
First, we write down the range-2 contribution under the most general form which preserves both $SU(2)$ (spin) symmetry and simultaneous translation or reflection along all the rungs : $$\begin{aligned}
H_{r=2}=N a_0 + \sum_{\langle i j \rangle}(b_0 \tau_i^z \tau_j^z
+c_0 (\tau_i^+ \tau_j^- + \tau_i^- \tau_j^+)) \nonumber\\
+ \vec{\sigma}_i\cdot
\vec{\sigma}_j (a_1+ b_1 \tau_i^z \tau_j^z+ c_1(\tau_i^+ \tau_j^- +
\tau_i^- \tau_j^+))\label{heff_3xL}\end{aligned}$$
In the perturbative regime given in (\[H3leg\_pert\]), the only non-vanishing coefficients are given by : $a_0=-1/4$, $a_1=\alpha/3$ and $c_1=4\alpha/3$.
The parameters of the effective Hamiltonian can be obtained and their dependence as a function of the inter-rung coupling $\alpha$ is shown in Fig. \[fig:3leg\_parameter\]. We immediately see some deviations from the perturbative result since coefficients in panel (i) and (ii) are non-zero and become as important as the other terms in the isotropic limit. Surprisingly, we observe that $c_1$ follows its perturbative expression on the whole range of couplings whereas $a_1$ deviates strongly as one goes to the isotropic case but does not change sign.
![CORE coefficients (see Eq. \[heff\_3xL\]) for two coupled triangles as a function of the inter-rung coupling $\alpha=J_\parallel/J_\perp$. The parameters were computed using range-2 CORE. The coefficients in panel (iii) have been divided by their values in the perturbative limit. They therefore all start at 1.[]{data-label="fig:3leg_parameter"}](Core3xL){width="0.9\linewidth"}
In order to study how the physical properties evolve as a function of $J_\parallel/J_\perp$, we have computed the GS energy and the spin gap both for a small-coupling case and in the isotropic limit, up to range 5 in the effective interactions.
[*Small interrung coupling :*]{} We have chosen $J_\parallel/J_\perp=0.25$ which corresponds to a case where perturbation theory should still apply. Using ED, we can solve the effective models on finite lattices and on Fig. \[fig:heisen\_3xL\_Jp4\], we plot the scaling of the GS energy and of the spin gap as a function of the system length $L$. Even for this rather small value of $J_\parallel/J_\perp$, our effective Hamiltonian can be considered as an improvement over the first order perturbation theory. Moreover, we observe a fast convergence with the range of interactions and already the range-3 approximation is almost indistinguishable from ED results.
![GS energy per site and spin gap for a $3\times L$ Heisenberg torus with $J_{\parallel}/J_\perp=0.25$. Results are obtained using the CORE method at various range $r$. \[fig:heisen\_3xL\_Jp4\] ](heisen_3xL_e0_Jp4){width="0.8\linewidth"}
![GS energy per site and spin gap for a $3\times L$ Heisenberg torus with $J_{\parallel}/J_\perp=0.25$. Results are obtained using the CORE method at various range $r$. \[fig:heisen\_3xL\_Jp4\] ](heisen_3xL_gap_Jp4){width="0.8\linewidth"}
The estimated gap is $0.16 J_\parallel$ and correspond to a lower bound since ultimately the gap should converge exponentially to its thermodynamic value. Our value is consistent with the DMRG one [@kawano97] ($\sim 0.2 J_\parallel$), and is already reduced compared to the strong coupling result [@kawano97] ($\Delta_S=0.28 J_\parallel$).
[*Isotropic case :*]{} We apply the same procedure in the isotropic limit. As expected, the convergence with the range of interactions is much slower than in the perturbative regime. We show on Fig. \[fig:heisen\_3xL\] that indeed the ground state energy converges slowly and oscillates around the correct value. These oscillations come from the fact that, in order to compute range-$r$ interactions, one has to study alternatively clusters with an even or odd number of sites. Since this system has a tendency to form dimers on nearest-neighbour bonds, it is better to compute clusters with an even number of sites.
For the spin gap, we find accurate results even with limited range interactions. In particular, we find that frustration induces a finite spin gap $\simeq 0.11~J_\parallel$ in that system. As in the previous case, this is a lower bound which is in perfect agreement with DMRG study [@kawano97].
![Same as Fig. \[fig:heisen\_3xL\_Jp4\] for the isotropic case $J_{\parallel}=J_{\perp}=1$.[]{data-label="fig:heisen_3xL"}](heisen_3xL_e0){width="0.8\linewidth"}
![Same as Fig. \[fig:heisen\_3xL\_Jp4\] for the isotropic case $J_{\parallel}=J_{\perp}=1$.[]{data-label="fig:heisen_3xL"}](heisen_3xL_gap){width="0.8\linewidth"}
Moreover, we find that the singlet gap vanishes in the thermodynamic limit as $1/L^2$ (data not shown). This singlet state at momentum $\pi$ along the chains corresponds to the state built in the generalized Lieb-Schultz-Mattis argument [@LSM]. Here, the physical picture is a two-fold degenerate GS due to the appearance of spontaneous dimerization.
[*Spinon dispersion relation for the spin tube :*]{} One of the advantages of this method is to be able to get information on some quantum numbers (number of particles, magnetization, momentum…) For example, the effective Hamiltonian $H_{\rm eff}$ still commutes with translations along the legs, with the total $S^{tot}_z$ and $\tau_z$ so that we can work in a given momentum sector $(k_x,k_y)$ with a fixed magnetization $S^{tot}_z$. By computing the energy in each sector, we can compute the dispersion relation.
![Spinon dispersion relation (see text) as a function of longitudinal momentum (in units of $\pi$). We only plot the lowest branch corresponding to $k_y=\pm 2\pi/3$. The odd lengths run from 5 to 13. The lines are guide to the eyes for an extrapolation on both sides of $\pi/2$.[]{data-label="fig:spinon"}](spinon){width="0.8\linewidth"}
In order to try to identify if the fundamental excitation is a spinon, we compute the energy difference between the lowest $S=1/2$ state when the length is odd ($L=2p+1$) and the extrapolated GS energy obtained from the data on systems with even length $2p$ and $2p+2$. The data are taken from CORE with range-4 approximation. On Fig. \[fig:spinon\], we plot this dispersion as a function of the longitudinal momentum, relative to the GS with $L=2p$.
We observe a dispersion compatible with a spinon-like dispersion, which is massive with a gap at $\pi/2$ $\simeq 0.05 \simeq \Delta_S/2$. This result is consistent with a picture in which the triplet excitation $\Delta_S$ is made of two elementary spinons. With our precision, it seems that the spinons are not bound but we cannot exclude a small binding energy.
We have on overall agreement with results obtained in the strong interchain coupling regime [@cabra98].
Therefore, with CORE method, we have both the advantage of working in the reduced subspace and not being limited to the perturbative regime. Amazingly, we have observed that for a very small effort (solving a small cluster), the effective Hamiltonian gives much better results (often less than 1% on GS energies) than perturbation theory. It also gives an easier framework to systematically improve the accuracy by including longer range interactions.
For these models, the good convergence of CORE results may be due to the fact that the GS in the isotropic limit is adiabatically connected to the perturbative one. In the following part we will therefore study 2D models where a quantum phase transition occurs as one goes from the perturbative to the isotropic regime.
Two dimensional spin models {#sec:2D}
===========================
In this section we would like to discuss the application of the numerical CORE method to two dimensional quantum spin systems. We will present spectra and observables and also discuss a novel diagnostic tool - the density matrix of local objects - in order to justify the truncation of the local state set.
One major problem in two dimension is the more elaborate cluster expansion appearing in the CORE procedure. Especially our approach based on numerical diagonalization of the resulting CORE Hamiltonian faces problems once the CORE interaction clusters wrap around the boundary of the finite size clusters. We therefore try to keep the range of the interactions minimal, but we still demand a reasonable description of low energy properties of the system. We will therefore discuss some ways to detect under what circumstances the low-range approximations fail and why.
![ (a) The plaquette lattice. Full lines denote the plaquette bonds $J$, dashed lines denote the inter-plaquette coupling $J'$. (b) The trimerized lattice. Full lines denote the up-triangle $J$ bonds, dashed lines denote the down-triangle coupling $J'$. The standard lattice is recovered for $J'/J=1$. \[fig:Lattices2D\] ](Lattices2D){width="0.9\linewidth"}
As a first example we discuss the plaquette lattice \[Fig. \[fig:Lattices2D\] (a)\], which exhibits a quantum phase transition from a gapped plaquette-singlet state with only short ranged order to a long range ordered antiferromagnetic state as a function of the interplaquette coupling [@KogaSigmaModelPlaquette; @KogaSeriesPlaquette; @AMLShastry; @VoigtED]. We will show that the CORE method works particularly well for this model by presenting results for the excitation spectra and the order parameter. It is also a nice example of an application where the CORE method is able to correctly describe a quantum phase transition, thus going beyond an augmented perturbation scheme.
The second test case is the highly frustrated lattice \[Fig. \[fig:Lattices2D\] (b)\] with non-integer spin, which has been intensively studied for $S=1/2$ during the last few years [@LeungElser; @Lecheminant; @Waldtmann; @MilaPRL; @MambriniMila]. Its properties are still not entirely understood, but some of the features are well accepted by now: There is no simple local order parameter detectable, neither spin order nor valence bond crystal order. There is probably a small spin gap present and most strikingly an exponentially growing number of low energy singlets emerges below the spin gap. We will discuss a convenient CORE basis truncation which has emerged from a perturbative point of view [@Subrahmanyam; @MilaPRL; @Raghu] and consider an extension of this basis for higher non-integer spin.
Plaquette lattice
-----------------
The CORE approach starts by choosing a suitable decomposition of the lattice and a subsequent local basis truncation. In the plaquette lattice the natural decomposition is directly given by the uncoupled plaquettes. Among the 16 states of an isolated plaquette we retain the lowest singlet \[$K=(0,0)$\] and the lowest triplet \[$K=(\pi,\pi)$\]. The standard argument for keeping these states relies on the fact that they are the lowest energy states in the spectrum of an isolated plaquette.
![ Density matrix weights of the two most important states on a strong ($J$-bonds) plaquette as a function of $J'/J$. These results were obtained by ED with the original Hamiltonian on a $4\times4$ cluster. \[fig:PlaquetteDensityMatrix\] ](PlaquetteDensityMatrix){width="0.9\linewidth"}
As discussed in appendix \[sec:DensityMatrix\], the density matrix of a plaquette in the fully interacting system gives clear indications whether the basis is suitably chosen. In Fig. \[fig:PlaquetteDensityMatrix\] we show the evolution of the density matrix weights of the lowest singlet and triplet as a function of the interplaquette coupling. Even though the individual weights change significantly, the sum of both contributions remains above 90% for all $J'/J\le 1$. We therefore consider this a suitable choice for a successful CORE application.
A next control step consists in calculating the spectrum of two coupled plaquettes, and one monitors which states are targeted by the CORE algorithm. We show this spectrum in Fig. \[fig:2Plaquettes\] along with the targeted states. We realize that the 16 states of our tensor product basis cover almost all the low energy levels of the coupled system. There are only two triplets just below the $S=2$ multiplet which are missed.
![ Low energy spectrum of two coupled plaquettes. The states targeted by the CORE algorithm are indicated by arrows together with their $SU(2)$ degeneracy. \[fig:2Plaquettes\] ](2Plaq){width="0.9\linewidth"}
In a first application we calculate the spin gap for different system sizes and couplings $J'/J$. The results shown in Fig. \[fig:PlaquetteGap\] indicate a reduction of the spin gap for increasing $J'/J$. We used a simple finite size extrapolation in $1/N$ in order to assess the closing of the gap. The extrapolation levels off to a small value for $J'/J\ge0.6$. The appearance of a small gap in this known gapless region is a feature already present in ED calculation of the original model [@VoigtED], and therefore [*not*]{} an artefact of our method. It is rather obvious that the triplet gap is not a very accurate tool to detect the quantum phase transition within our numerical approach. We will see later that order parameter susceptibilities are much more accurate.
![\[fig:PlaquetteGap\] Triplet Gap for effective system sizes between 20 and 52 sites, as a function of the interplaquette coupling $J'/J$. For $J'/J\ge 0.5$ a simple extrapolation in $1/N$ is also displayed. These results compare very well with ED results on the original model \[\]. ](PlaquetteGap){width="0.9\linewidth"}
It is well known that the square lattice $(J'/J=1)$ is Néel ordered. One possibility to detect this order in ED is to calculate the so-called [*tower of excitation*]{}, i.e. the complete spectrum as a function of $S(S+1)$, $S$ being the total spin of an energy level. In the case of standard collinear Néel order a prominent feature is an alignment of the lowest level for each $S$ on a straight line, forming a so called “Quasi-Degenerate Joint States” (QDJS) ensemble [@TowerTriangular], which is clearly separated from the rest of the spectrum on a finite size sample. We have calculated the tower of states within the CORE approach (Fig. \[fig:TowerSquare\]). Due to the truncated Hilbert space we cannot expect to recover the entire spectrum. Surprisingly however the CORE tower of states successfully reproduces the general features observed in ED calculations of the same model [@TowerEDSquare]: (a) a set of QDJS with the correct degeneracy and quantum numbers (in the folded Brillouin zone); (b) a reduced number of magnon states at intermediate energies, both set of states rather well separated from the high energy part of the spectrum. While the QDJS seem not to be affected by the CORE decimation procedure, clearly some the magnon modes get eliminated by the basis truncation.
![\[fig:TowerSquare\] Tower of states obtained with a range-2 CORE Hamiltonian on an effective $N=36$ square lattice (9-site CORE cluster) in different reduced momentum sectors. The tower of states is clearly separated from the decimated magnons and the rest of the spectrum. ](TowerSquare36){width="0.8\linewidth"}
In order to locate the quantum phase transition from the paramagnetic, gapped regime to the Néel ordered phase, a simple way to determine the onset of long range order is desireable. We chose to directly couple the order parameter to the Hamiltonian and to calculate generalized susceptibilities by deriving the energy with respect to the external coupling. This procedure is detailed in appendix \[sec:ObservablesNumericalCoreMethod\]. Its simplicity relies on the fact that only eigenvalue runs are necessary. Similar approaches have been used so far in ED and QMC calculations [@calandra00; @capriotti].
![ Staggered moment per site as a function of the rescaled applied staggered field for the plaquette lattice and different values of $J'/J$. Circles denote the approximate crossing point of curves for different system sizes. We take the existence of this crossing as a phenomenological indication for the presence of Néel LRO. In this way the phase transition is detected between $0.5<J'_c/J<0.6$, consistent with previous estimates. The arrows indicate curves for increasing system sizes. \[fig:MstaggeredPlaquette\] ](MstaggeredJp){width="0.9\linewidth"}
Our results in Fig. \[fig:MstaggeredPlaquette\] show the evolution of the staggered moment per site in a rescaled external staggered field for different inter-plaquette couplings $J'$ and different system sizes (up to $8\times8$ lattices). We note the appearance of an approximate crossing of the curves for different system sizes, once Néel LRO sets in. This approximate crossing relies on the fact that the slope of $m_L(hN)$ diverges with increasing $N$ in the ordered phase in our case [@capriotti]. We then consider this crossing feature as an indication of the phase transition and obtain a value of the critical point $J_c/J = 0.55 \pm 0.05$. This estimate is in good agreement with previous studies using various methods [@KogaSeriesPlaquette; @AMLShastry; @VoigtED]. We have checked the present approach by performing the same steps on the two leg ladder discussed in section \[sec:LadderGeometries\] and there was no long range magnetic order present, as expected.
systems with half-integer spins
-------------------------------
In the past 10 years many efforts have been devoted to understand the low energy physics of the antiferromagnet (KAF) for spins $1/2$ [@LeungElser; @Lecheminant; @Waldtmann; @MilaPRL; @MambriniMila]. At the theoretical level, the main motivation comes from the fact that this model is the only known example of a two-dimensional Heisenberg spin liquid. Even though many questions remain open, some very exciting low-energy properties of this system have emerged. Let us summarize them briefly: (i) the GS is a singlet ($S=0$) and has no magnetic order. Moreover no kind of more exotic ordering (dimer-dimer, chiral order, etc.) have been detected using unbiased methods; (ii) the first magnetic excitation is a triplet ($S=1$) separated from the GS by a rather small gap of order $J/20$; (iii) more surprisingly the spectrum appears as a continuum of states in all spin sectors. In particular the spin gap is filled with an exponential number of singlet excitations: ${\cal N}_{\rm singlets} \sim
1.15^N$; (iv) the singlet sector of the KAF can be very well reproduced by a short-range resonating valence bond approach involving only nearest-neighbor dimers.
From this point of view, the spin $1/2$ KAF with its highly unconventional low-energy physics appears to be a very sharp test of the CORE method. The case of higher half-integer spins $S=3/2, 5/2 \dots$ KAF is also of particular interest, since it is covered by approximative experimental realizations [@SCGO]. Even if some properties of these experimental systems are reminiscent of the spin $1/2$ KAF theoretical support is still lacking for higher spins due to the increased complexity of these models.
In this section we discuss in detail the range-two CORE Hamiltonians for spin 1/2 and 3/2 KAF considered as a set of elementary up-triangles with couplings $J$, coupled by down-triangles with couplings $J'$ \[see Fig. \[fig:Lattices2D\] (b)\]. The coupling ratio will be denoted by $\alpha=J'/J$. Before going any further into the derivation of the CORE effective Hamiltonian let us start with the conventional degenerate perturbation theory results. Note that in the perturbative regime these two approaches yield the same effective Hamiltonian.
As described in Appendix \[sec:TechnicalKagome\], the most general two-triangle effective Hamiltonian involving only the two spin 1/2 degrees of freedom on each triangle can be written in the following form: $$\begin{aligned}
\label{eq:general}
{\cal H} =
N a_0(\alpha)+
\sum_{\langle i,j \rangle}&& (
b_0(\alpha) \vec{\tau}_i.\vec{e}_{ij} \vec{\tau}_j.\vec{e}_{ij}
\label{eqn:GeneralTriangleCOREHamiltonian}\\
&&+ a_1(\alpha) \vec{\sigma}_i . \vec{\sigma}_j\nonumber \\
&&+ b_1(\alpha) \vec{\sigma}_i . \vec{\sigma}_j (\vec{\tau}_i.\vec{e}_{ij})
(\vec{\tau}_j.\vec{e}_{ij})\nonumber\\
&&+ c_1(\alpha) \vec{\sigma}_i . \vec{\sigma}_j
(\vec{\tau}_i.\vec{e}_{ij}+\vec{\tau}_j.\vec{e}_{ij}) )\nonumber.\end{aligned}$$
In the spirit of Mila’s approach [@MilaPRL] for spin $1/2$ the first order perturbative Hamiltonian in $\alpha$ can easily be extended to arbitrary half-integer spin $S$: $$\begin{aligned}
{\cal H}^{\rm pert.} &=& \frac{\alpha}{9} \vec{\sigma}_i . \vec{\sigma}_j
\label{eqn:PerturbativeTriangleCOREHamiltonian}\\
&&
\times \left ( 1-2(2S+1)\vec{\tau}_i.\vec{e}_{a} \right ) \left (
1-2(2S+1)\vec{\tau}_j . \vec{e}_{b} \right )\nonumber\end{aligned}$$ and the coefficients of (\[eq:general\]) in the perturbative limit are given as $a_1(\alpha)= \frac{\alpha}{9}$, $b_1(\alpha)= \frac{4\alpha}{9} (2S+1)^2$, $c_1(\alpha)=-\frac{2\alpha}{9} (2S+1)$ and $a_0(\alpha)=b_0(\alpha)=0$.
### Choice of the CORE basis
As discussed in the previous paragraph we keep the two degenerate $S=1/2$ doublets on a triangle for the CORE basis. In analogy to the the plaquette lattice we calculate the density matrix of a single triangle embedded in a 12 site lattice for both spin $S=1/2$ and $S=3/2$, in order to get information on the quality of the truncated basis. The results displayed in Fig. \[fig:TriangleDensityMatrix\] show two different behaviors: while the targeted states exhaust 95% for the $S=1/2$ case, they cover only $\approx$ 55% in the $S=3/2$ case. This can be considered a first indication that the range-two approximation in this basis might break down for $S>1/2$ half integer spin, while the approximation seems to work particularly well for $S=1/2$, thereby providing independent support for the adequacy of the basis chosen in a related mean-field study [@MilaPRL].
![ Density matrix weights of the different total spin states in a triangle of a 12 site cluster with $S=1/2$ and $S=3/2$ spins. These results are obtained for the homogeneous case $\alpha=1$. \[fig:TriangleDensityMatrix\] ](TriangleDMOneAndThreeHalf){width="0.9\linewidth"}
We continue the analysis of the CORE basis by monitoring the evolution of the spectra of two coupled triangles in the geometry (c.f. Fig. \[fig:TwoTriangles\]) as a function of the inter-triangle coupling $J'$, as well as the states selected by the range-two CORE algorithm. The spectrum for the spin $S=1/2$ case is shown in Fig. \[fig:SpectrumTwoTrianglesOneHalf\]. We note the presence of a clear gap between the 16 lowest states – correctly targeted by the CORE algorithm – and the higher lying bands. This can be considered an ideal case for the CORE method. Based on this and the results of the density matrix we expect the CORE range-two approximation to work quite well.
![\[fig:SpectrumTwoTrianglesOneHalf\] Spectrum of two coupled triangles in the geometry with $S=1/2$ spins. The entire lowest band containing 16 states is successfully targeted by the CORE algorithm. ](2TOneHalf){width="0.9\linewidth"}
We compare these encouraging results with the spectrum for the spin $S=3/2$ case displayed in Fig. \[fig:SpectrumTwoTrianglesThreeHalf\]. Here the situation is less convincing: very rapidly $(J'/J \gtrsim 0.45)$ the low energy states mix with originally higher lying states and the CORE method continues to target two singlets which lie high up in energy when reaching $J'/J=1$. We expect this to be a situation where the CORE method will probably not work correctly when restricted to range-two terms only.
![\[fig:SpectrumTwoTrianglesThreeHalf\] Spectrum of two coupled triangles in the geometry with $S=3/2$ spins. The 16 states targeted by the CORE algorithm are indicated by the arrows and their degeneracies. ](2TThreeHalf){width="0.9\linewidth"}
Based on the two-triangle spectra shown above we used the CORE algorithm to determine the coefficients of the general two-body Hamiltonian \[Eqn. (\[eqn:GeneralTriangleCOREHamiltonian\])\]. For an independent derivation, see Ref. \[\]. The coefficients obtained this way are shown in Figs. \[fig:CoreKagome12\] and \[fig:CoreKagome32\] for $S=1/2$ and $S=3/2$ respectively.
![\[fig:CoreKagome12\] Coefficients of the CORE range-two Hamiltonian for two coupled $S=1/2$ triangles. The coefficients in panel (iii) have been divided by their values in the perturbative limit.](CoreKagome12){width="0.9\linewidth"}
![\[fig:CoreKagome32\] Coefficients of the CORE range-two Hamiltonian for two coupled $S=3/2$ triangles. The coefficients in panel (iii) have been divided by their values in the perturbative limit.](CoreKagome32){width="0.9\linewidth"}
In the limit $\alpha \ll 1$ the coefficients can be obtained from the perturbative Hamiltonian \[Eqn. (\[eqn:PerturbativeTriangleCOREHamiltonian\])\]. There are two classes of coefficients in both cases: $a_0$ and $b_0$ are zero in the perturbative limit, i.e. they are at least second order in $\alpha$. The second class of coefficients ($a_1$, $b_1$, $c_1$) are linear in $\alpha$. For improved visualisation we have divided all the coefficients in the second class by their perturbative values. In this way we observe in Fig. \[fig:CoreKagome12\] that coefficients $b_1$ and $c_1$ change barely with respect to their values in the perturbative limit. However $a_1$ has a significant subleading contribution, which leads to a rather large reduction upon reaching the $\alpha=1$ point. It does however not change sign.
The situation for the $S=3/2$ case in Fig. \[fig:CoreKagome32\] is different: while the coefficients $b_1$ and $c_1$ decrease somewhat, it is mainly $a_1$ which changes drastically as we increase $\alpha$. Starting from 1 it rapidly goes through zero ($\alpha\approx0.07$) and levels off to roughly -7 times the value predicted by perturbation theory as one approaches $\alpha=1$. In this case it is rather obvious that this coefficient will dominate the effective Hamiltonian. We will discuss the implications of this behavior in the application to the $S=3/2$ magnet below.
Let us note that the behavior of the $a_1$ coefficient is mainly due to a rather large second order correction in perturbation theory. Indeed we find good agreement with the values obtained in the perturbative approach of Ref. \[\].
### Simulations for $S=1/2$
After having studied the CORE basis and the effective Hamiltonian at range two in some detail, we now proceed to the actual simulations of the resulting model. We perform the simulations for the standard lattice, therefore $\alpha=1$. We will calculate several distinct physical properties, such as the tower of excitations, the evolution of the triplet gap as a function of system size and the scaling of the number of singlets in the gap. These quantities have been discussed in great detail in previous studies of the $S=1/2$ antiferromagnet [@LeungElser; @Lecheminant; @Waldtmann; @MilaPRL; @MambriniMila].
First we calculate the tower of excitations for a $S=1/2$ system on a 27 sites sample. The data is plotted in Fig. \[fig:TowerKagome\]. The structure of the spectrum follows the exact data of Ref. \[\] rather closely; i.e there is no QDJS ensemble visible, a large number of $S=1/2$ states covering all momenta are found below the first $S=3/2$ excitations and the spectrum is roughly bounded from below by a straight line in $S(S+1)$. Note that the tower of states we obtain here is strikingly different from the one obtained in the Néel ordered square lattice case, see Fig. \[fig:TowerSquare\].
![\[fig:TowerKagome\] Tower of states obtained with a range-two CORE Hamiltonian on an effective $N=27$ lattice (9-site CORE cluster). There is a large number of low-lying states in each $S$ sector. The symbols correspond to different momenta. ](TowerKagome27){width="0.8\linewidth"}
Next we calculate the spin gap using the range-two CORE Hamiltonian. Results for system sizes up to 48 sites are shown in Fig. \[fig:CoreOneHalfGap\], together with ED data where available. In comparison we note two observations: (a) the CORE range-two approximation seems to systematically overestimate the gap, but captures correctly the sample to sample variations. (b) the gaps of the smallest samples (effective $N$=12,15) deviate strongly from the exact data. We observed this to be a general feature of very small clusters in the CORE approach. In order to improve the agreement with the ED data we calculated the two CORE range-three terms containing a closed loop of triangles. The results obtained with this extended Hamiltonian are shown as well in Fig. \[fig:CoreOneHalfGap\]. These additional terms improve the gap data somewhat. We now find the CORE gaps to be mostly smaller than the exact ones. The precision of the CORE gap data is not accurate enough to make a reasonable prediction on the spin gap in the thermodynamic limit. However we think that the CORE data is compatible with a finite spin gap.
![\[fig:CoreOneHalfGap\] Spin gap of the $S=1/2$ model on various samples, obtained with the CORE method (range-two and three). Exact diagonalization result are also shown for comparison where available.](CoreOneHalfGap){width="0.80\linewidth"}
Finally we determine the number of nonmagnetic excitations within the magnetic gap for a variety of system sizes up to 39 sites. Similar studies of this quantity in ED gave evidence for an exponentially increasing number of singlets in the gap [@Lecheminant; @Waldtmann]. We display our data in comparison to the exact results in Fig. \[fig:KagomeOneHalfSinglets\]. While the precise numbers are not expected to be recovered, the general trend is well described with the CORE results. For both even and odd $N$ samples we see an exponential increase of the number of these nonmagnetic states. In the case of $N=39$ for example, we find 506 states below the first magnetic excitation. These results emphasize again the validity of the two doublet basis for the CORE approach on the spin 1/2 system.
![\[fig:KagomeOneHalfSinglets\] Logarithm of the number of states within the magnetic gap. Results obtained with the CORE range-two Hamiltonian. For comparison exact data obtained in Refs. \[,\] are shown. The dashed lines are linear fits to the exact diagonalization data. ](KagomeSinglets){width="0.75\linewidth"}
### Simulations for $S=3/2$
We have also simulated the CORE Hamiltonian obtained above for $S=3/2$. While the energy per site is reproduced roughly, unfortunately the spectrum does not resemble an antiferromagnetic spin model, i.e. the groundstate is polarized in the spin variables. This fact is at odds with preliminary exact diagonalization data on the original $S=3/2$ model [@KagThreeHalf]. We therefore did not pursue the CORE study with this choice of the basis states any further. Indeed, as suggested by the analysis of the density matrix and by the evolution of the spectrum of two coupled triangles, we consider this a breakdown example of a naive range-two CORE approximation. It is important to stress that the method indicates its failure in various quantities throughout the algorithm, therefore offering the possibility of detecting a possible breakdown.
As a remedy in the present case we have extended the basis states to include all the $S=1/2$ and $S=3/2$ states on a triangle (i.e. keeping 20 out of 64 states). Computations within this basis set are more demanding, but give a better agreement with the exact diagonalization results. At the present stage we cannot decide whether the breakdown of the 4 states CORE basis is related only the CORE method or whether it implies that the $S=1/2$ and $S=3/2$ systems do not belong to the same phase.
Conclusions
===========
We have discussed extensively the use of a novel numerical technique - the so-called numerical Contractor Renormalization (CORE) method - in the context of low-dimensional quantum magnetism. This method consists of two steps: (i) building an effective Hamiltonian acting on the low-energy degrees of freedom of some elementary block; and (ii) studying this new model numerically on finite-size clusters, using a standard Exact Diagonalization or similar approach.
Like in other real-space renormalization techniques the effective model usually contains longer range interactions. The numerical CORE procedure will be most efficient provided the effective interactions decay sufficiently fast. We discussed the validity of this assumption in several cases.
For ladder type geometries, we explicitely checked the accuracy of the effective models by increasing the range of the effective interactions until reaching convergence. Both in the perturbative regime and in the isotropic case, our results on a 2-leg ladder and a 3-leg torus are in good agreement with previously established results. This rapid convergence might be due to the small correlation length that exists in these systems which both have a finite spin gap.
In two dimensions, we have used the density matrix as a tool to check whether the restricted basis gives a good enough representation of the exact states. When this is the case, as for the plaquette lattice or the $S=1/2$ lattice, the lowest order range-two effective Hamiltonian gives semi-quantitative results, even away from any perturbative regime. For example we can successfully describe the plaquette lattice, starting from the decoupled plaquette limit through the quantum phase transition to the Néel ordered state at homogeneous coupling. Furthermore we can also reproduce many aspects of the exotic low-energy physics of the $S=1/2$ lattice.
Therefore within the CORE method, we can have both the advantage of working in a strongly reduced subspace and not being limited to the perturbative regime in certain cases.
We thus believe that the numerical CORE method can be used systematically to explore possible ways of generating low-energy effective Hamiltonians. An important field is for example the doped frustrated magnetic systems, where it is not easy to decide which states are important in a low-energy description, and therefore the density matrix might be a helpful tool.
Density Matrix {#sec:DensityMatrix}
==============
In this appendix we introduce the density matrix of a basic building block in a larger cluster of the fully interacting problem as a diagnostic tool to validate or invalidate a particular choice of retained states on the basic building block in the CORE approach.
In previous applications of the CORE method, the choice of the states kept relied mostly on the spectrum of an isolated building block. While this usually gives reasonable results it is not a clear [*a priori* ]{} where to place the cut-off in the spectrum.
The density matrix of a “system block” embedded in a larger “super block” forms a key concept in the Density Matrix Renormalization Group (DMRG) algorithm invented by S.R. White in 1992 [@DMRG] and is at the heart of its success. Based on this and related ideas [@DMPhonons] we propose to monitor the density matrix of the basic building block embedded in a larger cluster and to retain these states exhausting a large fraction of the density matrix weight.
Consider now a subsystem $\mathcal{A}$ embedded in a larger system $\mathcal{B}$. Suppose that the overall system $\mathcal{B}$ is in state $|\Psi\rangle$ (e.g. the groundstate). We write the wavefunction as: $$|\Psi\rangle= \sum_{a,b}\ \psi_{a,b}\ |a\rangle\otimes|b\rangle,$$ where the sum index $a$ runs over all states in $\mathcal{A}$ and index $b$ over all states in $\mathcal{B} \setminus \mathcal{A}$. The density matrix $\rho^\mathcal{A}$ of the subsystem $\mathcal{A}$ is then defined as $$\rho^\mathcal{A}_{a,a'}=\sum_{b} \psi_{a,b} \psi_{a',b}^*$$ The eigenvalues of $\rho^\mathcal{A}$ denote the probability of finding a certain state $a$ in $\mathcal{A}$, given the overall system in state $|\Psi\rangle$.
Practically we calculate the groundstate of the fully interacting system on a medium size cluster by exact diagonalization, and then obtain the density matrix of a basic building block, e.g. a four site plaquette. The density matrix of a building block is a rather local object, so we expect that results on intermediate size clusters are already accurate on the percent level. The density matrix spectra shown in Figs. \[fig:PlaquetteDensityMatrix\] and \[fig:TriangleDensityMatrix\] have been obtained in this way. In the models considered, a density matrix weight of the retained states of at least 90% yielded reasonable results within a range-two CORE approximation. It is possible to allow for a lower overall weight, at the expense of increasing the range of the CORE interactions.
Observables in the numerical CORE method {#sec:ObservablesNumericalCoreMethod}
========================================
The calculation of observables beyond simple energy related quantities is not straightforward within the CORE method, as the observables need to be renormalized like the Hamiltonian in the first place [@core-bosonfermion; @dynamicCORE].
A somewhat simpler approach for measurements of symmetry breaking order parameters consists in adding a small symmetry breaking field to the Hamiltonian (for a review see Ref. \[\]).
Let us denote ${\cal \hat O}$ the extensive symmetry breaking operator, such that the order parameter is related to its GS average value $m=1/N \langle \psi_0|{\cal \hat O}|\psi_0\rangle$. The occurence of a symmetry broken phase can be detected by adding this operator to the Hamiltonian : $${\cal H}(\delta) ={\cal H} - \delta {\cal \hat O}$$ Since on a finite-size lattice the order parameter vanishes by symmetry for $\delta=0$, the ground-state energy per site varies quadratically for small $\delta$ : $$e(\delta)\simeq e_0 -\frac{1}{2}\chi_0 \delta^2,$$ where $\chi_0$ is termed the corresponding generalized susceptibility. In that way the second derivative of the energy with respect to $\delta$ at $\delta=0$ offers one possibility to detect a finite order parameter in the thermodynamic limit [@capriotti].
We found that another possibility to conveniently track the presence of a finite order parameter is to measure directly $m(\delta)$ in finite field $$m(\delta)=\langle\Psi_\delta|{\cal \hat O}|\Psi_\delta\rangle=
\mbox{d}e(\delta)/\mbox{d}\delta$$ by the Hellmann-Feynman theorem. When plotting $m(\delta)$ as a function of the rescaled field $N \delta$ for various system sizes we observe an approximate crossing of the curves if there is a finite order parameter and no crossing in the absence of the order parameter.
Gauge invariance on half-integer spins like systems {#sec:TechnicalKagome}
===================================================
In this appendix, we discuss half-integer spin Hamiltonians with triangles as the unit cell. The ground state manifold of each unit cell is generated by the four degenerate lowest states that can be built out of $3$ half-integer $S$ spins, namely the four $S_{\rm tot}=1/2$ states. The idea of selecting these states as a starting point to describe the whole system low energy properties was originally introduced by Subrahmanyam for $S=1/2\ $ [@Subrahmanyam] on the lattice and later used by Mila [@MilaPRL]. More recently it was reintroduced by Raghu [*et al*]{} [@Raghu] for arbitrary half-integer $S$ in the context of a chain of triangles. All these approaches are pertubative and state that the triangle couplings $J$ is much larger than the inter-triangle one $J'$.
Here we would like to discuss some general properties of any effective Hamiltonian that can be derived either by perturbative methods or more sophisticated ones such as CORE. In particular, we would like to point out that a gauge invariance appears as a direct consequence of the state selection.
![\[fig:OneTriangle\] Definition of chirality $\varepsilon$ (see text for details).](OneTriangle){width="0.7\linewidth"}
To be more specific, let us label $1$,$2$,$3$ the sites of the triangle (see Fig. \[fig:OneTriangle\]). In order to build a total spin $1/2$ out of the three $S$, spins $2$ and $3$ couple into a $S+\varepsilon(1/2)$ with $\varepsilon=\pm 1$. The coupling with the remaining site $1$ produces a spin $1/2$ with chirality $\varepsilon=\pm 1$. Note that this definition of chirality is equivalent to Eqs. (\[eqn:TorusChirality\]) for spin $S=1/2$ up to a global unitary transform which is just a redefinition of the chirality quantification axis.
In the following, the four selected spin-chirality states on a triangle $i$ will be denoted as $\vert \vert \varepsilon_i , \nu_i \rangle \rangle$. These states are the eigenstates of the $z$ components of spin $\vec{\sigma}$ and chirality $\vec{\tau}$ (both are spin 1/2 like operators) with $\tau_z \vert \vert
\varepsilon_i , \nu_i \rangle \rangle = (\varepsilon_i/2) \vert \vert
\varepsilon_i , \nu_i \rangle \rangle$ and $\sigma_z \vert \vert
\varepsilon_i , \nu_i \rangle \rangle = \nu_i \vert \vert
\varepsilon_i , \nu_i \rangle \rangle$.
Let us now turn to the two-triangle problem. As it can be seen in Fig. \[fig:TwoTriangles\], the Hamiltonian is invariant under reflections with respect to the $(xx')$ axis. Moreover, the reflection can be taken independently on each triangle. As a consequence, both chiralities ($\tau_i^z$ and $\tau_j^z$) are conserved by the effective Hamiltonian and the $\tau$ part is of the form $1+a
(\tau_i^z+\tau_j^z)+ b \tau_i^z \tau_j^z$. For any fixed value of $(\varepsilon_i,\varepsilon_j)$, the total spin of the system is conserved and thus the spin part is $SU(2)$ invariant. As a conclusion the most general two-triangle Hamiltonian allowed is of the form: $$(\vec{\sigma}_i \cdot \vec{\sigma}_j + c )(1+a
(\tau_i^z+\tau_j^z)+ b \tau_i^z \tau_j^z).$$
![\[fig:TwoTriangles\] The two-triangle problem. $\alpha$ is the coupling ratio $J'/J$.](TwoTriangles){width="0.7\linewidth"}
[*Gauge transformation:*]{} The form of the above Hamiltonian is the consequence of the particular choice we made for labeling the sites of the triangle (see Fig. \[fig:TwoTriangles\]): site $1$ of triangle $t_1$ couples to site $1$ of triangle $t_2$. Although this gauge was convenient for the calculation, in general this choice can not be made simultaneously on all couples of triangles of the lattice. So, it is essential to derive the form of the Hamiltonian in a generic situation where site $i=1,2,3$ of triangle $t_1$ couples to site $j=1,2,3$ of triangle $t_2$.
The unitary transformations involved in the redefinition of the coupling sequence (see Fig. \[fig:3j\_symbols\]) are covered by the $3j$ symbols of elementary quantum mechanics. The problem of 3 half-integer spins $S$ coupled into a total spin $1/2$ occurs to be particularly simple and independent of $S$. The form of the general effective Hamiltonian then reads: $$\begin{aligned}
{\cal H}_{ij}^{a,b}(\alpha) &=& \left ( \vec{\sigma}_i \cdot \vec{\sigma}_j +
c(\alpha) \right )\\
&& \times
[ 1+a(\alpha)(\vec{\tau}_i\cdot\vec{e}_{a}+\vec{\tau}_j\cdot\vec{e}_{b})\\
& &\quad + b(\alpha)
(\vec{\tau}_i\cdot\vec{e}_{a})(\vec{\tau}_j\cdot\vec{e}_{b})],\end{aligned}$$ where $\vec{e}_a$, $a=1,2,3$ are three coplanar normalized vectors in a $120^\circ$ configuration (for example, $\vec{e}_1=(0,1)$, $\vec{e}_2=(-\sqrt{3}/2,-1/2)$ and $\vec{e}_3=(\sqrt{3}/2,-1/2)$ in the $x-z$ plane) and $a$, $b$ are the labels of the original spins coupling triangles $t_i$ and $t_j$.
[*The lattice:*]{} In the particular geometry of the lattice \[see Fig. \[fig:Lattices2D\] (b)\], each triangular unit cell is coupled to six other triangular cells, each corner being coupled twice. As a consequence, for each cell the contribution involving only $\vec{\tau}_i\cdot\vec{e}_{\alpha}$ factorizes into $2 \vec{\tau}_i \cdot (\vec{e}_1+\vec{e}_2+\vec{e}_3) = 0$. The corresponding terms are then not relevant in the Hamiltonian and thus we denote the most general two-triangle Hamiltonian for the lattice as: $$\begin{aligned}
{\cal H}^ =
N a_0(\alpha)+
\sum_{\langle i,j \rangle}&& [
b_0(\alpha) \vec{\tau}_i\cdot\vec{e}_{ij} \vec{\tau}_j\cdot\vec{e}_{ij}\\
&&+ a_1(\alpha) \vec{\sigma}_i \cdot \vec{\sigma}_j \\
&&+ b_1(\alpha) \vec{\sigma}_i \cdot \vec{\sigma}_j (\vec{\tau}_i\cdot\vec{e}_{ij})
(\vec{\tau}_j\cdot\vec{e}_{ij})\\
&&+ c_1(\alpha) \vec{\sigma}_i \cdot \vec{\sigma}_j
(\vec{\tau}_i\cdot\vec{e}_{ij})+(\vec{\tau}_j\cdot\vec{e}_{ij}) ]\end{aligned}$$ which is the form used in the text.
![\[fig:3j\_symbols\] Three ways of coupling the three spins $S$ on a triangle into a total spin $1/2$ state. Each construction is related to the two others by the 3$j$ symbols (see text). ](3j_symbols){width="0.9\linewidth"}
We thank F. Alet, A. Auerbach, F. Mila and D. Poilblanc for fruitful discussions. Furthermore we are grateful to F. Alet for providing us QMC data. We thank M. Körner for his very useful `Mathematica` spin notebook. A.L. acknowledges support from the Swiss National Fund. We thank IDRIS (Orsay) and the CSCS Manno for allocation of CPU time.
C.J. Morningstar and M. Weinstein, Phys. Rev. Lett. [**73**]{}, 1873 (1994); C.J. Morningstar and M. Weinstein, Phys. Rev. D [**54**]{} 4131 (1996). M. Weinstein, Phys. Rev. B [**63**]{} 174421 (2001). E. Altman and A. Auerbach, Phys. Rev. B [**65**]{}, 104508 (2002). E. Berg, E. Altman, and A. Auerbach, Phys. Rev. Lett. [**90**]{}, 147204 (2003). J. Piekarewicz and J.R. Shepard, Phys. Rev. B [**56**]{}, 5366 (1997). J. Piekarewicz and J.R. Shepard, Phys. Rev. B [**57**]{}, 10260 (1998). S. Capponi and D. Poilblanc, Phys. Rev. B [**66**]{}, 180503(R) (2002). J.-P. Malrieu and N. Guihéry, Phys. Rev. B [**63**]{}, 085110 (2001). E. Dagotto and T. M. Rice, Science [**271**]{}, 618 (1996) and references therein. T. Barnes, E. Dagotto, J. Riera, and E. S. Swanson, Phys. Rev. B [**47**]{}, 3196 (1993); S. R. White, R. M. Noack, and D. J. Scalapino, Phys. Rev. Lett. [**73**]{}, 886 (1994); B. Frischmuth, B. Ammon, and M. Troyer, Phys. Rev. B [**54**]{}, R3714 (1996). M. Greven, R. J. Birgeneau, and U.-J. Wiese, Phys. Rev. Lett. [**77**]{}, 1865 (1996). K. Kawano and M. Takahashi, J. Phys. Soc. Jpn [**66**]{}, 4001 (1997). D.C. Cabra, A. Honecker and P. Pujol, Phys. Rev. B [**58**]{}, 6241 (1998). , edited by T. Martin, G. Montambaux, and J. Trân Thanh Vân, Editions Frontières, Gif-sur-Yvette, France, 1996 (cond-mat/9605075). E. Lieb, T. Schultz, D. Mattis, Ann. Phys. [**16**]{}, 407 (1961); I. Affleck, Phys. Rev. B [**37**]{}, 5186 (1988). A. Koga, S. Kumada and N. Kawakami, J. Phys. Soc. Jpn. [**68**]{}, 642 (1999). A. Koga, S. Kumada and N. Kawakami, J. Phys. Soc. Jpn. [**68**]{}, 2373 (1999). A. Läuchli, S. Wessel and M. Sigrist, Phys. Rev. B [**66**]{}, 014401 (2002). A. Voigt, Computer Physics Communication [**146**]{}, 125 (2002). P.W. Leung and V. Elser, Phys. Rev. B [**47**]{}, 5459 (1993). P. Lecheminant , Phys. Rev. B [**56**]{}, 2521 (1997). C. Waldtmann , Eur. Phys. J. B [**2**]{}, 501 (1998). F. Mila, Phys. Rev. Lett. [**81**]{}, 2356, (1998). M. Mambrini and F. Mila, Eur. Phys. J. B [**17**]{}, 651 (2000). V. Subrahmanyam, Phys. Rev. B [**52**]{}, 1133 (1995). C. Raghu, I. Rudra, S. Ramasesha and D. Sen, Phys. Rev. B [**62**]{}, 9484 (2000). B. Bernu, C. Lhuillier, and L. Pierre, Phys. Rev. Lett. [**69**]{}, 2590 (1992). P. Sindzingre, C. Lhuillier, J.B. Fouet, Int. J. Mod. Phys B [**17**]{} 5031 (2003); (cond-mat/0110283). M. Calandra and S. Sorella, Phys. Rev. B [**61**]{}, R11894 (2000). L. Capriotti, Int. J. Mod. Phys. B [**15**]{}, 1799 (2001). L. Limot, P. Mendels, G. Collin, C. Mondelli, B. Ouladdiaf, H. Mutka, N. Blanchard, and M. Mekata, Phys. Rev. B [**65**]{}, 144447 (2002), and references therein. R. Budnik and A. Auerbach, unpublished; R. Budnik, M.Sc. thesis (Technion, Haifa). S. Dommange, A. Läuchli, J.-B. Fouet, B. Normand and F. Mila, unpublished. S.R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992). C. Zhang, E. Jeckelmann, and S.R. White, Phys. Rev. Lett. [**80**]{}, 2661 (1998).
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'Superconducting persistent-current qubits are quantum-coherent artificial atoms with multiple, tunable energy levels. In the presence of large-amplitude harmonic excitation, the qubit state can be driven through one or more of the constituent energy-level avoided crossings. The resulting Landau-Zener-Stückelberg (LZS) transitions mediate a rich array of quantum-coherent phenomena. We review here three experimental works based on LZS transitions: Mach-Zehnder-type interferometry between repeated LZS transitions, microwave-induced cooling, and amplitude spectroscopy. These experiments exhibit a remarkable agreement with theory, and are extensible to other solid-state and atomic qubit modalities. We anticipate they will find application to qubit state-preparation and control methods for quantum information science and technology.'
author:
- 'W. D. Oliver'
- 'S. O. Valenzuela'
title: |
Large-amplitude driving of a superconducting artificial atom:\
Interferometry, cooling, and amplitude spectroscopy
---
Introduction {#intro}
============
Superconducting qubits are solid-state artificial atoms, comprising lithographically defined Josephson tunnel junctions and superconducting interconnects. When cooled to milli-Kelvin temperatures, these qubits exhibit quantized states of charge, flux, or junction phase depending on the circuit design parameters [@Clarke88; @Clarke08a]. Associated with these quantized states is a spectrum of energy levels, tunable via an external control parameter, e.g., an applied electric or magnetic field. Although generally only the lowest two energy eigenstates are utilized for quantum information science applications, the energy spectrum indeed extends to higher-energy levels corresponding to higher-excited states of the circuit. The separation between pairs of energy levels typically falls in the radio frequency and microwave regimes, and resonantly driving the artificial atom with a harmonic field can couple and induce quantum-state transitions.
Due to their relatively large size, superconducting artificial atoms can be strongly coupled to their external control fields. It is this feature, along with their quantum coherence, that we utilize in the present article. A large-amplitude harmonic control field can drive an artificial atom throughout its energy-level spectrum. When driven through an avoided level crossing, a Landau-Zener-Stückelberg (LZS) transition occurs. This is a coherent process akin to a beamsplitter for photons, taking an input state of the atom and outputting a superposition of states. Repeated passages through an avoided crossing act as an atom interferometer, causing the atomic superposition state to interfere quantum mechanically with itself. Since the weighting of the superposition state depends sensitively on the size of the avoided crossing and the velocity (change in relative energy between levels per unit time) with which it is traversed, the quantum interference reflects the energy spectrum of the artificial atom. In turn, the quantum interference can be leveraged to facilitate non-adiabatic quantum control.
We begin this article with an introduction to the superconducting persistent-current qubit [@Mooij99a; @Orlando99a] and an overview of Landau-Zener-Stückelberg (LZS) transitions. We then present three experimental works that utilize LZS transitions in a strongly-driven persistent-current qubit. The first is Mach-Zehnder-type interferometry between repeated LZS transitions [@Oliver05a], for which we observed quantum interference fringes in n-photon transition rates [@Berns06a], with n=1 …50. The second is microwave-induced cooling [@Valenzuela06a], by which we achieved effective qubit temperatures less than 3 mK, a factor 10-100 times lower than the environmental temperature. The third is amplitude spectroscopy [@Berns08a; @Rudner08a], a spectroscopy technique that monitors the system response to amplitude rather than frequency. Amplitude spectroscopy allowed us to probe the energy spectra of our artificial atom from 0.01 - 120 GHz, while driving it at a fixed frequency near 0.2 GHz. Finally, we conclude by considering the application of LZS transitions to quantum information science and technology.
Persistent-Current Qubit: Superconducting Artificial Atom {#sec:PCQB}
=========================================================
Superconducting artificial atoms exhibit a high degree of quantum coherence, and there have been numerous proposals and demonstrations of quantum phenomena in these systems, many derived from the fields of atomic physics and quantum optics. A few examples include: coherent superpositions of macroscopic states [@Friedman00a; @Wal00a; @Berkley03a], Rabi oscillations [@Nakamura99a; @Nakamura01; @Vion02a; @Yu02a; @Martinis02a; @Chiorescu03a; @Claudon04a; @Plourde05a; @Saito06a; @Lisenfeld07a], Landau-Zener transitions [@Izmalkov04a], Stückelberg oscillations [@Oliver05a; @Berns06a; @Sillanpaa06a; @Wilson07a; @Izmalkov08a], microwave cooling [@Valenzuela06a; @Niskanen07b; @You08a; @Grajcar08a], electromagnetically induced transparency [@Murali04a; @Dutton06a], geometrical phase [@Leek07a], and cavity quantum electrodynamics [@Chiorescu04a; @Wallraff04a; @Johansson06a; @Deppe08a; @Fink08a; @Fragner08a; @Hofheinz08a]. Significant progress has also been made toward their application to quantum information science [@Makhlin01a; @Mooij05], including state initialization [@Valenzuela06a], tunable [@Hime06a; @Ploeg07a; @Niskanen07a; @Kerman08a] and long-distance [@Sillanpaa07a; @Majer07a] coupling, quantum control [@Pashkin03a; @Yamamoto03a; @McDermott05a; @Plantenberg07a], quantum state [@Steffen06a; @Steffen06b] and process [@Neeley08a] tomography, and measurement [@Siddiqi04a; @Katz06a; @Lupascu07a; @Yamamoto08a; @Beltran08a; @Naaman08a]. For a recent review of superconducting qubits, see Ref. .
{width="5in"}
Here, we use a superconducting persistent-current qubit to implement an artificial atom [@Mooij99a; @Orlando99a]. The persistent-current qubit is a superconducting loop interrupted by three Josephson junctions (Fig. \[fig:1\]A). When biased with a static magnetic flux $f_{\textrm{dc}} \sim \Phi_0/2$, where $\Phi_0$ is the superconducting flux quantum, the system assumes a double-well potential profile (Fig. \[fig:1\]B). The diabatic ground state of the left (right) well corresponds to a persistent current $I_{\textrm{q}}$ with clockwise (counterclockwise) circulation. These two diabatic energy levels (Fig. \[fig:2\]A) have a separation $\varepsilon = 2I_{\textrm{q}} \delta f_{\textrm{dc}}$ linear in the flux detuning $\delta f_{\textrm{dc}} \equiv f_{\textrm{dc}} - \Phi_0/2$. Higher-excited states of the double-well potential (see Fig. \[fig:5\]C) will be considered in Sections \[sec:Cooling\] and \[sec:AS\].
The two-level Hamiltonian for the lowest two diabatic states is shown in Eq. \[Eq:Hamiltonian\]. At detuning $\delta f_{\textrm{dc}}=0$, the double-well potential is symmetric and the diabatic-state energies are degenerate. At this “degeneracy point,” resonant tunneling between the diabatic states opens an avoided level crossing of energy $\Delta$. Here, the qubit states are $\sigma_x$ eigenstates, corresponding to symmetric and anti-symmetric combinations of diabatic circulating-current states. Detuning the flux away from this point tilts the double well, allowing us to tune the eigenstates and eigenenergies of the artificial atom. Far from the degeneracy point the qubit states are approximately $\sigma_z$ eigenstates, the diabatic states with well-defined circulating current. The qubit is read out using a hysteretic DC SQUID (superconducting quantum interference device), a sensitive magnetometer that can distinguish the flux generated by circulating current states.
In addition to the static flux biases, the artificial atom is controlled and read out using the pulses illustrated in Fig. \[fig:1\]C. As we describe below, the qubit is first prepared in its ground state using a harmonic cooling pulse with amplitude $V_{\textrm{c}}$ and frequency $\nu_{\textrm{c}}$. Quantum-state transitions are then driven using a harmonic RF pulse with amplitude $V$ and frequency $\nu$. These fields are mutually coupled to the qubit through a small antenna. This is followed by a SQUID readout current pulse using the “sample and hold” technique [@Oliver05a; @Chiorescu03a]. If the sample current exceeds the SQUID switching current, a voltage pulse will appear at the output during the hold phase. Threshold detection looks for the presence or absence of a SQUID voltage, and this constitutes a digital measurement of the qubit state. Alternatively, although not used in these experiments, we have incorporated the SQUID into a resonant circuit and realized qubit readout via the shift in resonance frequency and phase for both the linear and non-linear resonance regimes [@Lee05a; @Lee07a].
The “qubit step” is shown in Fig. \[fig:1\]D as a function of the SQUID sample current and the flux detuning. The diabatic states $|L \rangle$ and $|R \rangle$ correspond to different levels of sample current (dashed lines) located symmetrically about the degeneracy point $\delta f_{\textrm{dc}}$=0. This plot constitutes a cumulative switching current distribution of the SQUID for each $\delta f_{\textrm{dc}}$ value. Additionally, a 3-$\mu$s RF pulse at 1.2 GHz is applied to the qubit, and resonant transitions can be observed as fingers extending down (up) from state $|L \rangle$ ($|R \rangle$) when $\textrm{n}\times$1.2 GHz becomes resonant with the energy-level separation. A best-estimator (dashed-dotted line) can be determined to provide the best statistical discrimination between states $|L \rangle$ and $|R \rangle$. The resulting qubit step with its saturated n-photon resonances along the best estimator line is shown in Fig. \[fig:1\]E.
The device used in this work was fabricated at Lincoln Laboratory using a fully-planarized niobium trilayer process and optical lithography. It has a critical current density $J_{\rm c} \approx 160 \,{\rm A/cm^2}$, and the characteristic Josephson and charging energies are $E_{\textrm{J}} \approx (2\pi\hbar)300\,{\rm{GHz}}$ $E_{\textrm C} \approx (2\pi\hbar)0.65\,{\textrm{GHz}}$ respectively. The ratio of the qubit Josephson junction (JJ) areas is $\alpha \approx 0.84$, and $\Delta \equiv \Delta_{0,0} \approx (2\pi\hbar)10\,{\textrm{MHz}}$. Although dependent on the flux detuning, the approximate values for coherence times are: interwell relaxation time $T_1 \sim 100 \,\mu \textrm{s}$, intrawell relaxation time $T_1 \sim 0.05 \,\mu \textrm{s}$, homogeneous transverse decay time $T_2 \sim 20 \, \textrm{ns}$, inhomogeneous transverse decay time $T_2^* \sim 10 \, \textrm{ns}$. The experiments were performed in a dilution refrigerator at a base temperature of 20 mK. The device was magnetically shielded, and all electrical lines were carefully filtered and attenuated to reduce noise (see Ref. for details).
Landau-Zener-Stückelberg Transitions {#sec:LZS}
====================================
Landau-Zener-Stückelberg (LZS) transitions occur when a quantum system is driven through an energy-level avoided crossing [@Landau32a; @Zener32a; @Stueckelberg32a]. The resulting quantum dynamics of the LZS mechanism in driven systems [@Grifoni98a] have been developed [@Kayanuma93a; @Kayanuma94a] within a two-level coherent scattering formalism [@Kayanuma97a; @Kayanuma00a; @Shytov03; @Ashab07a] with potential application to quantum information science and technology [@Shytov03; @Ashab07a; @Wubs06a; @KSaito06a; @KSaito07a].
{width="5in"}
We model a LZS transition at a single crossing using the two-level Hamiltonian $$\mathcal{H} =
-\frac{1}{2} \left(
\begin{matrix}
\varepsilon(t) & \Delta \\
\Delta & -\varepsilon(t)
\end{matrix}
\right),
\label{Eq:Hamiltonian}$$ in which $\varepsilon$ is the energy difference between diabatic-state energy levels (dashed lines in Fig. \[fig:2\]a), and $\Delta$ is the size of the avoided crossing and corresponds to the coupling strength between diabatic states $| \textrm{L} \rangle$ and $| \textrm{R} \rangle$.
The system is prepared in state $| R \rangle$ at a static value $\varepsilon_0 \gg \Delta$ far from the avoided crossing (blue dot, Fig. \[fig:2\]A) and driven longitudinally from that point by making $\varepsilon(t)$ time dependent. We distinguish here between “weak driving,” in which the resulting excursion is not large enough to reach the avoided crossing, and “strong driving” for which the avoided crossing is traversed. Under strong-driving conditions, the asymptotic probability $P_{\textrm{D-LZS}}$ of a transition between diabatic states, $$P_{\textrm{D-LZS}} \equiv 1 - P_{\textrm{E-LZS}} = 1 - \exp{
\frac{- \pi \Delta^2}{2 \hbar \zeta_i}},
\label{Eq:P_LZS}$$ is governed by the relative-energy sweep rate $\zeta_i$ $$\zeta_i = \frac{d\varepsilon(t)}{dt} |_{t=t_i}$$ evaluated at the time $t=t_i$ at which the system is swept through the avoided crossing. In the original LZS formulation, the system was driven with a fixed sweep rate, whereas we will later consider a harmonically driven system. We also note that we have elected to monitor the probability $P_{\textrm{D-LZS}}$ of a transition between the diabatic states $| L \rangle$ and $| R \rangle$, because our readout detector is sensitive to changes in diabatic state. This can be written with no loss of generality in terms of the probability $P_{\textrm{E-LZS}}$ of a transition between the system eigenstates.
There are two strong-driving limits characterized by the relative sizes of the sweep velocity $\zeta$ and the avoided crossing $\Delta$. In the adiabatic limit (Fig. \[fig:2\]A), the sweep velocity is small ($h \zeta \ll \Delta^2$) and the probability of a transition from state $| R \rangle$ to state $| L \rangle$ approaches unity, $P_{\textrm{D-LZS}}\rightarrow 1$. In this case, the system dynamics are slow enough that the system adiabatically follows the ground eigenstate through the avoided crossing. In the nonadiabatic limit (Fig. \[fig:2\]B), the sweep velocity is large ($h \zeta \gg \Delta^2$) and the probability of a transition approaches zero $P_{\textrm{D-LZS}}\rightarrow 0$. In this case, the dynamics are too fast for the system to follow; the system remains in diabatic state $| R \rangle$ and, thereby, effectively jumps the energy gap at the avoided crossing.
More generally, a superposition state $\alpha | L \rangle + \beta | R \rangle$ results from an excursion through the avoided crossing, as illustrated in Fig. \[fig:2\]C. Following an idea discussed by Shytov [*et al.*]{} [@Shytov03], the LZS transition acts as a beamsplitter for the atomic state. The amplitudes $\alpha$ and $\beta$ are determined by a unitary transformation $U_{\textrm{D-LZS}}$, effectively a $2 \times 2$ “beamsplitter matrix” comprising complex reflection $r$ and transmission $t$ coefficients related to the adiabaticity parameter [@H_Nakamura01a] $\Delta^2 / \hbar \zeta$ present in Eq. \[Eq:P\_LZS\] such that $|t|^2= 1-P_{\textrm{D-LZS}}$, $|r|^2= P_{\textrm{D-LZS}}$, and $U^{\dagger}_{\textrm{D-LZS}} U_{\textrm{D-LZS}} = I$. Note that we have defined $|t|^2$ and $|r|^2$ from the perspective of a beamsplitter, which “transmits” (“reflects”) an input state to the same (opposite) diabatic state, respectively. After the transition, a relative phase $\Delta \theta_{12}$ accrues due to the energy difference between the states $| L \rangle$ and $| R \rangle$. If the drive $\varepsilon(t)$ then returns the system through the avoided crossing a second time, the atomic state collides and quantum mechanically interferes with itself during the second LZS transition (Fig. \[fig:2\]D). The cumulative result is an atom-state interferometer whose output state depends on the LZS transition amplitudes and the interference phase.
An analogy can be made to an optical Mach-Zehnder interferometer: the atomic states play the role of the photon modes, the LZS transitions play the role of the photon beamsplitters, and the energy-level splitting, which determines the interference phase, plays the role of optical path length difference. The quantum interference is robust provided the evolution time of the state through the interferometer is short compared with the atom’s coherence times. In addition to superconducting artificial atoms [@Oliver05a; @Berns06a; @Nakamura01; @Sillanpaa06a; @Wilson07a; @Saito06a], this concept is generally applicable to other solid-state artificial atoms [@vanderWiel03a; @Hanson07a] and generalized spin systems [@Tannoudji92a] (e.g., molecular magnets [@Friedman96a; @Thomas96a; @Wernsdorfer99a], natural atoms [@Baruch92a; @Yoakum92a], and molecules [@Mark07a; @Mark07b; @Lang08a]) that exhibit avoided level crossings, and it is extensible to multiple energy levels as we demonstrate in Section \[fig:5\].
Mach-Zehnder-type Interferometry {#sec:MZ}
================================
The structure of the n-photon spectroscopy peaks seen in Fig. \[fig:1\]E consists of regularly-spaced resonances positioned according to the condition $\textrm{n}\times$1.2 GHz being resonant with the energy level separation. Notably, however, for this particular value of driving amplitude, the n=1, 3, and 6 photon peaks are missing. As we describe in this section, the presence and absence of these peaks arise from Mach-Zehnder-type quantum interference at a level crossing. The interference phase is tunable via the driving amplitude, leading to a “Bessel staircase” interference pattern in the observed spectroscopy. The interference oscillations are known as Stückelberg oscillations [@Stueckelberg32a], and they have been observed in both artificial [@Oliver05a; @Berns06a; @Sillanpaa06a; @Wilson07a] and natural [@Baruch92a; @Yoakum92a; @Mark07a] atomic systems.
{width="5in"}
In a conventional Mach-Zehnder interferometer, an optical signal is passed through two beamsplitters. The first beamsplitter coherently divides the signal into two output paths, which may have different effective optical path lengths. These paths are then recombined at the second beamsplitter, where the signal waves interfere and exit the interferometer through the two output ports. An intensity measurement at either output port exhibits interference fringes depending on the relative path length.
Here, we instead utilize LZS transitions at level crossings as beamsplitters for the atomic state [@Shytov03]. We drive the persistent-current qubit with a harmonic driving field of the form $$\varepsilon(t) = \varepsilon_0 + A \sin \omega t$$ with $\omega = 2\pi \nu $ the driving frequency and $A>\varepsilon_0$ the field amplitude (in units of energy), which is proportional to the RF voltage at the source. As illustrated in Fig. \[fig:2\]D, the qubit state undergoes two LZS transitions during one period of the driving field. The first LZS transition at time $t_1$ splits the qubit into a superposition of excited and ground states. A relative phase $\Delta \theta_{12}$ accumulates $$\Delta\theta_{12}=\frac1{\hbar}\int_{t_1}^{t_2} \epsilon(t)dt ,\quad \epsilon(t)=\epsilon_{|R\rangle}(t)-\epsilon_{|L\rangle}(t)$$ until the second LZS transition at time $t_2$, at which point the qubit state collides with itself and interferes. Interference fringes appear at half-integer and integer values of $\Delta\theta_{12}/2 \pi$, which is tunable via the drive amplitude.
It is clear, however, that a second phase must also play a role in this problem, since the qubit state continues to evolve for the remainder of the driving period. It is physically meaningful to consider the total phase $\theta$ accumulated over a single period: $$\theta = \frac1{\hbar} \oint \epsilon(t)dt=2\pi\epsilon_0/\hbar\omega,$$ which is [*independent*]{} of the driving amplitude. Over many periods of the driving field, the cascaded pairs of LZS transitions (cascaded interferometers) will constructively interfere provided $\theta = 2\pi n$. One can view this as a “time-domain” formulation for the familiar n-photon resonance condition $$\varepsilon_{0,n} = n h \nu$$ where n is the number of photons involved in the transition. It is only when the product $n h \nu$ equals the energy separation $\varepsilon_{0,n}$ that the cascaded LZS transitions lead to a net buildup of state population and, as a function of $\Delta \theta_{12}$, the observed interference fringes. These oscillations are related to photoassisted transport [@Tien63a; @Kouwenhoven94; @NakamuraTsai99] and Rabi oscillations [@Nakamura01; @Saito06a] in the multiphoton regime.
Mach-Zehnder-type interference in the discrete-resonance limit driven towards saturation is shown in Fig. \[fig:3\]A for driving frequency $\nu=270$ MHz. This frequency is larger than the resonance linewidth (Fig. \[fig:3\]B), and so individual n-photon resonances can be resolved. There are two main features observable in this plot. The first is the presence of equally-spaced n-photon transitions as a function of flux detuning, symmetrically located about the qubit step at $\delta f_{\textrm{dc}}=0$. As one might expect, the onset of the higher-photon transitions requires larger driving amplitudes. Remarkably, we observe up to 50-photon transitions in this scan. The second main feature is that for each n-photon resonance, the spectroscopy appears and disappears as a function of amplitude, which sweeps the interference phase $\Delta \theta_{12}$. For example, 14 oscillation lobes are visible for the 1-photon transition. A vertical slice of the spectroscopy (dashed white line in Fig. \[fig:3\]A) is plotted in Fig. \[fig:3\]C; at this particular amplitude, one can see the enhancement and suppression of the spectroscopy peaks as was observed in Fig. \[fig:1\]E.
The Mach-Zehnder interference for an n-photon transition yields a modified amplitude-dependent matrix element [@Oliver05a; @Nakamura01; @Tannoudji92a] $$\Delta_n = \Delta J_n(\lambda)$$ where $J_n(\lambda)$ is the nth-order Bessel-function, and its argument $\lambda = A / h \nu$ is the dimensionless driving amplitude. Intuitively, the Bessel-function dependence arises, because the qubit is driven harmonically through energy levels that are linear in flux detuning and, as a result, the interference phase accumulates with a harmonic time dependence. The transition rate and, therefore, the population transfer, at each n-photon resonance is related to the matrix element squared, $\Delta_n^2$. At specific amplitudes, despite driving the system resonantly, no net transition occurs due to a complete destructive Mach-Zehnder interference corresponding to the zeros of $J_n(\lambda)$; this is the coherent destruction of tunneling [@Grossmann91a; @Llorente92a; @Kayanuma08a] condition for driven n-photon transitions [@Oliver05a; @Berns06a]. More generally, as a function of the driving amplitude, we have observed the continuous evolution between regions of enhanced and suppressed tunneling as dictated by $J_n(\lambda)$ over several Bessel-function lobes (Fig. \[fig:3\]A).
In the discrete resonance limit, n-photon resonances are distinguishable, because the coherence time of the qubit is sufficiently longer than the driving period. By reducing the frequency to $\nu = 90$ MHz (Fig. \[fig:3\]A, right panel), we effectively made the drive period comparable with the linewidth, $\nu T_2^* \sim 1$ (Fig. \[fig:3\]B). In this spectroscopy plot, the individual n-photon resonances are no longer resolvable. However, the Mach-Zehnder interference fringes (vertical slice, Fig. \[fig:3\]C) are clearly visible, because the qubit remains coherent during the critical fraction of the drive period during which the phase $\Delta \theta_{12}$ accumulates and the Mach-Zehnder interference occurs. Intuitively, provided $\nu T_1 \gg 1$, the output populations of each quantum interferometer are then preserved (frozen) until the subsequent interferometer is reached in the following period; therefore, although the resonance condition is lost, the Mach-Zehnder quantum interference remains. This behavior can be contrasted with driven Rabi oscillations, for which there would be no signature of coherence in the limit $\nu T_2^* \sim 1$.
{width="5in"}
Microwave Cooling {#sec:Cooling}
=================
The previous discussion involved driving transitions in the lowest two energy levels in the double-well potential of our artificial atom (Fig. \[fig:4\]A), which constitute the two-level qubit subsystem of a more complex energy level diagram (Fig. \[fig:5\]C). When higher-excited states are accessed, the driven system behavior can be markedly different from the population saturation observed when only two levels are involved. For example, at least three levels are required to achieve population inversion, and such a multi-level artificial atom coupled to a microwave cavity has been used to demonstrate masing (microwave lasing) [@Astafiev07a]. In that work, Josephson quasi-particle states were driven to achieve inversion. Alternatively, population inversion can be established by accessing an ancilliary excited state via direct or LZS transitions. This will be briefly discussed in the next section.
Here, by reversing the cycle that leads to population inversion, we show that one can pump population from the qubit excited state $|0L \rangle$ to the qubit ground state $|0R\rangle$ (Fig. \[fig:4\]A) via an ancillary energy level $|1R\rangle$. In the case where the population in $|0L \rangle$ results from thermal excitation, the transfer of population to $|0R\rangle$ effectively cools the qubit by lowering its effective temperature. This kind of active cooling represents a means to initialize and reset qubits with high fidelity, key elements for quantum information science and technology. Alternatively, the pumping mechanism can be used to refrigerate environmental degrees of freedom [@Niskanen07b], or to cool neighboring quantum systems [@You08a; @Grajcar08a].
More explicitly, for a qubit in equilibrium with its environment, the population in $|0L\rangle$ that is thermally excited from $|0R\rangle$ follows the Boltzmann relation $$p_{0L}/p_{0R} = \exp[- \varepsilon /k_{\rm B} T_{\rm bath}],
\label{Boltzmann}$$ where $p_{0L,0R}$ are the qubit populations for energy levels $\varepsilon_{0L,0R}$, $\varepsilon = \varepsilon_{0L}- \varepsilon_{0R}$, $k_{\rm B}$ is the Boltzmann constant, and $T_{\rm bath}$ is the bath temperature. To cool the qubit subsystem below $T_{\rm bath}$, a microwave magnetic flux of amplitude $A$ and frequency $\nu$ targets the $|0L\rangle \to |1R\rangle$ transition, driving the state-$|0L\rangle$ thermal population to state $|1R\rangle$, from which it quickly relaxes to the ground state $|0R\rangle$. Efficient cooling only occurs when the driving-induced population transfer to $|0R\rangle$ is faster than the thermal repopulation of $|0L\rangle$. The hierarchy of relaxation and absorption rates required, $ \Gamma_{0R,1R}\gg \Gamma_{0L,1R},\Gamma_{0L,0R}$, is achieved in our system owing to a relatively weak tunneling between wells, which inhibits the interwell relaxation and absorption processes $|1R\rangle \to |0L\rangle$ and $|0R\rangle \to |0L\rangle$, compared with the relatively strong intrawell relaxation process $|1R\rangle \to |0R\rangle$.
Figure \[fig:4\]B shows the qubit step at $T_{\rm bath}$ = 150 mK in equilibrium with the bath (top) and after a 3-$\mu$s cooling pulse at 5 MHz (bottom). Under equilibrium conditions, the average level populations exhibit a thermally-broadened qubit step about $\delta f_{\rm {dc}} = 0$, with a width proportional to $T_{\rm bath}$. The presence of microwave excitation targeting the $|0L\rangle \to |1R\rangle$ transition, followed by relaxation, acts to increase the ground-state population and, thereby, sharpens the qubit step. Cooling can thus be quantified in terms of an effective temperature $T_{\rm eff} < T_{\rm bath}$, a signature that is evident from the narrowing of the qubit steps in Fig. \[fig:4\]B after cooling. More precisely, using the notation from Fig. \[fig:2\], the effective qubit temperature is obtained by fitting an effective temperature that would have been required in equilibrium to achieve the observed qubit population $p_{0R}$, $$p_{0R} = \frac{ \varepsilon}{\sqrt{ \varepsilon^2 + \Delta^2}}
\left[ \tanh \left( \frac{\sqrt{ \varepsilon^2 + \Delta^2}}{2 k_{\textrm{B}} T_{\textrm{eff}}} \right)
+ 1 \right].$$
Universal cooling (cooling that is independent of flux detuning) occurs near an optimal driving amplitude $A^*$ (Fig. \[fig:4\]C). This is demonstrated in Fig. \[fig:4\]D where we present the $|0R\rangle$ state population $P_{\rm sw}$ measured as a function of the microwave amplitude $A$ and flux detuning $\delta f_{\rm dc}$ for $\nu=5\,{\rm MHz}$. Cooling and the diamond feature can be understood in terms of the energy level diagram (Fig. \[fig:5\]C). As the amplitude of the microwave pulse is increased from $V=0$, population transfer first occurs when the $\Delta_{0,0}$ avoided crossing is reached, [*i.e.*]{} $A>|\delta f_{\rm {dc}}|$; this defines the front side of the observed diamond, symmetric about the qubit step (see also Fig. \[fig:3\]A). For amplitudes $A^*/2 \leq A \leq A^*$, the $\Delta_{0,1}$ ($\Delta_{1,0}$) side avoided crossing dominates the dynamics, resulting in a second pair of thresholds $A = A^*-|\delta f_{\rm {dc}}|$, which define the back side of the diamond. In the region outside of the diamond’s backside, the qubit is cooled. As the diamond narrows to the point $A = A^*$, the narrowest qubit step is observed. This is the universal cooling condition: only one of the two side avoided crossings ($\Delta_{0,1}$ or $\Delta_{1,0}$) is reached and, thereby, strong transitions with relaxation to the ground state result for all $\delta f_{\rm dc}$. In contrast, for $A > A^*$, both side avoided crossings ($\Delta_{0,1}$ and $\Delta_{1,0}$) are reached simultaneously for $|\delta f_{\rm{dc}}| < A-A^*$, leading once again to a large population transfer between $|0R\rangle$ and $|0L\rangle$, and opening the second diamond feature (see Fig. \[fig:5\]).
The cooling exhibits a rich structure as a function of driving frequency and detuning, resulting from the manner in which state $|1R\rangle$ is accessed (Fig. \[fig:4\]C). Transitions occur via a (multiphoton) resonant or adiabatic passage process when the driving frequency is high or low enough, respectively [@Berns06a; @Valenzuela06a]. At high frequencies (800 and 400 MHz in Fig. \[fig:4\]E) well-resolved resonances of n-photon transitions are observed and cooling is thus maximized near resonances. At intermediate frequencies (400 and 200 MHz), Mach-Zehnder interference at the side avoided crossing $\Delta_{01}$ becomes more prominent and modulates the intensity of the n-photon resonances [@Oliver05a; @Berns06a]. Below $\nu=200\,{\rm MHz}$, individual resonances are no longer discernible, but as in Fig. \[fig:3\]C, the modulation envelope persists [@Berns06a]. At the lowest frequencies ($\nu < 10\,{\rm MHz}$), state $|1R\rangle$ is reached via adiabatic passage through the $\Delta_{01}$ crossing (Fig. \[fig:4\]C), and the population transfer and cooling become conveniently independent of detuning (see $\nu=5 ~ {\rm MHz}$ in Fig. \[fig:4\]E). As shown in Fig. \[fig:4\]E, we achieve an effective qubit temperature $T_{\rm eff}$ = 3 mK, even for $T_{\rm bath}$ = 300 mK. In our qubit, our determination of $T_{\rm eff}$ was limited primarily by decoherence (linewidth), which limited the resolution with which we could distinguish the states $|0R\rangle$ and $|0L\rangle$ near degeneracy. Nonetheless, we can estimate the ideally resolvable cooling factor $\alpha_{\rm c}$ for this type of cooling process using Eq. \[Boltzmann\], $$\alpha_{\rm c} \equiv \frac{T_{\rm bath}}{T_{\rm eff}} =
\frac{\varepsilon_{1R \to 0R}}{\Delta},$$ where $\varepsilon_{1R \to 0R} \approx h \times 25$ GHz is the energy separation where the relaxation $|1R\rangle \to |0R\rangle$ occurs and $\Delta \approx h \times 0.01$ GHz for our qubit, yielding a cooling factor $\alpha_{\rm c} \sim 2500$. For a bath temperature $T_{\rm bath} = 50$ mK, this would correspond to an effective temperature $T_{\rm eff} = 20$ $\mu$K in our qubit. Cooling a qubit in equilibrium with the bath requires a characteristic cooling time. In turn, a cooled qubit will thermalize to the environmental bath temperature over a characteristic equilibration time. The relationship between these two times determines if it is possible to drive the qubit while it is still cold. We found in this qubit that equilibration times are at least one order of magnitude larger than cooling times at $T_{\rm bath} < 250$ mK and up to three order of magnitude larger at $T_{\rm bath} < 100$ mK [@Valenzuela06a]. This allowed us ample time to drive the qubit after cooling it. The implementation of an active cooling pulse prior to a generic driving pulse is highly advantageous. On the one hand, it sensibly shortens measurement times, enabling us to acquire data at repetition rates that far exceed the intrinsic equilibration rate due to interwell relaxation after each measurement trial. By adopting active cooling, we gained a factor 50 in data acquisition speed, limited by the bandwidth of our readout circuit. On the other hand, by analogy to the cooling of trapped ions and atoms, active cooling greatly reduces thermal smearing, allowing us to analyze features in the data that would have been hidden otherwise. This type of active cooling protocol was required to obtain the detailed, clean data in Fig. \[fig:3\], where we could clearly resolve resonances separated by only 270 MHz (13 mK). In fact, the necessity of active cooling becomes even more evident in the next section where, without the cooling pulse, the observed level of detail could not have been resolved over such a large parameter space in practical acquisition times.
{width="5in"}
Amplitude Spectroscopy {#sec:AS}
======================
Frequency-dependent absorption and emission spectroscopy has long played a fundamental role in the characterization of quantum systems. The development of coherent microwave (maser) and optical (laser) sources, high-intensity radiation with tunable, narrow spectral linewidth, has further enabled targeted absorption spectroscopy of atoms and molecules with high frequency resolution [@Schawlow82a; @Thompson85a]. However, the application of broadband frequency spectroscopy is not universally straightforward. This is particularly relevant for artificial atoms engineered from solid-state materials which, when cooled to cryogenic temperatures, assume quantized energy levels that extend into microwave and millimeter wave regimes (10-300 GHz). Although certainly not an impossible task, a broadband frequency-based spectroscopic study of these atoms in excess of around 50 GHz becomes extremely challenging and expensive to implement due to numerous frequency-dependent effects ([*e.g.*]{}, frequency dispersion and the requisite tolerances to control impedance).
Amplitude spectroscopy is a technique that allows broadband spectroscopic characterization of a quantum system. With amplitude spectroscopy, spectroscopic information is obtained from the system response to driving-field amplitude at a [*fixed*]{} frequency. The resulting spectroscopic interference patterns, “spectroscopy diamonds,” are mediated by multilevel LZS transitions and Mach-Zehnder-type interferometry, and they serve as a fingerprint of the artificial atom’s multilevel energy spectrum (Figs. \[fig:5\]A and \[fig:5\]B). The energy spectrum is then determined by analyzing the atomic fingerprint. In this way, the amplitude spectroscopy technique complements frequency spectroscopy: although a less direct approach, it allows one to probe the energy level structure of a quantum system over extraordinarily large (even practically prohibitive) bandwidths by circumventing many of the challenges associated with a frequency-based approach.
In general, the spectroscopy diamonds arise due to an interplay between the static flux detuning $\delta f_{\rm {dc}}$ and driving amplitude $V$. As described in Sections \[sec:LZS\] and \[sec:MZ\], transitions occur when an avoided crossing is reached for a particular set of values ($\delta f_{\rm {dc}}$,$V$). For example, at a flux detuning $\delta f^*_{\rm {dc}}$, the diamond boundaries occur at $V = V_1, V_2, V_3....$ (Fig. \[fig:5\] A and C). The Mach-Zehnder interference due to a phase accumulation $\Delta \theta_{q,q'}$ at a given avoided crossing $\Delta_{q,q'}$ can be modulated by varying both $\delta f_{\rm {dc}}$ and $V$.
There are two important contributions to the diamond spectroscopy patterns: LZS-mediated transitions and intrawell relaxation. In Fig. \[fig:5\]D, we show a subset of the diamond interference pattern in Fig. \[fig:5\]A. Arrows indicate lines of constant phase accumulation $\Delta \theta_{q,q'}=(2N-1)\pi$ in ($\delta f_{\rm {dc}}$,$V$)-space that leads to LZS transitions at each of the three listed avoided crossings, $\Delta_{0,0}$, $\Delta_{0,1}$, and $\Delta_{1,0}$. Where these lines cross, competition (coordination) between avoided crossings act to suppress (enhance) the net transition rate between pairs of energy levels. The arrangement of these crossing lines leads to the checkerboard patterns observed inside and outside the diamonds.
The second contribution, intrawell relaxation, provides another means to connect energy levels and results in both cooling and population inversion. In Section \[sec:Cooling\], Fig. \[fig:4\]C, $\Delta_{0,1}$ mediated the transition $|0L\rangle \to |1R\rangle$, and intrawell relaxation then mediated the transition $|1R\rangle \to |0R\rangle$; the net result was cooling, since the flux $\delta f_{\rm {dc}}$ was positive, making $|0R\rangle$ the ground state. However, in Fig. \[fig:5\]D $\delta f_{\rm {dc}}$ is negative. Furthermore, in the upper-right corner of Fig. \[fig:5\]D, both crossings $\Delta_{0,1}$ and $\Delta_{1,0}$ are accessed. In the bright red regions, $\Delta_{0,1}$ still tends to cause transitions $|0L\rangle \to |1R\rangle$, and relaxation puts that population in state $|0R\rangle$. However, the interference condition at $\Delta_{1,0}$ on the other side of the energy level diagram (Fig. \[fig:5\]C) tends to keep the population in $|0R\rangle$. Therefore, population builds in state $|0R\rangle$, the first excited state for negative flux detuning, resulting in strong population inversion. Varying the interference conditions at $\Delta_{0,1}$ and $\Delta_{1,0}$ by changing ($\delta f_{\rm {dc}}$,$V$) causes the observed modulation between population inversion and cooling.
We have developed several techniques for extracting information about the energy levels from the spectroscopy interference patterns [@Berns08a; @Rudner08a]. The key metrics are the positions of the avoided crossings in flux, the values of the splittings $\Delta_{q,q'}$, and the slopes of the energy levels. With this information, one can reconstruct a large portion of the energy level diagram.
The positions of the avoided crossings can be determined precisely from the diamond boundaries, because the onset of each diamond is associated with LZS transitions at a particular level crossing. The splitting of each avoided crossing can be obtained essentially by fitting the LZS formula in Eq. \[Eq:P\_LZS\] to the Mach-Zehnder interference patterns. Alternatively, one can study the dynamical population transfer between states using the pulsed, short-time implementation of amplitude spectroscopy [@Berns08a].
The energy-level slopes can be determined by two means. The first is by relating the separation between Mach-Zehnder interference nodes to the expected phase accumulation $\Delta \theta_{12}$, which depends sensitively on the energy-level slope. Alternatively, we show with Rudner [*et al.*]{} in Ref. that the two-dimensional Fourier transform of the diamond patterns yields a family of [*one-dimensional*]{} sinusoids in Fourier space; the periods of these sinusoids are related to the energy-level slopes. The intuition is that the Fourier transform inverts the energy domain of the spectroscopy to the time domain (scaled by $\hbar$). This means that the sinusoids in Fourier space image the time-dependence of the quantum phase of the qubit, which, in turn, depends sensitively on the energy-level slopes.
Using amplitude spectroscopy, we were able to scan the energy level diagram continuously beyond the fourth energy-level avoided crossing ($\Delta_{0,0}\ldots\Delta_{0,4}$, $\Delta_{4,0}$) with splitting values ranging from $\Delta_{0,0}\approx 0.01$ GHz to $\Delta_{0,4}, \Delta_{4,0}\approx 2.2$ GHz [@Berns08a]. The equivalent information obtained using frequency spectroscopy would have required scanning frequencies from 0.01 GHz out to beyond 100 GHz (in this device, each avoided crossing is separated by approximately 25 GHz, and so each subsequent crossing in Fig. \[fig:5\] is raised an additional 25 GHz above the ground state). Remarkably, with amplitude spectroscopy, the entire scan performed in Fig. \[fig:5\]A was performed at a fixed frequency $\nu=185$ MHz. The scan in Fig. \[fig:5\]B shows the amplitude spectroscopy of the same system for a fixed frequency $\nu=700$ MHz, clearly in the discrete resonance limit. The resonance condition adds another constraint, making a more complex and interesting checkerboard pattern.
Summary and Conclusions {#sec:Summary}
=======================
Strongly driving a superconducting artificial atom through an avoided level crossing results in a Landau-Zener-Stückelberg transition, which, in general, creates a superposition of atomic states whose weighting depends on the size of the avoided crossing and the velocity with which it is traversed. In this sense, as we discussed in Section \[sec:LZS\], the LZS mechanism acts as a beamsplitter for artificial atoms.
In Section \[sec:MZ\], we described how harmonically driving the system cascades two LZS transitions per driving period, resulting in an atomic analog to a Mach-Zehnder interferometer. The relative phase acquired between LZS transitions is the interference phase of the interferometer, and it is tunable by the driving amplitude. The buildup of population over many driving periods exhibits Stückelberg oscillations as a function of the driving amplitude (interference phase) due to the cascaded Mach-Zehnder-type interference effect. To observe these oscillations, the coherence time must only be longer than the small portion of the drive period during which the interference phase accrues, and the relaxation time must be long enough to maintain the population until readout.
The total phase accumulated over one period, in contrast, is amplitude independent. For coherence times longer than the drive period, cascaded interferometers yield a net population change when this round-trip phase accumulation is $2 \pi n$ per driving period, a condition which can be viewed as the “time-domain” counterpart to the $n$-photon resonance condition $n h \nu = \varepsilon_{0,n}$. By making the drive period commensurate with the coherence time, we showed that we could still observe the Stückelberg oscillations, even though the discrete resonances were no longer discernable.
We utilized strong driving and the LZS mechanism with higher-energy levels to achieve both cooling and population inversion in our artificial atom. In Section \[sec:Cooling\], we described using a microwave pumping scheme to cool the atomic degrees of freedom a factor 10-100 times colder than the ambient dilution refrigerator temperature. The scheme involved pumping unwanted thermal population to an ancillary excited state, from which it relaxed to the ground state. In Section \[sec:AS\], we showed that by reversing the order, we could pump population through an ancillary state to achieve inversion.
The energy level structure can be probed over extraordinarily large bandwidth using the amplitude spectroscopy approach presented in Section \[sec:AS\]. Since the LZS mechanism and Mach-Zehnder interference are sensitive to the defining features in the energy level diagram (energy band slopes, level splittings, and their locations), the interference patterns that result from sweeping the amplitude are a “fingerprint” of the artificial atom’s energy spectrum. Using amplitude spectroscopy at a fixed driving frequency of only 185 MHz, we could access continuously multiple energy levels from about 10 MHz out to beyond 120 GHz.
Large-amplitude driving and the LZS mechanism have application to quantum information science and technology. Active cooling has utility in state initialization and refreshing ancilla qubits in quantum error correction protocols. Amplitude spectroscopy provides a means to ascertain over large bandwidth the energy level structure of a qubit system beyond the lowest two levels. The Mach-Zehnder-type interference can facilitate nonadiabatic control schemes, in which the quantum interference at an avoided crossing is used to achieve state transitions that approach the intrinsic coupling rate $\Delta$. In cold atoms and molecules, this kind of non-adiabatic control has been used to drive transitions that would otherwise be challenging to realize in a weak-driving limit.
We thank T.P. Orlando, D.M. Berns, L.S. Levitov, M.S. Rudner, A.V. Shytov, and K.K. Berggren with whom we collaborated closely on this work; J. Bylander, S. Gustavsson, B. Turek, and J. Sage for carefully reading the manuscript; V. Bolkhovsky, G. Fitch, T.R Jordan, E. Macedo, P. Murphy, K. Parrillo, R. Slattery, and T. Weir at MIT Lincoln Laboratory for technical assistance. This work was supported by AFOSR and LPS (F49620-01-1-0457) under the DURINT program, and by the U.S. Government. The work at Lincoln Laboratory was sponsored by the US DoD under Air Force Contract No. FA8721-05-C-0002.
J. Clarke, A. N. Cleland, M. H. Devoret, D. Esteve, J. H. Martinis, Quantum mechanics of a macroscopic variable: the phase difference of a Josephson junction, *Science* **239**, 992-997 (1988).
J. Clarke and F.K. Wilhelm, Superconducting quantum bits, *Nature* **453**, 1031-1042 (2008).
J. E. Mooij, T. P. Orlando, L. S. Levitov, L. Tian, C. H. [van der Wal]{}, S. Lloyd, Josephson persistent-current qubit, *Science* **285**, 1036-1039 (1999).
T. P. Orlando, J. E. Mooij, L. Tian, C. H. [van der Wal]{}, L. S. Levitov, S. Lloyd, and J. J. Mazo, Superconducting persistent-current qubit, *Phys. Rev. B* **60**, 15398-15413 (1999).
W. D. Oliver, Y. Yu, J. C. Lee, K. K. Berggren, L. S. Levitov, T. P. Orlando, Mach-Zehnder interferometry in a strongly driven superconducting qubit, *Science* 310, 1653-1657 (2005).
D.M. Berns, W.D. Oliver, S.O. Valenzuela, A.V. Shytov, K.K. Berggren, L.S. Levitov, T.P. Orlando, Coherent Quasiclassical Dynamics of a Persistent Current Qubit, *Phys. Rev. Lett.* 97, 150502 (2006).
S. O. Valenzuela, W. D. Oliver, D. M. Berns, K. K. Berggren, L. S. Levitov, T. P. Orlando, Microwave-induced cooling of a superconducting qubit, *Science* 314, 1589-1592 (2006).
D.M. Berns, M.S. Rudner, S.O. Valenzuela, K. K. Berggren, W. D. Oliver, L. S. Levitov, T. P. Orlando, Amplitude spectroscopy of a solid-state artificial atom, *Nature* 455, 51-57 (2008).
M.S. Rudner, A.V. Shytov, L. S. Levitov, D. M. Berns, W. D. Oliver, S. O. Valenzuela, T. P. Orlando, Quantum phase tomography of a strongly driven qubit, *Phys. Rev. Lett.* 101, 190502 (2008).
J. R. Friedman, V. Patel, W. Chen, S. K. Tolpygo, J. E. Lukens, Quantum superposition of distinct macroscopic states, *Nature* **406**, 43-46 (2000).
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, J. E. Mooij, Quantum superposition of macroscopic persistent-current states, *Science* **290**, 773-777 (2000).
A. J. Berkley, H. Xu, R. C. Ramos, M. A. Gubrud, F. W. Strauch, P. R. Johnson, J. R. Anderson, A. J. Dragt, C. J. Lobb, F. C. Wellstood, Entangled macroscopic quantum states in two superconducting qubits, *Science* **300**, 1548-1550 (2003).
Y. Nakamura, Y. A. Pashkin, J. S. Tsai, Coherent control of macroscopic quantum states in a single-Cooper-pair box, *Nature* **398**, 786-788 (1999).
Y. Nakamura, Y. A. Pashkin, J. S. Tsai, Rabi oscillations in a large Josephson-junction charge two-level system, *Phys. Rev. Lett.* **87**, 246601 (2001).
D. Vion, A. Aassime, A. Cottet, P. Joyez, H. Pothier, C. Urbina, D. Esteve, and M. H. Devoret, Manipulating the quantum state of an electrical circuit, *Science* **296**, 886-889 (2002).
Y. Yu, S. Han, X. Chu, S.-I. Chu, Z. Wang, Coherent temporal oscillations of macroscopic quantum states in a Josephson junction, *Science* **296**, 889-892 (2002).
J. M. Martinis, S. Nam, J. Aumentado, C. Urbina, Rabi oscillations in a large Josephson-junction qubit, *Phys. Rev. Lett.* **89**, 117901 (2002).
I. Chiorescu, Y. Nakamura, C. J. P. M. Harmans, J. E. Mooij, Coherent quantum dynamics of a superconducting flux qubit, *Science* **299**, 1869-1871 (2003).
J. Claudon, F. Balestro, F. W. J. Hekking, O. Buisson, Coherent oscillations in a superconducting multilevel quantum system, *Phys. Rev. Lett.* **93**, 187003 (2004).
B. L. T. Plourde, T.L. Robertson, P.A. Reichardt, T. Hime, S. Linzen, C.-E. Wu, and John Clarke, Flux qubits and readout device with two independent flux lines, *Phys. Rev. B* **72**, 060506(R) (2005).
S. Saito, T. Meno, M. Ueda, H. Tanaka, K. Semba, and H. Takayanagi, Parametric control of a superconducting flux qubit, *Phys. Rev. Lett.* **96**, 107001 (2006).
J. Lisenfeld, A. Lukashenko, M. Ansmann, J.M. Martinis, and A.V. Ustinov, Temperature dependence of coherent oscillations in josephson phase qubits, *Phys. Rev. Lett.* **99**, 170504 (2007).
A. Izmalkov, M. Grajcar, E. Il’ichev, N. Oukhanski, Th. Wagner, H.-G. Meyer, W. Krech, M. H. S. Amin, A. Maassen van den Brink and A. M. Zagoskin, Observation of macroscopic Landau-Zener transitions in a superconducting device, *Europhys. Lett.* 65, 844-849 (2004).
M. Sillanpaa, T. Lehtinen, A. Paila, Yu. Makhlin, P. Hakonen, Continuous-time monitoring of Landau-Zener interference in a Cooper-pair box, *Phys. Rev. Lett.* 96, 187002 (2006).
C.M. Wilson, T. Duty, F. Persson, M. Sandberg, G. Johansson, and P. Delsing, Coherent times of dressed states of a superconducting qubit under extreme driving, *Phys. Rev. Lett.* 98, 257003 (2007).
A. Izmalkov, S.J.W. van der Ploeg, S.N. Shevchenko, M. Grajcar, E. Il’ichev, U. Hübner, A.N. Omelyanchouk, and H.-G.Meyer, Consistency of ground state and spectroscopic measurements on flux qubits, *Phys. Rev. Lett.* 101, 017003 (2008).
A.O. Niskanen, Y. Nakamura, and J.P. Pekola, Information entropic superconducting microcooler, *Phys. Rev. B* [**76**]{}, 174523 (2007).
J.Q. You, Yu-xi Liu, and Franco Nori, Simultaneous cooling of an artificial atom and its neighboring quantum system, *Phys. Rev. Lett.* 100, 047001 (2008).
M. Grajcar, S.H.W. van der Ploeg, A. Izmalkov, E. Il’ichev, H.-G. Meyer, A. Fedorov, A. Shnirman, and G. Schön, Sisyphus cooling and amplification by a superconducting qubit, *Nature Phys.* 4, 612-616 (2008).
K.V.R.M. Murali, Z. Dutton, W.D. Oliver, D.S. Crankshaw, and T.P. Orlando, Probing decoherence with electromagnetically induced transparency in superconductive quantum circuits, *Phys. Rev. Lett.* **93**, 087003 (2004).
Z. Dutton, K.V.R.M. Murali, W.D. Oliver, and T.P. Orlando, Electromagnetically induced transparency in superconducting quantum circuits: effects of decoherence, tunneling, and multilevel crosstalk, *Phys. Rev. B* **73**, 104516 (2006).
P.J. Leek, J.M. Fink, A. Blais, R. Bianchetti, M. Goppl, J.M. Gambetta, D.I. Schuster, L. Frunzio, R.J. Schoelkopf, and A. Wallraff, Observation of Berry’s phase in a solid-state qubit, *Science* [**318**]{}, 1889-1892 (2007).
I. Chiorescu, P. Bertet, K. Semba, Y. Nakamura, C.J.P.M. Harmans, and J.E. Mooij, Coherent dynamics of a flux qubit coupled to a harmonic oscillator, *Nature* [**431**]{}, 159-162 (2004).
A. Wallraff, D. I. Schuster, A. Blais, L. Frunzio, R.-S. Huang, J. Majer, S. Kumar, S. M. Girvin and R. J. Schoelkopf, Strong coupling of a single photon to a superconducting qubit using circuit quantum electrodynamics, *Nature* [**431**]{}, 162-167 (2004).
J. Johansson, S. Saito, T. Meno, H. Nakano, M. Ueda, H. Tanaka, K. Semba, and H. Takayanagi, Vacuum rabi oscillations in a macroscopic superconducting qubit LC oscillator system, *Phys. Rev. Lett.* **96**, 127006 (2006).
F. Deppe, M. Mariantoni, E.P. Menzel, A. Marx, S. Saito, K. Kakuyanagi, H. Tanaka, T. Menon, K. Semba, H. Takayanagi, E. Solano, and R. Gross Two-photon probe of the Jaynes-Cummings model and controlled symmetry breaking in circuit QED, *Nature Phys.* [**4**]{}, 686-691 (2008).
J.M. Fink, M. Goppl, M. Baur, R. Bianchetti, P.J. Leek, A. Blais, and A. Wallraff, Climbing the Jaynes-Cummings ladder and observing its root n nonlinearity in a cavity QED system, *Nature* [**454**]{}, 315-318 (2008).
A. Fragner, M. Goppl, J.M. Fink, M. Baur, R. Bianchetti, P.J. Leek, A. Blais, and A. Wallraff, Resolving vacuum fluctuations in an electrical circuit by measuring the Lamb shift, *Science* [**322**]{}, 1357-1360 (2008).
M. Hofheinz, E.M. Weig, M. Ansmann, R.C. Bialczak, E. Lucero, M. Neeley, A.D. O’Connell, H. Wang, J.M. Martinis, and A.N. Cleland, Generation of Fock states in a superconducting quantum circuit, *Nature* [**454**]{}, 310-314 (2008).
Y. Makhlin, G. Schön, A. Shnirman, Quantum-state engineering with Josephson-junction devices, *Rev. Mod. Phys.* **73**, 357-400 (2001).
J. E. Mooij, The road to quantum computing, *Science* [**307**]{}, 1210-1211 (2005)
T. Hime, P.A. Reichart, B.L.T. Plourde, T.L. Robertson, C.-E. Wu, A.V. Ustinov, and J. Clarke, Solid-state qubits with current-controlled coupling, *Science* [**314**]{} 1427-1429 (2006)
S.H.W. van der Ploeg, A. Izmalkov, A. Maassen van den Brink, U. Hübner, M. Grajcar, E. Il’ichev, H.-G. Meyer, and A.M. Zagoskin, Controllable coupling of superconducting flux qubits, *Phys. Rev. Lett.* [**98**]{}, 057004 (2007).
A.O. Niskanen, K. Harrabi, F. Yoshihara, Y. Nakamura, S. Lloyd, and J.S. Tsai, Quantum coherent tunable coupling of superconducting qubits, *Science* [**316**]{}, 723-726 (2007).
A.J. Kerman and W.D. Oliver, High-fidelity quantum operations on superconducting qubits in the presence of noise, *Phys. Rev. Lett.* [**101**]{}, 070501 (2008).
M.A. Sillanpaa, J.I. Park, and R.W. Simmonds, Coherent quantum state storage and transfer between two phase qubits via a resonant cavity, *Nature* [**449**]{}, 438-442 (2007).
J. Majer, J.M. Chow, J.M. Gambetta, J. Koch, B.R. Johnson, J.A. Schreier, L. Frunzio, D.I. Schuster, A.A. Houck, A. Wallraff, A. Blais, M.H. Devoret, S.M. Girvin, and R.J. Schoelkopf Coupling superconducting qubits via a cavity bus, *Nature* [**449**]{}, 443-447 (2007).
Y. A. Pashkin, T. Yamamoto, O. Astafiev, Y. Nakamura, D. V. Averin, and J. S. Tsai, Quantum oscillations in two coupled charge qubits, *Nature* **421**, 823-826 (2003).
T. Yamamoto, Yu. A. Pashkin, O. Astafiev, Y. Nakamura, J. S. Tsai, Demonstration of conditional gate operation using superconducting charge qubits, *Nature* **425**, 941-944 (2003).
R. McDermott, R. W. Simmonds, M. Steffen, K. B. Cooper, K. Cicak, K. D. Osborn, S. Oh, D. P. Pappas, J. M. Martinis, Simultaneous state measurement of coupled Josephson phase qubits, *Science* **307**, 1299-1302 (2005).
J.H. Plantenberg, P.C. de Groot, C.J.P.M. Harmans, and J.E. Mooij, Demonstration of controlled-NOT quantum gates on a pair of superconducting quantum bits, *Nature* **447**, 836-839 (2007).
M. Steffen, A. Ansmann, R.C. Bialczak, N. Katz, E. Lucero, R. McDermott, M. Neeley, E.M. Weig, A.N. Cleland, and J. M. Martinis, Measurement of the entanglement of two superconducting qubits via state tomography, *Science* **313**, 1423-1425 (2006).
M. Steffen, A. Ansmann, R. McDermott, N. Katz, R.C. Bialczak, E. Lucero, M. Neeley, E.M. Weig, A.N. Cleland, and J. M. Martinis, State tomography of capacitively shunted phase qubits with high fidelity, *Phys. Rev. Lett.* **97**, 050502 (2006).
M. Neeley, M. Ansmann, R.C. Bialczak, M. Hofheinz, N. Katz, E. Lucero, A. O’Connell, H. Wang, A.N. Cleland, and J. M. Martinis, Process tomography of quantum memory in a Josephson-phase qubit coupled to a two-level state, *Nature Phys.* **4**, 523-526 (2008).
I. Siddiqi, *et al.*, RF-driven Josephson bifurcation amplifier for quantum measurement, *Phys. Rev. Lett.* [**93**]{}, 207002 (2004)
N. Katz, M. Ansmann, R.C. Bialczak, E. Lucero, R. McDermott, M. Neeley, M. Steffen, E.M. Weig, A.N. Cleland, J.M. Martinis, A.N. Korotkov Coherent state evolution in a superconducting qubit from partial-collapse measurement, *Science* [**312**]{}, 1498-1500 (2006)
A. Lupascu, S. Saito, T. Picot, P.C. de Groot, C.J.P.M. Harmans, J.E. Mooij, Quantum non-demolition measurement of a superconducting two-level system, *Nature Physics* [**3**]{}, 119 (2007)
T. Yamamoto, K. Inomata, M. Watanabe, K. Matsuba, T. Miyazaki, W.D. Oliver, Y. Nakamura, and J.S. Tsai, Flux-driven Josephson parametric amplifier, *Appl. Phys. Lett.* [**93**]{}, 042510 (2008)
M.A. Castellanos-Beltran, K.D. Irwin, G.C. Hilton, L.R. Vale, and K.W. Lehnert, Amplification and squeezing of quantum noise with a tunable Josephson metamaterial, *Nature Physics* [**4**]{}, 929-931 (2008)
O. Naaman, J. Aumentado, L. Friedland, J.S. Wurtele, I. Siddiqi, Phase-locking transition in a chirped superconducting Josephson resonator, *Phys. Rev. Lett.* [**101**]{}, 117005 (2008)
J.C. Lee, W.D. Oliver, T.P. Orlando, and K.K. Berggren, Resonant readout of a persistent current qubit, *IEEE Trans. Appl. Supercond.* 15, 841-844 (2005).
J.C. Lee, W.D. Oliver, K.K. Berggren, and T.P. Orlando, Nonlinear resonant behavior of a dispersive readout circuit for a superconducting flux qubit, *Phys. Rev. B* 75, 144505 (2007).
L.D. Landau, *Phys. Z. USSR* [**2**]{}, 46-51 On the theory of transfer of energy at collisions II, (1932).
C. Zener, Non-adiabatic crossing of energy levels, *Proc. R. Soc. Lond. A* [**137**]{}, 696-702 (1932).
E. C. G. Stueckelberg, Theorie der un elastischen Stösse zwischen Atomen, *Helv. Phys. Acta* [**5**]{}, 369-422 (1932).
H. Nakamura, *Nonadiabatic Transition* (London, England: World Scientific, 2001).
M. Grifoni and P. Hänggi, Driven quantum tunneling, *Phys. Rep.* [**304**]{}, 229-354 (1998).
Y. Kayanuma, Phase coherence and nonadiabatic transition at a level crossing in a periodically driven two-level system, *Phys. Rev. B* [**47**]{}, 9940-9943 (1993).
Y. Kayanuma, Role of phase coherence in the transition dynamics of a periodically driven two-level system, *Phys. Rev. B* [**50**]{}, 843-845 (1994).
Y. Kayanuma, Stokes phase and geometrical phase in a driven two-level system, *Phys. Rev. A* [**55**]{}, R2495-R2498 (1997).
Y. Kayanuma and Y. Mizumoto, Landau-Zener transitions in a level-crossing system with periodic modulation of the diagonal energy, *Phys. Rev. A* [**62**]{}, 061401(R) (2000).
A. V. Shytov, D. A. Ivanov, M. V. Feigel’man, Landau-Zener interferometry for qubits, *Eur. Phys. J. B* [**36**]{}, 263-269 (2003)
S. Ashhab, J.R. Johansson, A.M. Zagoskin, and Franco Nori, Two-level systems driven by large-amplitude fields, *Phys. Rev. A* [**75**]{}, 063414 (2007).
M. Wubs, K. Saito, S. Kohler, P. Hänggi, and Y. Kayanuma, Gauging a quantum heat bath with dissipative Landau-Zener Transitions, *Phys. Rev. Lett.* [**97**]{}, 200404 (2006).
K. Saito, M. Wubs, S. Kohler, P. Hänggi, and Y. Kayanuma, Quantum state preparation in circuit QED via Landau-Zener tunneling, *Europhys. Lett.* [**76**]{}, 22-28 (2006).
K. Saito, M. Wubs, S. Kohler, Y. Kayanuma, and P. Hänggi, Dissipative Landau-Zener transitions of a qubit: bath-specific and universal behavior, *Phys. Rev. B* [**75**]{}, 214308 (2007).
F. Grossmann, T. Dittrich, P. Jung, and P. Hänggi Coherent destruction of tunneling, *Phys. Rev. Lett.* [**67**]{}, 516-519 (1991).
J.M. Gomez Llorente and J. Plata, Tunneling control in a two-level system, *Phys. Rev. A* [**45**]{}, R6958-R6961 (1992).
Y. Kayanuma and K. Saito, Coherent destruction of tunneling, dynamic localization, and the Landau-Zener formula, *Phys. Rev. A* [**77**]{}, 010101(R) (2008).
W. G. van der Wiel, *et al.*, Electron transport through double quantum dots, *Rev. Mod. Phys.* **75**, 1-22 (2003).
R. Hanson, L.P. Kouwenhoven, J.R. Petta, S. Tarucha, and L.M.K. Vandersypen Spins in few-electron quantum dots, *Rev. Mod. Phys.* **79**, 1217-1265 (2007).
C. Cohen-Tannoudji, J. Dupont-Roc, and G. Grynberg, *Atom-Photon Interactions: Basic Processes and Applications* Ch. 6 (Wiley, 1992).
F.R. Friedman, *et al.*, Macroscopic measurement of resonant magnetization tunnelling in high-spin molecules, *Phys. Rev. Lett.* [**76**]{}, 3830-3833 (1996).
L. Thomas, *et al.*, Macroscopic quantum tunnelling of magnetization in a single crystal of nanomagnets, *Nature* [**383**]{}, 145-147 (1996).
W. Wernsdorfer and R. Sessoli, Quantum phase interference and parity effects in magnetic molecular clusters, *Science* [**284**]{}, 133-135 (1999).
M. C. Baruch, T. F. Gallagher, Ramsey interference fringes in single pulse microwave multiphoton transitions, *Phys. Rev. Lett.* [**68**]{}, 3515-3518 (1992).
S. Yoakum, L. Sirko, P. M. Koch, Stückelberg oscillations in the multiphoton excitation of helium Rydberg atoms: Observation with a pulse of coherent, field and suppression by additive noise *Phys. Rev. Lett.* [**69**]{}, 1919-1922 (1992).
M. Mark, T. Kraemer, P. Waldburger, J. Herbig, C.Chin, H.-C. Nägerl, and R. Grimm “Stückelberg interferometry” with ultracold molecules, *Phys. Rev. Lett.* [**99**]{}, 113201 (2007).
M. Mark, F. Ferlaino, S. Knoop, J.G. Danzl, T. Kraemer, C.Chin, H.-C. Nägerl, and R. Grimm, Spectroscopy of ultracold trapped cesium Feshbach molecules, *Phys. Rev. A* [**76**]{}, 042514 (2007).
F. Lang, P.V.D. Straten, B. Brandstätter, G. Thalhammer, K. Winkler, P.S. Julienne, R. Grimm, and J. Hecker Denschlag, Cruising through molecular bound-state manifolds with radiofrequency, *Nature Physics* [**4**]{}, 223-226 (2008). P. K. Tien and J. P. Gordon, Multiphoton process observed in the interaction of microwave fields with the tunneling between superconductor films *Phys. Rev.* **129**, 647 (1963).
L. P. Kouwenhoven, S. Jauhar, J. Orenstein, P. L. McEuen, Y. Nagamune, J. Motohisa, and H. Sakaki, Observation of Photon-Assisted Tunneling through a Quantum Dot, *Phys. Rev. Lett.* [**73**]{}, 3443 (1994).
Y. Nakamura, J. S. Tsai, A coherent two-level system in a superconducting single-electron transistor observed through photon-assisted cooper-pair tunneling, *J. Supercond.* [**12**]{}, 799 (1999).
O. Astafiev, K. Inomata, A.O. Niskanen, T. Yamamoto, Yu. A. Pashkin, Y. Nakamura, and J.S. Tsai, Single artificial-atom lasing, *Nature* [**449**]{}, 588-590 (2007).
A.L. Schawlow, Spectroscopy in a new light, *Rev. Mod. Phys.* **54**, 697-707 (1982).
R.C. Thompson, High resolution laser spectroscopy of atomic systems, *Rep. Prog. Phys.* **48**, 531-578 (1985).
|
{
"pile_set_name": "ArXiv"
}
|
---
author:
- 'Christopher J. Leininger and Saul Schleimer'
bibliography:
- 'hyp\_teich.bib'
title: 'Hyperbolic spaces in Teichmüller spaces[^1]'
---
Introduction
============
We denote the Teichmüller space of a surface $S$ by $\T(S)$, and the $\epsilon$–thick part by $\T_\epsilon(S)$; see [Section \[S:teichmuller\]]{}. An [*almost-isometric embedding*]{} of one metric space into another is a $(1,C)$–quasi-isometric embedding, for some $C \geq 0$; see [Section \[S:hyperbolic\]]{}. Let ${\mathbb{H}}^n$ denote hyperbolic $n$–space. The main result of this paper is the following.
\[T:main\] For any $n \geq 2$, there exists a surface of finite type $S$ and an almost-isometric embedding $${\mathbb{H}}^n \to \T(S).$$ Moreover, the image is quasi-convex and lies in $\T_\epsilon(S)$ for some $\epsilon > 0$.
According to [Proposition \[P:isometric-teich\]]{} below, Theorem \[T:main\] remains true if we replace “surface of finite type” with “closed surface”. Our work is motivated, in part, by the following open question (see [@FMcc] for the case $n = 2$).
Does there exist a closed surface $S$ of genus at least $2$, a closed hyperbolic $n$–manifold $B$ with $n \geq 2$, and an $S$–bundle $E$ over $B$ for which $\pi_1(E)$ is Gromov hyperbolic?
To explain the relationship with our theorem, suppose that $$S \to E \to B$$ is an $S$–bundle over $B ={\mathbb{H}}^n/\Gamma$, for some closed surface $S$ and some torsion free cocompact lattice $\Gamma < \Isom({\mathbb{H}}^n)$. The monodromy is a homomorphism to the mapping class group of $S$, $\rho {\colon}\pi_1(B)= \Gamma \to \Mod(S)$. The mapping class group $\Mod(S)$ acts on $\T(S)$ by isometries with respect to the Teichmüller metric, and according to work of Farb-Mosher [@FMcc] and Hamenstädt [@hamenstadt], $\pi_1(E)$ is $\delta$-hyperbolic if and only if we can construct a $\Gamma$–equivariant quasi-isometric embedding $$f {\colon}{\mathbb{H}}^n \to \T(S)$$ with quasi-convex image lying in $\T_\epsilon(S)$ for some $\epsilon >
0$; see also [@mjsardar]. (In fact the $\Gamma$–equivariance and quasi-isometric embedding assumptions imply that the image lies in $\T_\epsilon(S)$.)
Our main theorem states that if we drop the assumption of equivariance, then quasi-isometric embeddings with all the remaining properties exist. On the other hand, as was shown in [@clm], one can find cocompact lattices $\Gamma < \Isom({\mathbb{H}}^2)$ and $\Gamma$–equivariant quasi-isometries into $\T(S)$ with image in $\T_\epsilon(S)$—for these examples the image is not quasi-convex.
The main theorem for $n = 2$ also contrasts with the situation of isometrically embedding hyperbolic planes in $\T(S)$. More precisely, every geodesic in $\T(S)$ is contained in an isometrically embedded hyperbolic plane (with the Poincaré metric) called a Teichmüller disk. However, it is well-known that [*no*]{} Teichmüller disk lies in any thick part—this follows from [@masurclosedtrajectories] which guarantees that along a dense set of geodesic rays in the Teichmüller disk the hyperbolic length of some curve on $S$ tends to zero.
The curve complex of $S$ is a metric simplicial complex $\C(S)$ whose vertices are isotopy classes of essential simple closed curves, and for which $k+1$ distinct isotopy classes of curves span a $k$–simplex if they can be realized disjointly. In [@MM1], Masur and Minsky proved that $\C(S)$ is $\delta$–hyperbolic. One of the key ingredients in their proof is the construction of a coarsely Lipschitz map $\T(S) \to \C(S)$. The restriction of this map to any quasi-convex subset of $\T_\epsilon(S)$ is a quasi-isometry (see for example [@rafischleimer Lemma 4.4] or [@shadows Theorem 7.6]). Composing the almost-isometry of Theorem \[T:main\] with the map $\T(S) \to \C(S)$ we have the following corollary.
\[C:C(S)\] For every $n \geq 2$, there exists a surface of finite type $S$ and a quasi-isometric embedding $${\mathbb{H}}^n \to \C(S).$$
The case of $n=2$ here can be compared to the result of Bonk and Kleiner [@bonkkleiner] in which it is shown that every $\delta$–hyperbolic group which is not virtually free contains a quasi-isometrically embedding hyperbolic plane. The assumption that the group is not virtually free implies the existence of an arc in the boundary. According to [@gabai] (see also [@leiningerschleimer; @lms]) with the exception of a few small surfaces, there are indeed arcs in the boundary of $\C(S)$. In [@bonkkleiner] however, essential use is made of the fact that there is an action of the group, and so even in the case $n = 2$, Corollary \[C:C(S)\] does not follow from [@bonkkleiner].
We now explain the idea for the construction in the case $n = 2$. Given a closed Riemann surface $Z$ and a point $z \in Z$, the Teichmüller space $\T(Z,z)$ is naturally a ${\mathbb{H}}^2$–bundle over $\T(Z)$; see [Section \[S:forget\]]{}. Given a biinfinite geodesic $\tau$ in $\T(Z)$, the preimage of $\tau$ in $\T(Z,z)$ is a $3$–manifold. The parameterization $t \mapsto \tau(t)$ lifts to a flow on the preimage of $\tau$ for which the flow lines are geodesics in $\T(Z,z)$. The fiber over $\tau(0)$ admits a pair of transverse $1$–dimensional singular foliations—these are naturally associated to the vertical and horizontal foliations of the quadratic differential defining $\tau$. Any two flow lines meeting the same nonsingular leaf of the vertical foliation are forward asymptotic. Therefore, we have a $1$–parameter family of forward asymptotic geodesics in $\T(Z,z)$. We use this to define a map from ${\mathbb{H}}^2$ to $\T(Z,z)$: we pick a horocycle $C \subset {\mathbb{H}}^2$ and send the pencil of geodesics perpendicular to $C$ to our set of forward asymptotic geodesics in $\T(Z,z)$.
At the beginning of [Section \[S:H\^3\]]{} we give a brief explanation of how this can be modified to give the construction for $n = 3$. The idea for $n \geq 4$ is then a straightforward inductive construction.\
[**Acknowledgements.**]{} We thank Richard Kent for useful conversations as well as having originally asked about the existence of quasi-isometric embeddings of hyperbolic planes into $\C(S)$. We thank the referee for their comments.
Hyperbolic geometry {#S:hyperbolic}
===================
Suppose that $(X, d_X)$ and $(Y, d_Y)$ are metric spaces.
\[Def:AlmostIsom\] A map $F {\colon}X \to Y$ is a [*$K$–almost-isometric embedding*]{} if for all $x, x' \in X$ we have $$|d_X(x,x') - d_Y(F(x),F(x'))| \leq K.$$
We use the [*exponential model*]{} for hyperbolic space: ${\mathbb{H}}^n =
{\mathbb{R}}^{n-1} \times {\mathbb{R}}$ with length element $$ds^2 = e^{-2t}{\mathopen{}\mathclose\bgroup\originalleft}(dx_1^2 + \ldots + dx_{n-1}^2{\aftergroup\egroup\originalright}) + dt^2.$$ For two points $p, q \in {\mathbb{H}}^n$ we use $d_{\mathbb{H}}(p,q)$ to denote the distance between them. The exponential model of hyperbolic space is related to the upper-half space model $U = {\mathbb{R}}^{n-1} \times
(0,\infty)$ by the map ${\mathbb{H}}^n \to U$ given by $(x,t) \mapsto (x,e^t)$. In the exponential model, for every $x \in {\mathbb{R}}^{n-1}$ the path $\eta_x(t) = (x,t)$ is a [*vertical geodesic*]{} and is parameterized by arc-length.
\[L:criteria\] Suppose $(X,d_X)$ is a geodesic metric space and $\delta, \epsilon, R
> 0$ are constants. Suppose $F {\colon}{\mathbb{H}}^n \to X$ is a function with the following properties.
1. \[L:crit:geod\] $F \circ \eta_x$ is a geodesic for all $x \in {\mathbb{R}}^{n-1}$.
2. \[L:crit:slim\] For distinct $x, x' \in {\mathbb{R}}^{n-1}$ the geodesics $F \circ \eta_x$ and $F \circ \eta_{x'}$ are two sides of an ideal $\delta$–slim triangle in $(X,d_X)$.
3. \[L:crit:unif\] For any $x, x' \in {\mathbb{R}}^{n-1}$ if $e^{-t}|x-x'| < \epsilon$ then $d_X(F(x,t),F(x',t)) \leq R$.
4. \[L:crit:prop\] If $(x_k,t_k),(x_k',t_k) \in {\mathbb{H}}^n$ satisfy $\displaystyle{\lim_{k \to
\infty}e^{-t_k}|x_k - x_k'| = \infty}$, then\
$\displaystyle{\lim_{k
\to \infty} d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x_k, t_k), F(x_k', t_k){\aftergroup\egroup\originalright}) = \infty}$.
Then there exists a constant $K$ so that $F$ is a $K$–almost isometric embedding.
A useful consequence of Property \[L:crit:unif\] is that for any $x, x', t \in {\mathbb{R}}$ we have $$\label{E:alt3}
d{\mathopen{}\mathclose\bgroup\originalleft}(F(x,t),F(x',t){\aftergroup\egroup\originalright}) \leq \frac{R}{\epsilon}e^{-t}|x-x'| + R.$$
The remainder of this section gives the proof of [Lemma \[L:criteria\]]{}. We begin by controlling how $F$ moves the centers of ideal triangles. To be precise: Suppose that $T = {\mathcal{P}}\cup {\mathcal{Q}}\cup {\mathcal{R}}\subset
{\mathbb{H}}^n$ is an ideal triangle where ${\mathcal{P}}$ and ${\mathcal{Q}}$ are distinct vertical geodesics. Let $r$ denote the point of ${\mathcal{R}}$ with maximal $t$–coordinate. We call $r$ the [*midpoint*]{} of ${\mathcal{R}}$. Thus $r$ serves as a center for $T$. Define $x = x({\mathcal{P}}), x' = x({\mathcal{Q}})$.
Observe, say from the upper-half space model, that for all $t \geq
t(r)$ we have $$\label{E:horocycledist}
d_{\mathbb{H}}{\mathopen{}\mathclose\bgroup\originalleft}((x,t), (x',t){\aftergroup\egroup\originalright}) \leq e^{-t}|x-x'| \leq e^{-t(r)}|x-x'| = 2.$$ Thus, by Inequality (\[E:alt3\]) we have $d_X(F(x,t), F(x',t)) \leq
2R/\epsilon+R$. Define $\Delta = \max \{ 3\delta, 2R/\epsilon + R \}$ and define the [*displaced height*]{} of $T$ to be $$h_T = h(T) = \min \Big\{ t \in {\mathbb{R}}\,\, \Big| \,\,
d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x,t), F({\mathcal{Q}}){\aftergroup\egroup\originalright}) \leq \Delta
\,\, \mbox{or} \,\,
d_X{\mathopen{}\mathclose\bgroup\originalleft}(F({\mathcal{P}}), F(x',t){\aftergroup\egroup\originalright}) \leq \Delta
\Big\}.$$ It follows that $h(T) \leq t(r)$. Note that for any vertical triangle $T$, Property \[L:crit:slim\] implies that $h(T) > -\infty$.
\[C:SlimAbove\] For any vertical triangle $T = {\mathcal{P}}\cup {\mathcal{Q}}\cup {\mathcal{R}}\subset {\mathbb{H}}^n$, $$d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x, h_T), F(x', h_T){\aftergroup\egroup\originalright}) \leq 3 \Delta,$$ where $x = x({\mathcal{P}})$, $x' = x({\mathcal{Q}})$.
Breaking symmetry, in this setting, allows us to assume that there is some $s \in {\mathbb{R}}$ so that $d_X(F(x',s), F(x,h_T)) \leq \Delta$. Let $t' = \max\{s, t(r)\}$. Using the triangle inequality, Inequality \[E:alt3\] and Property \[L:crit:geod\] we have $$\begin{aligned}
t' - h_T & = d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x,t'),(x,h_T){\aftergroup\egroup\originalright}) \\
& \leq d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x,t'), F(x',t'){\aftergroup\egroup\originalright}) + d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x',t'), F(x',s){\aftergroup\egroup\originalright})
+ d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x',s), F(x,h_T){\aftergroup\egroup\originalright}) \\
& \leq (2R/\epsilon + R) + (t' - s) + \Delta \\
\intertext{and similarly}
t' - s & \leq 2R/\epsilon + R + t' - h_T + \Delta. \end{aligned}$$ Thus $|h_T - s| \leq 2R/\epsilon + R + \Delta$. Another application of the triangle inequality and Property \[L:crit:geod\] implies that $d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x,h_T), F(x',h_T){\aftergroup\egroup\originalright}) \leq 2R/\epsilon + R + 2\Delta \leq
3\Delta$, as desired.
As mentioned above, for every vertical triangle $T$ we have $h(T) >
-\infty$ and hence $t(r) - h(T) < \infty$. We now obtain a uniform bound on this quantity.
\[C:UniformBoundOnDisplacement\] There is a constant $C_0 = C_0(F)$ so that $t(r) - h(T) \leq C_0$ for all vertical triangles $T \subset {\mathbb{H}}^n$.
Suppose not. Then we are given a sequence of vertical triangles $T_k
= {\mathcal{P}}_k \cup {\mathcal{Q}}_k \cup {\mathcal{R}}_k$ where $t(r_k) - h(T_k)$ tends to infinity with $k$. Here $r_k$ is the midpoint of ${\mathcal{R}}_k$, the non-vertical side. Define $t_k = t(r_k)$, $h_k = h(T_k)$. Define $x_k
= x({\mathcal{P}}_k)$, $x_k' = x({\mathcal{Q}}_k)$ to be the horizontal coordinates of the vertical sides of $T_k$.
Note that by Equation (\[E:horocycledist\])
&& e\^[-t\_k]{}|x\_k - x\_k’| & = 2 &\
&& e\^[-h\_k]{} |x\_k - x\_k’| & = e\^[-h\_k]{} 2e\^[t\_k]{} = 2e\^[t\_k-h\_k]{}. &
Thus $e^{-h_k} |x_k - x_k'|$ tends to infinity with $k$. From Property \[L:crit:prop\] we deduce that the quantity $d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(x_k, h_k), F(x_k', h_k){\aftergroup\egroup\originalright})$ also tends to infinity with $k$. This last, however, contradicts [Claim \[C:SlimAbove\]]{}.
We give the proof of [Lemma \[L:criteria\]]{}. Fix any $p, q \in {\mathbb{H}}^n$. If $x(p) = x(q)$ then we are done by Property \[L:crit:geod\]. Suppose instead that $x(p) \neq x(q)$. Let ${\mathcal{P}}\cup {\mathcal{Q}}\cup {\mathcal{R}}$ denote the vertical triangle having vertical sides ${\mathcal{P}}$ and ${\mathcal{Q}}$ so that $x({\mathcal{P}}) = x(p)$, $x({\mathcal{Q}}) = x(q)$; let $r \in {\mathcal{R}}$ be the midpoint of the non-vertical side. Define $C_1 = 2C_0 + 5\Delta + 1$. There are now two cases to consider.
Suppose that $t(p) \geq h(T) - C_1$.
Let $p' \in {\mathcal{P}}$ and $q' \in {\mathcal{Q}}$ be the points with $t(p') =
t(q') = \max \{t(p), t(r)\}$. Then by the triangle inequality and Equation (\[E:horocycledist\]) we have $$\begin{aligned}
d_{\mathbb{H}}(p, q') & \leq d_{\mathbb{H}}(p, p') + d_{\mathbb{H}}(p', q') \\
& \leq t(p') - t(p) + 2 \\
& \leq t(r) - h(T) + C_1 + 2 \\
& \leq C_0 + C_1 + 2.\end{aligned}$$ It follows that $d_{\mathbb{H}}(p,q)$ is estimated by $d_{\mathbb{H}}(q',q) = |t(q') -
t(q)|$ up to an additive error at most $C_0 + C_1 + 2$. Appealing to Property \[L:crit:geod\], Inequality (\[E:alt3\]), and the triangle inequality we similarly have $$\begin{aligned}
d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p),F(q'){\aftergroup\egroup\originalright})
& \leq d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p),F(p'){\aftergroup\egroup\originalright}) + d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p'),F(q'){\aftergroup\egroup\originalright})\\
& \leq t(p') - t(p) + 2R/\epsilon + R \\
& \leq C_0 + C_1 + 2R/\epsilon + R.\end{aligned}$$ Thus $d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p),F(q){\aftergroup\egroup\originalright})$ is estimated by $d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(q'),F(q){\aftergroup\egroup\originalright}) = d_{\mathbb{H}}(q',q)$ with an additive error at most $C_0 + C_1 + 2R/\epsilon + R$. This completes the proof in this case.
Suppose that $t(p), t(q) \leq h(T) - C_1$.
In this case, since the triangle $T = {\mathcal{P}}\cup {\mathcal{Q}}\cup {\mathcal{R}}$ is slim in ${\mathbb{H}}^n$, we find that that $d_{\mathbb{H}}(p,q)$ is estimated by $t(r)
- t(p) + t(r) - t(q)$ up to an additive error of at most $2$. We now show that $d_X(F(p), F(q))$ is also estimated by the latter quantity, with a uniformly bounded error. Using Property \[L:crit:geod\] and Inequality (\[E:alt3\]) deduce $$d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p),F(q){\aftergroup\egroup\originalright})
\leq t(r) - t(p) + 2R/\epsilon + R + t(r) - t(q).$$
We now give a lower bound for $d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p), F(q){\aftergroup\egroup\originalright})$. Recall that $F({\mathcal{P}})$ and $F({\mathcal{Q}})$ are two sides of a $\delta$–slim triangle in $X$. Let ${\mathcal{R}}_X$ be the third side of this triangle. Since $$d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p), F({\mathcal{Q}}){\aftergroup\egroup\originalright}),d_X{\mathopen{}\mathclose\bgroup\originalleft}(F({\mathcal{P}}), F(q){\aftergroup\egroup\originalright})
> \Delta \geq \delta$$ it follows that there are points $p_X, q_X \in {\mathcal{R}}_X$ so that $d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p), p_X{\aftergroup\egroup\originalright}), d_X{\mathopen{}\mathclose\bgroup\originalleft}(q_X, F(q){\aftergroup\egroup\originalright}) \leq \delta$. Thus the distance $d_X(p_X, q_X)$ is an estimate for $d_X{\mathopen{}\mathclose\bgroup\originalleft}(F(p),
F(q){\aftergroup\egroup\originalright})$ with an additive error at most $2\delta$.
Define $a = (x, h_T), b = (x', h_T)$. Again, as in the previous paragraph, there are points $a_X, b_X \in {\mathcal{R}}_X$ within distance $\delta$ of $F(a), F(b)$. Since $d_{\mathbb{H}}(a,b) \leq 2(t(r) - h(T)) + 2$ we find $$\begin{aligned}
d_X(a_X, b_X) & \leq 2\delta + 2(t(r) - h(T)) + 2R/\epsilon + R\\
& \leq 2\delta + 2C_0 + 2R/\epsilon + R.\end{aligned}$$ Note that the geodesic segments $[p_X, a_X], [b_X, q_X] \subset
{\mathcal{R}}_X$ have length at least $h(T) - t(p) - 2\delta$ and $h(T) - t(q)
- 2\delta$ respectively. Each of these is greater than $C_1 -
2\delta$.
If $p_X \in [a_X, b_X]$ then $C_1 - 2\delta \leq 2\delta + 2C_0 +
2R/\epsilon + R$ and this is a contradiction. Similarly, deduce $q_X
\not\in [a_X, b_X]$. If $p_X = q_X$ then $d_X(F(p), F(q)) \leq
2\delta < \Delta$, contradicting our assumption that $t(p) < h(T)$. Finally, if $p_X \in (b_X, q_X)$ then an intermediate value argument using the fact that ${\mathcal{R}}_X$ is a geodesic implies $d_X(F(p),
F({\mathcal{Q}})) \leq 3\delta$, again a contradiction. Similarly $q_X$ is not in $(p_X, a_X)$. Thus, $[p_X, a_X] \cap [b_X, q_X]$ is either empty or is equal to $[a_X, b_X]$. We deduce that $$\begin{aligned}
d_X(p_X, q_X) & \geq 2h(T) - t(p) - t(q) - 4\delta - 2\delta - 2C_0 - 2R/\epsilon - R\\
& \geq 2t(r) - t(p) - t(q) - 7\Delta - 4C_0.\end{aligned}$$ The proof of [Lemma \[L:criteria\]]{} is complete.
Foliations and projections {#S:subsurfaces}
==========================
Let $Z$ be a closed surface of genus at least $2$ and ${\bf z}$ a set of marked points. A [*measured singular foliation*]{} ${\mathcal{F}}$ on $(Z,{\bf z})$ is a singular topological foliation so that
- ${\mathcal{F}}$ has only prong-type singularties,
- all one-prong singularties of ${\mathcal{F}}$ appear at points of ${\bf z}$, and
- ${\mathcal{F}}$ is equipped with a transverse measure of full support.
We refer the reader to [@FLP; @masurinterval] for a detailed discussion of measured foliations. Two measured (respectively, topological) foliations are [*measure equivalent*]{} (respectively, [*topologically equivalent*]{}) if they differ by isotopy and Whitehead moves. We will only be concerned with those foliations which appear as the vertical foliation for some meromorphic quadratic differential on $Z$ (see Section \[S:definitions\]). Every measured singular foliation is measure equivalent to such a foliation for a fixed complex structure on $Z$; see [@hubbardmasur].
The space of measure classes of measured foliation on $(Z,{\bf z})$ is denoted by $\MF(Z,{\bf z})$ and its projectivization by $\PMF(Z,{\bf
z})$. A measured foliation ${\mathcal{F}}\in \MF(Z,{\bf z})$ is [*arational*]{} if it has no closed leaf cycles. We say that ${\mathcal{F}}$ is [*uniquely ergodic*]{} if whenever ${\mathcal{F}}' \in \MF(Z,{\bf z})$ is topologically equivalent to ${\mathcal{F}}$, then ${\mathcal{F}}$ and ${\mathcal{F}}'$ project to the same point in $\PMF(Z,{\bf z})$. Both of these notions depend only on the topological classes of the foliations, and not the transverse measures.
If ${\mathcal{F}}$ is a measured foliation representing an element of $\MF(Z)$, and ${\bf z} \subset Z$ is a set of marked points, then ${\mathcal{F}}$ also determines an element of $\MF(Z,{\bf z})$. We note that it is important in this case that ${\mathcal{F}}$ be a foliation, and not an equivalence class of foliations. If ${\mathcal{F}}$ is arational as an element of $\MF(Z)$, and if ${\bf z} = \{z\}$ is a single point, then ${\mathcal{F}}$ is also arational as an element of $\MF(Z,z)$; see [@leiningerschleimer].
By a [*strict subsurface*]{} $Y \subset Z - {\bf z}$ we mean a properly embedded surface with [*nonempty*]{} boundary and a set of punctures, possibly empty, such that every component of $\partial Y$ is an [*essential curve*]{} in $Z - {\bf z}$; that is, homotopically nontrivial and nonperipheral. We also assume that $Y$ is not a sphere with $k$ punctures and $j$ boundary components where $k + j = 3$. We will only refer to subsurfaces in one context, and that is as follows. Given a pair of arational measured foliation ${\mathcal{F}}, {\mathcal{G}}\in
\MF(Z,{\bf z})$ and a proper subsurface $Y \subset Z - {\bf z}$, we have the [*projection distance*]{} $$d_Y({\mathcal{F}},{\mathcal{G}}) \in \mathbb Z_{\geq 0}$$ between ${\mathcal{F}}$ and ${\mathcal{G}}$ in $Y$. This is the distance in the arc-and-curve complex of $Y$ between the the subsurface projections of ${\mathcal{F}}$ and ${\mathcal{G}}$ to $Y$. For a detailed discussion, see [@MM1; @MM2]. All we use is that $d_Y$ satisfies a triangle inequality $$d_Y({\mathcal{F}}_1,{\mathcal{F}}_2)
\leq d_Y({\mathcal{F}}_1,{\mathcal{G}}) + d_Y({\mathcal{G}},{\mathcal{F}}_2)$$ for all arational measured foliations ${\mathcal{F}}_1,{\mathcal{F}}_2,{\mathcal{G}}\in
\MF(Z,{\bf z})$. This relates to Teichmüller geometry by [Theorem \[T:Rafi\]]{} below.
Teichmüller spaces {#S:teichmuller}
==================
Here we set notation and recall some basic properties of Teichmüller space. For background on Teichmüller space, we refer the reader to any of [@ahlforsqc; @gardiner; @abikoff; @IT].
Teichmüller space, quadratic differentials and geodesics {#S:definitions}
--------------------------------------------------------
Given a closed Riemann surface $Z$ with a finite (possibly empty) set of marked points ${\bf z} \subset Z$, let $\T(Z,{\bf z})$ denote the [*Teichmüller space*]{} of equivalence classes of marked Riemann surfaces $$\T(Z,{\bf z}) = {\mathopen{}\mathclose\bgroup\originalleft}\{ [f {\colon}(Z,{\bf z}) \to (X,{\bf x})] \, {\mathopen{}\mathclose\bgroup\originalleft}| \,
\begin{array}{l} f \mbox{ is an orientation preserving homeo-}\\
\mbox{morphism to the Riemann surface } X \end{array}
{\aftergroup\egroup\originalright}. {\aftergroup\egroup\originalright}\}.$$ The equivalence relation is defined by $$\big( f {\colon}(Z,{\bf z}) \to (X,{\bf x}) \big)
\sim \big( g {\colon}(Z,{\bf z}) \to (Y,{\bf y}) \big)$$ if $f \circ g^{-1} {\colon}(Y,{\bf y}) \to (X,{\bf x})$ is isotopic (rel marked points) to a conformal map. If ${\bf z} = \emptyset$, then we write $\T(Z) = \{[f {\colon}Z \to X]\}$.
The [*Teichmüller distance*]{} on $\T(Z,{\bf z})$ is defined by $$d_\T \big( [f {\colon}(Z,{\bf z}) \to (X,{\bf x})],
[g {\colon}(Z,{\bf z}) \to (Y,{\bf y})] \big)
= \inf {\mathopen{}\mathclose\bgroup\originalleft}\{ {\mathopen{}\mathclose\bgroup\originalleft}. \frac{1}{2} \log {\mathopen{}\mathclose\bgroup\originalleft}( K_h {\aftergroup\egroup\originalright})
\,{\aftergroup\egroup\originalright}|\, h \simeq f \circ g^{-1} {\aftergroup\egroup\originalright}\}$$ where $K_h$ is the dilatation of $h$ and where $h {\colon}(Y,{\bf y}) \to
(X,{\bf x})$ ranges over all quasi-conformal maps isotopic (rel marked points) to $f \circ g^{-1}$.
Given $\epsilon > 0$, the [*$\epsilon$–thick part of Teichmüller space*]{} $\T_\epsilon(Z,{\bf z}) \subset \T(Z,{\bf z})$ is the set of points $[f {\colon}(Z,{\bf z}) \to (X,{\bf x})] \in \T(Z,{\bf z})$ where the unique complete hyperbolic surface in the conformal class of $X-{\bf x}$ has its shortest geodesic of length at least $\epsilon$. When $\epsilon$ is understood from context we will simply refer to $\T_\epsilon(Z,{\bf z})$ as the [*the thick part of Teichmüller space*]{}.
Let $\T(Z,{\bf z}) \to \M(Z,{\bf z})$ denote the projection to moduli space obtained by forgetting the marking $$[
f {\colon}(Z,{\bf z}) \to (X,{\bf x})] \mapsto [(X,{\bf x})]$$ or, equivalently, by taking the quotient by the mapping class group. Mumford’s compactness criterion [@bersmumford] now implies: For any $\epsilon > 0$, the thick part $\T_\epsilon(Z,{\bf z})$ projects to a compact subset of $\M(Z,{\bf z})$. Conversely, the preimage of any compact subset of $\M(Z,{\bf z})$ is contained in $\T_\epsilon(Z,{\bf z})$ for some $\epsilon > 0$.
Suppose $(X,{\bf x})$ is a closed Riemann surface with marked points and $q \in \Q(X,{\bf x})$ is a unit norm, meromorphic quadratic differential with all poles simple and contained in ${\bf x}$. We also use $q$ to denote the associated Euclidean cone metric on $X$. We note that $\Q(X) \subset \Q(X,{\bf x})$, for any set of marked point ${\bf x} \subset X$. Given $q \in \Q(X)$ we view it as an element of $\Q(X,{\bf x})$ whenever it is convenient.
Given $q \in \Q(X,{\bf x})$ and $t \in {\mathbb{R}}$, let $g_t {\colon}(X,{\bf
x}) \to (X_t,g_t({\bf x}))$ denote the $e^{2t}$–quasi-conformal Teichmüller mapping defined by $(q,t)$. Let $q_t \in
\Q(X_t,g_t({\bf x}))$ denote the terminal quadratic differential. For any point $p \in X$ which is not a zero or pole of $q$ we have a [*preferred coordinate*]{} $z_0$ for $(X,q)$ near $p$ and preferred coordinate $z_t$ for $(X_t,q_t)$ near $g_t(p)$. In these coordinates $q = dz_0^2$ and $q_t=dz_t^2$, and $g_t$ is given by $(u,v) \mapsto
(e^t u, e^{-t} v)$. If we mark $(X,{\bf x})$ by $f {\colon}(Z,{\bf z}) \to
(X,{\bf x})$, then setting $f_t = g_t \circ f$ we have $$\tau_q(t) = [f_t {\colon}(Z,{\bf z}) \to (X_t,g_t({\bf x}))]$$ being a Teichmüller geodesic through $[f {\colon}(Z,{\bf z}) \to (X,{\bf
x})]$; note that every Teichmüller geodesic can be described in this way. The Teichmüller geodesic $\tau$ is [*$\epsilon$–thick*]{} if the image of $\tau$ lies in $\T_\epsilon(Z,{\bf z})$. We also simply say a geodesic $\tau$ is [*thick*]{} if it is $\epsilon$–thick for some $\epsilon > 0$. A collection of geodesics $\{\tau_\alpha\}$ is [*uniformly thick*]{} if there is an $\epsilon > 0$ so that each $\tau_\alpha$ is $\epsilon$–thick.
Given $q \in \Q(X,{\bf x})$ we will let ${\mathcal{F}}(q), {\mathcal{G}}(q)$ denote the vertical and horizontal foliations respectively; that is, the preimage in preferred coordinates of the foliations of $\mathbb C$ by vertical and horizontal lines. For $q \in \Q(X,{\bf x})$ and $t \in
{\mathbb{R}}$ consider the associated Teichmüller mapping $g_t {\colon}(X,{\bf
x}) \to (X_t,g_t({\bf x}))$ as above. If $c {\colon}{\mathbb{R}}\to X$ is a nonsingular leaf of ${\mathcal{F}}(q)$ parameterized by arc-length with respect to the $q$–metric, then composing with $g_t$ we obtain a nonsingular leaf of the vertical foliation for the terminal quadratic differential ${\mathcal{F}}(q_t)$, $$g_t \circ c {\colon}{\mathbb{R}}\to X_t.$$ From the description of $g_t$ in local coordinates we see that this is parameterized [*proportional*]{} to arc-length and, in fact, the $q_t$–length is given by $$\label{E:exp-length}
\ell_{q_t}{\mathopen{}\mathclose\bgroup\originalleft}( g_t \circ c |_{[x,x']} {\aftergroup\egroup\originalright}) = e^{-t}|x' - x|.$$
Properties of Teichmüller geodesics {#S:geodesicprop}
-----------------------------------
Suppose $\tau = \tau_q$ is the Teichmüller geodesic determined by $[f {\colon}(Z,{\bf z}) \to (X,{\bf x})] \in \T(Z,{\bf z})$ and $q \in
\Q(X,{\bf x})$. The forward asymptotic behavior of $\tau$ is reflected in the structure of the vertical foliation ${\mathcal{F}}(q)$. For us, the most important instance of this is a result of Masur [@masurhaus].
\[T:Masur-ue\] If there exists $\epsilon > 0$ and $\{t_k\}_{k=1}^\infty$ such that
- $t_k \to \infty$ as $k \to \infty$ and
- $\tau_q(t_k) \in \T_\epsilon(Z,{\bf z})$ for all $k$
then ${\mathcal{F}}(q)$ is arational and uniquely ergodic.
In particular, if $\tau_q$ is thick then both ${\mathcal{F}}(q)$ and ${\mathcal{G}}(q)$ are uniquely ergodic. We say a pair of arational foliations ${\mathcal{F}}$ and ${\mathcal{G}}$ are [*$K$–cobounded*]{} if for all strict subsurfaces $Y \subset X - {\bf x}$ we have $d_Y({\mathcal{F}}, {\mathcal{G}})
\leq K$. A result of Rafi [@rafi Theorem 1.5] relates the thickness of a geodsic $\tau_q \subset \T$ to the coboundedness of the associated vertical and horizontal foliations.
\[T:Rafi\] For all $\epsilon > 0$ there exists $K > 0$ so that if $q \in
\Q(X,{\bf x})$ has $\tau_q$ being $\epsilon$–thick then ${\mathcal{F}}(q)$ and ${\mathcal{G}}(q)$ are $K$–cobounded.
Conversely, for all $K > 0$ there exists $\epsilon > 0$ so that if $q \in \Q(X,{\bf x})$ has ${\mathcal{F}}(q)$ and ${\mathcal{G}}(q)$ being $K$–cobounded then $\tau_q$ is $\epsilon$–thick.
Forgetting the marked point: the Bers fibration {#S:forget}
-----------------------------------------------
Suppose now that $Z$ is a closed surface and $z \in Z$ is a single marked point; we use $(Z, z)$ to denote $(Z, \{z\})$. Let $p {\colon}\widetilde Z \to Z$ denote the universal covering. Given $[f {\colon}(Z,z)
\to (X,f(z))]$ we can forget the marked point to obtain an element $[f {\colon}Z \to X] \in \T(Z)$. This defines a holomorphic map $$\Pi {\colon}\T(Z,z) \to \T(Z)$$ called the [*Bers fibration*]{} [@bersfiber]. The fiber of this map over $[f {\colon}Z \to X]$ is holomorphically identified with $\widetilde
X$, the universal covering of $X$. Moreover, this identification is canonical, up to the action of the covering group on $\widetilde X$.
The projection of Teichmüller spaces $\Pi {\colon}\T(Z,z) \to \T(Z)$ descends to a projection of moduli spaces $\hat \Pi {\colon}\M(Z,z) \to
\M(Z)$. The fiber of $\hat \Pi$ over $X \in \M(Z)$ is just $X/{\mathop{\rm Aut}}(X)$ and this is compact.
Recall that puncturing a closed surface once increases the hyperbolic systole. (Lift to universal covers and apply the Schwarz-Pick lemma.) It follows that the preimage of $\T_{\epsilon}(Z)$ by $\Pi^{-1}$ is contained in $\T_{\epsilon}(Z,z)$.
By a theorem of Royden [@royden] the Teichmüller metric agrees with the Kobayashi metric on Teichmüller space. Recall that the inclusion of the universal covering $\widetilde X \to \T(Z,z)$ is a holomorphic embedding [@bersfiber]. Thus, if we give $\widetilde
X$ the Poincaré metric $\rho_0$ — one-half the hyperbolic metric — then $(\widetilde X,\rho_0) \to (\T(Z,z), d_\T)$ is a contraction [@kobayashihyperbolic]. Kra [@kra] further proved the following.
\[T:kra\] There exists a homeomorphism $h {\colon}[0,\infty) \to [0,\infty)$ so that for any $[f {\colon}Z \to X] \in \T(Z)$, and any $\tilde x_1,\tilde x_2 \in
\widetilde X \subset \T(Z,z)$, we have $$h(\rho_0(\tilde x_1,\tilde x_1)) \leq d_\T(\tilde x_1,\tilde x_2)
\leq \rho_0(\tilde x_1,\tilde x_2).$$
The function $h$ can be described concretely in terms of the solution to a certain extremal mapping problem for the hyperbolic plane which was solved by Teichmüller [@teichmuller] and Gehring [@gehring]. We will extend $h$ to a nondecreasing function, $h
{\colon}{\mathbb{R}}\to [0,\infty)$ by declaring $h(t) = 0$ for all $t
\leq 0$.
Branched covers {#S:branched}
---------------
Here we use branched covers to induce maps on Teichmüller space.
Suppose $P {\colon}\Sigma \to Z$ is a branched cover, branched over some finite set of points ${\bf z} \subset Z$. Then any complex structure on $Z$ pulls back to a complex structure on $\Sigma$, and thus induces a map $P^* {\colon}\T(Z,{\bf z}) \to \T(\Sigma)$. Regarding Teichmüller space as the space of marked Riemann surfaces, $\T(Z,{\bf z}) = \{[f {\colon}(Z,{\bf z}) \to (X,{\bf x})]\}$, the embedding is described as follows. The branched covering $P {\colon}\Sigma \to (Z,{\bf z})$ induces a branched covering $U {\colon}\Omega
\to (X,{\bf x})$, for some Riemann surface $\Omega$, namely the branched cover induced by the subgroup $(f \circ P)_*(\pi_1(\Sigma -
P^{-1}({\bf z}))) < \pi_1(X - {\bf x})$. By construction, there is a lift of the marking homeomorphism $\phi {\colon}\Sigma \to \Omega$. This is described by the following commutative diagram. $$\xymatrix{
\Sigma \ar[d]_P \ar[r]^\phi & \Omega \ar[d]^U\\
(Z,z) \ar[r]^f & (X,x).}$$ Then, we have $$P^*([f {\colon}(Z,{\bf z}) \to (X,{\bf x})]) = [\phi {\colon}\Sigma \to \Omega].$$
We now give a well-known consequence of these definitions.
\[P:isometric-teich\] If $P {\colon}\Sigma \to Z$ is nontrivially branched at every point of $P^{-1}({\bf z})$, then $P^* {\colon}\T(Z,{\bf z}) \to \T(\Sigma)$ is an isometric embedding. Moreover, for all $\epsilon > 0$ there exists $\epsilon' > 0$ so that $P^*(\T_\epsilon(Z,{\bf z})) \subset
\T_{\epsilon'}(\Sigma)$.
When $P$ is a covering then $P^*$ is an isometric embedding; see [@rafischleimer Section 7]. The proof is identical in the presence of nontrivial branching, as a one-prong singularity at a point of ${\bf z}$ lifts to a regular point or to a three-prong or higher singularity.
Let $\widetilde \M(Z, {\bf z})$ be the quotient of $\T(Z,{\bf z})$ by the group of mapping classes of $(Z,{\bf z})$ that lift to $\Sigma$. Note that $\widetilde \M (Z,{\bf z}) \to \M(Z,{\bf z})$ is a finite sheeted (orbifold) covering. The embedding $P^* {\colon}\T(Z,{\bf z})
\to \T(\Sigma)$ descends to a map $\widetilde \M (Z,{\bf z}) \to
\M(\Sigma)$, giving a commutative square. $$\xymatrix{ \T(Z,{\bf z}) \ar[r]^{P^*} \ar[d] & \T(\Sigma) \ar[d]\\
\widetilde \M(Z,{\bf z}) \ar[r] & \M(\Sigma)}$$ By Mumford’s compactness criteria [@bersmumford], the image of $\T_\epsilon(Z,{\bf z})$ in $\widetilde \M(Z,{\bf z})$ is compact, and hence so is the image in $\M(\Sigma)$. Appealing to Mumford’s criteria again (for $\M(\Sigma)$), it follows that for some $\epsilon'
> 0$ we have $P^*(\T_\epsilon(Z,{\bf z})) \subset
\T_{\epsilon'}(\Sigma)$.
In general, for any branched cover $P {\colon}\Sigma \to Z$, branched over ${\bf z} \subset Z$, consider ${\bf \sigma} = P^{-1}({\bf z})$ as a set of marked points on $\Sigma$. Then again there is an isometric embedding $$P^* {\colon}\T(Z, {\bf z}) \to \T(\Sigma, {\bf \sigma}).$$
If ${\bf \omega} \subset {\bf \sigma}$ then define $\Pi_\omega {\colon}\T(\Sigma, {\bf \sigma}) \to \T(\Sigma,{\bf \omega})$ by forgetting the points of ${\bf \sigma}$ not in ${\bf \omega}$. When ${\bf
\omega}$ is empty we may omit the subscript. In this notation, the composition $\Pi \circ P^*$ gives the map of [Proposition \[P:isometric-teich\]]{}. So, if $P$ is non-trivially branched at all points of ${\bf \sigma}$ then $\Pi \circ P^*$ is an isometric embedding. If $P$ is not branched at all points of ${\bf \sigma}$ then $\Pi \circ P^*$ fails to be an isometric embedding; however it remains $1$–Lipschitz.
\[P:1-lip-teich\] If $P {\colon}\Sigma \to Z$ is branched over ${\bf z}$ and if ${\bf \omega} \subset {\bf \sigma} = P^{-1}({\bf z})$ is any subset then $$\Pi_\omega \circ P^* {\colon}\T(Z,{\bf z}) \to \T(\Sigma,{\bf \omega})$$ is $1$–Lipschitz.
The Bers fibration is a holomorphic map [@bersfiber] and, by forgetting the points of ${\bf \sigma} - {\bf \omega}$ one at a time, we see that $\Pi_{\bf \omega} {\colon}\T(\Sigma, {\bf \sigma}) \to
\T(\Sigma,{\bf \omega})$ is a composition of holomorphic maps, hence holomorphic. In particular, because the Teichmüller metric agrees with the Kobayashi metric [@royden], it follows that $\Pi_{\bf
\omega}$ is $1$–Lipschitz [@kobayashihyperbolic]. Since $P^*$ is an isometric embedding, the composition is $1$–Lipschitz.
An inductive construction {#S:proof}
=========================
The proof of Theorem \[T:main\] is constructive, but also appeals to an inductive procedure. We begin by constructing the required embedding of ${\mathbb{H}}^2$ into some Teichmüller space as the base case of the induction, then produce an embedding of ${\mathbb{H}}^3$ into some other Teichmüller space, then an embedding of ${\mathbb{H}}^4$, and so on. All the main ideas and technical difficulties are present in the construction of the embedding of ${\mathbb{H}}^2$ and then the embedding of ${\mathbb{H}}^3$ from that of ${\mathbb{H}}^2$. The only further complications which arise to describe the embedding of ${\mathbb{H}}^n$ from ${\mathbb{H}}^{n-1}$ for $n \geq 4$ are in the notation, which becomes increasingly messy as $n$ increases. This is due to the fact that the proof for $n$ really depends on the proof for all $2 \leq k < n$ (rather than just $n-1$). For this reason, we carefully describe the cases $n=2$ and $n=3$, and sketch the general inductive step indicating only those things that require modification.
The hyperbolic plane case {#S:H^2}
-------------------------
Let $Z$ be a closed hyperbolic surface. Let $q \in \Q(Z)$ be a nonzero holomorphic quadratic differential on $Z$ so that the associated Teichmüller geodesic $[g_t {\colon}Z \to Z_t]$ is thick. Write ${\mathcal{F}}= {\mathcal{F}}(q)$ and ${\mathcal{G}}= {\mathcal{G}}(q)$ for the vertical and horizontal foliations of $q$, respectively. Next, let $c {\colon}{\mathbb{R}}\to Z$ be a nonsingular leaf of ${\mathcal{F}}$ parameterized by arc-length with respect to $q$ and let $z = c(0)$ be a marked point on $Z$; see [Section \[S:teichmuller\]]{}.
Our goal is to construct an almost-isometric embedding $${\bf Z} {\colon}{\mathbb{H}}^2 \to \T(Z,z).$$ We consider an isotopy $Z \times {\mathbb{R}}\to Z$, written $(w,x)
\mapsto f^x(w)$, where $f^x {\colon}Z \to Z$ is a homeomorphism for all $x \in
{\mathbb{R}}$, $f^0$ is the identity and $f^x(z) = c(x)$ for all $x \in
{\mathbb{R}}$. We further assume that $f^x$ preserves ${\mathcal{F}}$ for all $x \in {\mathbb{R}}$.
We can construct such an isotopy by piecing together isotopies defined on small balls. More precisely, we start with some $\epsilon$–ball around $z$, and construct a vector field tangent to $\mathcal F$ supported in the ball with with norm identically equal to $1$ on the $\epsilon/2$ ball. The flow for time $t \in (-\epsilon/2,\epsilon/2)$ is an isotopy of the correct form. Now we repeat this for a ball around $c(\epsilon/2)$. Since the arc of $c$ from $z$ to any point $c(x)$ is compact, we can cover it with finitely many such balls to produce the required isotopy.
We think of the isotopy as “pushing $z$ along $c$”. This determines the [*horocyclic coordinate*]{} $$\widetilde c {\colon}{\mathbb{R}}\to \T(Z,z)$$ given by $$\widetilde c (x) = [f^x {\colon}(Z,z) \to (Z,c(x))].$$ The image of $\widetilde c$ lies in the Bers fiber over the basepoint $[{\mathop{\rm Id}}{\colon}Z \to Z] \in \T(Z)$; the fiber is identified with the universal cover $\widetilde Z$ of $Z$. As such, we can identify $\widetilde c$ with a lift of $c$ to $\widetilde Z$ and write $$\widetilde c {\colon}{\mathbb{R}}\to \widetilde Z \subset \T(Z,z).$$
Applying the Teichmüller mapping $g_t {\colon}Z \to Z_t$ determined by $q$ and $t \in {\mathbb{R}}$ gives the [*height coordinate*]{}. These coordinates together define ${\bf Z} {\colon}{\mathbb{H}}^2 \to \T(Z,z)$ where $${\bf Z}(x,t) = [g_t \circ f^x {\colon}(Z,z) \to (Z_t,g_t(c(x)))].$$ Here we are using the coordinates $(x,t)$ on ${\mathbb{H}}^2$ described in [Section \[S:hyperbolic\]]{}.
Since the marking homeomorphisms are determined by $x$ and $t$, we simplify notation and denote the values in Teichmüller space by $$\label{E:alternatemap}
{\bf Z}(x,t) = \widetilde c_t(x) = (Z_t,g_t(c(x))).$$ We also write $${\bf Z}(x,0) = \widetilde c(x) = (Z,c(x)).$$ As the notation suggests, $\widetilde c_t {\colon}{\mathbb{R}}\to
\widetilde Z_t \subset \T(Z,z)$ is a lift of $g_t \circ c {\colon}\mathbb
R \to Z_t$ to the universal cover $\widetilde Z_t$, thought of as the fiber over $[g_t {\colon}Z \to Z_t]$.
\[T:H2\_isometric\] The map ${\bf Z} {\colon}{\mathbb{H}}^2 \to \T(Z,z)$ is an almost-isometric embedding. Moreover, the image lies in the thick part and is quasi-convex.
We verify the hypothesis of Lemma \[L:criteria\] to prove that ${\bf
Z}$ is an almost-isometric embedding and, along the way, prove that the image is quasi-convex and lies in the thick part.
First, fix any $x \in {\mathbb{R}}$ so that $\eta_x(t) = (x,t)$ is a vertical geodesic in ${\mathbb{H}}^2$. Then $t \mapsto {\bf Z} \circ
\eta_x(t) = {\bf Z}(x,t) = (Z_t,g_t(c(x)))$ is a Teichmüller geodesic, and hence Property \[L:crit:geod\] of [Lemma \[L:criteria\]]{} holds. Furthermore, since $t \mapsto Z_t$ is a thick geodesic, we see that $\{{\bf Z} \circ \eta_x(t)\}_{x \in {\mathbb{R}}}$ are uniformly thick geodesics. That is, that the union of these geodesics, over all $x \in {\mathbb{R}}$, project into a compact subset of $\M(Z,z)$; namely, the preimage of the compact subset of $\M(Z)$ containing the image of $t \mapsto Z_t$ (see [Section \[S:forget\]]{}). In particular, the image of ${\bf Z}$ lies in the thick part of $\T(Z,z)$.
For each $x \in {\mathbb{R}}$, the geodesic ${\bf Z} \circ \eta_x$ is defined by the quadratic differential $q \in \Q(Z)$ viewed as a quadratic differential in $\Q(Z,c(x))$. We denote the vertical and horizontal foliations of $q \in \Q(Z,c(x))$ by ${\mathcal{F}}^x$ and ${\mathcal{G}}^x$, respectively, and consider them as measured foliations in $\MF(Z,z)$ by pulling them back via $f^x$. Since $f^x$ preserves ${\mathcal{F}}$, it follows that ${\mathcal{F}}^x = {\mathcal{F}}^0 \in
\MF(Z,z)$ for all $x \in {\mathbb{R}}$.
Now, since $t \mapsto Z_t$ is a thick geodesic, by [Theorem \[T:Masur-ue\]]{} the foliations ${\mathcal{F}}$ and ${\mathcal{G}}$ are arational. Puncturing an arational foliation once gives an arational foliation in the punctured surface. Hence ${\mathcal{F}}^x$ and ${\mathcal{G}}^x$ are also arational for all $x$. Since ${\mathcal{F}}^x = {\mathcal{F}}^0$ for all $x \in {\mathbb{R}}$ and since the geodesics $\{{\bf Z} \circ \eta_x\}_{x \in {\mathbb{R}}}$ are uniformly thick, [Theorem \[T:Rafi\]]{} implies that there exists $K > 0$ so that the pairs $({\mathcal{F}}^0,{\mathcal{G}}^x) = ({\mathcal{F}}^x, {\mathcal{G}}^x)$ are $K$–cobounded for all $x$. By the triangle inequality (applied to each subsurface $Y$) we see that for all $x, x' \in {\mathbb{R}}$ the pair $({\mathcal{G}}^x, {\mathcal{G}}^{x'})$ is $2K$–cobounded (to see that ${\mathcal{G}}^x$ and ${\mathcal{G}}^{x'}$ are different foliations, note that $({\mathcal{F}}^0,{\mathcal{G}}^x)$ and $({\mathcal{F}}^0,{\mathcal{G}}^{x'})$ define different geodesics ${\bf Z} \circ \eta_x$ and ${\bf Z} \circ
\eta_{x'}$, respectively).
Appealing to the other direction in [Theorem \[T:Rafi\]]{} the geodesic $\Gamma^{x,x'}$, determined by ${\mathcal{G}}^x$ and ${\mathcal{G}}^{x'}$ for distinct $x, x' \in {\mathbb{R}}$, is uniformly thick, independent of $x$ and $x'$. From this and [@shadows Theorem 4.4] it follows that there is a $\delta > 0$ so that ${\bf Z}\circ \eta_x$, ${\bf Z} \circ \eta_{x'}$ and $\Gamma^{x,x'}$ are the sides of a $\delta$–slim triangle for every pair of distinct points $x,x' \in {\mathbb{R}}$, and hence Property \[L:crit:slim\] of [Lemma \[L:criteria\]]{} holds. From this, it follows that ${\bf Z}({\mathbb{H}}^2)$ (is contained in and) has Hausdorff distance at most $\delta$ from the union of the geodesics $${\bf Z}({\mathbb{H}}^2)
\cup {\mathopen{}\mathclose\bgroup\originalleft}( \bigcup_{x \neq x' \in {\mathbb{R}}} \Gamma^{x,x'} {\aftergroup\egroup\originalright})
=
{\mathopen{}\mathclose\bgroup\originalleft}( \bigcup_{x \in {\mathbb{R}}} {\bf Z} \circ \eta_x {\aftergroup\egroup\originalright})
\cup {\mathopen{}\mathclose\bgroup\originalleft}( \bigcup_{x \neq x' \in {\mathbb{R}}} \Gamma^{x,x'}{\aftergroup\egroup\originalright}).$$ This is precisely the [*weak hull*]{} of $\{{\mathcal{G}}^x\}_{x \in
{\mathbb{R}}} \cup \{{\mathcal{F}}^0\} \subset \PMF(Z,z)$, and so according to [@shadows Theorem 4.5], this set, hence also ${\bf Z}(\mathbb
H^2)$, is quasi-convex (the assumption in [@shadows] that the subset of $\PMF(Z)$ be closed was not used in the proof).
Finally, we must prove that Properties \[L:crit:unif\] and \[L:crit:prop\] of Lemma \[L:criteria\] hold. For this we can appeal directly to Theorem \[T:kra\]. More precisely, observe that because $\{Z_t\}_{t \in {\mathbb{R}}}$ lies in the thick part, the pull-back of the flat metric on $\widetilde Z_t$ (which we also denote $q_t$) is uniformly quasi-isometric to the Poincaré metric $\rho_0$ on $\widetilde Z_t$. That is, there exist constants $A,B \geq 0$ so that $$\label{E:flathyperbolic}
\frac{1}{A} {\mathopen{}\mathclose\bgroup\originalleft}( d_{q_t} (\widetilde z,\widetilde z') - B {\aftergroup\egroup\originalright})
\leq \rho_0(\widetilde z,\widetilde z') \leq A \, d_{q_t}(\widetilde
z,\widetilde z') + B$$ for all $t \in {\mathbb{R}}$ and $\widetilde z,\widetilde z' \in Z_t$ (see for example [@FMcc Lemma 2.2]).
Applying , the upper bound of [Theorem \[T:kra\]]{}, and , in that order, we find $$\begin{aligned}
d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf Z}(x,t),{\bf Z}(x',t){\aftergroup\egroup\originalright})
& = d_\T(\widetilde c_t(x),\widetilde c_t(x'))\\
& \leq \rho_0(\widetilde c_t(x),\widetilde c_t(x'))\\
& \leq A \, d_{q_t}(\widetilde c_t(x),\widetilde c_t(x')) + B\\
& = A e^{-t} |x'-x| + B.\end{aligned}$$ So, setting $\epsilon = 1$ and $R = A + B$, Property \[L:crit:unif\] of [Lemma \[L:criteria\]]{} holds.
On the other hand, , the lower bound of Theorem \[T:kra\], monotonicity of $h$, and gives $$\begin{aligned}
d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf Z}(x,t),{\bf Z}(x',t){\aftergroup\egroup\originalright})
& = d_\T(\widetilde c_t(x),\widetilde c_t(x'))\\
& \geq h(\rho_0(\widetilde c_t(x),\widetilde c_t(x')))\\
& \geq h {\mathopen{}\mathclose\bgroup\originalleft}( \frac{1}{A} {\mathopen{}\mathclose\bgroup\originalleft}( d_{q_t}(\widetilde c_t(x),\widetilde c_t(x')) - B {\aftergroup\egroup\originalright}) {\aftergroup\egroup\originalright})\\
& = h {\mathopen{}\mathclose\bgroup\originalleft}( \frac{1}{A} (e^{-t}|x' - x| - B) {\aftergroup\egroup\originalright}).\end{aligned}$$ From this, and because $h$ is a homeomorphism on $[0,\infty)$ and hence proper, Property \[L:crit:prop\] of Lemma \[L:criteria\] also holds. This completes the proof of [Theorem \[T:H2\_isometric\]]{}.
Hyperbolic $3$–space {#S:H^3}
--------------------
Before diving into the construction, we explain the basic idea. Our embedding of the hyperbolic plane in [Section \[S:H\^2\]]{} sends $(x,t)$ to ${\bf Z}(x,t) \in \T(Z,z)$ by pushing the marked point $z$ distance $x$ along a leaf of the vertical foliation of a quadratic differential then travelling distance $t$ along the Teichmüller flow. There is a simple extension of this construction which produces a map of hyperbolic $3$–space into Teichmüller space $\T(Z,\{z,w\})$. Take $z$ and $w$ to lie on distinct leaves and send $(x,y,t)$ to the point of $\T(Z,\{z,w\})$ obtained by pushing $z$ a distance $x$ along its leaf, pushing $w$ a distance $y$ along its leaf, and applying the Teichmüller flow for time $t$.
The problem is that whenever $z$ and $w$ move close to each other on $Z$, the corresponding point in $\T(Z,\{z,w\})$ is in the thin part of Teichmüller space; if $z$ and $w$ are very close to each other then there is a simple closed curve surrounding $z$ and $w$ having an annular neighborhood of large modulus. This also shows that this map $(x,y,t) \mapsto \T(Z,\{z,w\})$ is not a quasi-isometric embedding. In fact the map is not even coarsely Lipschitz.
A more subtle construction is required. We first choose a branched cover $P {\colon}\Sigma \to Z$, nontrivially branched at each point of $P^{-1}(z)$. According to [Proposition \[P:isometric-teich\]]{}, this induces an isometric embedding of $\T(Z,z)$ into $\T(\Sigma)$. Fix a suitably generic point $w \in (Z, z)$ and pick a point $\sigma \in P^{-1}(w)$. Roughly, we map our three parameters $(x,y,t)$ into $\T(\Sigma,\sigma)$ as follows. The coordinates $(x,t)$ determine ${\bf Z}(x,t) \in \T(Z,z)$ as in [Section \[S:H\^2\]]{}. The map $P^*$ applied to ${\bf Z}(x,t)$ gives a point in $\T(\Sigma)$ as in [Section \[S:branched\]]{}. Finally use $y$ to determine a point ${\bf
\Sigma}(x,y,t) \in \T(\Sigma, \sigma)$, lying in the Bers fiber above $P^* \circ {\bf Z}(x,t)$. On its face, this new construction avoids the problem we had before. In $(Z, z)$ we have only one marked point; after taking the branched covering over $z$ [*we forget all of the branch points over $z$*]{}. The single image of $\sigma$ can now move freely enough so that we stay in the thick part of $\T(\Sigma,
\sigma)$. We now explain this construction in more detail and prove that the resulting map has all the required properties.
### The construction
The notation from Section \[S:H\^2\] carries over to this section without change. Let $P {\colon}\Sigma \to Z$ be a branched cover, branched over the marked point $z \in Z$ so that $P$ is nontrivially branched at every point of $P^{-1}(z)$. This determines an isometric embedding of Teichmüller spaces $$P^* {\colon}\T(Z,z) \to \T(\Sigma)$$ by Proposition \[P:isometric-teich\]. We write $$P^*([g_t \circ f^x {\colon}(Z,z) \to (Z_t,g_t(c(x)))]) = [\phi_t^x {\colon}\Sigma \to
\Sigma_t^x]$$ so that $\phi_t^x$ is a lift of the marking $g_t \circ f^x$, and $P_t^x$ is the induced branched cover making the following commute: $$\xymatrix{
\Sigma \ar[d]_P \ar[rr]^{\phi_t^x} & & \Sigma_t^x \ar[d]^{P_t^x}\\
(Z,z) \ar[rr]^{g_t \circ f^x} & & (Z_t,g_t(c(x))).}$$ The quadratic differentials $q_t$ pull back to quadratic differentials $\lambda_t^x$ on $\Sigma_t^x$, and $g_t$ lifts to Teichmüller mappings of the covers $$\psi_t^x {\colon}\Sigma_0^x \to \Sigma_t^x$$ so that $t \mapsto \Sigma_t^x$ is a Teichmüller geodesic for all $x$. The lifts satisfy $\phi_t^x = \psi_t^x \circ \phi_0^x$. We have another commutative diagram which may be helpful in organizing all the maps: $$\xymatrix{
\Sigma \ar[d]_P \ar[rr]^{\phi_0^x} & & \Sigma_0^x
\ar[rr]^{\psi_t^x} \ar[d]^{P_0^x} & & \Sigma_t^x \ar[d]^{P_t^x}\\
(Z,z) \ar[rr]^{f^x} & & (Z,c(x))
\ar[rr]^{g_t} & & (Z_t,g_t(c(x))).}$$
Denote the vertical foliation for $\lambda_t^x$ by $\Phi_t^x$. Each nonsingular leaf of $\Phi_t^x$ maps isometrically to a nonsingular leaf of the vertical foliation ${\mathcal{F}}_t$ for $q_t$ via the branched covering $\Sigma_t^x \to Z_t$ since $\lambda_t^x$ is the pull back of $q_t$. Choose any nonsingular leaf $\gamma_0^0 {\colon}{\mathbb{R}}\to
\Sigma_0^0 = \Sigma$, parameterized by arc-length. Observe that $\gamma^0_0$ maps isometrically by $P$ to a leaf $\gamma {\colon}\mathbb
R \to Z$ for ${\mathcal{F}}$. Note that $c$ and $\gamma$ are distinct leaves; the preimage of $c$ in $\Sigma$ consists entirely of singular leaves, namely the leaves that meet the branch points of $P$.
As we vary $x$, we can continuously choose lifts of $\gamma$ to leaves $\gamma_0^x {\colon}{\mathbb{R}}\to \Sigma_0^x$ which agrees with our initial leaf $\gamma_0^0$ when $x = 0$. Specifically, we define the lift to be $$\gamma_0^x
= \phi_0^x \circ {\mathopen{}\mathclose\bgroup\originalleft}( P|_{\gamma_0^0({\mathbb{R}})}{\aftergroup\egroup\originalright})^{-1}
\circ (f^x)^{-1} \circ \gamma.$$ Composing with the lifts $\psi_t^x$, we obtain leaves $\gamma_t^x =
\psi_t^x \circ \gamma_0^x$. Observe that via the branched covering $P_t^x {\colon}\Sigma_t^x \to Z_t$, $\gamma_t^x$ projects to the leaf $g_t
\circ \gamma$, independent of $x$. Furthermore, this shows that the $\lambda_t^x$–length of the arc $\gamma_t^x([y,y'])$ is the $q_t$–length of $g_t \circ \gamma$ which is $e^{-t}|y-y'|$.
We pick a basepoint $\sigma = \gamma_0^0(0) \in \Sigma$, and consider the surface $(\Sigma,\sigma)$, marked by the identity ${\mathop{\rm Id}}= \phi_0^0$ as a point in $\T(\Sigma,\sigma)$. Just as we constructed $f^x$ by pushing along $c$ to $c(x)$, we push $\sigma$ along $\gamma_t^x$ to $\gamma_t^x(y)$ to obtain maps $$\xi_t^{x,y} {\colon}(\Sigma,\sigma) \to (\Sigma_t^x,\gamma_t^x(y)).$$ Specifically, we take $\xi_0^{x,y}$ to be the composition of $\phi_0^x$ and a map isotopic to the identity on $\Sigma_0^x$ which preserves the foliation $\Phi_0^x$ and pushes $\phi_0^x(\sigma)$ along $\gamma_0^x$ to $\gamma_0^x(y)$. Then $\xi_t^{x,y} = \psi_t^x \circ \xi_0^{x,y}$ maps the foliation $\Phi_0^0$ to $\Phi_t^x$.
We denote the associated point in Teichmüller space $[\xi_t^{x,y} {\colon}(\Sigma,\sigma) \to (\Sigma_t^x,\gamma_t^x(y))] \in
\T(\Sigma,\sigma)$ simply by $(\Sigma_t^x,\gamma_t^x(y))$ as this point is uniquely determined in this construction by $(x,y,t)$.
We define $${\bf \Sigma} {\colon}{\mathbb{H}}^3 \to \T(\Sigma,\sigma)$$ in the coordinates $(x,y,t)$ for ${\mathbb{H}}^3$ from Section \[S:hyperbolic\] by $${\bf \Sigma}(x,y,t) = (\Sigma_t^x,\gamma_t^x(y)).$$
### Fibration over ${\mathbb{H}}^2$ case
We also require a slightly different description of the map ${\bf
\Sigma}$ to take advantage of the construction in the ${\mathbb{H}}^2$ case. Observe that $P^* \circ {\bf Z} {\colon}{\mathbb{H}}^2 \to \T(Z,z) \to
\T(\Sigma)$ is an almost-isometric embedding, and is given by $$P^* {\mathopen{}\mathclose\bgroup\originalleft}( {\bf Z}(x,t) {\aftergroup\egroup\originalright}) = \Sigma_t^x,$$ where $\Sigma_t^x$ denotes the point $[\phi_t^x {\colon}\Sigma \to
\Sigma_t^x]$. Recall that $$\Pi {\colon}\T(\Sigma,\sigma) \to \T(\Sigma).$$ is the Bers fibration. If we fix $(x,t) \in {\mathbb{H}}^2$, then for every $y$ we see that $(\Sigma_t^x,\gamma_t^x(y))$ is contained the fiber $\Pi^{-1}(\Sigma_t^x)$. Since $\Pi^{-1}(\Sigma_t^x)$ is identified with the universal covering $\widetilde \Sigma_t^x$ of $\Sigma_t^x$, just as in the case of ${\mathbb{H}}^2$ we see that $t
\mapsto (\Sigma_t^x,\gamma_t^x(y))$ is a lift of $\gamma_t^x$ to $\widetilde \Sigma_t^x \subset \T(\Sigma,\sigma)$. As such, we use the alternative notation $$\widetilde \gamma_t^x {\colon}{\mathbb{R}}\to \widetilde \Sigma_t^x \subset
\T(\Sigma,\sigma)$$ with $$\widetilde \gamma_t^x(y) = (\Sigma_t^x,\gamma_t^x(y))$$ when it is convenient to do so.
Finally we record the equation $$\label{E:fibration}
\Pi \circ {\bf \Sigma}(x,y,t) = P^* \circ {\bf Z}(x,t)$$ which holds for all $(x,y,t) \in {\mathbb{H}}^3$. The fact that $\Pi$ is $1$–Lipschitz and $P^* \circ {\bf Z}$ is an almost-isometric embedding provides us with useful metric information about ${\bf
\Sigma}$.
\[T:H3\_isometric\] The map ${\bf \Sigma} {\colon}{\mathbb{H}}^3 \to \T(\Sigma, \sigma)$ is an almost-isometric embedding. Moreover, the image lies in the thick part and is quasi-convex.
As before, we will verify the hypothesis of [Lemma \[L:criteria\]]{} to prove that ${\bf \Sigma}$ is an almost-isometry and, along the way, prove that the image is quasi-convex and lies in the thick part.
For all $(x,y) \in {\mathbb{R}}^2$, the geodesic $\eta_{(x,y)}(t)$ in ${\mathbb{H}}^3$ is sent to $${\bf \Sigma} \circ \eta_{(x,y)}(t) = {\mathopen{}\mathclose\bgroup\originalleft}( \Sigma_t^x,\gamma_t^x(y) {\aftergroup\egroup\originalright})
= {\mathopen{}\mathclose\bgroup\originalleft}( \psi_t^x(\Sigma_0^x),\psi_t^x(\gamma_0^x(y)) {\aftergroup\egroup\originalright}).$$ This is a geodesic in $\T(\Sigma,\sigma)$ because $\psi_t^x {\colon}\Sigma_0^x \to \Sigma_t^x$ is a Teichmüller mapping; thus Property \[L:crit:geod\] follows. Furthermore, note that ${\bf
\Sigma} \circ \eta_{(x,y)}(t)$ lies over $P^*\circ {\bf Z} \circ
\eta_x(t)$ for all $(x,y,t)$. Since $P$ is nontrivially branched over every point, the uniform thickness of the set of geodesics $\{{\bf Z}
\circ \eta_x(t)\}_{x \in {\mathbb{R}}}$ implies the same for $\{ P^* \circ {\bf
Z} \circ \eta_x(t)\}_{x \in {\mathbb{R}}}$ by Proposition \[P:isometric-teich\], and hence also for $\{ {\bf \Sigma} \circ
\eta_{(x,y)}(t) {\mathbin{\mid}}(x,y) \in {\mathbb{R}}^2 \}$ by as discussed in Section \[S:forget\]. That is, ${\bf \Sigma}({\mathbb{H}}^3)$ lies in the thick part.
By our choice of maps $\xi_0^{x,y}$, if we pull back the vertical foliation $\Phi_0^x$ of $\lambda_0^x$ to a foliation $\Phi_0^{x,y} \in
\MF(\Sigma,\sigma)$ the result is independent of $x$ and $y$. Furthermore, Theorem \[T:Masur-ue\] implies that these foliations, as well as the pull backs of the horizontal foliations, are arational. Thus all strict subsurface projection distances are defined. [Theorem \[T:Rafi\]]{} and the results of [@shadows] can be applied as in the ${\mathbb{H}}^2$ case to prove that Property \[L:crit:slim\] of [Lemma \[L:criteria\]]{} is satisfied for some $\delta > 0$. Furthermore, ${\bf \Sigma}({\mathbb{H}}^3)$ is quasi-convex.
We now come to the subtle point of the proof, which is verifying Properties \[L:crit:unif\] and \[L:crit:prop\] of Lemma \[L:criteria\]. We start with Property \[L:crit:unif\].
There exists $\epsilon > 0$ and $R > 0$ so that if $e^{-t}
{\mathopen{}\mathclose\bgroup\originalleft}| (x,y)-(x',y') {\aftergroup\egroup\originalright}| < \epsilon$ then $$d_\T {\mathopen{}\mathclose\bgroup\originalleft}( {\bf \Sigma}(x,y,t),{\bf \Sigma}(x',y',t) {\aftergroup\egroup\originalright}) < R.$$
Before we give the proof, we briefly explain the core technical difficulty. Fix $t$ and define $C_x = g_t(c(x))$ and $\Gamma_y =
g_t(\gamma(y))$. Observe that, as before, when we vary $y$ we are simply point pushing; thus the change in Teichmüller distance is controlled by Theorem \[T:kra\]. On the other hand, varying $x$ means that we are varying the conformal stucture on the closed surface $\Sigma_t^x$. This is obtained by varying $x$ in $(Z_t, C_x)$ (which is also point pushing) then taking a branched cover. However, while we vary $C_x$ in $Z_t$ we must also keep track of our $y$ coordinate, which means we should also project $\gamma_t^x(y)$ down to $Z_t$—this is precisely the point $\Gamma_y$. Now if $C_x$ and $\Gamma_y$ are close together and we vary $x$ so as to push these points apart, then this can result in a large distance in the “auxiliary” Teichmüller space $\T(Z, \{z, w\})$, even for small variation of $x$. The idea is therefore to first vary $y$, if necessary, to move $\gamma_t^x(y)$ in $\Sigma_t^x$ and so guaranteeing that $\Gamma_y$ is not too close to $C_x$. We can then vary $x$ as required, then vary $y$ back to its original value. Since the variation of $y$ can be carried out independent of $x$, this will result in a uniformly bounded change in Teichmüller distance.
Since the surfaces $\{\Sigma_t^x\}_{t,x \in {\mathbb{R}}}$ lie in the thick part, the (pulled back) metrics $\lambda_t^x$ and the Poincaré metric(s) $\rho_0$ on the universal cover $\widetilde \Sigma_t^x$ are uniformly comparable. That is, there exist constants $A$ and $B$ so that for all $\widetilde \sigma,\widetilde \sigma' \in \widetilde
\Sigma_t^x$ $$\label{E:flathyperbolic2}
\frac{1}{A} {\mathopen{}\mathclose\bgroup\originalleft}( d_{\lambda_t^x} (\widetilde \sigma,\widetilde \sigma') - B {\aftergroup\egroup\originalright})
\leq \rho_0 {\mathopen{}\mathclose\bgroup\originalleft}( \widetilde \sigma,\widetilde \sigma' {\aftergroup\egroup\originalright}) \leq A \,
d_{\lambda_t^x} {\mathopen{}\mathclose\bgroup\originalleft}( \widetilde \sigma,\widetilde \sigma' {\aftergroup\egroup\originalright}) + B.$$ Applying Theorem \[T:kra\], Equations and we have $$\begin{aligned}
\label{E:halfupper3}
d_\T {\mathopen{}\mathclose\bgroup\originalleft}( {\bf \Sigma}(x,y,t),{\bf \Sigma}(x,y',t) {\aftergroup\egroup\originalright})
& = d_\T {\mathopen{}\mathclose\bgroup\originalleft}( \widetilde \gamma_t^x(y),\widetilde \gamma_t^x(y') {\aftergroup\egroup\originalright}) \\ \nonumber
& \leq \rho_0 {\mathopen{}\mathclose\bgroup\originalleft}( \widetilde \gamma_t^x(y),\widetilde \gamma_t^x(y') {\aftergroup\egroup\originalright}) \\ \nonumber
& \leq A d_{\lambda_t^x} {\mathopen{}\mathclose\bgroup\originalleft}( \widetilde \gamma_t^x(y),\widetilde \gamma_t^x(y') {\aftergroup\egroup\originalright}) + B \\ \nonumber
& = A {\mathopen{}\mathclose\bgroup\originalleft}( e^{-t} {\mathopen{}\mathclose\bgroup\originalleft}| y - y' {\aftergroup\egroup\originalright}| {\aftergroup\egroup\originalright}) + B. \nonumber\end{aligned}$$
We now fix $t$ and the notation $C_x = g_t(c(x))$, $\Gamma_y =
g_t(\gamma(y))$. To understand the effect of varying $x$ we must consider the branched covering $P_t^x {\colon}\Sigma_t^x \to (Z_t,
C_x)$, but also keep track of the image of our marked point $\gamma_t^x(y) = \psi_t^x(\gamma_0^x(y))$ down in $(Z_t, C_x)$; that is, the point $\Gamma_y$. This results in the surface $Z_t$ with [*two marked points*]{}: $$(Z_t, \{C_x, \Gamma_y\}).$$ Appealing to Proposition \[P:1-lip-teich\] we have $$\label{E:branchedforgetbounds}
d_\T {\mathopen{}\mathclose\bgroup\originalleft}( {\bf \Sigma}(x,y,t),{\bf \Sigma}(x',y',t) {\aftergroup\egroup\originalright})
\leq d_\T {\mathopen{}\mathclose\bgroup\originalleft}( (Z_t, \{C_x,\Gamma_y\}), (Z_t,\{C_{x'},\Gamma_{y'}\}) {\aftergroup\egroup\originalright}).$$ This is because we are taking a branched covering, $\Sigma_t^x \to
Z_t$, and then forgetting all but one of the marked points in $\Sigma_t^x$.
Since $Z_t$ lies in some fixed thick part of $\T(Z)$ for all $t \in
{\mathbb{R}}$, there exists $\epsilon > 0$ so that the $2\epsilon$–ball about $C_x$ in the $q_t$ metric, $B_{q_t}(C_x, 2 \epsilon)$ is a disk for all $t,x \in {\mathbb{R}}$ (that is, we have a lower bound on the $q_t$–injectivity radius of $Z_t$, independent of $t$). Now suppose $\Gamma_y$ lies outside this ball $$\Gamma_y \not\in B_{q_t}(C_x, 2 \epsilon).$$ Using again the fact that $Z_t$ lies in some thick part of $\T(Z)$ for all $t \in {\mathbb{R}}$, it follows that there is some $R' > 0$ with the property that for any point $z' \in B_{q_t}(C_x, \epsilon)$ we have $$d_\T {\mathopen{}\mathclose\bgroup\originalleft}( (Z_t,\{C_x, \Gamma_y\}),(Z_t,\{z',\Gamma_y\}) {\aftergroup\egroup\originalright})
< R'.$$ Here the marking homeomorphism for $(Z_t,\{z',\Gamma_y\})$ is assumed to differ from that of $(Z_t,\{C_x,\Gamma_y\})$ by composition with a homeomorphism of $Z_t$ that is the identity outside $B_{q_t}(C_x,2\epsilon)$. In particular, if $e^{-t}|x-x'| < \epsilon$ and, crucially, $\Gamma_y \not\in B_{q_t}(C_x,2\epsilon)$ then deduce that $C_{x'} \in B_{q_t}(C_x,\epsilon)$ and, from Equation , that $$\begin{aligned}
\label{E:movexbound}
d_\T {\mathopen{}\mathclose\bgroup\originalleft}({\bf \Sigma}(x,y,t),{\bf \Sigma}(x',y,t) {\aftergroup\egroup\originalright})
& \leq d_\T {\mathopen{}\mathclose\bgroup\originalleft}((Z_t,\{C_x,\Gamma_y\}),(Z_t,\{C_{x'}, \Gamma_y\}) {\aftergroup\egroup\originalright})
< R'.\end{aligned}$$
On the other hand, because the leaves of ${\mathcal{F}}$ are geodesics for $q_t$ and because $B_{q_t}(C_x, 2 \epsilon)$ is a disk, if $\Gamma_y \in B_{q_t}(C_x, 2 \epsilon)$ then there exists $y' \in
{\mathbb{R}}$ so that $e^{-t}|y'-y| \leq 2 \epsilon$ and $$\Gamma_{y'} \not\in B_{q_t}(C_x, 2\epsilon).$$ Then, from we have $$\begin{aligned}
d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Sigma}(x,y',t),{\bf \Sigma}(x',y',t) {\aftergroup\egroup\originalright})
& \leq d_\T{\mathopen{}\mathclose\bgroup\originalleft}((Z_t, \{C_x,\Gamma_{y'}\}) , (Z_t, \{ C_{x'}, \Gamma_{y'}\}){\aftergroup\egroup\originalright})
< R'.\end{aligned}$$
Combining this, inequalities and , and the triangle inequality, it follows that for any $x,y,x',t$ with $e^{-t}|x-x'| \leq \epsilon$ there is some $y' \in
{\mathbb{R}}$ with $e^{-t}|y'-y| \leq 2 \epsilon$ such that $$\begin{aligned}
\label{E:bigmovexbound}
d_\T({\bf \Sigma}(x,y,t),{\bf \Sigma}(x',y,t))
& \leq d_\T({\bf \Sigma}(x,y,t),{\bf \Sigma}(x,y',t)) + d_\T({\bf \Sigma}(x,y',t),{\bf \Sigma}(x',y',t)) \\ \nonumber
& \phantom{\leq\,} + d_\T({\bf \Sigma}(x',y',t),{\bf \Sigma}(x',y,t)) \\ \nonumber
& \leq 2(A(e^{-t}|y - y'|) + B) + R'\\ \nonumber
& < 2(A2\epsilon + B)+R'\\ \nonumber
& \leq 4(A \epsilon + B) + R'. \end{aligned}$$ Now, let $\epsilon> 0$ be as above and set $R = 5 (A \epsilon + B) +
R'$. Given $(x,y,t),(x',y',t) \in {\mathbb{H}}^3$ with $e^{-t}|(x,y)-(x',y')| < \epsilon$, then we have $e^{-t}|x-x'|,e^{-t}|y-y'| \leq e^{-t}|(x,y)-(x',y')| < \epsilon $. Applying Equations and and the triangle inequality we obtain $$\begin{aligned}
d_\T {\mathopen{}\mathclose\bgroup\originalleft}({\bf \Sigma}(x,y,t),{\bf \Sigma}(x',y',t){\aftergroup\egroup\originalright})
& \leq d_\T {\mathopen{}\mathclose\bgroup\originalleft}({\bf \Sigma}(x,y,t),{\bf \Sigma}(x',y,t){\aftergroup\egroup\originalright})
+ d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Sigma}(x',y,t),{\bf \Sigma}(x'y',t){\aftergroup\egroup\originalright}) \\
& \leq 4(A \epsilon + B) + R' + A \epsilon + B \\
& < 5(A \epsilon + B) + R' = R.\end{aligned}$$ This completes the proof of the claim, and so verifies Property \[L:crit:unif\] of [Lemma \[L:criteria\]]{}.
All that remains to show is Property \[L:crit:prop\] of [Lemma \[L:criteria\]]{}. Suppose we have a sequence of pairs $\{(x_n,y_n,t_n),(x_n',y_n',t_n)\}_{n=1}^\infty$ with $e^{t_n}|(x_n,y_n)-(x_n',y_n')| \to \infty$ as $n \to \infty$. Then, up to subsequence, we must be in one of two cases.
$e^{t_n}|x_n - x_n'| \to \infty$ as $n \to \infty$.
Forgetting the marked point is $1$–Lipschitz, and so we have $$\begin{aligned}
d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Sigma}(x_n,y_n,t_n),{\bf \Sigma}(x_n',y_n',t_n){\aftergroup\egroup\originalright})
& \geq d_\T{\mathopen{}\mathclose\bgroup\originalleft}(\Sigma_{t_n}^{x_n},\Sigma_{t_n}^{x_n'}{\aftergroup\egroup\originalright}) \\
& = d_\T{\mathopen{}\mathclose\bgroup\originalleft}((Z_{t_n},g_{t_n}(x_n)),(Z_{t_n},g_{t_n}(x_n')){\aftergroup\egroup\originalright}) \\
& = d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf Z}(x_n,t_n),{\bf Z}(x_n',t_n){\aftergroup\egroup\originalright}).\end{aligned}$$ However, we have already verified that ${\bf Z} {\colon}{\mathbb{H}}^2 \to
\T(Z,z)$ satisfies Lemma \[L:criteria\]. Therefore the last expression tends to infinity, and hence $$\lim_{n \to \infty} d_\T({\bf \Sigma}(x_n,y_n,t_n),{\bf \Sigma}(x_n',y_n',t_n)) = \infty$$ as required.
$e^{t_n}|y_n - y_n'| \to \infty$ as $n \to \infty$.
If we also have $e^{t_n}|x_n - x_n'| \to \infty$, then we can appeal to the previous case and we are done. So we assume, as we may, that $e^{t_n}|x_n-x_n'| < M$, for some constant $M > 0$. Since we have already shown that there are $\epsilon,R > 0$ so that part \[L:crit:unif\] from Lemma \[L:criteria\] holds, it follows from that $$d_\T{\mathopen{}\mathclose\bgroup\originalleft}( {\bf \Sigma}(x_n,y_n',t_n),{\bf \Sigma}(x_n',y_n',t_n){\aftergroup\egroup\originalright})
\leq \frac{R}{\epsilon} {\mathopen{}\mathclose\bgroup\originalleft}( e^{-t_n}|x_n - x_n'| {\aftergroup\egroup\originalright}) + R
\leq \frac{R}{\epsilon} M + R.$$ Now, by the triangle inequality we have $$\begin{aligned}
\label{E:readyforkra}
d_\T {\mathopen{}\mathclose\bgroup\originalleft}( {\bf \Sigma}(x_n,y_n,t_n),{\bf \Sigma}(x_n',y_n',t_n) {\aftergroup\egroup\originalright})
& \geq d_\T {\mathopen{}\mathclose\bgroup\originalleft}( {\bf \Sigma}(x_n,y_n,t_n),{\bf \Sigma}(x_n,y_n',t_n) {\aftergroup\egroup\originalright}) \\ \nonumber
& \phantom{\geq\,} - d_\T {\mathopen{}\mathclose\bgroup\originalleft}( {\bf \Sigma}(x_n,y_n',t_n),{\bf \Sigma}(x_n',y_n',t_n) {\aftergroup\egroup\originalright}) \\ \nonumber
& \geq d_\T {\mathopen{}\mathclose\bgroup\originalleft}( \widetilde{\gamma}_{t_n}^{x_n}(y_n),\widetilde{\gamma}_{t_n}^{x_n}(y_n') {\aftergroup\egroup\originalright})
- \frac{R}{\epsilon}M - R \nonumber\end{aligned}$$ We can now appeal to Theorem \[T:kra\] as in our proof for ${\bf Z}
{\colon}{\mathbb{H}}^2 \to \T(Z,z)$ to find $A,B$ so that $$d_\T(\widetilde{\gamma}_{t_n}^{x_n}(y_n),\widetilde{\gamma}_{t_n}^{x_n}(y_n'))
\geq h {\mathopen{}\mathclose\bgroup\originalleft}( \frac{1}{A}e^{-{t_n}}|y_n' - y_n| - B{\aftergroup\egroup\originalright})$$ The right-hand side tends to infinity by the properness of $h$, so we can combine this with to obtain $$\lim_{n \to \infty} d_\T({\bf \Sigma}(x_n,y_n,t_n),{\bf
\Sigma}(x_n',y_n',t_n)) = \infty$$ as required. Therefore, Property \[L:crit:prop\] from [Lemma \[L:criteria\]]{} holds, and the proof of [Theorem \[T:H3\_isometric\]]{} is complete.
The general case
----------------
The previous arguments set up an inductive scheme for producing almost-isometric embeddings of ${\mathbb{H}}^n$ into Teichmüller spaces. The idea is as follows.
For $n-1 \geq 3$, induction gives us an almost-isometric embedding ${\bf W} {\colon}{\mathbb{H}}^{n-1} \to \T(W,w)$ satisfying all the hypotheses of Lemma \[L:criteria\] for some closed surface $W$ with a marked point $w$. We again take a branched cover $$P {\colon}\Omega \to W$$ nontrivially branched over each point in $P^{-1}(w) \subset \Omega$. This determines a map $$P^*\circ {\bf W} {\colon}{\mathbb{H}}^{n-1} \to \T(\Omega).$$ Using the coordinates $(x, t) = (x_1, x_2, \ldots, x_{n-2}, t) \in
{\mathbb{H}}^{n-1}$ we write $$P^* \circ {\bf W}(x,t) = \Omega_t^x.$$
Inductively, we assume that the foliation of ${\mathbb{H}}^{n-1}$ by asymptotic geodesics $\{\eta_x(t)\}_{x \in {\mathbb{R}}^{n-2}}$ are all mapped by ${\bf
W}$ to uniformly thick geodesics in $\T(W,w)$, so the same is true for $P^* \circ W$. These geodesics are obtained by applying the Teichmüller mapping $\psi_t^x {\colon}\Omega_0^x \to \Omega_t^x$ giving $$P^* \circ {\bf W} (x,t) = \psi_t^x \circ P^* \circ {\bf W}(x,0)$$ for all $x \in {\mathbb{R}}^{n-2}$ and $t \in {\mathbb{R}}$. Furthermore, the defining quadratic differentials all have the same vertical foliation.
We pick a leaf of this foliation $\gamma {\colon}{\mathbb{R}}\to \Omega$, and arguing as before, this determines a leaf in each surface $\gamma_t^x
{\colon}{\mathbb{R}}\to \Omega_t^x$ with $\gamma_0^0= \gamma {\colon}\mathbb R \to
\Omega_0^0 = \Omega$. Now, pick $\omega = \gamma_0^0(0)$ to be our marked point, add a factor of ${\mathbb{R}}$ to ${\mathbb{H}}^{n-1}$ with coordinate $y
= x_{n-1}$ to obtain ${\mathbb{H}}^n$ with coordinates $(x, y, t) = (x_1,
\ldots, x_{n-2}, x_{n-1}, t)$, and define $${\bf \Omega} {\colon}{\mathbb{H}}^n \to \T(\Omega,\omega)$$ by $${\bf \Omega}(x,y,t) = (\Omega_t^x,\gamma_t^x(y)).$$ So we are again pushing a point along a leaf of the vertical foliation.
\[T:Hn\_isometric\] The map ${\bf \Omega} {\colon}{\mathbb{H}}^n \to \T(\Omega, \omega)$ is an almost-isometric embedding. Moreover, the image lies in the thick part and is quasi-convex.
Again, we must verify the hypotheses of Lemma \[L:criteria\] and prove that the image of ${\bf \Omega}$ is quasi-convex in the thick part, assuming that this is true in all previous steps of the construction.
We can argue exactly as in the case of ${\mathbb{H}}^3$ to prove Properties \[L:crit:geod\] and \[L:crit:slim\] of [Lemma \[L:criteria\]]{} as well as the fact that the image of ${\bf \Omega}$ is quasi-convex in the thick part. Property \[L:crit:unif\] requires more care. However, once established, Property \[L:crit:prop\] follows formally, just as in the case of ${\mathbb{H}}^3$.
We elaborate on the proof that Property \[L:crit:unif\] holds for some $\epsilon$ and $R$. For this, we must give a more precise description of the construction. Write $\Omega_{n-1} = \Omega$, $\Omega_{n-2} = W$ and $$P_{n-2} = P {\colon}\Omega_{n-1} \to \Omega_{n-2}$$ for the branched cover used in the construction. Inductively, we have a tower of branched covers $$\xymatrix{ \Omega_{n-1} \ar[r]^{P_{n-2}} & \Omega_{n-2}
\ar[r]^{P_{n-3}} & \cdots \ar[r] & \Omega_2 \ar[r]^{P_1} & \Omega_1}$$ In this tower, $P_j$ is nontrivially branched at every point $P_j^{-1}(\omega_j)$ where $\omega_j \in \Omega_j$ is the marked point. To clarify, we note that $\Omega_1 = Z$, $\omega_1 = z$, $\Omega_2 = \Sigma$ and $\omega_2 = \sigma$ from the preceding discussion.
We also have a quadratic differential $\nu_1$ on $\Omega_1$ (this is $\nu_1 = q$ from before), which pulls back via all the branched covers to quadratic differentials $\nu_i = P_{i-1}^*(\nu_{i-1}) \in
\Q(\Omega_i)$. On $\Omega_1$, we have chosen $n-1$ distinct nonsingular leaves from the vertical foliation of $\nu_1$ which we denote $\{\zeta_i {\colon}{\mathbb{R}}\to \Omega_1\}_{i=1}^{n-1}$. These leaves are parametrized by arc-length so that $\zeta_j(0) = P_1 \circ
P_2 \circ \cdots \circ P_{j-1}(\omega_j)$.
Recall that $y = x_{n-1}$. We can now describe $\Omega(x, y, t) =
\Omega(x_1,\ldots,x_{n-2},x_{n-1},t)$ for any $(x,y,t) \in {\mathbb{H}}^n$. At the bottom of the tower we push $\omega_1$ along $\zeta_1$ to $\zeta_1(x_1)$, then take the branched cover $\Omega_2^{x_1} \to
(\Omega_1,\zeta_1(x_1))$ induced by $P_1$ (it is the [*induced*]{} branched cover since it branches over $\zeta_1(x_1)$ rather than over $\zeta_1(0) = \omega_1$; see Section \[S:branched\]). Next, the lifted marking identifies $\omega_2$ with a point in the preimage of $\zeta_2(0)$, and we push this along an appropriate lift $\zeta_2^{x_1}$ of $\zeta_2$ to a point $\zeta_2^{x_1}(x_2)$ in the preimage of $\zeta_2(x_2)$. At the next level, there is an branched cover $\Omega_3^{x_1,x_2} \to (\Omega_2^{x_1},\zeta_2^{x_1}(x_2))$ induced by $P_2$. The lifted marking identifies $\omega_3$ with a point in the preimage of $\zeta_3(0)$ in the composition of branched covers $\Omega_3^{x_1,x_2} \to \Omega_2^{x_1} \to \Omega_1$ and we push this along an appropriate lift $\zeta_3^{x_1,x_2}$ of $\zeta_3$ to a point $\zeta_3^{x_1,x_2}(x_3)$ in the preimage of $\zeta_3(x_3)$. We continue in this way to produce a tower of branched covers induced by $P_1, P_2, \ldots, P_{n-3}, P_{n-2}$: $$\xymatrix{ \Omega_{n-1}^{x_1,\ldots,x_{n-2}} \ar[r] &
\Omega_{n-2}^{x_1,\ldots,x_{n-3}} \ar[r] & \cdots \ar[r] &
\Omega_3^{x_1,x_2} \ar[r] & \Omega_2^{x_1} \ar[r] & \Omega_1 .}$$ The point $\omega_{n-1}$ is identified with a marked point in $\Omega_{n-1}^{x_1,\ldots,x_{n-2}}$ in the preimage of $\zeta_{n-1}(0)$, and then we push this point along an appropriate lift $\zeta_{n-1}^{x_1,\ldots,x_{n-2}}$ of $\zeta_{n-1}$ to the point $\zeta_{n-1}^{x_1,\ldots,x_{n-2}}(y) =
\zeta_{n-1}^{x_1,\ldots,x_{n-2}}(x_{n-1})$. With this notation $${\bf \Omega}(x,y,0) =
{\bf \Omega}(x_1,\ldots,x_{n-2},x_{n-1},0) =
(\Omega_{n-1}^{x_1,\ldots,x_{n-2}},\zeta_{n-1}^{x_1,\ldots,x_{n-2}}(x_{n-1})).$$ To find ${\bf \Omega}(x,y,t)$ for any $t$, we apply the appropriate Teichmüller deformation to ${\bf
\Omega}(x,y,0)$. This is the Teichmüller deformation determined by $t$ and the pull back of $\nu_1$ (via the composition of branched covers). We can pull back $\nu_1$ by any of the branched covers, and since the resulting quadratic differential depends only on the surface in this construction, we will simply write $\Phi_t$ for the associated Teichmüller deformation on any of the surfaces $\Omega_j^{x_1,\ldots,x_{j-1}}$. In particular, we have $${\bf \Omega}(x,y,t) = \Phi_t({\bf
\Omega}(x,y,0)).$$ Set $x' = (x_1', x_2', \ldots, x_{n-2}')$. We now must find an $\epsilon$ and $R$ so that if $$e^{-t}|(x,y) - (x',y')| \leq \epsilon$$ then $$d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Omega}(x,y,t),{\bf
\Omega}(x',y',t){\aftergroup\egroup\originalright}) \leq R.$$ As in the case of ${\mathbb{H}}^3$, appealing to the triangle inequality it suffices to find an $\epsilon$ and $R'$ so that if $(x_1,\ldots,x_{n-2},y)$ and $(x_1',\ldots,x_{n-2}',y')$ agree in all but one coordinate, and in that coordinate differ by at most $\epsilon$, then $$d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Omega}(x,y,t), {\bf \Omega}(x',y',t){\aftergroup\egroup\originalright}) \leq R'.$$
If $(x,y)$ and $(x',y')$ differ only in the last coordinate, then we can apply Theorem \[T:kra\] just as before to produce $\epsilon = 1$ and $R' = A+B$. Suppose instead that $y = y'$ and $x$ differs from $x'$ in the $n-2$–coordinate only. We start at the highest coordinate, $y = x_{n-1}$ and work two steps down to $x_{n-2}$. The idea is similar to what was done in varying $x$ in $(x,y,t) \in
{\mathbb{H}}^3$. We look on $\Phi_t(\Omega_{n-2}^{x_1,\ldots,x_{n-3}})$ as an “auxiliary” surface when it is equipped with the [*two*]{} marked points $\Phi_t(\zeta_{n-2}^{x_1,\ldots,x_{n-3}}(x_{n-2}))$ and the image of $\Phi_t(\zeta_{n-1}^{x_1,\ldots,x_{n-2}}(x_{n-1}))$ via the branched covering $$\Phi_t(\Omega_{n-1}^{x_1,\ldots,x_{n-2}}) \to
\Phi_t(\Omega_{n-2}^{x_1,\ldots,x_{n-3}}).$$ If these two points are not too close, then we can move from $\Phi_t(\zeta_{n-2}^{x_1,\ldots,x_{n-3}}(x_{n-2}))$ to $\Phi_t(\zeta_{n-2}^{x_1,\ldots,x_{n-3}}(x_{n-2}'))$ keeping the other marked point fixed, and the distance between these two points in the Teichmüller space of the auxiliary surface with two marked points is uniformly bounded. Since the branched cover induces a $1$–Lipschitz map (compare ), this means that $$d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Omega}(x_1,\ldots,x_{n-3},x_{n-2},x_{n-1},t),{\bf
\Omega}(x_1,\ldots,x_{n-3},x_{n-2}',x_{n-1},t){\aftergroup\egroup\originalright})$$ is uniformly bounded.
On the other hand, if the two marked points in $\Phi_t(\Omega_{n-2}^{x_1,\ldots,x_{n-3}})$ are close, we $$\mbox{move} \quad \Phi_t(\zeta_{n-1}^{x_1,\ldots,x_{n-2}}(x_{n-1})) \quad
\mbox{to} \quad \Phi_t(\zeta_{n-1}^{x_1,\ldots,x_{n-2}}(x_{n-1}')),$$ $$\mbox{move} \quad \Phi_t(\zeta_{n-2}^{x_1,\ldots,x_{n-3}}(x_{n-2})) \quad
\mbox{to} \quad \Phi_t(\zeta_{n-2}^{x_1,\ldots,x_{n-3}}(x_{n-2}')),$$ and then $$\mbox{move} \quad \Phi_t(\zeta_{n-1}^{x_1,\ldots,x_{n-2}'}(x_{n-1}')) \quad
\mbox{back to} \quad \Phi_t(\zeta_{n-1}^{x_1,\ldots,x_{n-2}'}(x_{n-1})).$$ By the triangle inequality, we obtain the desired uniform bound on $$d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Omega}(x_1,\ldots,x_{n-3},x_{n-2},x_{n-1},t),{\bf
\Omega}(x_1,\ldots,x_{n-3},x_{n-2}',x_{n-1},t){\aftergroup\egroup\originalright}).$$ Note that this required three point pushes in two different auxiliary surfaces. We varied the $(n-1)^{\rm st}$ coordinate twice, in the highest surface, and varied the $(n-2)^{\rm nd}$ coordinate once.
Now suppose that $x$ differs from $x'$ in the $(n-3)^{\rm rd}$ coordinate only. We view $\Phi_t(\Omega_{n-3}^{x_1,\ldots,x_{n-4}})$ as an auxiliary surface with [*three marked points*]{}: the images of the points $\Phi_t(\zeta_{n-1}^{x_1,\ldots,x_{n-2}}(x_{n-1}))$ and $\Phi_t(\zeta_{n-2}^{x_1,\ldots,x_{n-3}}(x_{n-2}))$ under the respective branched covers and the point $\Phi_t(\zeta_{n-3}^{x_1,\ldots,x_{n-4}}(x_{n-3}))$. We can move this last point a small amount, changing the Teichmüller distance a bounded amount, provided the other two points, higher in the tower, are not too close to it. If they are too close, we first move them out of the way (as in the first two pushes above), move the third point, then move the two higher points back. The triangle inequality together with the $1$–Lipschitz property of the branched cover map applied as before, implies a uniform bound on the change in Teichmüller distance $$d_\T{\mathopen{}\mathclose\bgroup\originalleft}({\bf \Omega}(x_1,\ldots,x_{n-3},x_{n-2},x_{n-1},t),{\bf
\Omega}(x_1,\ldots,x_{n-3}',x_{n-2},x_{n-1},t){\aftergroup\egroup\originalright}).$$ It follows that varying $x_{n-3}$ requires at most five point pushes in the three highest auxiliary surfaces.
In general, varying $x_{n-k}$ in this way requires $2k - 1$ point pushes in the $k$ highest auxiliary surfaces. Thus we can change any coordinate by a small amount $\epsilon$ and change the Teichmüller distance by a bounded amount $R'$, as required. This completes the sketch of the proof of [Theorem \[T:Hn\_isometric\]]{}.
[^1]: This work is in the public domain. The first author was supported by NSF grants DMS 0905748 and DMS 1207183. The second author was supported by EPSRC grant EP/I028870/1.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: |
In this paper, we give several results on area minimizing surfaces in strictly mean convex $3$-manifolds. First, we study the genus of absolutely area minimizing surfaces in a compact, orientable, strictly mean convex $3$-manifold $M$ bounded by a simple closed curve in $\partial M$. Our main result is that for any $g\ge 0$, the space of simple closed curves in $\partial M$ where all the absolutely area minimizing surfaces they bound in $M$ has genus $\geq g$ is open and dense in the space $\mathcal{A}$ of nullhomologous simple closed curves in $\partial M$. For showing this we prove a bridge principle for absolutely area minimizing surfaces. Moreover, we show that for any $g\ge 0$, there exists a curve in ${\mathcal{A}}$ such that the minimum genus of the absolutely area minimizing surfaces it bounds is exactly $g$.
As an application of these results, we further prove that the simple closed curves in $\partial M$ bounding more than one minimal surface in $M$ is an open and dense subset of $\mathcal{A}$. We also show that there are disjoint simple closed curves in $\partial M$ bounding minimal surfaces in $M$ which are not disjoint. This allows us to answer a question of Meeks, by showing that for any strictly mean convex $3$-manifold $M$, there exists a simple closed curve $\Gamma$ in $\partial M$ which bounds a stable minimal surface which is not embedded.
address:
- 'Frei University, Berlin, Germany'
- |
Koc University\
Department of Mathematics\
Sariyer, Istanbul 34450 Turkey
author:
- Theodora Bourni
- Baris Coskunuzer
title: 'Area minimizing surfaces in mean convex 3-manifolds'
---
\[section\] \[thm\][Lemma]{} \[thm\][Corollary]{}
\[thm\][Remark]{} \[thm\][**[Definition]{}**]{}
[^1]
Introduction
============
The Plateau problem concerns the existence of an area minimizing disk with boundary a given curve in an ambient manifold $M$. This problem was solved in the case when the ambient manifold is ${\mathbf{R}}^3$ by Douglas [@Do], and Rado [@Ra] in the early 1930s. Later, it was generalized by Morrey [@Mo] for Riemannian manifolds. In the 1980s, Meeks and Yau showed that if $M$ is a mean convex $3$-manifold, and $\Gamma$ is a simple closed curve in $\partial
M$, then any area minimizing disk with boundary $\Gamma$ is embedded [@MY1]. Later, White gave a generalization of this result to any genus [@Wh2].
In the early 1960s, the same question was studied for absolutely area minimizing surfaces, i.e. for surfaces that minimize area among all orientable surfaces with a given boundary (without restriction on the genus). Using techniques from geometric measure theory, Federer and Fleming [@FeF] were able to solve this problem by proving the existence of an absolutely area minimizing integral current (see also [@DG; @Giu] for the existence of a minimizing Caccioppoli set). In [@ASSi] Almgren, Schoen and Simon showed that this current is a smooth embedded surface away from its boundary and Hardt [@H] showed that it is also smooth at the boundary, provided that the prescribed boundary is smooth and lies on a convex set. Later, Hardt and Simon [@HSi], improved this boundary regularity result, by dropping the assumption that the prescribed boundary lies on a convex set.
It can be seen that there are two main versions of the Plateau problem; one of them is that of fixed genus (area minimizing in a fixed topological class), and the other one is with no restriction on the genus (absolutely area minimizing case). There have also been many important results on the [*a priori*]{} bounds on the genus of an absolutely area minimizing surface bounded by a given simple closed curve [@HSi].
In this paper, we study the genus of absolutely area minimizing surfaces $\Sigma$ in a compact, orientable, strictly mean convex $3$-manifold $M$, with boundary $\partial \Sigma$ a simple closed curve lying in $\partial M$. We consider the stratification of the space of nullhomologous (bounding a surface in $M$) simple closed curves in $\partial M$ with respect to the minimum genus of the absolutely area minimizing surfaces that they bound. In particular, we let $\mathcal{A}$ be the space of all nullhomologous simple closed curves in $\partial M$ with the $\mathcal C^0$ topology (see Definition \[c0rmk\]). By the previously mentioned results, any $\Gamma\in \mathcal {A}$ bounds an embedded absolutely area minimizing surface $\Sigma$ in $M$ with $\partial \Sigma = \Gamma$. Let $\mathcal{A}_g$ be the set of all the curves in $\mathcal{A}$ such that any embedded absolutely area minimizing surface in $M$ that they bound has genus $\geq g$. Then, clearly ${\mathcal{A}}={\mathcal{A}}_0\supset {\mathcal{A}}_1\supset {\mathcal{A}}_2 \supset ...\supset {\mathcal{A}}_n \supset
...$. We show that for any $g\geq 0$, ${\mathcal{A}}_g$ is an open subset of ${\mathcal{A}}$ (Lemma \[Agopen\]).
Then, we define an operation, which we call [*horn surgery*]{}, on a given simple closed curve $\Gamma$ in $\partial M$ which adds a [*thin handle*]{} to an absolutely area minimizing surface $\Sigma$ which $\Gamma$ bounds in $M$. In other words, by modifying the boundary curve, the operation increases the genus of an absolutely area minimizing surface (See Figure 4 and Figure 5). Moreover, this new curve can be made as close as we want to the original curve. Hence by using the horn surgery operation, it is easy to show that ${\mathcal{A}}_g$ is not only open, but also dense in ${\mathcal{A}}$ (Theorem \[Adense\]). On the other hand, to define the horn surgery operation, we prove [*a bridge principle for absolutely area minimizing surfaces*]{} (Lemma \[bridge principle\]).
Next, we study the the space ${\mathcal{B}}_g$ of all the curves in $\mathcal{A}$, such that the minimum genus of the embedded absolutely area minimizing surfaces in $M$ that they bound is exactly $g$, i.e. ${\mathcal{B}}_g={\mathcal{A}}_g \setminus {\mathcal{A}}_{g+1}$. Then, clearly ${\mathcal{A}}= \bigcup_{g=0}^{\infty} {\mathcal{B}}_g$ and ${\mathcal{B}}_g\cap {\mathcal{B}}_{g'} =
\emptyset$ for any $g\neq g'$. Again by using the horn surgery operation, we show that for any $g\geq 0$, ${\mathcal{B}}_g$ is not empty (Theorem \[Bgne\]). In other words, we show that for any $g\geq 0$, there is a simple closed curve $\Gamma_g$ in $\partial M$ that bounds an absolutely area minimizing surface $\Sigma$ with genus $g$. Also, we prove that for any $g$, ${\mathcal{B}}_g$ is nowhere dense in ${\mathcal{A}}$ (Corollary \[Bndense\]). Furthermore, we show that there exist simple closed curves $\Gamma$ in $\partial M$ which bound two absolutely area minimizing surfaces of different genus (Theorem \[diffgenus\]).
In Section \[minsurfaces section\], we further derive several interesting results about minimal surfaces in strictly mean convex $3$-manifolds by using the previously mentioned results. We show that some important properties for properly embedded area minimizing surfaces in such manifolds are not valid for properly embedded minimal surfaces. First, we show that curves that bound more that one minimal surface are generic for a strictly mean convex $3$-manifold $M$ (Theorem \[nonuniquecurves\]). In [@Co1], and [@CE], it is proven that a generic nullhomotopic simple closed curve in $\partial M$ bounds a unique area minimizing disk, and similarly, a generic nullhomologous simple closed curve in $\partial M$ bounds a unique absolutely area minimizing surface in $M$. However, here we show that when we relax the condition of being area minimizing to just minimal, the situation is completely opposite.
Finally, we generalize Peter Hall’s results [@Ha], which answers Meeks’ questions, to any strictly mean convex $3$-manifold $M$. As previously mentioned, in [@MY1], Meeks and Yau proved that any area minimizing disk in a mean convex $3$-manifold bounded by a simple closed curve in $\partial M$ must be embedded . After establishing this result, Meeks posed the question of whether or not the same holds for stable minimal surfaces. Then, Hall constructed an example of a simple closed curve $\Gamma$ in $S^2=\partial {\mathbf{B}}^3$, where ${\mathbf{B}}^3$ is the unit $3$-ball in $\R^3$, such that $\Gamma$ bounds a stable minimal disk $M$ in $B^3$ which is not embedded. This shows that if we relax the area minimizing condition to just being minimal again, the embeddedness result of Meeks and Yau is no longer valid. We generalize Hall’s example, and show that this is true for any strictly mean convex $3$-manifold (Theorem \[nonembeg\]).
To construct the nonembedded minimal examples, we first show that for any strictly mean convex $3$-manifold $M$, there are disjoint simple closed curves $\Gamma_1$ and $\Gamma_2$ in $\partial M$ that bound minimal surfaces, $M_1$ and $ M_2$ respectively, in $M$, such that $M_1\cap M_2 \neq
\emptyset$ (Theorem \[intersectingminsurfaces\]). This is very interesting since if $M$ is a mean convex $3$-manifold with trivial second homology, and $\Gamma_1,\Gamma_2$ are two disjoint simple closed curves in $\partial M$, then any two absolutely area minimizing surfaces $\Sigma_1, \Sigma_2$ with $\partial \Sigma_1 = \Gamma_1, \partial \Sigma_2 = \Gamma_2$ are also disjoint [@Co1]. The same holds for area minimizing disks, too. Hence, again the case of minimal surfaces, and that of area minimizing are very different. Then, by connecting these intersecting minimal surfaces, which have disjoint boundaries, with a [*bridge*]{}, we construct nonembedded examples of minimal surfaces in the strictly mean convex $3$-manifold $M$. Thus, we show that for any such $M$, there exists a simple closed curve $\Gamma$ in $\partial M$ which bounds a stable minimal surface, and in particular a stable minimal disk, in $M$, which is not embedded (Theorem \[nonembeg\]).
Acknowledgements: {#acknowledgements .unnumbered}
-----------------
We would like to thank Tolga Etgu and Brian White for very useful conversations.
Preliminaries
=============
In this section we give some standard definitions as well as an overview of the basic known results which we use in the following sections.
An [*area minimizing disk*]{} is a disk which has the smallest area among all disks with the same boundary. An [*area minimizing surface*]{} is a surface which has the smallest area among all surfaces with the same genus and the same boundary. An [*absolutely area minimizing surface*]{} is a surface which has the smallest area among all orientable surfaces (with no topological restriction) with the same boundary.
Let $M$ be a compact Riemannian $3$-manifold with boundary. Then $M$ is called [*mean convex*]{} (or sufficiently convex) if the following conditions hold:
- $\partial M$ is piecewise smooth.
- Each smooth subsurface of $\partial M$ has nonnegative mean curvature with respect to the inward normal.
- There exists a Riemannian manifold $N$ such that $M$ is isometric to a submanifold of $N$ and each smooth subsurface $S$ of $\partial M$ extends to a smooth embedded surface $S'$ in $N$ such that $S' \cap M = S$.
If in (ii) of the definition we require that each smooth subsurface has strictly positive mean curvature, then $M$ is called [*strictly mean convex*]{}.
A simple closed curve is called an [*extreme curve*]{} if it is on the boundary of its convex hull. A simple closed curve is called an [*$H$-extreme curve*]{} if it is a curve in the boundary of a mean convex manifold $M$.
An embedded surface $S$ in a $3$-manifold $M$ is called [*properly embedded*]{} if $S\cap \partial M = \partial S$.
We now state the main facts which we will be using in the following sections.
\[meeksyau\][@MY2; @MY3] Let $M$ be a compact, mean convex $3$-manifold, and $\Gamma\subset\partial M$ be a nullhomotopic simple closed curve. Then, there exists an area minimizing disk $ D\subset M$ with $\partial D = \Gamma$. Moreover, all such disks are properly embedded in $M$ and they are pairwise disjoint. Furthermore, if $\Gamma_1, \Gamma_2 \subset \partial M$ are two disjoint simple closed curves, then two area minimizing disks $ D_1$ and $ D_2$ spanning $\Gamma_1$ and $
\Gamma_2$ respectively are also disjoint.
There is an analogous fact for area minimizing surfaces, too.
[@FeF; @ASSi; @H]\[absmincur\] Let $M$ be a compact, strictly mean convex $3$-manifold and $\Gamma\subset\partial M$ a nullhomologous simple closed curve. Then there exists $\Sigma\subset M$ an absolutely area minimizing surface with $\partial \Sigma = \Gamma$ and each such $\Sigma$ is smooth away from its boundary and it is smooth around points of the boundary where $\Gamma$ is smooth.
The regularity results in [@ASSi; @H] extend from the ambient space being $\R^n$ to $M$, a manifold as in the Theorem \[absmincur\], as explained in [@Wh2 Theorem 6.3].
\[white2\] [@Wh2] Let $M$ be a compact, orientable, strictly mean convex $3$-manifold and $\Gamma$ a simple closed curve in $\partial M$. For each $g\geq 0$, define $$a(g)=\inf|\Sigma|$$ where the infimum is taken over all piecewise smooth embedded surfaces $\Sigma$ in $M$, with genus $g$ and boundary $\Gamma$ and where $|\Sigma|$ denotes the area of a surface $\Sigma$. Then, if $a(g)< a(g-1)$, the infimum $a(g)$ is attained.
\[white2c\][@Wh2] Let $M$ be a compact, orientable, strictly mean convex $3$-manifold, $\Gamma$ a simple closed curve in $\partial M$ and $\Sigma$ a stable surface in $M$ with $\partial \Sigma=\Gamma$. Assume that for a geodesic ball ${\mathcal{B}}(x, R)$ in $M$ we have $$\sup_{{\mathcal{B}}(y,r)\subset{\mathcal{B}}(x,R)}\frac{|\Sigma\cap{\mathcal{B}}(y,r)|}{\pi r^2}=C<\infty.$$ Then the following hold:
- If ${\mathcal{B}}(x,R)\cap\Gamma=\emptyset$, then the principal curvatures of $M\cap {\mathcal{B}}(x, R/2)$ are bounded by a constant that depends only on $R, C$.
- If ${\mathcal{B}}(x,R)\cap\Gamma$ is $C^{2,\a}$ , then the principal curvatures of $M\cap {\mathcal{B}}(x, R/2)$ are bounded by a constant that depends only on $R, C$ and $\Gamma$.
Finally, we state two results on the generic uniqueness of area minimizing disks, and absolutely area minimizing surfaces for $H$-extreme curves.
[@Co1] Let $M$ be a compact, orientable, mean convex $3$-manifold with $H_2(M,{\mathbf{Z}})=0$. Then the following hold:
- For a generic nullhomotopic (in $M$) simple closed curve $\Gamma$ in $\partial M$, there exists a unique area minimizing disk $D$ in $M$ with $\partial D = \Gamma$.
- For a generic nullhomologous (in $M$) simple closed curve $\Gamma$ in $\partial M$, there exists a unique absolutely area minimizing surface $\Sigma$ in $M$ with $\partial \Sigma =
\Gamma$.
[**Convention:**]{} Throughout the paper, all the manifolds will be assumed to be compact, orientable, and strictly mean convex unless otherwise stated. We will also assume that all the surfaces are orientable.
Stratification of the space of simple closed curves in $\partial M$
===================================================================
In this section, we study the space of nullhomologous simple closed curves in $\partial M$, where $M$ is a compact, orientable, strictly mean convex $3$-manifold. We stratify this space with respect to the minimum genus of the absolutely area minimizing surfaces that its curves bound.
Let ${\mathcal{A}}$ denote the space of nullhomologous simple closed curves in $\partial
M$, equipped with the $\mathcal{C}^0$ topology. We give here the precise definition of the neighborhoods defining this topology.
\[c0rmk\] Let $\Gamma\in{\mathcal{A}}$, then there exists a continuous function $$f: S^1\to\partial M$$ such that $f(S^1)=\Gamma$, where $S^1=\partial {\mathbf{B}}^2$ and ${\mathbf{B}}^2$ is the unit 2-ball in $\R^2$.
An $\e$-neighborhood of $\Gamma$, ${\mathcal{U}}_\Gamma(\e)$ in the $\mathcal{C}^0$ topology consists of all $\wt\Gamma\in{\mathcal{A}}$ for which there exists a continuous function $$g: S^1\to\partial M$$ such that $g(S^1)=\wt\Gamma$ and $\displaystyle\sup_{x\in S^1}|f(x)-g(x)|<\e$.
Note that, for $\e$ small enough, so that the $\e$-neighborhood of $\Gamma$ in $\partial M$, defined by $$N_\e(\Gamma)=\{x\in\partial M:\operatorname{dist}(x, \Gamma)<\e\},$$ is topologically an annulus, we have that $\wt\Gamma\in\mathcal {\mathcal{U}}_\Gamma(\e)$ is equivalent to $\wt\Gamma\subset N_\e(\Gamma)$ and $\wt\Gamma\sim\Gamma$, i.e. $\wt\Gamma$ is homotopic to $\Gamma$.
We define a relation $\varphi: {\mathcal{A}}\to {\mathbf{N}}$, by $$\varphi(\Gamma) = \min_\Sigma \operatorname{genus}(\Sigma),$$ where the minimum is taken over all absolutely area minimizing surfaces $\Sigma$ in $M$ with $\partial \Sigma = \Gamma$ and naturally $\operatorname{genus}(\Sigma)$ denotes the genus of a surface $\Sigma$.
Now, define ${\mathcal{A}}_g = \{ \ \Gamma\in {\mathcal{A}}\ | \ \varphi(\Gamma)\geq g \ \}$ and ${\mathcal{B}}_g = \varphi^{-1}(g)= {\mathcal{A}}_g \setminus{\mathcal{A}}_{g+1}$. In particular, ${\mathcal{A}}_g$ denotes the space of simple closed curves in $\partial M$ such that the minimum genus of the absolutely area minimizing surfaces that they bound is greater than or equal to $g$. ${\mathcal{B}}_g$ denotes the space of simple closed curves in $\partial M$ such that the minimum genus of the absolutely area minimizing surfaces that they bound is exactly $g$. Clearly, ${\mathcal{A}}={\mathcal{A}}_0\supset {\mathcal{A}}_1\supset {\mathcal{A}}_2 \supset ...\supset {\mathcal{A}}_n \supset ...$.
\[Agopen\] For any $g\geq 0$, ${\mathcal{A}}_g$ is open in ${\mathcal{A}}$.
Let $\Gamma_0\in {\mathcal{A}}_g$ be a simple closed curve in $\partial M$. Since $\Gamma_0\in
{\mathcal{A}}_g$, we have $\varphi(\Gamma_0)\geq g$, i.e. if $\Sigma$ is an absolutely area minimizing surface in $M$ with $\partial \Sigma = \Gamma_0$, then $\operatorname{genus}(\Sigma) \geq g$.
Assume now that the Lemma is not true. Then for any $\e>0$, there exists $\Gamma_\e\in{\mathcal{U}}_{\Gamma_0}(\e)$, such that $\Gamma_\e\notin{\mathcal{A}}_g$, where ${\mathcal{U}}_{\Gamma_0}(\e)$ is the $\e$-neighborhood of $\Gamma_0$ in the ${\mathcal{C}}^0$ topology, as in Definition \[c0rmk\]. Note that for $\e$ small enough, $\Gamma_\e\sim\Gamma$ (cf. Definition \[c0rmk\]). Hence, for $\e$ small enough, there exists a region $A_\e$ in $\partial M$ between $\Gamma_\e$ and $\Gamma$, i.e. $\partial A_\e=\Gamma_\e\cup\Gamma$, for which we have $|A_\e|\stackrel{\e\to 0}{\longrightarrow} 0$, where $|\cdot|$ denotes the area of a surface. This implies that for any absolutely area minimizing surface $\Sigma_\e$, with $\partial\Sigma_\e=\Gamma_\e$, we have $|\Sigma_\e|\le |\Sigma|+|A_\e|$, where $\Sigma$ is any absolutely area minimizing surface with boundary equal to $\Gamma_0$. Therefore there exists a sequence $\{\Gamma_i\}_{i\in\N}\subset{\mathcal{A}}$, such that $\Gamma_i\stackrel{i\to\infty}{\longrightarrow}\Gamma_0$ in the ${\mathcal{C}}^0$ topology and such that for each $i$, there exists an absolutely area minimizing surface $\Sigma_i$, with $\partial\Sigma_i=\Gamma_i$, $\operatorname{genus}(\Sigma_i)\le g-1$ and such that their areas satisfy $|\Sigma_i|\stackrel{i\to\infty}{\longrightarrow}|\Sigma|$.
Let now $S_i=[\![\Sigma_i]\!]$, be the corresponding integral currents. Then, since the areas $|\Sigma_i|$ are uniformly bounded, by the Federer-Fleming compactness theorem for integral currents [@FeF], after passing to a subsequence $S_i\to S_0$, in the sense of currents, where $S_0$ is an integral current, with $\mass(S_0)\le \lim_{i\to\infty} \mass(S_i)=|\Sigma|$; here $\mass$ denotes the mass of the current (see [@Sim]). Hence $\mass(S_0)=|\Sigma_0|$ and thus (using Theorem \[absmincur\]) $S_0=[\![\Sigma_0]\!]$, where $\Sigma_0$ is an absolutely area minimizing surface with $\partial \Sigma_0=\Gamma_0$ and since $\Gamma_0\in{\mathcal{A}}_g$, we have that $\operatorname{genus}(\Sigma_0)\ge g$.
Note now that, since $S_i$ are absolutely area minimizing, in the above convergence, the corresponding radon measures converge $\mu_{S_i}\to \mu_S$ (for a proof of this see [@Sim Theorem 34.5]). Furthermore, since each $\Sigma_i$ is a minimal surface, the monotonicity formula holds [@Sim], i.e. for each $i$, each $x\in \ov M\setminus \Gamma_i$ and for $r$ small enough (so that ${\mathcal{B}}(x, r)\cap\Gamma_i=\emptyset$), $r^{-2}|\Sigma_i|$ is an increasing function of $r$. Therefore we can apply the principal curvature bound of White, as given in Theorem \[white2c\], to conclude that the convergence $\Sigma_i\to\Sigma_0$ is actually smooth in compact sets of $\overline M\setminus\Gamma_0$. Let now $K$ be a compact subset of $M$, such that $\Sigma_0\cap K$ has genus $\ge g$. Then the previous convergence, along with the Gauss-Bonnet theorem, implies that the same should be true for the $\Sigma_i$’s, for $i$ big enough, i.e. that $\operatorname{genus}(\Sigma_i\cap K)\ge g$. But this contradicts the fact that $\operatorname{genus}(
\Sigma_i)\le g-1$.
A bridge principle for Absolutely Area Minimizing Surfaces {#bridgeprinciplesubsection}
----------------------------------------------------------
In this section we define a surgery operation on a simple closed curve $\Gamma$ in $\partial M$, by gluing a “bridge” on $\Gamma$. Let $\alpha$ be a simple path in $\partial M$ such that $\a$ intersects $\Gamma$ transversely and $\Gamma \cap \alpha = \{x,y\}$ are the endpoints of $\alpha$. Let $\epsilon_0 >0$ be sufficiently small so that for almost every $0<\epsilon<\epsilon_0$, $\partial N_{\epsilon}(\alpha)\cap \Gamma$ consists of exactly 4 points, namely $\{x^+_{\epsilon}, x^-_{\epsilon}, y^+_{\epsilon}, y^-_{\epsilon}\}$, where $N_{\epsilon}(\alpha)$ is the $\epsilon$ neighborhood of $\alpha$ in $\partial M$, i.e. $$N_\e(\alpha)=\{x\in\partial M:\operatorname{dist}(x,\alpha)<\e\}.$$ Then, $\Gamma$ divides $N_{\epsilon}(\alpha)$ into three components, which we call $C_x, C_y$ and $S^\epsilon_\alpha$, where $S^\epsilon_\alpha$ is the component containing $\alpha$ and $C_x, C_y$ are the remaining caps of $N_{\epsilon}(\alpha)$ near $x$ and $y$ respectively (see Figure 1).
For orientation issues, that are explained in Remark \[OI\], we further assume that the path $\alpha$ is such that $C_x$ and $C_y$ are in the same side of $\Gamma$. In particular we will be considering paths $\a$ satisfying the following definition.
\[a\] Let $\Gamma\in{\mathcal{A}}$ and let $\alpha$ be a simple path in $\partial M$ such that $\a$ intersects $\Gamma$ transversely and $\Gamma
\cap \alpha = \{x,y\}$ are the endpoints of $\alpha$. We say that $\a$ is a [*$\Gamma$-admissible path*]{} if the following holds:
If $N_{\epsilon '} (\Gamma)$ is a small annular neighborhood of $\Gamma$ in $\partial M$, and $N_{\epsilon '} (\Gamma) \setminus \Gamma =
H^+\cup H^-$ where $H^+$ is an annulus, say the “positive side" of $\Gamma$, and $H^-$ is the other annulus, say the “negative side" of $\Gamma$, then either both $C_x$ and $C_y$ intersect $H^+$ or they both intersect $H^-$, where $C_x, C_y$ are as above (see Figure 1).
$S^\epsilon_\alpha$, as constructed above is a tiny strip (“bridge”) of width $\e$, around $\alpha$. For its boundary we have $\partial
S^\epsilon_\alpha = (\alpha^+_\epsilon \cup \alpha^-_\epsilon) \cup (\beta^x_\epsilon \cup \beta^y_\epsilon)$ where $\beta^x_\epsilon$ is the arc segment of $\Gamma$ between $x^+_\epsilon$ and $x^-_\epsilon$ containing $x$, and similarly $\beta^y_\epsilon$ is the arc segment of $\Gamma$ between $y^+_\epsilon$ and $y^-_\epsilon$ containing $y$. Also, $\alpha^+_\epsilon$ is a simple path in $\partial M$ connecting $x^+_\epsilon$ to $y^+_\epsilon$, and $\alpha^-_\epsilon$ is a simple path in $\partial M$ connecting $x^-_\epsilon$ to $y^-_\epsilon$ in $\partial M$ (See Figure 2). “Gluing” this strip $S^\e_\a$ on $\Gamma$ results to the following definition.
\[wtG\] Let $\Gamma\in{\mathcal{A}}$ and let $\a$ be a $\Gamma$- admissible path. Then we define a new curve $$\widehat{\Gamma}^\epsilon_\alpha=\Gamma\sharp S_\a^\e,$$ obtained by gluing on $\Gamma$ a bridge $S^\e_\a$ of width $\e$ around $\a$. With the above notation $\Gamma\sharp S_\a^\e$ is the curve $$\widehat{\Gamma}^\epsilon_\alpha =\Gamma\sharp S_\a^\e=(\Gamma \setminus (\beta^x_\epsilon \cup \beta^y_\epsilon)) \cup (\alpha^+_\epsilon \cup \alpha^-_\epsilon).$$ Since $\alpha$ is $\Gamma$-admissible, $\widehat{\Gamma}^\epsilon_\alpha$ is the union of two simple closed curves in $\partial M$, which we will denote by $\widehat{\Gamma}^{\epsilon,1}_\alpha$ and $\widehat{\Gamma}^{\epsilon,2}_\alpha$ (See Figure 2).
The above gluing construction is also possible when instead of a single simple closed curve, we have two disjoint curves and $\a$ is a path joining them. In particular, let $\Gamma_1, \Gamma_2\in {\mathcal{A}}$ be such that $\Gamma_1\cap\Gamma_2=\emptyset$. Then the above gluing operation goes through in exactly the same way as described, with $\Gamma$ replaced by $\Gamma_1\cup\Gamma_2$. In this case we have no orientation issues but we want $\a$ to be a path “connecting” the two curves. In particular we will be considering paths $\a$ satisfying the following definition.
\[a2\] Let $\Gamma_1, \Gamma_2\in {\mathcal{A}}$ be such that $\Gamma_1\cap\Gamma_2=\emptyset$, and let $\alpha$ be a simple path in $\partial M$ such that $\a$ intersects $\Gamma_1$ and $\Gamma_2$ transversely and $(\Gamma_1\cup\Gamma_2) \cap \alpha = \{x,y\}$ are the endpoints of $\alpha$. We say that $\a$ is a path [ *connecting* ]{} $\Gamma_1$ and $\Gamma_2$ if $x\in\Gamma_1$ and $y\in \Gamma_2$.
As in the case of a simple closed curve we can consider $S_\a^\e$, a tiny strip (“bridge”) of width $\e$ around $\a$ with $\partial
S^\e_\a=(\a^+_\e\cup\a^-_\e) \cup(\beta_\e^x\cup\beta_\e^y)$, where $S_\a^\e, a^+_\e, \a^-_\e, \beta_\e^x$ and $\beta_\e^y$ are defined exactly as before. Connecting $\Gamma_1 $ and $\Gamma_2$ via this stip $S^\e_\a$ results to the following definition.
\[wtG2\] Let $\Gamma_1, \Gamma_2\in{\mathcal{A}}$ be such that $\Gamma_1\cap\Gamma_2=\emptyset$ and let $\a$ be a path connecting them. Then we define a new curve $$\widehat{\Gamma}^\epsilon_\alpha=\Gamma\sharp S_\a^\e,$$ obtained by gluing on $\Gamma$ a bridge $S_\a^\e$ of width $\e$ around $\a$, where $\Gamma=\Gamma_1\cup\Gamma_2$. With the previous notation $\Gamma\sharp S_\a^\e$ is the curve $$\widehat{\Gamma}^\epsilon_\alpha =\Gamma\sharp S_\a^\e=(\Gamma \setminus (\beta^x_\epsilon \cup \beta^y_\epsilon)) \cup (\alpha^+_\epsilon \cup \alpha^-_\epsilon).$$ Since $\alpha$ is a path connecting $\Gamma_1$ and $\Gamma_2$, $\widehat{\Gamma}^\epsilon_\alpha$ is a simple closed curve.
\[smoothglue\] We remark here that if the curve $\Gamma$ is smooth around $\Gamma\cap\a$, where $\a$ is a $\Gamma$-admissible path, then we can consider gluing the bridge $S^\e_\a$ smoothly along $\Gamma$, so that the resulting curve $\widehat{\Gamma}_\a^\e=\widehat{\Gamma}_\a^{\e,1}\cup
\widehat{\Gamma}_\a^{\e,2}$, as in Definition \[wtG\], is smooth around $x^\pm, y^\pm$.
Similarly, if $\Gamma_1$ and $\Gamma_2$ are smooth around $(\Gamma_1\cup\Gamma_2)\cap\a$, where $\a$ is a path connecting $\Gamma_1$ and $\Gamma_2$, then we consider gluing the bridge $S^\e_\a$ smoothly along $\Gamma_1$ and $\Gamma_2$, so that the resulting curve $\widehat{\Gamma}_\a^\e$, as in Definition \[wtG2\], is smooth around $x^\pm, y^\pm$.
In the rest of the paper and as long as these smoothness assumptions are satisfied we will always consider constructing the curve $\widehat{\Gamma}_\a^\e$ to be smooth around $\a$.
[**(A bridge principle for absolutely area minimizing surfaces)**]{}\[bridge principle\]
Assume that either
- $\Gamma\in{\mathcal{A}}$ and $\alpha$ is a $\Gamma$-admissible simple path in $\partial M$ (as in *Definition \[a\]*) or
- $\Gamma=\Gamma_1\cup\Gamma_2$, where $\Gamma_1,\Gamma_2\in{\mathcal{A}}$ are such that $\Gamma_1\cap\Gamma_2=\emptyset$ and $\alpha$ is a simple path in $\partial M$ connecting them (as in *Definition \[a2\]*).
Let $\widehat{\Gamma}^\epsilon_\alpha=\Gamma\sharp S^\e_\a$ be the curve obtained by attaching on $\Gamma$ a strip of width $\e$ around $\a$ (see *Definition \[wtG\]* and *\[wtG2\]* for cases *(a)* and *(b)* respectively). Then, for any $\eta>0$ there exists $\epsilon_\alpha>0$ such that for almost every $0<\epsilon<\epsilon_\alpha$, any absolutely area minimizing surface in $M$, $\Sigma^\e_\alpha$, with $\partial
\Sigma^\e_\alpha=\widehat{\Gamma}^\epsilon_\alpha$ is $\eta-$close in Hausdorff distance and in area, to the surface obtained by attaching an $\e$-strip around $\alpha$ to one of the absolutely area minimizing surface bounded by $\Gamma$, i.e. $$\sup_{x\in \Sigma^\e_\alpha}\operatorname{dist}(x, \Sigma\cup S^\e_\alpha)<\eta$$ and $$\left||\Sigma_\a^\e|-|\Sigma|\right|<\eta$$ for some absolutely area minimizing surfaces $\Sigma$, with $\partial \Sigma=\Gamma$.
We prove this by contradiction. Assume that the lemma is not true, then there exist $\Gamma, \a$ satisfying either the assumption (a) or the assumption (b) of the lemma, and some $\eta>0$, such that the following holds: For any $\e_\a>0$, there exists $\e<\e_\a$ and an absolutely area minimizing surface $\Sigma^\e_\alpha$ in $M$, with $\partial \Sigma^\e_\alpha=\widehat{\Gamma}^\epsilon_\alpha$, such that for any absolutely area minimizing surface $\Sigma$, with $\partial \Sigma=\Gamma$ either $$\label{supcontr}
\sup_{x\in \Sigma^\e_\alpha}\operatorname{dist}(x, \Sigma\cup S^\e_\alpha)\ge\eta$$ or $$\label{supcontr2}
\left||\Sigma_\a^\e|-|\Sigma|\right|\ge\eta$$ Hence there exists a sequence $\e_i\downarrow 0$, and surfaces $\Sigma_\alpha^{\e_i}$ as above such that for any absolutely area minimizing surface $\Sigma$, with $\partial \Sigma=\Gamma$, either or holds, with $\e$ replaced by $\e_i$.
Let $T_i=[\![\Sigma_\a^{\e_i}]\!]$ be the corresponding currents. Note that for any absolutely area minimizing current $T_0$, with $\partial T_0=[\![\Gamma]\!]$, we have that $$\partial T_i=\partial(T_0+[\![S^{\e_i}_\a]\!])$$ and thus $$\underline{\underline M}(T_i)\le \underline{\underline M}(T_0+[\![S^{\e_i}_\a]\!])\le \underline{\underline M}(T_0)+C\e_i,$$ where $\underline{\underline M}$ denotes the mass and $C$ is a constant that depends on $\a$ (independent of $i$). Hence, the currents $T_i$ have uniformly bounded masses, and thus by the Federer-Fleming compactness theorem [@FeF], after passing to a subsequence, $T_i\to T$, in the sense of currents, where $T$ is an integral current. Furthermore (cf. [@Sim Theorem 34.5]) $T$ is absolutely area minimizing, $\partial T=\Gamma$ and the corresponding Radon measures also converge $\mu_{T_i}\to \mu_T$. Therefore, by Theorem \[absmincur\], $T=[\![\Sigma_0]\!]$, where $\Sigma_0$ is an absolutely area minimizing surface and it is also smooth and embedded away from $\Gamma$. The measure convergence implies that
$$\left||\Sigma_\a^{\e_i}|-|\Sigma_0|\right|\stackrel{i\to\infty}{\longrightarrow}0$$ and so cannot hold. Therefore we assume that for every $i$, holds. In particular, for each $i$, there exists $x_i\in \Sigma_\a^{\e_i}$ such that $$\operatorname{dist}(x_i, \Sigma\cup S^{\e_i}_\alpha)\ge\eta$$ for any absolutely area minimizing surface $\Sigma$, with $\partial \Sigma=\Gamma$.
By passing to a further subsequence if necessary we also have that $$x_i\to x_0$$ for some $x_0\in \ov M$. Note that by the assumption on $x_i$, we have that $\operatorname{dist}(x_i,\Sigma_0\cup\alpha)\ge \eta$ and hence we also have that $\operatorname{dist}(x_0,
\Sigma_0\cup\alpha)\ge \eta$.
By the above convergence, there exists $i_0$, such that $\forall i\ge i_0$ we have $${\mathcal{B}}(x_i,{\eta/4})\subset {\mathcal{B}}(x_0,{\eta/2})\subset {\mathcal{B}}(x_i, \eta),$$ where ${\mathcal{B}}(x,\rho)$ denotes a geodesic ball of radius $\rho$ and centered at $x$, in $M$. Hence, using the monotonicity formula (cf. [@Sim]) and the fact that $x_i\in\operatorname{spt}T_i\setminus \operatorname{spt}\partial T_i$ $$\mu_{T_i}({\mathcal{B}}(x_0, {\eta/2}))\ge\mu_{T_i}({\mathcal{B}}(x_i,{\eta/4}))\ge \pi\left(\frac\eta 4\right)^2.$$ Using now the measure convergence, we have that $$\mu_T({\mathcal{B}}(x_0,{\eta/2}))=\lim_i\mu_{T_i}({\mathcal{B}}(x_0,{\eta/2}))\ge \pi\left(\frac\eta 4\right)^2$$ and therefore $$\mu_T({\mathcal{B}}(x_i,{\eta}))\ge\mu_T({\mathcal{B}}(x_0,{\eta/2}))\ge \pi\left(\frac\eta 4\right)^2$$ which implies that $$\operatorname{spt}T\cap {\mathcal{B}}(x_i,\eta)\ne \emptyset\Rightarrow \operatorname{dist}(x_i, \Sigma_0\cup\alpha)<\eta.$$ This contradicts the choice of $x_i$ and thus cannot hold for every $i$. This finishes the proof of the lemma.
\[C2convrmk\] Let $\{\e_i\}_{i\in\N}$, be a sequence with $\e_i\downarrow 0$ and for each $i$ let $\Sigma_\a^{\e_i}$ be an area minimizing surface with $\partial \Sigma_\a^{\e_i}=\widehat\Gamma_{\a}^{\e_i}$ as in Theorem \[bridge principle\]. Then the proof of Theorem \[bridge principle\] shows that after passing to a subsequence, $\Sigma_\a^{\e_i}\to\Sigma$, where $\Sigma$ is an absolutely area minimizing surface with $\partial\Sigma=\Gamma$ and the convergence is with respect to Hausdorff distance and with respect to measure. Therefore we can argue as in the proof of Lemma \[Agopen\], using White’s curvature bound, Theorem \[white2c\] (which is applicable because of the measure convergence and the monotonicity formula [@Sim]), to conclude that this convergence is smooth in compact sets of $\ov M\setminus (\Gamma\cup\a)$ and in fact it is also smooth around points of $\Gamma\setminus\a$, where $\Gamma$ is smooth. In particular, if $\Gamma$ is smooth, then it is a smooth convergence in compact sets of $\ov M\setminus \a$.
Note that the result of Lemma \[bridge principle\] is not an extension of the classical bridge principle to absolutely area minimizing surfaces. In the classical bridge principle, one starts with a fixed stable minimal surface $T$ with boundary $\Gamma$, and after adding a thin bridge $S^\epsilon_\alpha$ to $\Gamma$, one gets a new surface $T \sharp S^\epsilon_\alpha$ with boundary $\widehat{\Gamma}^\epsilon_\alpha$ which is close to the original surface $T$ [@MY3], [@Wh4].
In our case, if $\Gamma$ bounds more than one absolutely area minimizing surface, say $\{\Sigma_1, \Sigma_2,..., \Sigma_k\}$, then the absolutely area minimizing surface $\Sigma^\epsilon_\alpha$ bounded by $\widehat{\Gamma}^\epsilon_\alpha$ is close to $\Sigma_{i_0} \cup S^\epsilon_\alpha$ for some $1\leq i_0 \leq k$. Therefore, one may not get a modified absolutely area minimizing surface $\Sigma_\a^\e$ which is close to $\Sigma_i$ for each $i$. The absolutely area minimizing surface $\Sigma_{i_0}$ which is close to $\Sigma^\epsilon_\alpha$ depends on the choice of the simple path $\alpha$ and then $\Sigma_{i_0}$ is the limit of the sequence of absolutely area minimizing surfaces $\{\Sigma^{\epsilon_i}_\alpha\}$ in $M$ as it appears in the proof of Theorem \[bridge principle\]. At this point, we conjecture that $\Sigma_{i_0}$ must be one of the two canonical absolutely area minimizing surfaces $\Sigma^+,\Sigma^-$ bounding $\Gamma$ as in Lemma 4.2 of [@Co1], since they are the extremal absolutely area minimizing surfaces with boundary $\Gamma$.
On the other hand, if one relax the condition of area minimizing to just minimal for the new surface, then the classical bridge principle for extreme curves is valid by using the techniques in Section 4. There, we show how one can apply [@Wh4] (see Theorem \[white4\] below) to get a [*minimal surface*]{} $S$ with $\partial S = \widehat{\Gamma}^\epsilon_\alpha$, which is close to $\Sigma_{i_0} \cup S^\epsilon_\alpha$ for a specified area minimizing surface $\Sigma_{i_0}$ with $\partial \Sigma_{i_0} = \Gamma$.
\[OI\] (Orientation Issues) If $\Sigma$ is an oriented surface whose boundary is a simple closed curve $\Gamma$, gluing a strip $S\sim I\times
I$ “trivially" along the boundary $\Gamma$ will give another oriented surface whose boundary consists of $2$ simple closed curves. However, if the strip $S$ glued to $\Sigma$ along $\Gamma$ has a “twist", then the new surface will not be oriented anymore, and the boundary will be a simple closed curve again. The $\Gamma$-admissibility condition on $\alpha$ expresses this difference. Therefore, the surgery along a $\Gamma$-admissible path $\alpha$ gives an oriented surface with $2$ boundary components, whereas the surgery along a non-$\Gamma$-admissible path $\beta$ would be gluing a twisted strip along $\Gamma$ to $\Sigma$, and give a nonorientable surface whose boundary is a simple closed curve again (See Figure 3).
Next we show in Corollary \[genuscor\] how one can relate the genus of the surfaces $\Sigma^\e_\a$ to that of $\Sigma$, where $\Sigma^\e_\a,
\Sigma$ are as in Lemma \[bridge principle\]. For the proof, we need White’s bridge principle for stable minimal surfaces [@Wh4]. For the convenience of the reader, we state it here in the form that we will apply it, using our notation so far.
\[white4\][@Wh4] Let $\Gamma, \a$ satisfy assumption *(a)* or *(b)* of *Theorem \[bridge principle\]*. We further assume that both $\Gamma$ and $\a$ are smooth. Let $\Sigma$ be a smooth and strictly stable surface with $\partial\Sigma=\Gamma$. Let also $\Sigma_n$ be a sequence of minimal surfaces converging smoothly to $\Sigma$ and such that for $\Gamma_n:=\partial\Sigma_n$, we have that $W\cap \Gamma_n= W\cap\Gamma$, for an open set $W$ containing $\a$. Then:
- There is an open set $U$ containing $\Sigma\cup\a$ such that $\Sigma$ is the unique area minimizing surface in $\ov{U}$.
- If for a sequence $\e_n\downarrow 0$, $T_n$ minimizes area among all surfaces in $\ov{U}$ with boundary ${\widehat\Gamma^{\e_n}_{n\a}}$, where ${\widehat\Gamma^{\e_n}_{n\a}}$ are the smooth curves given in *Definition \[wtG\]* or *\[wtG2\]* (see also *Remark \[smoothglue\]*), using now $\Gamma_n$ instead of $\Gamma$, then
- $|T_n|\stackrel{n\to\infty}{\longrightarrow}|\Sigma|$.
- $T_n$ converges smoothly to $\Sigma$ on compact subsets of $\ov M\setminus \a$.
- For sufficiently large $n$, $T_n$ and $\Sigma_n\cup S_\a^\e$ are diffeomorphic.
- For sufficiently large $n$, $T_n$ is unique and strictly stable.
\[rmkonwhite4\] In [@Wh4], part (i) of the above theorem is actually proven without any assumptions on the regularity of $\Gamma$ and furthermore the Theorem holds, without necessarily $\Gamma$ lying in the boundary of $M$. Also, in [@Wh4] it is actually proven (and we will need it later) that Theorem \[white4\] still holds if $\Sigma_0$ is an immersed surface. Note also that although $\Gamma$ might not lie in $\partial M$, as long as $\Gamma\cap
W\subset\partial M$ and thus also $\Gamma_n\cap W\subset\partial M$, the construction of the curves ${\widehat\Gamma^{\e_n}_{n\a}}$ given in Definitions \[wtG\] and \[wtG2\] still makes sense.
\[genuslem\] Let $\Gamma, \a$ satisfy assumption *(a)* or *(b)* of *Theorem \[bridge principle\]*. We further assume that $\a$ is smooth, $\Gamma$ is smooth in an open set $W$ containing $\a$ and for any absolutely area minimizing surface $\Sigma$ with $\partial\Sigma=\Gamma$, $\operatorname{genus}(\Sigma)<\infty$. Then there exists $\e_\a>0$ such that for almost every $0<\e<\e_a$ and for any absolutely area minimizing surface $\Sigma_\a^{\e}$, with $\partial\Sigma_\a^{\e}=\widehat\Gamma_\a^\e$, where $\widehat\Gamma_\a^\e$ is as in *Definition \[wtG\]* or *\[wtG2\]* in case assumption *(a)* or *(b)* is satisfied respectively, the following holds:
- there exists a surface $\Sigma_{0\a}^{\e}\subset\Sigma_\a^{\e}$ such that $$\chi(\Sigma_{0\a}^{\e})=\chi(\Sigma\cup\alpha),$$ for some absolutely area minimizing surface $\Sigma$ with $\partial\Sigma=\Gamma$ and where $\chi$ denotes the Euler characteristic, and furthermore
- under assumption *(a)*, $\Sigma_{0\a}^{\e}$ has two boundary components and
- under assumption *(b)*, $\Sigma_{0\a}^{\e}$ has a unique boundary component.
- In particular, if $\Gamma$ is smooth then the surfaces $\Sigma_{0\a}^{\e}$ in (ii) can be taken to be equal to $\Sigma_{\a}^{\e}$ and thus we have that $$\chi (\Sigma_\a^{\e})=\chi(\Sigma\cup\a).$$
We prove this lemma by contradiction. Assume the lemma is not true, then for any $\e_\a$, there exists $0<\e<\e_a$ such that the conclusion (i) of the lemma (and in particular conclusion (ii), in case $\Gamma$ is smooth) is not true. Hence there exists a sequence $\{\e_i\}_{i\in\N}$, with $\e_i\downarrow 0$ such that for all $i$, the lemma with $\e=\e_i$ fails.
Let $\Sigma_\a^{\e_i}$ be a sequence of absolutely area minimizing surfaces with $\partial\Sigma_\a^{\e_i}=\widehat\Gamma_\a^{\e_i}$. Then by Lemma \[bridge principle\] and in particular Remark \[C2convrmk\], there exists an absolutely area minimizing surface $\Sigma$, with $\partial\Sigma=\Gamma$, such that after passing to a subsequence, $\Sigma_\a^{\e_i}\to\Sigma$, with the convergence being with respect to the Hausdorff distance, the measure and also it is a smooth convergence in compact sets of $\ov M\setminus(\Gamma\cup \a)$ and it is also smooth around points of $\Gamma\setminus\alpha$, where $\Gamma$ is smooth. Finally we let $g_0=\operatorname{genus}(\Sigma)<\infty$
Let now $\Gamma_0\subset\Sigma$ be a smooth curve on $\Sigma$, such that $\Gamma_0=\Gamma$ in $W\supset\a$, $\Gamma_0\sim\Gamma$ and such that the part of $\Sigma$ bounded by $\Gamma_0$, which we call $\Sigma_0$, has genus $g_0$. Note that $\Sigma_0$ is absolutely area minimizing and it is smooth (by Theorem \[absmincur\] and [@HSi]). We furthermore take $\Gamma_0$, so that it is not identical to $\Gamma$, so that $\Sigma_0$ is strictly contained in $\Sigma$ and therefore it is strictly stable (because $\Sigma$ is absolutely area minimizing), (cf. [@FiSc]).
Now, because of the fact that the convergence $\Sigma^{\epsilon_i}_\alpha\to \Sigma$ is smooth in compact sets of $\overline M\setminus (\Gamma\cup\a)$ and also around points of $\Gamma\setminus \a$ where $\Gamma$ is smooth, we can find curves $\widehat\Gamma_{0\a}^{\e_i}\subset \Sigma_\a^{\e_i}$ such that
- Under hypothesis (a), i.e. when $\Gamma$ is a single simple closed curve and $\a$ a $\Gamma$-admissible path, $\widehat\Gamma_{0\a}^{\e_i}\subset \Sigma_\a^{\e_i}$ consists of two connected components and under hypothesis (b), i.e. when $\Gamma=\Gamma_1\cup\Gamma_2$ and $\a$ is a path connecting them, $\widehat\Gamma_{0\a}^{\e_i}\subset \Sigma_\a^{\e_i}$ consists of a single connected component.
- $\widehat\Gamma_{0\a}^{\e_i}=\widehat\Gamma_{\a}^{\e_i}$ in $W\supset\widehat\Gamma_{\a}^{\e_i}\cap S_\a^{\e_i}$
- $\widehat\Gamma_{0\a}^{\e_i}$ is smooth away from $S_\a^{\e_i}$ (hence $\widehat\Gamma_{0\a}^{\e_i}$ is smooth) and
- $\widehat\Gamma_{0\a}^{\e_i}\to \Gamma_0$ smoothly away from $\a$.
Let $\Sigma^{\e_i}_{0\a}$ be the part of $\Sigma^{\e_i}_{\a}$ bounded by $\widehat\Gamma_{0\a}^{\e_i}$. Then, because of the convergences $\Sigma_\a^{\e_i}\to\Sigma$ and $\widehat\Gamma_{0\a}^{\e_i}\to \Gamma_0$, as described above, we have that $\Sigma^{\epsilon_i}_{0\alpha}$ converges to $\Sigma_0$ and the convergence $\Sigma^{\epsilon_i}_{0\alpha}\to \Sigma_0$ is smooth in compact sets of $\overline M\setminus \a$.
Furthermore, letting $\Gamma_i=\partial(\Sigma^{\e_i}_{0\a}\cup S_\a^{\e_i})$, we have that $\Gamma_i$ is smooth, $\widehat\Gamma_{0\a}^{\e_i}$ is the curve that we obtain by attaching a bridge around $\a$ on the curve $\Gamma_i$ as described in Definitions \[wtG\] and \[wtG2\] and $\Gamma_i\to\Gamma_0$ smoothly.
Note, that since $\Sigma_0$ is strictly stable, by White’s bridge principle, and in particular by (i) of Theorem \[white4\] (see also Remark \[rmkonwhite4\]), there exists an open set $U$ containing $\Sigma_0\cup\a$, such that $\Sigma_0$ is the unique area minimizing surface with boundary $\Gamma_0$ in $\overline U$. Since the convergence $\Sigma^{\epsilon_i}_{0\alpha}\to\Sigma_0$ is also with respect to the Hausdorff distance, for any $\eta>0$, there exists $i_0$ such that for any $i\ge i_0$, $\Sigma^{\epsilon_i}_{0\alpha}$ is in $\eta$-neighborhood of $\Sigma_0\cup\alpha$. Taking $\eta$ small enough (so that the $\eta$-neighborhood of $\Sigma_0\cup\alpha$ is contained in $U$) we have that there exists $i_0$ such that for any $i\ge i_0$, $\Sigma^{\epsilon_i}_{0\alpha}\subset U$. Since $\Sigma^{\epsilon_i}_{0\alpha}$ are absolutely area minimizing they also minimize area among all surfaces with the same boundary in $\overline U$.
Let $\Sigma_i$ now be an area minimizing surface in $U$, such that $\partial \Sigma_i=\Gamma_i$. Then $\operatorname{Area}(\Sigma_i)\le
\operatorname{Area}(\Sigma^{\e_i}_{0\a})+\operatorname{Area}(S_\a^{\e_i})\le \operatorname{Area}(\Sigma^{\e_i}_{\a})+\operatorname{Area}(S_\a^{\e_i})$, so after passing to a subsequence, $\Sigma_i$ converge to an area minimizing surface in $U$ with boundary equal to $\Gamma_0$, and since $
\Sigma_0$ is the unique such surface, we have that $\Sigma_i\to\Sigma_0$. Arguing as in the proof of Lemma \[bridge principle\] and in particular as in Remark \[C2convrmk\], we can apply White’s curvature estimate, Theorem \[white2c\], to conclude that the convergence is smooth.
Note now that $\Sigma^{\e_i}_{0\a}$ is absolutely area minimizing and thus by White’s bridge principle, in particular by (ii) of Theorem \[white4\], for $i$ large enough $\Sigma_{0\a}^{\e_i}$ and $\Sigma_i\cup S_\a^{\e_i}$ are diffeomorphic. Since $\Sigma_i\to\Sigma_0$ smoothly and $(\Sigma_0\cup
S_\a^{\e_i})\sim(\Sigma_0\cup\a)$ we have that $\Sigma_{0\a}^{\e_i}$ is homotopic to $\Sigma_0\cup\a$, which by construction is homotopic to $\Sigma\cup\alpha$, and thus $\chi(\Sigma_{0\a}^{\e_i})=\chi(\Sigma\cup\a)$. This leads to a contradiction, since we have assumed that (i) of the lemma with $\e=\e_i$ fails for any $i$ and thus finishes the proof of part (i) of the lemma.
If $\Gamma$ is smooth, then we can pick $\Gamma_0$, so that $\Sigma\setminus\Sigma_0$ is topologically a disk. Furthermore, since $\Gamma_0=\Gamma$ in $W\supset \a$, there exists an open set $\mathcal O\subset M\setminus \a$, such that $\mathcal O\supset \Sigma\setminus\Sigma_0$, $\mathcal O\supset \Sigma_{\a}^{\e_i}\setminus\Sigma_{0\a}^{\e_i}$ for $i$ large enough and the convergence $\Sigma_\a^{\e_i}\to\Sigma$ is smooth in $\mathcal O$. This, along with the convergence $\Sigma_{0\a}^{\e_i}\to\Sigma_0$ and the fact that $\Sigma\setminus\Sigma_0$ is topologically a disk, implies that for large $i$, $\Sigma_{\a}^{\e_i}\setminus\Sigma_{0\a}^{\e_i}$ is also topologically a disk and therefore $\chi(\Sigma_{\a}^{\e_i})=\chi(\Sigma_{0\a}^{\e_i})=\chi(\Sigma\cup\a)$. This leads also to a contradiction, since we have assumed that if $\Gamma$ is smooth, then (ii) of the lemma with $\e=\e_i$ fails for any $i$ and thus finishes the proof of the lemma.
\[genuscor\] Let $\Gamma, \a$ satisfy assumption *(a)* or *(b)* of *Theorem \[bridge principle\]*. We further assume that $\a$ is smooth, $\Gamma$ is smooth in an open set $W$ containing $\a$ and for any absolutely area minimizing surface $\Sigma$, with $\partial\Sigma=\Gamma$, $\Sigma$ is connected and $\operatorname{genus}(\Sigma)<\infty$. Then, for almost every $0<\e<\e_a$ , where $\e_a$ is as in *Lemma \[genuscor\]*, and for any absolutely area minimizing surface $\Sigma_\a^{\e}$, with $\partial\Sigma_\a^{\e}=\widehat\Gamma_\a^\e$, where $\widehat\Gamma_\a^\e$ is as in *Definition \[wtG\]* or *\[wtG2\]* in case assumption *(a)* or *(b)* is satisfied respectively, the following holds:
- - under assumption *(a)* $$\operatorname{genus}(\Sigma_\a^{\e})\ge\operatorname{genus}(\Sigma),$$
- under assumption *(b)* $$\operatorname{genus}(\Sigma_\a^{\e})\ge\operatorname{genus}(\Sigma)+1$$
for some absolutely area minimizing surface $\Sigma$ with $\partial\Sigma=\Gamma$ .
- If in particular $\Gamma$ is smooth, then
- under assumption *(a)* $$\operatorname{genus}(\Sigma_\a^{\e})=\operatorname{genus}(\Sigma),$$
- under assumption *(b)* $$\operatorname{genus}(\Sigma_\a^{\e})=\operatorname{genus}(\Sigma)+1.$$
Recall that by the classification of orientable surfaces, $S$ is a genus $g$ surface with $k$ boundary components if and only if $\chi(S)= (2-k)-2g$.
Let $\Sigma^{\e}_{0\alpha}\subset \Sigma^{\e}_0$ be as in (ii) of Lemma \[genuslem\], so that $$\chi(\Sigma^{\e}_{0\alpha})=\chi(\Sigma\cup\alpha)$$ for some absolutely area minimizing surface $\Sigma$ with $\partial\Sigma=\Gamma$ and so that $\Sigma_{0\alpha}^\e$ has two boundary components in case assumption (a) is satisfied and a unique boundary component in case assumption (b) is satisfied. Let also $g_0=\operatorname{genus}(\Sigma)<\infty$.
By realizing the arc $\alpha$ as an edge in the triangulation, we have $\chi(\Sigma\cup{\alpha}) = \chi(\Sigma) -1$.
Now under hypothesis (a), $\Sigma$ is a genus $g_0$ surface with one boundary component, therefore $\chi(\Sigma\cup{\alpha}) = \chi(\Sigma)
-1=-2g_0$. Since, $\Sigma^{\epsilon}_{0\alpha}$ has two boundary components and $\chi(\Sigma_{0\a}^{\e})=\chi(\Sigma\cup\a)=-2g_0$, we have that $\Sigma^{\epsilon}_{0\alpha}$ is a genus $g_0$ surface with two boundary components.
On the other hand, under hypothesis (b), as $\Sigma$ is a connected surface with two boundary components, $\chi(\Sigma\cup{\alpha}) = \chi(\Sigma)
-1=-2g_0-1$ where $g_0$ is the genus of $\Sigma$. Since, $\Sigma^{\epsilon}_{0\alpha}$ has a unique boundary component and $\chi(\Sigma_{0\a}^{\e})=\chi(\Sigma\cup\a)=-2g_0-1$, we have that $\Sigma^{\epsilon}_{0\alpha}$ is a genus $g_0+1$ surface with one boundary component.
Note now that since $\Sigma_{0\a}^{\e}\subset\Sigma_\a^{\e}$, $\operatorname{genus}(\Sigma_\a^{\e})\ge\operatorname{genus}(\Sigma_{0\a}^{\e})$. This finishes the proof of (i) of the lemma.
Part (ii) is immediate from (ii) of Lemma \[genuslem\], since if $\Gamma$ is smooth, then $\Sigma_{0\a}^{\e}=\Sigma_\a^{\e}$.
Horn Surgery and Thin Handles {#hssection}
-----------------------------
In the previous part, by adding a bridge near a $\Gamma$-admissible path $\alpha$ to the simple closed curve $\Gamma$, we get $\widehat{\Gamma}^\epsilon_\alpha=\Gamma\sharp S_\a^\e$ (as in Definition \[wtG\]) which is the union of two simple closed curves $\widehat{\Gamma}^{\epsilon,1}_\alpha$ and $\widehat{\Gamma}^{\epsilon,2}_\alpha$ in $\partial M$. Now, we repeat this process one more time, by connecting the two simple closed curves $\widehat{\Gamma}^{\epsilon,1}_\alpha$ and $\widehat{\Gamma}^{\epsilon,2}_\alpha$ via a bridge around a new path $\tau$, connecting the two curves. Thus, we get a simple closed curve $\widehat{\Gamma}^{\epsilon,\delta}_{\alpha,\tau}=\widehat\Gamma_{\a}^\e\sharp S_\tau^\delta$ (as in Definition \[a2\] and where $\d$ here denotes the width of the new bridge) which bounds an absolutely area minimizing surface whose genus, we will show, is strictly greater than the genus of the absolutely area minimizing surface $\Sigma$ in $M$ with $\partial \Sigma = \Gamma$. In other words, by adding two bridges $S_\epsilon(\alpha)$ and $S_\delta(\tau)$ successively to a simple closed curve $\Gamma$, we modify the absolutely area minimizing surface it bounds by adding a thin handle. We will refer to this operation on $\Gamma$ as [*the horn surgery*]{} (See Figure 4), and to the resulting additional handle in the new absolutely area minimizing surface as [*the thin handle*]{} (See Figure 5). Now, let’s give the formal construction and definitions.
We will use the notation introduced in Section \[bridgeprinciplesubsection\]. Let $\Gamma$, $\alpha$ and $\widehat{\Gamma}^\epsilon_\alpha$ be as in Definition \[wtG\]. Let $z$ be the midpoint of $\alpha$. Let, $z^+_\epsilon, z^-_\epsilon$ be the midpoints of $\alpha^+_\epsilon,\alpha^-_\epsilon$ respectively. Let $\tau_\epsilon$ be a small path from $z^+_\epsilon$ to $z^-_\epsilon$ through $z$ in $\partial M$. For $0<\delta\ll \epsilon$, let $S^\delta_\tau= N_\delta(\tau_\epsilon)\cap S^\epsilon_\alpha$ be a tiny strip (“bridge") of width $\d$ around $\tau_\epsilon$ connecting the simple closed curves $\widehat{\Gamma}^{\epsilon,1}_\alpha$ and $\widehat{\Gamma}^{\epsilon,2}_\alpha$ where $N_\delta(\tau_\epsilon)$ is the $\delta$ neighborhood of $\tau_\epsilon$ in $\partial M$. Here, we assume $\delta$ is sufficiently small so that for any $0<\delta'<\delta$, $\partial S^{\delta'}_\tau \cap \operatorname{int}(S^\epsilon_\alpha)$ consists of exactly two arc segments, which we call $\tau^+_\epsilon$ and $\tau^-_\epsilon$. Then $\partial S^\delta_\tau = (\tau^+_\epsilon \cup \tau^-_\epsilon) \cup (\gamma^1_\delta \cup
\gamma^2_\delta)$ where $\gamma^i_\delta \subset \widehat{\Gamma}^{\epsilon,i}_\alpha$, $i=1,2$.
\[hs\] We define [*the horn surgery*]{} on $\Gamma\in{\mathcal{A}}$ along a $\Gamma$-admissible path $\alpha$ to be the curve $$\widetilde{\Gamma}_\alpha = \Gamma\sharp S^\epsilon_\alpha\sharp S^\delta_\tau$$ with $\tau$ being a path passing through the midpoint of $\a$ as described above (recall Definitions \[wtG\], \[wtG2\]). With the above notation this curve is given $$\widetilde{\Gamma}_\alpha=(\widehat{\Gamma}^\epsilon_\alpha \setminus
(\gamma^1_\delta \cup \gamma^2_\delta)) \cup (\tau^+_\epsilon \cup \tau^-_\epsilon).$$ Hence, $\widetilde{\Gamma}_\alpha$ is a simple closed curve in $\partial M$ (See Figure 4).
\[smoothglueII\] By Remark \[smoothglue\], we note here that if $\a$ is smooth and the curve $\Gamma$ is smooth around $\a$, then $\widehat\Gamma_\a^\e$ is smooth around $\widehat\Gamma_\a^\e\cap S^\e_\a$ and in particular around $\widehat\Gamma_\a^\e\cap\tau$. Then we can take the path $\tau$ to be smooth and we can consider gluing the bridge $S^\d_\tau$ smoothly along $\widehat\Gamma_\a^\e$, so that the resulting curve $\widetilde{\Gamma}_\a$ is smooth around $\widehat\Gamma_\a^\e\cap S^\d_\tau$ and hence it is also smooth around $\widehat\Gamma_\a^\e\cap S^\e_\a$.
In the rest of the paper and as long as these smoothness assumptions are satisfied we will always consider constructing the curve $\wt{\Gamma}_\a$ to be smooth around $\a$.
Let $\Gamma\in {\mathcal{A}}_g$, and $\alpha$ be a $\Gamma$-admissible path. Now, we claim that after the horn surgery on $\Gamma$ along $\alpha$, $\widetilde{\Gamma}_\alpha$ is in ${\mathcal{A}}_{g+1}$. In other words, the horn surgery on the boundary curve $\Gamma$ will add a thin handle to an area minimizing surface $\Sigma$ with boundary $\Gamma$ and thus increase its genus (See Figure 5).
\[thinhandle\] [**(Thin handle)**]{} Let $\Gamma\in{\mathcal{A}}_g$, and $\alpha$ be a $\Gamma$-admissible smooth path, such that $\a$ is smooth, $\Gamma$ is smooth in an open set $W$ containing $\a$ and for any absolutely area minimizing surface $\Sigma$ with $\partial\Sigma=\Gamma$, $\operatorname{genus}(\Sigma)<\infty$. Then there exists $\e_a$, such that for almost every $0<\epsilon<\epsilon_\alpha$ the following holds:
- There is a sufficiently small $\delta_{\tau}>0$, such that and for any $0<\delta<\delta_{\tau}$, after the horn surgery on $\Gamma$ along $\alpha$ (see *Definition \[hs\]*), $\widetilde{\Gamma}_\alpha\in{\mathcal{A}}_{g+1}$. In particular, for any absolutely area minimizing surface $\wt\Sigma_\a$ with $\partial\wt\Sigma_\a=\wt\Gamma_\a$, $\operatorname{genus}(\wt\Sigma_\a)\ge\operatorname{genus}(\Sigma)+1$, where $\Sigma$ is an absolutely area minimizing surface with $\partial\Sigma=\Gamma$.
- If $\Gamma$ is smooth, then $\widetilde{\Gamma}_\alpha$ is a smooth curve in ${\mathcal{A}}_{g+1}$ and in particular, for any absolutely area minimizing surface $\wt\Sigma_\a$ with $\partial\wt\Sigma_\a=\wt\Gamma_\a$, $\operatorname{genus}(\wt\Sigma_\a)=\operatorname{genus}(\Sigma)+1$, where $\Sigma$ is an absolutely area minimizing surface with $\partial\Sigma=\Gamma$.
Recall (Remarks \[smoothglue\], \[smoothglueII\]) that $\widehat \Gamma_\a^\e$ and $\widetilde{\Gamma}_\alpha$ are both smooth around $S^\e_\a$. Note first that for any $\e<\e_a$, with $\e_\a$ as in Corollary \[genuscor\], and for any absolutely area minimizing surface $\Sigma_\a^\e$ with $\partial\Sigma_\a^\e=\widehat\Gamma_\a^\e$ $$\operatorname{genus}(\Sigma_\a^\e)\ge \operatorname{genus}(\Sigma)$$ for an absolutely area minimizing surface $\Sigma$ with $\partial\Sigma=\Gamma$ and the above inequality is actually an equality if $\Gamma$ is smooth.
We apply now again Corollary \[genuscor\] with $\Gamma$, $\a$, $\Sigma_\a^\e$ and $\Sigma$ replaced by $\widehat\Gamma_\a^\e$, $\tau$, $\wt\Sigma_\a$ and $\Sigma_\a^\e$ (note that here $\Sigma_\a^\e$, $\wt\Sigma_\a$ denote [*[any]{}*]{} absolutely area minimizing surfaces with boundary equal to $\widehat\Gamma_\a^\e$ and $\wt\Gamma_\a$ respectively). Since now we use the hypothesis (b) of Corollary \[genuscor\], we get that there exists a $\delta_\tau>0$ such that for any $0<\d<\d_\tau$ and for any absolutely area minimizing surface $\wt\Sigma_\a$, with $\partial\wt\Sigma_\a=\wt\Gamma_\a$ $$\operatorname{genus}(\wt\Sigma_\a)\ge\operatorname{genus}(\Sigma_\a^\e)+1\ge \operatorname{genus}(\Sigma)+1$$ for an absolutely area minimizing surface $\Sigma_\a^\e$ with boundary equal to $\widehat\Gamma_\a^\e$ and an absolutely area minimizing surface $\Sigma$ with boundary equal to $\Gamma$. And in the case when $\Gamma$ is smooth (cf. Remark \[smoothglueII\]) the above inequalities are actually equalities and thus $$\operatorname{genus}(\wt\Sigma_\a)=\operatorname{genus}(\Sigma_\a^\e)+1= \operatorname{genus}(\Sigma)+1.$$
In the horn surgery construction, the $\Gamma$-admissible path $\alpha$ determines the $\epsilon_\alpha>0$, and the path $\tau$ determines the $\delta_\tau>0$. Since $\tau$ depends on the $0<\epsilon<\epsilon_\alpha$, $\delta_\tau$ also depends on $\epsilon_\alpha$. From now on, we will assume the choices for the horn surgery on $\Gamma$ along $\alpha$ are canonical. In other words, we always consider $0<\epsilon<\epsilon_\a$, $\tau$ and $0<\delta<\delta_\tau$, so that the horn surgery is well defined. Furthermore from now on we assume that $\e_\a$ and $\d_\tau$ are small enough so that Lemma \[thinhandle\] holds, i.e. if $\Gamma$ is smooth around $\a$, then for $\widetilde{\Gamma}_\alpha = \Gamma\sharp S^\epsilon_\alpha\sharp S^\delta_\tau$ we can apply Lemma \[thinhandle\] and in particular if $\Gamma$ is smooth, then we can apply (ii) of Lemma \[thinhandle\].
Now, let’s consider the space ${\mathcal{B}}_g$. We show that ${\mathcal{B}}_g$ is nonempty for any $g\geq 0$. As noted before, ${\mathcal{B}}_g = {\mathcal{A}}_g \setminus {\mathcal{A}}_{g+1}$. In Lemma \[Agopen\] we proved that for any $g$, ${\mathcal{A}}_g$ is open, but this does not imply that ${\mathcal{B}}_g$ is nonempty. In order to show this, we need to rule out the case ${\mathcal{A}}_g =
{\mathcal{A}}_{g+1}$. In other words, we will show that for any $g\geq 0$, there is a simple closed curve $\Gamma_g\in {\mathcal{A}}$ such that the minimum genus among all absolutely area minimizing surface with boundary $\Gamma_g$ is exactly $g$.
\[Bgne\] For any $g\geq 0$, ${\mathcal{B}}_g$ is nonempty.
Since ${\mathcal{B}}_g={\mathcal{A}}_g\setminus{\mathcal{A}}_{g+1}$, it suffices to show the existence of a simple closed curve $\Gamma_g\in {\mathcal{A}}_g$ such that $\Gamma_g$ bounds an absolutely area minimizing surface $\Sigma$ of genus $g$.
Clearly, ${\mathcal{B}}_0$ is not empty, as we can take a sufficiently small smooth simple closed curve in $\partial M$, so that the absolutely area minimizing surface bounded by this curve is a smooth disk in $M$, say $D_0$, and since $M$ is strictly mean convex, $D_0$ is also properly embedded, i.e. $D_0\cap\partial M=\partial D$.
Now, let $\Gamma_0$ be a smooth curve in ${\mathcal{B}}_0$ and take a $\Gamma_0$-admissible and smooth path $\alpha_1$. Then, by (ii) of Lemma \[thinhandle\], the horn surgery on $\Gamma_0$ along $\alpha_1$ will give us a simple closed and smooth curve $\widetilde{\Gamma_0}_{\alpha_1}$, say $\Gamma_1$, such that $\Gamma_1 \in {\mathcal{B}}_1$.
We can argue now by induction, that for any $k$ there exists a smooth curve in ${\mathcal{B}}_k$ as follows: Assume that there exists a smooth curve $\Gamma_{k-1}\in{\mathcal{B}}_{k-1}$ and take a $\Gamma_{k-1}$-admissible and smooth path $\alpha_k$. The horn surgery on $\Gamma_{k-1}$ along $\alpha_k$ (Lemma \[thinhandle\]) gives a new smooth curve $\Gamma_k$ with $\Gamma_k \in {\mathcal{B}}_k$ and thus ${\mathcal{B}}_k\ne\emptyset$.
\[Bgnermk\] Intuitively, this gives a construction of adding a handle to an absolutely area minimizing surface by modifying the boundary curve. Note that, since the handles can be as small as we want, the area of the new surfaces can be as close as we want to the area of the original surface (cf. Lemma \[bridge principle\]). Furthermore, by the proof of Theorem \[Bgne\], we actually conclude that not only can we always find a simple closed curve in ${\mathcal{B}}_g$, but also we can find a smooth simple closed curve in ${\mathcal{B}}_g$.
In Lemma \[Agopen\], we proved that for any $g>0$, ${\mathcal{A}}_g$ is open in ${\mathcal{A}}$, and in Theorem \[Bgne\], we showed that ${\mathcal{B}}_g={\mathcal{A}}_g\setminus{\mathcal{A}}_{g+1}$ is nonempty. Now, by using the handle construction in Lemma \[thinhandle\], we will prove that ${\mathcal{A}}_g$ is not only open in ${\mathcal{A}}$, but also dense in ${\mathcal{A}}$.
\[Adense\] For any $g\geq 0$, ${\mathcal{A}}_g$ is dense in ${\mathcal{A}}$, i.e. $\overline{{\mathcal{A}}}_g = {\mathcal{A}}$.
Fix $g>0$. It suffices to show that for any $\Gamma_0 \in {\mathcal{A}}$, and for any $\mu>0$, there exists a simple closed curve $\Gamma\in {\mathcal{A}}_g$ with $d(\Gamma,\Gamma_0)<\mu$, where by $d(\Gamma,\Gamma_0)<\mu$ we mean that $\Gamma$ is in a $\mu$-neighborhood of $\Gamma_0$ in the $\mathcal C^0$ topology (cf. Definition \[c0rmk\]).
Let $\Gamma_0\in {\mathcal{A}}$. If $\Gamma_0\in{\mathcal{A}}_g$, then we are done as we can take $\Gamma=\Gamma_0$. Hence we can assume that $\Gamma_0\notin {\mathcal{A}}_g$, which implies that for some $g_0$, with $0\leq g_0\leq g-1$, $\Gamma_0$ belongs to ${\mathcal{B}}_{g_0}\subset{\mathcal{A}}_{g_0}$. Since ${\mathcal{A}}_{g_0}$ is open we can take $\Gamma_1\in {\mathcal{A}}_{g_0}$, so that $\Gamma_1$ is piecewise smooth and $d(\Gamma_0, \Gamma_1)<\mu/2$.
If $\Gamma_1\in{\mathcal{A}}_g$, then we can take $\Gamma=\Gamma_1$ and we are done. Therefore, we assume that $\Gamma_1\in{\mathcal{B}}_{g_1}$, with $g>g_1\ge g_0$ and we will construct $\Gamma$ by $h=g-g_1$ horn surgeries on $\Gamma_1$.
Recall the construction in Lemma \[thinhandle\]. Since $\Gamma_1$ is piecewise smooth, we can find a $\Gamma_1$-admissible and smooth path $\alpha_1$ satisfying the hypothesis of Lemma \[thinhandle\], i.e. so that $\Gamma_1$ is smooth in an open neighborhood of $\a_1$. This implies that since $\Gamma_1\in {\mathcal{A}}_{g_1}$, $\widetilde{\Gamma_1}_\alpha \in {\mathcal{A}}_{g_1+1}$. Let $\Gamma_2= \widetilde{\Gamma_1}_\alpha$, so that $\Gamma_2$ is obtained by horn surgery on $\Gamma_1$ along the $\Gamma_1$-admissible path $\alpha_1$. In general we let $\Gamma_k$ to be the curve obtained by horn surgery on $\Gamma_{k-1}$ along a $\Gamma_{k-1}$-admissible smooth path ${\alpha_k}$, satisfying the hypothesis of Lemma \[thinhandle\]. Then $\Gamma_k\in{\mathcal{A}}_{g_1+k-1}$, $\Gamma_{k}\sim\Gamma_{k-1}$ and since we are free to choose the admissible paths $\a_k$ as short as we want, we can take the $\Gamma_k$’s so that, $d(\Gamma_{k-1}, \Gamma_k) < \displaystyle \frac{1}{2h}\mu$, for any given $\mu>0$. Let $\Gamma=\Gamma_{h+1}$. Then, by construction $\Gamma_h \in {\mathcal{A}}_{g_1+h}={\mathcal{A}}_{g}$, and $d(\Gamma_0,\Gamma) < \frac{\mu}{2}+h\frac{\mu}{2h}=\mu$. The proof follows.
\[Bndense\] For any $g\geq 0$, ${\mathcal{B}}_g$ is nowhere dense in ${\mathcal{A}}$.
${\mathcal{B}}_g={\mathcal{A}}_g\setminus{\mathcal{A}}_{g+1}$ and thus $\overline{\mathcal{B}}_g = \overline{\mathcal{A}}_g\setminus\operatorname{int}(A_{g+1}) ={\mathcal{A}}\setminus A_{g+1}$, since ${\mathcal{A}}_g$ is dense in ${\mathcal{A}}$ by Theorem \[Adense\] and ${\mathcal{A}}_{g+1}$ is open in ${\mathcal{A}}$, by Lemma \[Agopen\]. Since ${\mathcal{A}}_{g+1}$ is also dense in ${\mathcal{A}}$, $\operatorname{int}(\overline{{\mathcal{B}}_g})=\emptyset$.
\[genusinfty\] Note that ${\mathcal{B}}_\infty:=\bigcap_g {\mathcal{A}}_g \neq \emptyset$. To see this, even though ${\mathcal{B}}_\infty=\bigcap_g {\mathcal{A}}_g$, where all the ${\mathcal{A}}_g$ are open and dense in ${\mathcal{A}}$, the classical Baire Category Theorem argument ($\bigcap_g {\mathcal{A}}_g$ is dense in ${\mathcal{A}}$) would not work since ${\mathcal{A}}$ is not a complete metric space. However, it is possible to construct a simple rectifiable curve $\Gamma_\infty$ in $\partial M$ which bounds an absolutely area minimizing surface $\Sigma$ with infinite genus. To see this, take a simple closed curve $\Gamma$ in $\partial M$, and do the horn surgery infinitely many times by choosing sufficiently small admissible paths $\{\alpha_1, \alpha_2, ..\}$ successively (See Figure 6).
Note also that since the above construction can be done on any piecewise smooth $\Gamma\in{\mathcal{A}}$, we actually have that ${\mathcal{B}}_\infty$ is dense in ${\mathcal{A}}$.
To summarize, so far in this section, we have shown that ${\mathcal{A}}={\mathcal{A}}_0\varsupsetneq{\mathcal{A}}_1\varsupsetneq{\mathcal{A}}_2\varsupsetneq ...\varsupsetneq A_g \varsupsetneq ...$ where ${\mathcal{A}}_g$ is an open dense subset of ${\mathcal{A}}$ for any $g\geq 0$. Moreover, we have shown that ${\mathcal{B}}_g={\mathcal{A}}_g\setminus{\mathcal{A}}_{g+1}$ is not empty, and nowhere dense in ${\mathcal{A}}$ for any $g\geq 0$.
We finish this section with a result which states that there are simple closed curves bounding absolutely area minimizing surfaces with different genus. In other words, we show that the relation $\mathfrak{g}:{\mathcal{A}}\to {\mathbf{N}}$ such that $\mathfrak{g}(\Gamma)$ is defined to be the genus of an absolutely area minimizing surface $\Sigma_\Gamma$ in $M$ with $\partial \Sigma_\Gamma=\Gamma$, is not a function.
\[diffgenus\] Let $M$ be a compact, orientable, strictly mean convex $3$-manifold. Then, for any $g\geq 0$, there exists a simple closed curve $\Gamma$ in $\partial M$ such that $\Gamma$ bounds two different absolutely area minimizing surfaces $\Sigma_1$ and $\Sigma_2$ in $M$ with $g=\text{genus}(\Sigma_1)\ne \text{genus}(\Sigma_2)$.
Let $\Gamma$ be a smooth curve in ${\mathcal{B}}_g$ (cf. Theorem \[Bgne\], Remark \[Bgnermk\]). Let $\alpha$ be a smooth and $\Gamma$-admissible path in $\partial M$. Then, by Lemma \[bridge principle\], for any $\eta_0>0$, there is an $\epsilon_0>0$ such that any absolutely area minimizing surface $\Sigma_\alpha$ with boundary $\widehat{\Gamma}^{\epsilon_0}_\alpha$, is $\eta_0$-close to the surface $\Sigma\cup S^{\epsilon_0}_\alpha$, for some absolutely area minimizing surface $\Sigma$, with $\partial\Sigma=\Gamma$ (where $\widehat\Gamma_\a^{\e_0}, S_\a^{\e_0}$ are as in Definition \[wtG\]). Recall that since $\Gamma$ is smooth, we consider this bridge construction so that $\widehat\Gamma_\a^{\e_0}$ is smooth (cf. Remark \[smoothglue\]). Now foliate the strip $S^{\epsilon_0}_\alpha$ with arcs $\{\tau_{\delta}\}_{|\delta|\leq c}$ where $\tau_\delta$ is parametrized by the arclength of $\alpha$, with $|\alpha|=2c$, such that $\tau_0 = \tau$ and $\tau_{\pm c} \subset \Gamma$. We also construct the arcs $\tau_\d$, so that they are smooth and hence after the horn surgery, the curves $\widetilde{\Gamma}^\d_\alpha =\Gamma\sharp S_\a^{\e_0}\sharp
S_\tau^\delta=(\widehat{\Gamma}^{\epsilon_0}_\alpha \setminus (\gamma^1_\delta \cup \gamma^2_\delta)) \cup (\tau_{\d} \cup \tau_{-\d})$ are smooth, where $S^\delta_\tau$ represents the strip in $S^{\epsilon_0}_\alpha$ separated by the arcs $\tau_\delta$ and $\tau_{-\delta}$ and $\partial
S_\tau^\delta=\gamma^1_\delta \cup \gamma^2_\delta\cup\tau_{\d}\cup\tau_{-\d}$ (cf. Definition \[hs\], Remark \[smoothglueII\]). Then $\widetilde{\Gamma}^c_\alpha = \Gamma$ and $\widetilde{\Gamma}^0_\alpha =\widehat{\Gamma}^{\epsilon_o}_\alpha$.
Now we note the following:
- Let $\d\in(0,c]$, $\{\d_i\}_{i\in\N}$ be a sequence in $[0, c]$ such that $\d_i\to\d$ and let $\wt\Sigma_\a^{\d_i}$ be a sequence of absolutely area minimizing surfaces with $\partial\wt\Sigma_\a^{\d_i}=\wt\Gamma_\a^{\d_i}$. Then, arguing as in Lemma \[bridge principle\] (see in particular Remark \[C2convrmk\]), with the use of the compactness theorem in [@FeF] and White’s curvature estimate, Theorem \[white2c\], we have that after passing to a subsequence, $\wt\Sigma_\a^{\d_i}$ converges smoothly to $\wt\Sigma_\a^{\d}$, where $\wt\Sigma_\a^{\d}$ is an absolutely area minimizing surface with $\partial\wt\Sigma_\a^{\d}=\wt\Gamma_\a^{\d}$. This, along with the Gauss Bonnet theorem implies that $\lim_i\operatorname{genus}(\wt\Sigma_\a^{\d_i})=\operatorname{genus}(\wt\Sigma_\a^{\d})$.
- By Lemma \[thinhandle\], there is a $\delta_0>0$ such that for any $0<\delta<\delta_0$, $\widetilde{\Gamma}^\delta_\alpha$ is a smooth curve in ${\mathcal{B}}_{g+1}$ and on the other hand we know that $\widetilde{\Gamma}^c_\alpha=\Gamma\in{\mathcal{B}}_g$. Moreover, by (i) we have that $\d_0<c$.
Let $\d_1$ be the infimum of all $\d\in[0,c]$, such that $\widetilde{\Gamma}^\d_\alpha\in{\mathcal{B}}_g$. By (i) and (ii) above we know that $\d_1\in[\d_0,c)\subset(0,c)$. Let now $\Gamma_0=\wt{\Gamma}^{\d_1}_\alpha$. Then $\Gamma_0$ is a smooth curve and $\Gamma_0\ne\Gamma$. Applying (i), with a sequence $\d_i\downarrow \d_1$ and a sequence $\wt\Sigma_\a^{\d_i}$ of absolutely area minimizing surfaces with genus equal to $g$ (which we can do since we can pick the $\d_i$’s, so that for each $i$, $\wt\Gamma^{\d_i}_\a\in{\mathcal{B}}_g$), we conclude that there exists an absolutely area minimizing surface $\Sigma_1$ with $\partial\Sigma_1=\Gamma_0$ and $\operatorname{genus}(\Sigma_1)=g$. Applying now (i), with a sequence $\d_i\uparrow \d_1$ and a sequence $\wt\Sigma_\a^{\d_i}$ of absolutely area minimizing surfaces with genus not equal to $g$ (which we can do since for each $i$, $\wt\Gamma^{\d_i}_\a\notin{\mathcal{B}}_g$), we conclude there exists an absolutely area minimizing surface $\Sigma_2$ with $\partial\Sigma_2=\Gamma_0$ and $\operatorname{genus}(\Sigma_2)\ne g$. Hence, this curve $\Gamma_0$, satisfies the hypothesis of the Theorem.
Results on Minimal Surfaces in Mean Convex Domains {#minsurfaces section}
==================================================
In this section, we will give interesting results on uniqueness, embeddedness and the genus of minimal surfaces which are properly embedded in a mean convex $3$-manifold. In particular, we generalize the examples of Peter Hall [@Ha] in the unit $3$-ball ${\mathbf{B}}^3$, addressing the questions by Bill Meeks, to any strictly mean convex $3$-manifold.
First, by using the results of the previous section, we show that the space ${\mathfrak{C}}$ of simple closed curves in $\partial M$ bounding more than one minimal surface in $M$ is generic in the set of simple nullhomologous closed curves in $\partial M$, where $M$ is a strictly mean convex $3$-manifold.
Before giving the rigorous proof of this theorem, we give an outline of the main argument used. If $\partial M$ is homeomorphic to $S^2$, then the theorem follows directly from the fact that ${\mathcal{A}}_1$ is open dense in ${\mathcal{A}}$, and ${\mathcal{A}}_1\subset {\mathfrak{C}}$. This is because any simple closed curve $\Gamma$ would be nullhomotopic in $M$ as $\partial M \simeq S^2$, and by the result of Meeks and Yau (see Theorem \[meeksyau\]), $\Gamma$ bounds an area minimizing disk $D$ in $M$. Hence, any simple closed curve $\Gamma$ in ${\mathcal{A}}_1$ automatically bounds two different minimal surfaces where one of them is the absolutely area minimizing surface $\Sigma$ with genus $\geq 1$, and the other one is the area minimizing disk $D$. The same argument works in general with a slight modification. In particular, we will show that for an open dense subset ${\mathfrak{C}}$, any simple closed curve $\Gamma$ in this subset bounds an absolutely area minimizing surface $\Sigma$ in $M$ where $\Sigma$ is not the smallest genus surface which $\Gamma$ bounds in $M$. By a Theorem of White, which we have stated in Theorem \[white2\], there is an area minimizing surface $S$ of smallest genus. This implies $\Gamma$ bounds at least two different minimal surfaces, where one of them is the absolutely area minimizing surface $\Sigma$, and the other one is the area minimizing representative $S$ of the smallest genus surfaces which $\Gamma$ bounds in $M$.
\[nonuniquecurves\] [**(Curves bounding more than one minimal surface are generic)**]{}
Let $M$ be a strictly mean convex $3$-manifold. Let ${\mathcal{A}}$ be the space of nullhomologous simple closed curves in $\partial M$ equipped with the $\mathcal C^0$ topology. Let $\mathfrak{C}\subset {\mathcal{A}}$ represent the simple closed curves in $\partial M$ bounding more than one embedded stable minimal surface in $M$. Then, $\mathfrak{C}$ is generic in ${\mathcal{A}}$ in the sense that $\mathfrak{C}$ contains an open dense subset of ${\mathcal{A}}$.
\[Crmk\] In particular we show the following: Let AAM$(\Gamma)$, AM$(\Gamma)$ be the set of Absolutely Area Minimizing and that of Area Minimizing surfaces respectively bounded by $\Gamma$. Then the set $$\{\Gamma\in{\mathcal{A}}:\exists \Sigma_1\in \text{AAM}(\Gamma), \Sigma_2\in \text{AM}(\Gamma):\Sigma_1\ne\Sigma_2\}$$ is open and dense in ${\mathcal{A}}$.
We first show that $\mathfrak{C}$ is dense in ${\mathcal{A}}$. Let $\Gamma_0$ be any nullhomologous simple closed curve in $\partial M$. It suffices to show that for any $\e>0$, there exists $\Gamma\in\mathfrak C$, such that $d(\Gamma, \Gamma_0)<\e$, where by $d(\Gamma,\Gamma_0)<\e$ we mean that $\Gamma$ is in a $\e$-neighborhood of $\Gamma_0$ in the $\mathcal C^0$ topology (cf. Definition \[c0rmk\]).
Assume that $g_0$ is the minimum genus of a surface bounded by $\Gamma_0$. By Theorem \[white2\], there exists an area minimizing surface $\Sigma_{g_0}$ of genus $g_0$. If $\Gamma_0\notin {\mathcal{B}}_{g_0}$, then $\Gamma_0$ bounds an absolutely area minimizing surface $\Sigma$ which has to be different from $\Sigma_0$. In this case $\Gamma_0$ bounds at least two different stable minimal surfaces, namely $\Sigma_{g_0}$ and $\Sigma$, so we can take $\Gamma=\Gamma_0$. Hence we can assume that $\Gamma_0\in{\mathcal{B}}_{g_0}$.
Given $g>g_0$, by Theorem \[Adense\], there exists $\Gamma\in {\mathcal{A}}_g$, so that $d (\Gamma_0, \Gamma)<\e$ and for $\e$ small enough $\Gamma$ is homotopic to $\Gamma_0$. We denote the annulus in $\partial M$ between $\Gamma$ and $\Gamma_0$ by $A_\e$, as we will need it later.
We claim that $\Gamma$ bounds two different stable minimal surfaces. Note first that since $\Gamma\in {\mathcal{A}}_g$ and $g> g_0$, it bounds a stable minimal surface $\Sigma_1$ of genus bigger than $g_0$ (for example take an absolutely area minimizing surface that has minimum genus). To show the existence of the second surface, assume first that $\Gamma$ is nullhomotopic. Then, by the theorem of Meeks and Yau, Theorem \[meeksyau\], $\Gamma$ bounds a stable minimal disk $\Sigma_2$. Assume now that $\Gamma$ is not nullhomotopic, i.e. it bounds no disk. Let $\wt g_\Gamma$ be the minimum genus of a surface bounded by $\Gamma$. i.e. $$\wt g_\Gamma =\min\{g \ | \ \Gamma \text{ bounds a surface of genus g}\}.$$ We have that $0<\wt g_\Gamma \le g_0$; the second inequality being true because $\partial(\Sigma\cup A_\e)=\Gamma$, where $\Sigma$ is an area minimizing surface of genus $g_0$ and with $\partial\Sigma=\Gamma_0$. This implies that the minimum genus of a surface bounded by $\Gamma$ is less than or equal to $g_0$. By Theorem \[white2\], there exists an area minimizing surface $\Sigma_2$ of genus $\wt g_\Gamma$, which is therefore a stable minimal surface .
We have actually showed that the set $$\mathfrak{C}'=\displaystyle\bigcup_{g>0}\{\Gamma\in{\mathcal{A}}_g \ | \ \Gamma \text{ bounds an area minimizing surface of genus }\le g-1\}\subset\mathfrak C$$ is dense in ${\mathcal{A}}$. Next we show that $\mathfrak{C}'$ is also open.
Let $\Gamma_0\in\mathfrak C'$. Then for some $g>0$, $\Gamma_0\in{\mathcal{A}}_g$ and it bounds a minimal surface of genus $\le g-1$. Since ${\mathcal{A}}_g$ is open, there is an $\e>0$ such that for any $\Gamma\in{\mathcal{A}}$ with $d(\Gamma_0, \Gamma)<\e$, $\Gamma$ is homotopic to $\Gamma_0$ and belongs in ${\mathcal{A}}_g$.
Now, we can argue as before to show that any $\Gamma$ with $d(\Gamma, \Gamma_0)<\e$ bounds two minimal surfaces. In particular we have the following: since $\Gamma\in {\mathcal{A}}_g$, it bounds an area minimizing surface $\Sigma_1$ of genus $\ge g$. If $\Gamma$ is nullhomotopic in $M$, then, as before, by Theorem \[meeksyau\], it bounds an area minimizing disk $\Sigma_2$. If $\Gamma$ is not nullhomotopic, then we let $$\wt g_\Gamma=\min\{g \ | \ \Gamma \text{ bounds a surface of genus g}\}.$$ We have that $0<\wt g_\Gamma\le g-1$ as before, and by Theorem \[white2\], there exists an area minimizing surface $\Sigma_2$ of genus $\wt g_\Gamma$.
This shows $\mathfrak{C}'$ is open and dense in ${\mathcal{A}}$. Since $\mathfrak{C}'\subset \mathfrak{C}$, the proof follows.
Note that by [@Co1] and [@CE], a generic simple closed curve in $\partial M$ bounds a unique area minimizing disk, and similarly, a generic simple closed curve in $\partial M$ bounds a unique absolutely area minimizing surface in $M$. The result above shows that when we relax the condition being area minimizing to just minimal, the situation is opposite. In other words, a generic simple closed curve bounds a unique absolutely area minimizing surface, while a generic simple closed curve bounds more than one minimal surface. Hence the situations for minimal and absolutely area minimizing surfaces are quite different.
Now, we generalize Peter Hall’s examples to a more general setting. In [@Ha], Peter Hall gave examples of nonembedded stable minimal surfaces in the unit ball ${\mathbf{B}}^3$ of ${\mathbf{R}}^3$. To generalize this result to any strictly mean convex $3$-manifold $M$, we first show that there are disjoint simple closed curves $\Gamma_1$ and $\Gamma_2$ in $\partial M$ and stable minimal surfaces $S_1$ and $S_2$ in $M$ with $\partial S_i = \Gamma_i$, $i=1,2$, that are not disjoint, i.e. $S_1\cap S_2 \neq \emptyset$.
\[intersectingminsurfaces\][**(Disjoint extreme curves may not bound disjoint minimal surfaces)**]{} Let $M$ be a strictly mean convex $3$-manifold. Then, there are stable minimal surfaces $S_1$ and $S_2$ in $M$ with $S_1\cap S_2 \neq
\emptyset$ and $\partial S_1 \cap \partial S_2 = \emptyset$. Moreover, these surfaces can be chosen to be disks.
We argue by contradiction using the technique in [@Co1 Theorem 3.2], along with Theorem \[nonuniquecurves\]
Assume that the theorem is not true. By Theorem \[nonuniquecurves\], there exists $\Gamma_0\in{\mathcal{A}}$ and $\e>0$, such that any $\Gamma\in {\mathcal{A}}$ in an $\e$-neighborhood of $\Gamma_0$ is homotopic to $\Gamma_0$ and bounds two distinct stable minimal surfaces. Taking $\e$ small enough, $A_{\Gamma_0}=\{x\in\partial M: \operatorname{dist}(x, \Gamma_0)<\e\}$ is an annulus and we foliate this annulus by curves $\Gamma_t\in{\mathcal{A}}$, $t\in [-\e, \e]$. Each $\Gamma_t$, by the choice of $\e$, bounds 2 stable minimal surfaces $\Sigma_t^+, \Sigma_t^-$. If $H_2(M,{\mathbf{Z}})=0$, then both $\Sigma_t^+$ and $\Sigma_t^-$ are separating in $M$ and so there is a region [*between*]{} them. We let $N_t$ be that region, i.e. $N_t$ is such that $\partial
N_t=\Sigma_t^+\cup \Sigma_t^-$ as in [@Co1]. Then for $t_1< t_2$, $N_{t_1}\cap N_{t_2}=\emptyset$. This is because either $N_{t_1}\subset
N_{t_2}$, which is impossible since $\Gamma_{t_1}\cap\Gamma_{t_2} =\emptyset$, or else $\partial N_{t_1}\cap
\partial N_{t_2}\neq \emptyset$ which cannot hold because of our assumption that the theorem is false. Let $N=\cup_t N_t$. Then, for any $t\in [-\e, \e]$, $|N_t|>0$ where $|N_t|$ now represents the volume of $N_t$, and $\sum |N_t|=| N| \le |M| <\infty$. This implies that $|N_t|\neq 0$ for at most countably many $t$. But when $|N_t|=0$ and since $\Sigma^\pm_t$ are smooth, this implies $\Sigma_t^+=\Sigma_t ^-$. This is a contradiction as we have assumed that for any $t$, $\Gamma_t$ bounds two distinct stable minimal surfaces $\Sigma_t^+$ and $\Sigma_t^-$. Hence for some $t_1<t_2$, we must have $\partial N_{t_1}\cap \partial N_{t_2}\neq \emptyset$, i.e. $(\Sigma_{t_1}^+\cup\Sigma_{t_1}^-)\cap (\Sigma_{t_2}^+\cup\Sigma_{t_2}^-)\neq
\emptyset$ even though $\Gamma_{t_1}\cap\Gamma_{t_2} =\emptyset$. The proof follows.
For the case $H_2(M,{\mathbf{Z}}) \neq 0$, a similar argument from [@CE] works with a slight modification. If $H_2(M,{\mathbf{Z}}) \neq 0$, then the surfaces $\Sigma_t^+$ and $\Sigma_t^-$ may not be separating in $M$, and we cannot talk about the region $N_t$ [*between*]{} them. However, in this case, we can restrict ourselves to a small neighborhood of $\Gamma_0$ in $M$, say $T_{\Gamma_0}$, which is a very thin solid torus with $T_{\Gamma_0} \cap
\partial M = A_{\Gamma_0}$. Now, the argument in the previous paragraph works, if we replace our ambient manifold $M$ with $T_{\Gamma_0}$ and the surfaces $\Sigma^\pm_t$ with $S^\pm_t = \Sigma^\pm_t \cap T_{\Gamma_0}$. Then for any $t$, $S^+_t$ and $S^-_t$ are separating in $T_{\Gamma_0}$, and if we assume that the theorem is false, then $(S^+_t\cup S^-_t)\cap (S^+_s\cup S^-_s) = \emptyset$ for any $s\neq t$, since $S^\pm_t \subset
\Sigma^\pm_t$. Hence, by defining $N_t$ as the region between $S^+_t$ and $S^-_t$ in $T_{\Gamma_0}$, the whole summation argument goes through, and we get a contradiction in this case, too. This proves the existence of intersecting stable minimal surfaces with disjoint boundaries.
Finally, we need to show that these surfaces can be chosen to be disks. For this we work as follows. We want to prove that there are stable minimal disks $D_1$ and $D_2$ in $M$ with $D_1\cap D_2 \neq \emptyset$ and $\partial D_1 \cap \partial D_2 = \emptyset$. Let $p$ be a point in the smooth part of $\partial M$. Let $\epsilon>0$ be sufficiently small so that $\overline{B}_\epsilon(p)$ is strictly mean convex (i.e. $\epsilon< \rho$ where $\rho$ is the convexity radius at $p$). Let $\gamma\in {\mathcal{B}}_0$ with $\gamma \subset \partial M\cap B_\epsilon(p)$ and let $E$ be the absolutely area minimizing surface, which is a disk, with $\partial E=\gamma$. By construction, $E$ separates $M$ into two parts and we let $\Omega$ be the “small” part containing $p$. Then, $\Omega$ is topologically a $3$-ball, and any properly embedded surface would be separating in $\Omega$. Let $\Gamma_0$ be in ${\mathcal{A}}_1$ and such that it has an annular neighborhood $A_{\Gamma_0}$, for which $A_{\Gamma_0} \subset \Omega \cap \partial M$. Since ${\mathcal{A}}_1$ is open by Theorem \[Agopen\] (and by taking the annular neighborhood small enough), we can foliate $A_{\Gamma_0}$ with a collection $\{\Gamma_t\}\in {\mathcal{A}}_1$ with $t\in [-\e, \e]$. Then for any $t\in [-\e, \e]$, the absolutely area minimizing surface $\Sigma_t$ with boundary $\Gamma_t$ lies in $\Omega$, because $E$ is absolutely area minimizing surface and by [@Co1], $E\cap\Sigma_t=\emptyset$. $\Sigma_t$ separates $\Omega$ into two parts, say $\Omega_t^+$ (the part containing $p$) and $\Omega_t^-$. By construction, both $\Omega_t^+$ and $\Omega_t^-$ are mean convex. Moreover, $\Gamma_t$ is nullhomotopic in both $\Omega_t^+$ and $\Omega_t^-$ as $\partial \Omega$ is a sphere. Hence, by the theorem of Meeks and Yau, Theorem \[meeksyau\], there are area minimizing disks $D^+_t$ and $D^-_t$ in $\Omega_t^+$ and $\Omega_t^-$ respectively. Since $\Sigma_t$ has genus greater than or equal to one, $D^\pm_t\neq \Sigma_t$, and hence $D^+_t\neq D^-_t$. This shows that for any $t\in [-\e, \e]$, $\Gamma_t$ bounds two stable minimal disks $D^+_t$ and $D^-_t$ in $M$.
Assume that $D_t^\pm \cap D^\pm_s = \emptyset$ for any $t\neq s$. Then, we define $N_t$ as the region between $D^+_t$ and $D^-_t$ as in the first paragraph. Notice that by assumption, $N_t\cap N_s =\emptyset$ and by construction, $|N_t|>0$ for any $t$. Then, the argument in the first paragraph gives a contradiction as before and thus the proof follows.
The result of Theorem \[intersectingminsurfaces\] is interesting even for the case when $M$ is a strictly mean convex $3$-manifold with $H_2(M,{\mathbf{Z}}) = 0$. For such $M$, any two properly embedded area minimizing surfaces $\Sigma_1$ and $\Sigma_2$ are disjoint, provided that their boundaries $\Gamma_1$ and $\Gamma_2$ in $\partial M$ are disjoint by the Meeks-Yau exchange roundoff trick [@Co1 Lemma 4.1]. However, this theorem shows that if one relax the condition being area minimizing to just being minimal, this is no longer true.
\[manyintms\] Note that the proof of Theorem \[intersectingminsurfaces\] also shows that for any simple closed curve $\Gamma$ in $\partial M$, it is possible to find [*nearby*]{} curves $\Gamma_1$ and $\Gamma_2$ such that $\Gamma_1\cap \Gamma_2 =\emptyset$ while $S_1\cap S_2 \neq \emptyset$ where $S_i$, $i=1,2$ is a stable minimal surface with boundary $\Gamma_i$. In particular, we can find such curves $\Gamma_1$ and $\Gamma_2$ in an $\e$-neighborhood of $\Gamma$, for any $\e$, such that the $\e$-neighborhood of $\Gamma$ in $\partial M$ is an annulus and moreover if $\Gamma$ is smooth, then we can find such curves $\Gamma_1, \Gamma_2$ so that they are also smooth. Furthermore, by the proof of Theorem \[nonuniquecurves\], it is easy to see that we can take $S_2$ to be absolutely area minimizing and $S_1$ to be area minimizing of genus $g$, where $g$ is less than or equal to the minimum genus of a surface bounded by $\Gamma$.
Now, we need a version of a bridge principle to construct a nonembedded stable minimal surface in a strictly mean convex $3$-manifold whose boundary is a simple closed curve.
\[bridgeprincipleforextremecurves\] [**(Bridge Principle for Extreme Curves)**]{}
Let $M$ be a strictly mean convex $3$-manifold, and $\Gamma_1$ and $\Gamma_2$ be two disjoint smooth simple closed curves bounding two stable minimal surfaces $S_1$ and $S_2$ respectively. Let $\alpha$ be a smooth path in $\partial M$ connecting $\Gamma_1$ and $\Gamma_2$, as in *Definition \[a2\]*. Then, for any $\e>0$, there is a stable minimal surface $S_1\sharp_\alpha S_2$ which is $\e$-close to $S_1\cup
S_2 \cup \alpha$.
This lemma is a direct application of the Meeks and Yau Bridge principle [@MY3 Theorem 7], noting only that in their proof one can pick the bridge to lie on $\partial M$, since $M$ is strictly convex.
Finally, by using the previous result, and by the bridge principle above, we generalize Peter Hall’s examples of nonembedded stable minimal surfaces in ${\mathbf{B}}^3$ to any strictly mean convex $3$-manifold $M$. In particular, we show that for any strictly mean convex $3$-manifold $M$, there exists a simple closed curve $\Gamma$ in $\partial M$ which bounds a nonembedded stable minimal surface $S$ in $M$.
\[nonembeg\] [**(Nonembedded Stable Minimal Surfaces)**]{}
Let $M$ be a strictly mean convex $3$-manifold. Then, there is a simple closed curve $\Gamma$ in $\partial M$ such that $\Gamma$ bounds a nonembedded stable minimal surface $S$ in $M$. Furthermore there is a simple closed curve $\Gamma$ in $\partial M$ such that $\Gamma$ bounds a nonembedded stable minimal disk.
Take $S_1, S_2$ as in Theorem \[intersectingminsurfaces\], so that $\partial S_1,\partial S_2$ are smooth (cf. Remark \[manyintms\]). Then $S_1,
S_2$ intersect transversally, (using a local graphical representation and the the maximum principle one sees that they cannot be tangent). Now, apply Lemma \[bridgeprincipleforextremecurves\] to construct a minimal surface that is $\e$-close to $S_1\cup S_2\cup\alpha$, where $\alpha$ is a curve connecting $\partial S_1$ and $\partial S_2$. Since $S_1, S_2$ intersect transversally, for $\e$ small enough the new surface has at least one self intersection. For the disk case, take the surfaces as disks as in Theorem \[intersectingminsurfaces\]. The proof follows.
\[manycurvesremark\] Note that using Remark \[manyintms\], we actually have many curves satisfying Theorem \[nonembeg\]. In particular, given a smooth curve $\Gamma_0\in{\mathcal{A}}$ and $\e$, such that the $\e$-neighborhood of $\Gamma_0$ in $\partial M$ is an annulus, there exists $\Gamma$ in the $\e$-neighborhood of $\Gamma_0$, satisfying Theorem \[nonembeg\].
Again, when we compare this result with the area minimizing case, we see that this cannot happen in the area minimizing case. By the regularity results of [@ASSi], any area minimizing surface $\Sigma$ in $M$ bounded by a simple closed curve $\Gamma$ in $\partial M$ must be smooth in the interior. Hence, if we relax the condition being area minimizing to being just minimal, we see that this is no longer true.
Applications to Curves in ${\mathbf{R}}^3$
==========================================
In this section we apply our results to simple closed curves in ${\mathbf{R}}^3$. So far, we discussed the simple closed curves in the boundary of a strictly mean convex $3$-manifold, and proved several results about the area minimizing surfaces in the manifold that these curves bound. Now, we change our focus to the simple closed curves in ${\mathbf{R}}^3$ and give extensions of these results to this case.
We use the notation from the previous sections. In particular, ${\mathcal{A}}$ denotes the space of simple closed curves in ${\mathbf{R}^3}$. ${\mathcal{A}}_g\subset {\mathcal{A}}$ denotes the space of simple closed curves in ${\mathbf{R}}^3$ such that the minimum genus of the absolutely area minimizing surfaces that they bound is greater than or equal to $g$. ${\mathcal{B}}_g$ denotes the space of simple closed curves in $\partial M$ such that the minimum genus of the absolutely area minimizing surfaces that they bound is exactly $g$. i.e. ${\mathcal{B}}_g={\mathcal{A}}_g-{\mathcal{A}}_{g+1}$. Note that ${\mathcal{A}}_0={\mathcal{A}}$ because of the existence of absolutely area minimizing surfaces with boundary any given simple closed curve in ${\mathbf{R}}^3$ [@Fe]. Furthermore we let ${\mathcal{A}}^k$ denote the ${\mathcal{C}}^k$ smooth simple closed curves in ${\mathbf{R}}^3$ and we let ${\mathcal{A}}^k_g$ be ${\mathcal{A}}^k\cap{\mathcal{A}}_g$.
First, we extend Lemma \[Agopen\] and Theorem \[Adense\] to the case of curves in ${\mathbf{R}}^3$.
\[AgopenR3\] For any $g\geq 0$, ${\mathcal{A}}_g$ is open and dense in ${\mathcal{A}}$ equipped with the ${\mathcal{C}}^0$ topology. Moreover, for any $k\geq 0$, ${\mathcal{A}}^k_g$ is open and dense in ${\mathcal{A}}^k$ equipped with ${\mathcal{C}}^0$ topology.
First, we show that ${\mathcal{A}}_g$ is open in ${\mathcal{A}}$. The proof of this fact is the same as the proof of Lemma \[Agopen\], with the only difference that here $\mathcal U _{\Gamma_0}(\e)$ is an $\e$-neighborhood of $\Gamma_0$ in ${\mathbf{R}}^3$ with respect to the $\mathcal C^0$ topology; whereas in the proof of Lemma \[Agopen\] we restricted this neighborhood in $\partial M$; the boundary of the mean convex manifold $M$. Restricting ourselves now to ${\mathcal{C}}^k$ curves shows that ${\mathcal{A}}^k_g$ is open in ${\mathcal{A}}^k$ with respect to the ${\mathcal{C}}^0$ topology.
Now, we show that ${\mathcal{A}}_g$ is dense in ${\mathcal{A}}$. Let $\Gamma$ be a simple closed curve in ${\mathbf{R}^3}$. Let $\Sigma$ be the absolutely area minimizing surface in ${\mathbf{R}^3}$ with $\partial \Sigma = \Gamma$. Then, by [@MY3 Page 159, Corollary 1] , for a sequence $\Sigma_i$ of proper subsurfaces ($\Sigma_i \varsubsetneq \Sigma$) with $\Sigma_i\to \Sigma$, there are strictly mean convex neighborhoods $N_i$ of $\Sigma_i$ in ${\mathbf{R}^3}$ such that $\Gamma_i :=\partial \Sigma_i \subset \partial N_i$.
For a set $X$, let ${\mathcal{A}}(X)$ denote the space of simple closed curves in $\partial X$ and let ${\mathcal{A}}_g(X)$ denote the space of simple closed curves in $\partial X$ such that the minimum genus of the absolutely area minimizing surfaces that they bound in $X$ is greater than or equal to $g$. With this notation, ${\mathcal{A}}_g={\mathcal{A}}_g({\mathbf{R}}^3)$. Since $N_i$ is strictly mean convex, by Theorem \[Adense\], ${\mathcal{A}}_g(N_i)$ is dense in ${\mathcal{A}}(N_i)$, the space of simple closed curves in $\partial N_i$. Hence, for each $i$, there is a sequence $\{\alpha^j_i\}_{j\in\N}$ of simple closed curves in ${\mathcal{A}}_g(N _i)$ such that $\alpha^j_i \stackrel{j\to\infty}\longrightarrow \Gamma_i$, with respect to the $\mathcal C^0$ topology, where recall that $\Gamma_i=\partial \Sigma_i\subset \partial
N_i$.
Let $S^j_i$ be an absolutely area minimizing surface in $N_i$ with boundary $\partial S^j_i=\alpha^j_i$ and with $\operatorname{genus}(S^j_i)\geq g$. Notice that $S^j_i$ is absolutely area minimizing in $N_i$, and this does not necessarily imply that it is also absolutely area minimizing in ${\mathbf{R}}^3$. We claim however that for any $i>0$, there is a $j_0>0$, such that for any $j\ge j_0$, $S^j_i$ is also absolutely area minimizing in ${\mathbf{R}}^3$. Hence, when we prove the claim, the density of ${\mathcal{A}}_g$ in ${\mathcal{A}}$ follows.
First, note that $\Sigma_i$ is the unique absolutely area minimizing surface in ${\mathbf{R}}^3$ with $\partial \Sigma_i = \Gamma_i$. This because if there was another absolutely area minimizing surface $\Sigma'_i$ in ${\mathbf{R}}^3$ with $\partial \Sigma_i' = \Gamma_i$, then $\Sigma'=(\Sigma\setminus\Sigma_i)\cup
\Sigma_i'$ would be an absolutely area minimizing surface with $\partial \Sigma'=\Gamma$, since $|\Sigma_i|=|\Sigma_i'|$. However, $\Sigma'$ then would have a singularity along $\Gamma_i$, and this contradicts to the regularity theorem for absolutely area minimizing surfaces [@Fe].
Now, fix $i_0>0$. Assume that the above claim is not true and hence there is a sequence $\{j_k\}_{k\in\N}$ with $j_k\stackrel{k\to\infty}{\longrightarrow}\infty$, such that $S^{j_k}_{i_0}$ is not absolutely area minimizing surface in ${\mathbf{R}}^3$ for any $k\in\N$. Let $T^{j_k}_{i_0}$ be the absolutely area minimizing surface in ${\mathbf{R}}^3$ with $\partial T^{j_k}_{i_0}=\alpha^{j_k}_{i_0}$. Since $\alpha^{j_k}_{i_0}\stackrel{k\to\infty}{\longrightarrow}\Gamma_{i_0}$, and $\Sigma_{i_0}$ is the unique absolutely area minimizing surface in ${\mathbf{R}}^3$ with $\partial \Sigma_{i_0}=\Gamma_{i_0}$, then after passing to a subsequence if necessary, $T^{j_k}_{i_0}\stackrel{k\to\infty}{\longrightarrow}\Sigma_{i_0}$, where the convergence is with respect to measure and to Hausdorff distance. Since $N_{i_0}$ is strictly mean convex, there is a neighborhood $N^\delta$ of $N_{i_0}$ in ${\mathbf{R}}^3$ such that the nearest point projection map $\pi:N^\delta \to N_{i_0}$ does not increase the area, i.e. $|\pi(S)|\leq |S|$ for a surface $S$ in $N^\delta$. Since, $T^{j_k}_{i_0}\stackrel{k\to\infty}{\longrightarrow}\Sigma_{i_0}$ in Hausdorff distance, there exists a $k_0>0$ such that for any $k\ge k_0$, $T^{j_k}_{i_0}\subset N^\delta$. Hence, $|\pi(T^{j_k}_{i_0})|\leq |T^{j_k}_{i_0}|$, and since $\pi(T^{j_k}_{i_0})\subset N_{i_0}$ with $\partial \pi(T^{j_k}_{i_0})=\alpha^{j_k}_{i_0}$, $|S^{j_k}_{i_0}|\leq
|\pi(T^{j_k}_{i_0})|\leq |T^{j_k}_{i_0}|$. This shows that $S^{j_k}_{i_0}$ is also absolutely area minimizing in ${\mathbf{R}}^3$ for $k\ge k_0$, and the claim follows with $j_0=j_{k_0}$.
This shows that for any $i>0$, there is a $j_0>0$ such that $\alpha^j_i\in {\mathcal{A}}_g({\mathbf{R}}^3)$ for all $j\ge j_0$. Hence, ${\mathcal{A}}_g$ is not only open, but also dense in ${\mathcal{A}}$ in ${\mathcal{C}}^0$ topology. Notice that if the initial curve $\Gamma$ is smooth (${\mathcal{C}}^k$ smooth), then we can also take the curves $\alpha^j_i$ to be as smooth as $\Gamma$ by Lemma \[thinhandle\]. This implies that ${\mathcal{A}}^k_g$ is dense in ${\mathcal{A}}^k$ in the ${\mathcal{C}}^0$ topology.
\[rmkAgopenR3\] Note that from the proof of Theorem \[AgopenR3\] we can conclude that ${\mathcal{A}}^k_g$ is open in ${\mathcal{A}}^k$ with respect to the ${\mathcal{C}}^k$ topology. However, the statement that ${\mathcal{A}}^k_g$ is dense in ${\mathcal{A}}^k$ with respect to the ${\mathcal{C}}^k$ topology is far from being true. To see that, let $\gamma$ be a ${\mathcal{C}}^2$ simple closed curve in ${\mathbf{R}}^3$, that is furthermore contained in a plain. Then, by [@Me], there is a ${\mathcal{C}}^2$ neighborhood $N(\gamma)$ of $\gamma$ such that for any $\gamma'\in N(\gamma)$, there exists a unique minimal surface $S$ in ${\mathbf{R}}^3$ with $\partial S
=\gamma'$, and $S$ is a graph over the plane (hence a disk). This shows that $\gamma\in \operatorname{int}({\mathcal{B}}_0)$, and thus ${\mathcal{A}}_1$ (or any ${\mathcal{A}}_g$ for $g\ge 1$) is not dense in ${\mathcal{A}}$ in the ${\mathcal{C}}^2$ topology. The main reason why the techniques of this paper do not generalize to the smooth topology is that the horn surgery produces curves that are “close” to the original curve in the ${\mathcal{C}}^0$ topology but not in any ${\mathcal{C}}^k$ topology for $k\ge 2$.
Next, we apply the generic non-uniqueness result of Section \[minsurfaces section\] to curves in ${\mathbf{R}}^3$. Let $\mathfrak C$ be the space of simple closed curves in ${\mathbf{R}}^3$ bounding more than one embedded stable minimal surface. We show that $\mathfrak{C}$ is dense in ${\mathcal{A}}$ with respect to the ${\mathcal{C}}^0$ topology.
\[nonuniqueR3\][**(Curves bounding more than one minimal surface are dense)**]{} Let ${\mathcal{A}}$ be the space of simple closed curves in ${\mathbf{R}}^3$ equipped with the ${\mathcal{C}}^0$ topology. Let $\mathfrak{C}\subset {\mathcal{A}}$ represent the simple closed curves in $\partial M$ bounding more than one embedded stable minimal surface in ${\mathbf{R}}^3$. Then, $\mathfrak{C}$ is dense in ${\mathcal{A}}$ with respect to the ${\mathcal{C}}^0$ topology. Moreover, for any $k\geq 0$, ${\mathfrak{C}}\cap{\mathcal{A}}^k$ is dense in ${\mathcal{A}}^k$ with respect to the ${\mathcal{C}}^0$ topology.
Let $\wh{{\mathfrak{C}}}=\{\Gamma\in{\mathcal{A}}:\exists \Sigma_1\in \text{AAM}(\Gamma), \Sigma_2\in \text{AM}(\Gamma):\Sigma_1\ne\Sigma_2\}$, where similarly to Theorem \[nonuniquecurves\], we let AAM$(\Gamma)$ and AM$(\Gamma)$ be the set of Absolutely Area Minimizing and Area minimizing surfaces respectively bounded by $\Gamma$. Clearly $\wh{{\mathfrak{C}}}\subset{\mathfrak{C}}$. Hence, if we show that $\wh{{\mathfrak{C}}}$ is dense in ${\mathcal{A}}$, we are done.
Let $\Gamma , \Sigma , \Sigma_i$ and $N_i$ be as in the proof of Theorem \[AgopenR3\], i.e. $\Gamma\in{\mathcal{A}}$, $\Sigma\in $AAM$(\Gamma)$, $\Sigma_i\subsetneq \Sigma$, with $\Sigma_i\to\Sigma$ and $N_i$ are strictly mean convex neighborhoods of $\Sigma_i$ such that $\Gamma_i:=\partial\Sigma_i\subset\partial N_i$. Analogously with ${\mathcal{A}}(X)$, as defined in the proof of Theorem \[AgopenR3\], for a set $X$, we let ${\mathfrak{C}}(X)$ represent the curves in ${\mathcal{A}}(X)$ that bound more than one stable minimal surface in $X$. Since $N_i$ is strictly mean convex, Theorem \[nonuniquecurves\] implies that ${\mathfrak{C}}(N_i)$ is open and dense in ${\mathcal{A}}(N_i)$. The embedded minimal surfaces in $N(\Sigma_i)$ are also minimal in ${\mathbf{R}^3}$. The smooth case is similar. The proof follows.
Notice that Theorem \[nonuniquecurves\] says not only density, but also genericity (containing open dense subset) of ${\mathfrak{C}}$ in ${\mathcal{A}}$, when the ambient manifold $M$ is strictly mean convex and the curves are in $\partial M$. Here, genericity of ${\mathfrak{C}}$ for the curves in ${\mathbf{R}}^3$ cannot be proved with the techniques of Theorem \[nonuniquecurves\] as White’s result in [@Wh2] (cf. Theorem \[white2\]) is for strictly mean convex manifolds; this theorem gives us an embedded minimal surface of smallest genus when the manifold is strictly mean convex.
On the other hand, instead of Theorem \[white2\], one might want to consider [@Jo Corollary 2.1] to get a minimal representative of the smallest genus surface, and generalize these arguments as it applies to ${\mathbf{R}}^n$. However, Jost’s result gives us an (not necessarily embedded) least area representative of the smallest genus surface the curve bounds (which is always a disk in our case). Hence, we cannot use it here, as we are looking for embedded minimal surfaces.
Finally, we will generalize Theorem \[nonembeg\] to simple closed curves in ${\mathbf{R}}^3$.
For any simple closed curve $\Gamma$ in ${\mathbf{R}^3}$ and for any $\e>0$, there is a smooth simple closed curve $\Gamma'$ in $N_\e(\Gamma)$; an $\e$-neighborhood of $\Gamma$, such that $\Gamma'$ bounds a nonembedded stable minimal surface in ${\mathbf{R}^3}$.
Let $\Gamma ,\Gamma_i, \Sigma , \Sigma_i$ and $N_i$ be as in the proof of Theorems \[AgopenR3\] and \[nonuniqueR3\], i.e. $\Gamma\in{\mathcal{A}}$, $\Sigma\in $AAM$(\Gamma)$, $\Sigma_i\subsetneq \Sigma$, with $\Sigma_i\to\Sigma$ and $N_i$ are strictly mean convex neighborhoods of $\Sigma_i$ such that $\Gamma_i:=\partial\Sigma_i\subset\partial N_i$. Since $N_i$ is strictly mean convex, Theorem \[nonembeg\] (see also Remark \[manycurvesremark\]) implies that for any $\e>0$, there is a simple closed curve $\Gamma'_i\subset N_\e(\Gamma_i)\cap\partial
N_i$, where $N_\e(\Gamma_i)$ is an $\e$-neighborhood of $\Gamma_i$ in the ${\mathcal{C}}^0$ topology, such that $\Gamma'_i$ bounds a nonembedded stable minimal surface in $N_i$. Since any minimal surface in $N_i$ is also minimal in ${\mathbf{R}^3}$ and since, by construction, the curves can be taken smooth, the proof follows.
The result above shows that the simple close curves in ${\mathbf{R}}^3$ bounding nonembedded minimal surfaces, say $\mathfrak{D}$, are very abundant. Note that our result does not imply the density of $\mathfrak{D}$ in ${\mathcal{A}}$. Even though the curve $\Gamma'$ is in $\e$ neighborhood of $\Gamma$, it is not ${\mathcal{C}}^0$-close to $\Gamma$ (see proof of Theorem \[nonembeg\]).
On the other hand, even though we cannot say whether $\mathfrak{D}$ is dense in ${\mathcal{A}}$ in the ${\mathcal{C}}^0$ topology or not, it is easy to see that $\mathfrak{D}\cap{\mathcal{A}}^2$ is not dense in ${\mathcal{A}}^2$ in the ${\mathcal{C}}^2$ topology. By [@EWW], any minimal surface $\Sigma$, bounded by a simple closed curve $\Gamma$ in ${\mathbf{R}}^3$ with total curvature $k(\Gamma)\leq 4\pi$ must be embedded. Hence, a simple closed curve $\Gamma$, satisfying $k(\Gamma)<4\pi$ has a ${\mathcal{C}}^2$-neighborhood $N(\Gamma)$ such that for any $\Gamma'\in N(\Gamma)$ the following holds: any minimal surface $\Sigma'$ in ${\mathbf{R}}^3$ bounded by $\Gamma'$ is embedded. Similarly, the result by Meeks [@Me] mentioned in the previous Remark \[rmkAgopenR3\] gives such a neighborhood of a curve $\Gamma$. This shows that $\mathfrak{D}$ is not dense in ${\mathcal{A}}^2$ in the ${\mathcal{C}}^2$ topology.
Final Remarks
=============
In the last section, for curves in $\R^3$, we considered the questions of openness and density not only for simple closed curves but also for smooth curves. The same thing can be done for curves on the boundary of a mean convex manifold $M$ and similar questions can be considered in the smooth category. Recall that ${\mathcal{A}}_g$ denotes the space of simple closed curves in $\partial M$ such that the minimum genus of the absolutely area minimizing surfaces that they bound is greater than or equal to $g$. ${\mathcal{B}}_g$ denotes the space of simple closed curves in $\partial M$ such that the minimum genus of the absolutely area minimizing surfaces that they bound is exactly $g$. Also, ${\mathcal{A}}$ denotes the space of nullhomologous simple closed curves in $\partial M$ equipped with the $\mathcal C^0$ topology. A natural extension is to consider the results of this paper in the smooth category. In other words, let ${\mathcal{A}}^k$ be equal to ${\mathcal{A}}\cap C^k$, the space of nullhomologous curves that are ${\mathcal{C}}^k$ smooth, considered again with the $\mathcal{C}^0$ topology.
Let also ${\mathcal{A}}^k_g ={\mathcal{A}}_g\cap C^k$ and ${\mathcal{B}}^k_g ={\mathcal{B}}_g\cap C^k$ be the ${\mathcal{C}}^k$ smooth curves in ${\mathcal{A}}_g$ and in ${\mathcal{B}}_g$ respectively. Then, it is easy to show that ${\mathcal{A}}^k_g$ is open in ${\mathcal{A}}^k$ by using the techniques of this paper (Lemma \[Agopen\]). Furthermore, if $\partial M$ is smooth then the fact that ${\mathcal{A}}_g$ is open and dense in ${\mathcal{A}}$ (Lemma \[Adense\]), and ${\mathcal{A}}^k$ is dense in ${\mathcal{A}}$ implies that ${\mathcal{A}}^k_g$ is dense in ${\mathcal{A}}$. If $\partial M$ is not smooth we only get that ${\mathcal{A}}^k_g$ is dense in ${\mathcal{A}}^k$.
Another question would be whether the similar results are true in the mean convex setting instead of strictly mean convex manifolds. Unfortunately, most of the results in Section 3, like ${\mathcal{A}}_g$ is open dense in ${\mathcal{A}}$, are not true for mean convex $3$-manifolds. The main obstacle is the following. Let $M$ be a mean convex $3$-manifold and $S$ be a subsurface in $\partial M$ which is a minimal surface. Take a sufficiently small disk $D$ in $S$ such that $D$ is the absolutely area minimizing surface for its boundary. Then, for any simple closed curve in the interior of $D$, the absolutely area minimizing surface would be the disk it bounds in $D$. Hence, any simple closed curve $\Gamma$ in $\operatorname{int}(D) \subset \partial M$ would automatically be in ${\mathcal{B}}_0$ and therefore there is no way to approximate this $\Gamma$ with the curves in ${\mathcal{A}}_g$, for $g\geq 1$, since there is an open neighborhood of $\Gamma$ in $A$ which lies entirely in ${\mathcal{B}}_0$. This example shows that Theorem \[Adense\] (${\mathcal{A}}_g$ is dense in ${\mathcal{A}}$) is not true for mean convex $3$-manifolds whose boundary contains a minimal subsurface. For an extreme example, if $M$ is a mean convex manifold whose boundary is an absolutely area minimizing sphere, then clearly ${\mathcal{A}}={\mathcal{B}}_0$. The reason for this is that the horn surgery lemma fails when the absolutely area minimizing surface it bounds is in the boundary of the manifold. What in particular fails, is that we cannot apply White’s bridge principle, Theorem \[white4\]. In [@Wh4], there is no restriction on $M$ being strictly mean convex, however it is required that the arc $\a$ is not tangential at its endpoints to the tangent halfspace of the area minimizing surface. Of course this is not the case when both $\a$ and the absolutely area minimizing surface lie in $\partial M$. Hence, the relevant results depending on Theorem \[white4\], would not be valid for mean convex $3$-manifolds.
In Section 4, we showed that the case of absolutely area minimizing surfaces and that of minimal surfaces are quite different when bounding extreme curves. For a strictly mean convex $3$-manifold $M$, while generically a simple closed curve in $\partial M$ bounds a unique absolutely area minimizing surface in $M$ [@Co1], [@CE], the situation is opposite for minimal surfaces, as generically a simple closed curve in $\partial M$ bounds more than one minimal surface in $M$ (Theorem 4.1). Also, for $H_2(M,Z)=0$, while absolutely area minimizing surfaces with disjoint boundaries in $\partial M$ are also disjoint, this is no longer true for minimal surfaces. In Theorem 4.2, we proved that for any strictly mean convex $3$-manifold $M$, there are disjoint simple closed curves in $\partial M$, bounding intersecting stable minimal surfaces.
It might be interesting to study the space of curves bounding nonembedded minimal surfaces as described in Section 4. In other words, we showed the existence of such extreme curves for any strictly mean convex $3$-manifold $M$ in Theorem \[nonembeg\]. However, a condition on the curve to guarantee the embeddedness of all minimal surfaces it bounds would be very interesting. In our construction for the curves bounding nonembedded minimal surfaces, we used two intersecting minimal surfaces in $M$ which have disjoint boundaries in $\partial M$. Then, we used a bridge between them to get a nonembedded minimal surface. This implies that for the final boundary curve $\Gamma$ and near the bridge, the ratio between the extrinsic distance (in $\partial M$) and the intrinsic distance (in $\partial M$) gets very large. Maybe, a bound (of course, depending on $M$) on this ratio might be a good condition to guarantee the embeddedness of all minimal surfaces which $\Gamma$ bounds in $M$.
Furthermore, for any $g>0$, Almgren and Thurston showed existence of unknotted simple closed curves $\Gamma_g$ in ${\mathbf{R}}^3$ with the property that the absolutely area minimizing surface $\Gamma_g$ bounds has genus at least $g$ [@AT]. Indeed, they showed that for any $g>0$, there are curves $\Gamma_g$ such that any embedded surface they bound in their convex hull of the curve must have genus at least $g$. Since any minimal surface with boundary $\Gamma$ must be in the convex hull of $\Gamma$, and any absolutely area minimizing surface is embedded, this automatically implies that the absolutely area minimizing surface $\Gamma_g$ bound must have genus at least $g$. In this paper’s terminology, this means Almgren and Thurston showed that for any $g>0$, ${\mathcal{A}}_g \neq \emptyset$. However, their examples are not extreme curves. Here, one of the corollaries of Theorem \[Adense\] is that for any convex body $\Omega$ in ${\mathbf{R}}^3$ and for any $g>0$, there is a curve $\Gamma_g \in \partial \Omega$ such that the absolutely area minimizing surface, which $\Gamma_g$ bounds, has genus at least $g$.
[MSY]{} W. K. Allard [*On the first variation of a varifold*]{}, Ann. of Math. (2) [**95**]{} (1972), 417–491.
W. K. Allard [*On the first variation of a varifold: boundary behavior*]{}, Ann. of Math. (2) [**101**]{} (1975), 418–446.
F.J. Almgren, L. Simon, [*Existence of embedded solutions of Plateau’s problem*]{}, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) [**6**]{} (1979) no. 3, 447–495.
F.J. Almgren, W.P. Thurston, [*Examples of unknotted curves which bound only surfaces of high genus within their convex hulls*]{}, Ann. of Math. (2) [**105**]{} (1977), no. 3, 527–-538.
F.J. Almgren, R. Schoen, L. Simon [*Regularity and singularity estimates on hypersurfaces minimizing parametric elliptic variational integrals*]{}, Acta Math. [**139**]{} (1977) 217–265.
T. Bourni, [*$C^{1,\alpha}$ Theory for the Prescribed Mean Curvature Equation with Dirichlet Data*]{}, J. Geom. Anal. [**21**]{} (2011) 982–1035.
B. Coskunuzer, [*Generic uniqueness of area minimizing disks for extreme curves*]{}, Amer. J. Math. [**132**]{} (2010) 1091–1104.
B. Coskunuzer, T. Etgu, [*Generic uniqueness of area minimizing disks for extreme curves II*]{}, in preparation.
E. De Giorgi, [*Frontiere orientate di misura minima*]{}, Sem. Mat. Scuola Norm. Sup. Pisa, 1960-61.
J. Douglas, [*Solution of the problem of Plateau*]{}, Trans. Amer. Math. Soc. [**33**]{} (1931) 263–321.
T. Ekholm, B. White, D. Wienholtz, [*Embeddedness of minimal surfaces with total boundary curvature at most $4\pi$*]{}, Ann. of Math. (2) [**155**]{} (2002), no. 1, 209–234.
H. Federer, [*Geometric measure theory*]{}, Springer-Verlag, New York 1969.
H. Federer, W. Fleming, [*Normal and integral currents*]{}, Ann. of Math. (2) [**72**]{} (1960) 458-520.
D. Fischer-Colbrie, R. Schoen, [*The structure of complete stable minimal surfaces in 3-manifolds of nonnegative scalar curvature*]{}, Comm. Pure Appl. Math. 33 (1980) 199Ð211.
E. Giusti, [*Minimal surfaces and functions of bounded variation*]{}, Birkhauser, Basel, 1984.
P. Hall, [*Two topological examples in minimal surface theory*]{}, J. Differential Geom. [**19**]{} (1984) 475–481.
R. Hardt [*On boundary regularity for integral currents or flat chains modulo two minimizing the integral of an elliptic integrand*]{}, Comm. Partial Differential Equations [**2**]{} (1977), no. 12, 1163–1232.
R. Hardt, L. Simon, [*Boundary regularity and embedded solutions for the oriented Plateau problem*]{}, Ann. of Math. (2) [**110**]{} (1979), no. 3, 439–486.
J. Hass, P. Scott, [*The existence of least area surfaces in $3$-manifolds*]{}, Trans. Amer. Math. Soc. [**310**]{} (1988) no. 1, 87–114.
J. Jost, [*Conformal mappings and the Plateau-Douglas problem in Riemannian manifolds*]{}, J. Reine Angew. Math. 359 (1985) 37–54.
C.B. Morrey, [*The problem of Plateau on a Riemannian manifold*]{}, Ann. of Math. (2) [**49**]{} (1948) 807–851.
W. Meeks, [*Uniqueness theorems for minimal surfaces*]{}, Illinois J. Math. [**25**]{} (1981) 318–336.
W. Meeks and S.T. Yau, [*Topology of three-dimensional manifolds and the embedding problems in minimal surface theory*]{}, Ann. of Math. (2) [**112**]{} (1980) 441–484.
W. Meeks and S.T. Yau, [*The classical Plateau problem and the topology of three manifolds*]{}, Topology [**21**]{} (1982) 409–442.
W. Meeks and S.T. Yau, [*The existence of embedded minimal surfaces and the problem of uniqueness*]{}, Math. Z. [**179**]{} (1982) no. 2, 151–168.
T. Rado, [*On Plateau’s problem*]{}, Ann. of Math. (2) [**31**]{} (1930) no. 3, 457–469.
L. Simon [*Lectures on geometric measure theory*]{}, Proceedings of the Centre for Mathematical Analysis, Canberra, 1983.
B. White, [*The space of $m$-dimensional surfaces that are stationary for a parametric elliptic functional*]{} Indiana Univ. Math. J. [**36**]{} (1987), no. 3, 567–602.
B. White, [*Existence of smooth embedded surfaces of prescribed genus that minimize parametric even elliptic functionals on $3$-manifolds*]{}, J. Diff. Geom. [**33**]{} (1991) no. 2, 413–443.
B. White, [*On the topological type of minimal submanifolds*]{}, Topology [**31**]{} (1992) 445–448.
B. White, [*The bridge principle for stable minimal surfaces*]{}, Calc. Var. PDE [**2**]{} (1994) no. 4, 405–425.
-
[^1]: The second author is partially supported by EU-FP7 Grant IRG-226062, TUBITAK Grant 109T685 and TUBA-GEBIP Award.
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We use the LMO invariant to find constraints for a knot to admit a purely or reflectively cosmetic surgery. We also get a constraint for knots to admit a Lens space surgery, and some information for characterizing slopes.'
address: |
Department of Mathematics, Graduate School of Science, Osaka University\
1-1 Machikaneyama Toyonaka, Osaka 560-0043, JAPAN
author:
- Tetsuya Ito
title: 'On LMO invariant constraints for cosmetic surgery and other surgery problems for knots in $S^{3}$'
---
Introduction
============
For a knot $K$ in $S^{3}$ and $r =p/q \in {\mathbb{Q}}\cup \{\infty\}$, let $S^{3}(K,r)$ be the oriented closed 3-manifold obtained by the Dehn surgery on $K$ along the slope $r$. We denote the 3-manifold $M$ with opposite orientation by $-M$, and we write $M \cong M'$ if two 3-manifolds are homeomorphic by an orientation preserving homeomorphism.
Two Dehn surgeries along a knot $K$ with different slopes $r$ and $r'$ are *purely cosmetic* if $S^{3}(K,r) \cong S^{3}(K,r')$ and *reflecitively cosmetic* (or, *chirally cosmetic*) if $S^{3}(K,r) \cong -S^{3}(K,r')$.
A famous cosmetic surgery conjecture (for a knot in $S^{3}$) [@kir Problem 1.81] states that a non-trivial knot $K$ does not admit purely cosmetic surgeries. The case one of the slope $r=\infty$ was shown in [@gl], which particularly says that two knots in $S^{3}$ are equivalent if and only if they have the homeomorphic complements.
There are various constraints for two Dehn surgeries are purely cosmetic. Among them, using Heegaard Floer homology theory in [@nw Theorem 1.2] the following strong restrictions are shown. $$\label{eqn:NW}
\begin{split}
&
\mbox{If } S^{3}(K,p\slash q) \cong S^{3}(K,p'\slash q'), \mbox{ then } p' \slash q'=\pm p\slash q \mbox{ and } q^{2} \equiv -1 \pmod p.\\ &\mbox{In particular, } L(p,q) \cong L(p',q')
\end{split}$$
The cosmetic surgery conjecture can be regarded as a statement saying that when we fix a knot $K$ then the Dehn surgery gives an injective map $S^{3}(K,\ast):\{\mbox{Slopes}\}={\mathbb{Q}}\cup \{\infty\} \rightarrow \{\mbox{(oriented) 3-manifolds}\}$ except the case $K$ is the unknot. In this point of view, it is natural to ask the injectivity of the Dehn surgery map when we fix a slope; Is the Dehn surgery map $S^{3}(\ast,r):\{\mbox{Knots}\} \rightarrow \{\mbox{(oriented) 3-manifolds}\}$ injective ? A slope $r$ is a called *characterizing slope* of $K$ if the answer is affirmative, that is, $S^{3}(K,r) \cong S^{3}(K',r)$ implies $K = K'$.
Compared with purely cosmetic surgeries, a situation for characterizing slopes is more complicated. There are various examples of non-characterizing slopes. Among them, in [@bamo] hyperbolic knots with infinitely many (integral) non-characterizing slopes are given. On the other hand, if $K$ is the unknot [@kmos; @os2], trefoil, or the figure-eight knot [@os1] then all the slopes are characterizing. Moreover, if $K$ is a torus knot, a slope $r$ is characterizing provided $r$ is sufficiently large [@nz Theorem 1.3], whereas some small slopes are not characterizing.
In this paper we use the LMO invariant to study a structure of Dehn surgery along knots. We obtain various constraints for a knot to admit a purely or reflectively cosmetic surgery, or, a slope $r$ to be characterizing.
The LMO invariant is an invariant of closed oriented 3-manifolds which takes value in certain graded algebra $\mathcal{A}(\emptyset)$. The degree one part $\lambda_1$ of the LMO invariant is equal to the Casson-Walker invariant [@lmmo] that satisfies the following surgery formula $$\label{eqn:LMO1} \lambda_{1}(S^{3}(K,p \slash q))=\frac{1}{2}a_{2}(K) \frac{q}{p} + \lambda_{1}(L(p,q)).$$ Here $a_{2}(K)$ denotes the coefficients of $z^{2}$ of the Conway polynomial, and $L(p,q) = S^{3}(\mathsf{Unknot},p\slash q)$ denote the $(p,q)$-Lens space.
Using (\[eqn:NW\]) and the surgery formula (\[eqn:LMO1\]) we immediately get the following constraint for cosmetic surgery and characterizing slopes. (In [@boli], this is proved without using (\[eqn:NW\]) – instead they used Casson-Gordon invariant to get an additional constraint.)
[@boli Proposition 5.1] \[theorem:LMO1\] Let $K$ and $K'$ be knots in $S^{3}$ and $r,r' \in {\mathbb{Q}}\setminus \{0\}$ with $r\neq r'$.
1. If $S^{3}(K,r)\cong S^{3}(K',r)$ then $a_{2}(K)=a_{2}(K')$.
2. If $S^{3}(K,r)\cong S^{3}(K,r')$ then $a_{2}(K)=0 \ (=a_{2}({\sf Unknot}))$.
Our purpose is to get further constraints that generalize Theorem \[theorem:LMO1\] by looking at higher order part of the LMO invariants.
Two knots $K$ and $K'$ are called *$C_{n+1}$-equivalent* if $v(K)=v(K')$ for all finite type invariant $v$ whose degree is less than or equal to $n$. A knot which is $C_{n+1}$ equivalent to the unknot is called a *$C_{n+1}$-trivial* knot. In [@gu; @ha] it is shown that two knots are $C_{n+1}$-equivalent if and only if they are moved each other by certain local moves called $C_{n+1}$-moves.
In this terminology, Theorem \[theorem:LMO1\] can be understood that Dehn surgery characterizes a knot or a slope *up to $C_3$-equivalence*: (i) says that if Dehn surgeries of two knots $K$ and $K'$ along the same slope are homeomorphic then $K$ and $K'$ are $C_{3}$-equivalent, and (ii) says that the cosmetic surgery conjecture is true unless $K$ is $C_3$-trivial.
In [@bl] Bar-Natan and Lawrence gave a rational surgery formula of the LMO invariant. First we write down a rational surgery formula for degree two and three part of the (primitive) LMO invariants of $S^{3}(K,r)$.
\[theorem:LMO23\] Let $K$ be a knot in $S^{3}$. $$\begin{aligned}
\lambda_2(S^{3}(K,p/q))\!\!&\!\! = \!&\!\!\!
\left(\!v_2(K)^{2} + \frac{1}{24}v_2(K)+\frac{5}{2}v_{4}(K)\!\right)\frac{q^{2}}{p^{2}}-v_3(K)\frac{q}{p} + \frac{v_2(K)}{24}\left( \frac{1}{p^{2}}-1\right)\\
& & \hspace{0.4cm}+ \lambda_{2}(L(p,q))\\
& \!\!=\!&\!\!\! \left(\!\frac{7a_2(K)^2-a_2(K)-10a_{4}(K)}{8}\! \right)\frac{q^{2}}{p^{2}} -v_{3}(K)\frac{q}{p} + \frac{a_{2}(K)}{48}
\left(1-\frac{1}{p^{2}}\right)\\
& & \hspace{0.4cm}+ \lambda_{2}(L(p,q))\\\end{aligned}$$ $$\begin{aligned}
\lambda_{3}(S^{3}(K,p/q)) \!
&\!\! = \!& \!
-\left(\!\frac{35}{4}v_6(K)\!+\!\frac{5}{24}v_4(K)\!+\!10v_2(K)v_4(K)\!+\!\frac{4}{3}v_2(K)^{3}\!+\! \frac{1}{12}v_2(K)^{2}\!\right)\frac{q^{3}}{p^{3}}\\
& &\! - \left( \frac{5}{24}v_4(K) +\frac{1}{288}v_2(K)+\frac{1}{12}v_2(K)^{2}\right)\frac{q}{p^{3}} \\
& &\! + \left(\frac{5}{2}v_{5}(K)+ 2v_3(K)v_2(K)+\frac{1}{24}v_3(K)\right)\frac{q^2}{p^2} + \frac{v_3(K)}{24}\left(\frac{1}{p^2}-1\right)\\
& &\! - \left(w_4(K)-\frac{1}{12}v_2(K)^{2}-\frac{1}{288}v_2(K)-\frac{5}{24}v_4(K)\right)\frac{q}{p} +\lambda_{3}(L(p,q))\end{aligned}$$
Here $v_2(K),v_3(K),v_4(K),w_4(K),v_5(K)$ and $v_6(K)$ are certain canonical finite type invariant of the knot $K$ (see Section \[section:LMO\] for details – as we will see in Lemma \[lemma:AJ\], except $v_5$ they are determined by the Alexander and the Jones polynomial). Also, $a_{2n}(K)$ is the coefficient of $z^{2n}$ in $\nabla_{K}(z)$, the Conway polynomial of $K$.
The degree two part of the LMO invariant (combined with (\[eqn:NW\])) gives rise to the followings.
\[corollary:LMO2\] Let $K$ and $K'$ be knots in $S^{3}$, and $r,r' \in {\mathbb{Q}}\setminus \{0\}$ with $r \neq r'$.
1. If $S^{3}(K,r) \cong S^{3}(K,r')$ then $v_3(K)=0$.
2. If $S^{3}(K,r) \cong -S^{3}(K,-r)$ then $v_3(K)=0$.
3. If $S^{3}(K,r) \cong -S^{3}(K,r')$ for $r' \neq \pm r$ then either
- $v_{3}(K)=0$, or,
- $v_{3}(K) \neq 0$ and $\displaystyle \frac{rr'}{r+r'} = \frac{7a_2(K)^2-a_2(K)-10a_{4}(K)}{8v_{3}(K)}. $
4. If $S^{3}(K,r) \cong S^{3}(K',r)$ then either
- $a_{4}(K)=a_{4}(K')$, $v_{3}(K)=v_{3}(K')$, or,
- $a_{4}(K) \neq a_{4}(K')$, $v_{3}(K)\neq v_{3}(K')$, and $\displaystyle r = \frac{5(a_{4}(K)-a_{4}(K'))}{4(v_{3}(K)-v_{3}(K'))}. $
\(i) was proven in [@iw] by a similar argument using Lescop’s surgery formula of the Kontsevich-Kuperberg-Thurston invariant [@ko2; @kt] (see Remark \[remark:KKT\]).
We note that the degree two part gives the following constraint for a knot to admit a Lens space surgery.
\[corollary:Lens\] If $S^{3}(K,p\slash q)$ is a Lens space, then $$\left( \frac{7a_2(K)^2-a_2(K)-10a_{4}(K)}{8} \right)\frac{q^{2}}{p^{2}} -v_{3}(K)\frac{q}{p} + \frac{a_{2}(K)}{48}
\left(1-\frac{1}{p^{2}}\right) =0.$$
By cyclic surgery theorem [@cgls], if $K$ is not a torus knot, then $q=1$ hence we get $$\label{eqn:Lens-obstruction}
a_{2}(K)p^{2} -48v_{3}(K)p+\left(42a_2(K)^2-7a_2(K)-60a_{4}(K) \right) =0.$$
Combined with the fact that $a_{2}(K),4v_{3}(K)$ and $a_{4}(K)$ are integers, (\[eqn:Lens-obstruction\]) brings some interesting informations. For example, a non-torus knot $K$ admitting lens surgery, $576v_{3}(K)^{2}-a_{2}(K)(42a_2(K)^2-7a_2(K)-60a_{4}(K))$ is a square number. If a non-torus knot $K$ admits more than one Lens space surgeries, the surgery slopes are successive integers [@cgls Corollary 1] so such a knot has $a_{2}(K) \neq \pm 1$.
The formula of degree three part is more complicated. Fortunately, as for cosmetic surgery, using (\[eqn:NW\]) we get the following simple constraints.
\[corollary:LMO3\] Let $K$ and $K'$ be a knot in $S^{3}$ and $r= p\slash q \in {\mathbb{Q}}\setminus\{0\}$.
1. If $S^{3}(K,r) \cong S^{3}(K,r')$ for $r'\neq r$, then $$p^{2}(24w_{4}(K)-5v_{4}(K)) + 5v_{4}(K) + q^{2}(210v_{6}(K)+5v_{4}(K)) = 0.$$
2. If $S^{3}(K,r) \cong -S^{3}(K,-r)$, then $v_{5}(K)=0$.
Corollary \[corollary:LMO3\] (i) leads to the following.
\[cor:c11\] The cosmetic surgery conjecture is true for all knots with less than or equal to 11 crossings, possibly except $10_{118}$.
Using higher degree part of the LMO invariant, adding suitable $C_n$-equivalence assumptions we prove the following more direct generalizations of Theorem \[theorem:LMO1\].
\[theorem:LMOhigh1\] Let $K$ and $K'$ be a knot in $S^{3}$ and $r, r' \in {\mathbb{Q}}\setminus\{0\}$ with $r \neq r'$.
1. Assume that $K$ and $K'$ are $C_{2m+2}$-equivalent. If $S^{3}(K,r) \cong S^{3}(K',r)$ then $a_{2m+2}(K) = a_{2m+2}(K')$.
2. Assume that $K$ is $C_{4m+2}$-trivial. If $S^{3}(K,r) \cong S^{3}(K,r')$ then $a_{4m+2}(K) = 0$.
We say that $K$ and $K'$ are *odd $C_{n+1}$*-equivalent if $v(K)=v(K')$ for all *odd degree* canonical finite type invariant $v$ of degree $\leq n$. We call a knot is *odd $C_{n+1}$-trivial* if it is odd $C_{n+1}$-equivalent to the unknot.
Let $(Z^{\sigma}(K) \sqcup \Omega^{-1})_{e,k}$ denotes the bigrading $(e,k)$ of the Kontsevich invariant of $K$, normalized so that the unknot takes value $1$. Let $K_{e,k}$ denotes the kernel of the Aarhus integration (diagram pairing) $\langle \ast , { \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{\frac{k}{2}}\rangle: \mathcal{B}_{e,k} \rightarrow \mathcal{A}(\emptyset)_{e + \frac{k}{2}}$ (See Section \[section:LMO\]).
\[theorem:LMOhigh2\] Let $K$ and $K'$ be a knot in $S^{3}$ and $r \in {\mathbb{Q}}\setminus\{0\}$.
1. Assume that $K$ and $K'$ are $C_{2m+1}$-equivalent. If $S^{3}(K,r) \cong S^{3}(K',r)$ and $a_{2m+2}(K) = a_{2m+2}(K')$, then $(Z^{\sigma}(K) \sqcup \Omega^{-1})_{1,2m}- (Z^{\sigma}(K) \sqcup \Omega^{-1})_{1,2m} \in K_{1,2m}$.
2. Assume that $K$ is odd $C_{4m+1}$-trivial. If $S^{3}(K,r) \cong - S^{3}(K,-r)$ then $(Z^{\sigma}(K) \sqcup \Omega^{-1})_{1,4m} \in K_{1,4m}$.
3. Assume that $K$ is odd $C_{4m+3}$-trivial. If $S^{3}(K,r) \cong \pm S^{3}(K,-r)$ then $(Z^{\sigma}(K) \sqcup \Omega^{-1})_{1,4m+2} \in K_{1,4m+2}$.
As a corollary, we prove a vanishing of certain finite type invariants that come from colored Jones polynomial (Quantum $\mathfrak{sl}_2$ invariant). Let $V_{n}(K;t)$ be the $n$-colored Jones polynomial, normalized $V_{n}(\textsf{Unknot};t) = 1$. The colored Jones polynomials have the following expansion called the *loop expansion*, or *Melvin-Morton expansion* [@mm]. $$V_{n}(K;e^{h}) = \sum_{e\geq 0} \Bigl(\sum_{k\geq 0} j_{e,k}(K) (nh)^{k}\Bigr) h^{e}.$$ Here the coefficient $j_{e,k}(K) \in {\mathbb{Q}}$ is a canonical finite type invariant of degree $e+k$.
\[corollary:MM\] Let $K$ and $K'$ be a knot in $S^{3}$ and $r \in {\mathbb{Q}}\setminus\{0\}$.
1. Assume that $K$ and $K'$ are $C_{2m+1}$-equivalent. If $S^{3}(K,r) \cong S^{3}(K',r)$ and $a_{2m+2}(K) = a_{2m+2}(K')$, then $j_{1,2m}(K)=j_{1,2m}(K')$.
2. Assume that $K$ is odd $C_{4m+1}$-trivial. If $S^{3}(K,r) \cong - S^{3}(K,-r)$ then $j_{1,4m}(K)=0$.
3. Assume that $K$ is odd $C_{4m+3}$-trivial. If $S^{3}(K,r) \cong \pm S^{3}(K,-r)$ then $j_{1,4m+2}(K)=0$.
For a canonical finite type invariant $v$ of degree $n$, $v(\mbox{mirror of }K) = (-1)^{n}v(K)$. Thus for an amphicheiral knot $K$, $v(K)=0$ for all *odd* degree canonical finite type invariant $v$. It is conjectured that converse is true (this is related to a more familiar conjecture that finite type invariants do not detect the orientation of knots [@kir Problem 1.89]). Corollary \[corollary:LMO2\] (i), Corollary \[corollary:LMO3\] (ii), Corollary \[corollary:MM\] (ii),(iii) says that if $S^{3}(K,r) \cong -S^{3}(K,-r)$ then various canonical odd degree finite type invariants of $K$ vanish. Thus, they bring a supporting evidence for an affirmative answer to the following question. $$\mbox{\emph{If} } S^{3}(K,r) \cong -S^{3}(K,-r) \mbox{\emph{ for some }} r \neq 0,\infty, \mbox{\emph{ then is }} K \mbox{\emph{ amphicheiral?}}$$
Acknowledgements {#acknowledgements .unnumbered}
================
The author would like to thank K. Ichihara for stimulating talk and discussion which inspires the author to work on this subject, and to thank to Z. Wu for pointing out inaccuracy of reference in the first version of the paper. This work was partially supported by JSPS KAKENHI 15K17540,16H02145.
LMO invariants and rational surgery formula {#section:LMO}
===========================================
In this section we briefly review the basics of Kontsevich and LMO invariants. We use Aarhus integral construction of the LMO invariant developed [@bgrt1; @bgrt2; @bgrt3] and a rational surgery formula of the LMO invariant due to Bar-Natan and Lawrence [@bl]. For basics of the Kontsevich and the LMO invariants we refer [@oht].
Open Jacobi diagrams
--------------------
An *(open) Jacobi diagram* or *(vertex-oriented) uni-trivalent graph* is a graph $D$ whose vertex is either univalent or trivalent, such that at each trivalent vertex $v$ a cyclic order on three edges around $v$ is equipped. The *degree* of $D$ is the half of the number of vertices. We will often call a univalent vertex a *leg*, and denote the number of the legs of a Jacobi diagram $D$ by $k(D)$. For a Jacobi diagram $D$, let $e(D)=-\chi(D)$ be the minus of the euler characteristic of $D$. We call $e(D)$ the *euler degree* of $D$. Then $\deg(D)=e(D)+k(D)$.
Let $\mathcal{B}$ (resp. $\mathcal{A}(\emptyset)$) be the vector space over ${\mathbb{C}}$ spanned by Jacobi diagrams (resp. Jacobi diagrams without univalent vertex), modulo the AS and IHX relations given in Figure \[fig:IHX\].
![The AS and IHX relation: we understand that at each trivalent vertex, cyclic order is defined by counter-closkcwise direction.[]{data-label="fig:IHX"}](IHX.eps){width="100mm"}
By taking the disjoint union $\sqcup$ as the product, both $\mathcal{B}$ and $\mathcal{A}(\emptyset)$ have the structure of graded algebras. Since the IHX and the AS relations and the disjoint union product respect both $k(D)$ and $e(D)$, we view $\mathcal{B}$ as a bi-graded algebra. For $X \in \mathcal{B}$ we denote by $X_{e,k}$ the part of $X$ whose bigrading is $(e,k)$. Strictly speaking, we will use the completion of $\mathcal{B}$ and $\mathcal{A}(\emptyset)$ with respect to degrees which we denote by the same symbol $\mathcal{B}$ and $\mathcal{A}(\emptyset)$ by abuse of notations.
Let $\exp_{\sqcup}:\mathcal{B} \rightarrow \mathcal{B}$ (or, $\mathcal{A}(\emptyset) \rightarrow \mathcal{A}(\emptyset)$) be the exponential map with respect to $\sqcup$ product operation, defined by $$\exp_{\sqcup}(D) = 1 + D + \frac{1}{2}D \sqcup D + \cdots = \sum_{n=0}^{\infty} \frac{1}{n!}\underbrace{(D \sqcup\cdots \sqcup D)}_{n}.$$ We will simply denote $\underbrace{(D \sqcup\cdots \sqcup D)}_{n}$ by $D^{n}$. For a Jacobi diagram $C$, let $\partial_{C}:\mathcal{B} \rightarrow \mathcal{B}$ be the differential operator defined by $$\partial_{C} (D) =\begin{cases} 0 & k(C)>k(D) \\
\sum(\mbox{glue} \textit{ all } \mbox{the legs of } \Omega \mbox{ to some legs of }D) & k(C)\leq k(D)
\end{cases}$$ In a similar manner, we define the pairing $\langle C,D\rangle \in\mathcal{A}(\emptyset)$ of $C,D \in \mathcal{B}$ by $$\langle C,D \rangle =
\begin{cases}
0 & k(C) \neq k(D)\\
\sum \mbox{(glue the legs of } C \mbox{ to the legs of } D) & k(C)=k(D)
\end{cases}$$ Thus $\partial_{C}(D)=\langle C,D\rangle$ if $k(C)=k(D)$. In both cases, the summation runs all the possible ways to gluing *all* the legs of $C$ to *some* legs of $D$. We denote this summation by using box, as Figure \[fig:pairing\]. It is known that $\partial_{C \sqcup C'} = \partial_{C'}\circ\partial_{C}$. Thus if $C \in \mathcal{B}$ is invertible (with respect to $\sqcup$ product) then $\partial_{C}$ is invertible with $\partial_{C}^{-1}=\partial_{C^{-1}}$ (see [@bgrt2; @blt; @bl] for details).
![(i) Differential operator $\partial_{C}(D)$ (ii) Pairing $\langle C,D \rangle$.[]{data-label="fig:pairing"}](zu_pairing_new.eps){width="75mm"}
Let $b_{2i}$ be the modified Bernoulli numbers, defined by $$\label{eqn:modBer} \frac{1}{2}\log \frac{\sinh(x\slash2)}{x \slash 2} = \sum_{i=0}^{\infty} b_{2i} x^{2i} = 1+ \frac{1}{48}x^{2}-\frac{1}{5760}x^{4}+ \frac{1}{362880}x^{6}+ \cdots.$$ For $q \in {\mathbb{Z}}\setminus\{0\}$, let $$\Omega_{q} = \exp_{\sqcup} \Bigl(\sum_{n=1}^{\infty} \frac{b_{2n}}{q^{2n}}
\overbrace{\raisebox{-4.5mm}{\includegraphics*[width=10mm]{D_12m.eps}}}^{2n} \;\Bigr) = 1 + \frac{1}{48q^{2}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}- \frac{1}{5760q^{4}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}+ \frac{1}{4608q^{4}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+\cdots.$$
The element $\Omega=\Omega_{1}$ is called the *wheel element*.
Wheeled Kontsevich invariant
----------------------------
The *Kontsevich invariant* $Z(K)$ is an invariant of a framed knot, which takes value in $\mathcal{A}(S^{1})$, the space of Jacobi diagram over $S^{1}$ [@ko1; @bar]. Throughout the paper, we will always assume that the knot $K$ is zero-framed. The target space $\mathcal{A}(S^{1})$ is isomorphic to $\mathcal{B}$ as a graded vector space, by Poincaré-Birkoff-Witt isomorphism $\chi: \mathcal{B} \rightarrow \mathcal{A}(S^{1})$. Let $\sigma:\mathcal{A}(S^{1}) \rightarrow \mathcal{B}$ be the inverse of $\chi$. In the rest of the paper, we will always view the Kontsevich invariant takes value in $\mathcal{B}$, by defining $$Z^{\sigma}(K) = \sigma(Z(K)) \in \mathcal{B}$$ We will denote by $Z^{\sigma}(K)_{e,k}$ the bigrading $(e,k)$ part of the Kontsevich invariant. See [@gr] for a topological meaning of this bigrading.
Let $V_n$ be the vector space spanned by finite type invariants of degree $\leq n$. The Kontsevich invariant gives a map $\mathcal{Z}:(\mathcal{B}_{\deg =n})^{*} \rightarrow V_{n}$, by $\mathcal{Z}(w)(K)= w(Z^{\sigma}(K)_{\deg=n})$. Here $w: \mathcal{B}_{\deg=n} \rightarrow {\mathbb{C}}$ is an element of $(\mathcal{B}_{\deg =n})^{*}$, the dual space of degree $n$ part of $\mathcal{B}$ and $Z^{\sigma}(K)_{\deg=n} \in \mathcal{B}_{\deg=n}$ denotes the degree $n$ part of $Z^{\sigma}(K)$. On the other hand, there is a map called *symbol* $\mathsf{Symb}: V_{n} \rightarrow (\mathcal{B}_{\deg = n})^{*}$ (see [@bar; @bg] for definition). A finite type invariant $v \in V_n$ is called
- *canonical*, if $v= \mathcal{Z}(\mathsf{Symb}(v))$.
- *primitive*, if $v(K\# K')=v(K)+v(K')$, which is equivalent to saying that $v$ lies in the image of the primitive subspace of $\mathcal{B}$.
The *wheeled Kontsevich invariant* $Z^{\sf Wheel}(K) \in \mathcal{B}$ is a version of the Kontsevich invariant defined as follows.
Let $\partial_{\Omega}= 1 + \frac{1}{48}\partial_{\includegraphics*[width=3mm]{D_12.eps}}+\cdots$ be the differential operator defined by the wheel element $\Omega$. The *wheeling map* $\Upsilon = \chi \circ \partial_{\Omega}:\mathcal{B} \rightarrow \mathcal{A}(S^{1})$ is the composite of $\partial_{\Omega}$ and the Poincaré-Birkoff-Witt isomorphism $\chi$. The wheeling map $\Upsilon$ gives an isomorphism of *algebra* [@blt Wheeling theorem], whereas Poincaré-Birkoff-Witt isomorphism $\chi$ only gives an isomorphism of vector space.
The wheeled Kontsevich invariant is the image of the Kontsevich invariant under the inverse of the wheeling map $\Upsilon$: $$Z^{\sf Wheel}(K) = \Upsilon^{-1}(Z(K)) = (\partial_{\Omega})^{-1}\circ\sigma(Z(K)) = \partial_{\Omega^{-1}} Z^{\sigma}(K).$$
The wheel element $\Omega$ is equal to the Kontsevich invariant of the unknot [@blt Wheel theorem]: $Z^{\sigma}({\sf Unknot})=\Omega$. Therefore instead of $Z^{\sigma}$ or $Z^{\sf Wheel}$, it is often useful to use $Z^{\sigma}(K) \sqcup \Omega^{-1}$ since $Z^{\sigma}({\sf Unknot}) \sqcup \Omega^{-1} = 1$.
The Kontsevich invariant is group-like, so $Z^{\sigma}(K) = \exp_{\sqcup}(z^{\sigma}(K))$, where $z^{\sigma}(K)$ denotes the primitive part of $Z^{\sigma}(K)$. We express low degree part of the primitive Kontsevich invariant as $$\begin{aligned}
Z^{\sigma}(K) \sqcup \Omega^{-1}&= &\exp_{\sqcup}\Bigl( v_2(K){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+ v_3(K){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ v_{4}(K){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}+ w_4(K) {\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}\\
& & \hspace{1.8cm}+v_5(K){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_24.eps}}}+v_6(K){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_16.eps}}}+(\mbox{higher degree parts}) \Bigr).\end{aligned}$$ Here $v_{2}(K),v_3(K),v_4(K),w_4(K),v_5(K),v_6(K)$ are canonical, primitive finite type invariants of degree $2,3,4,4,5,6$, respectively.
Thus the bigrading $(e,k)$ part of $Z^{\sigma}(K)$ with $e+\frac{k}{2} \leq 3$ are explicitly written by $$\begin{aligned}
Z^{\sigma}(K)&\!\!=\!\!&1 +\bigl(v_2(K) + b_2\bigr) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+ v_3(K){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ \frac{1}{2}\bigl(v_2(K) +b_{2}\bigr)^{2} {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+(v_{4}(K)+b_{4}) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}\\
& &+ w_4(K) {\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}+ v_3(K)(v_2(K)+b_2){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ v_5(K) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_24.eps}}}+ \frac{1}{6}\bigl(v_2(K) +b_{2}\bigr)^{3}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}\\
& &+\bigl( v_2(K)+b_2\bigr)\bigl( v_4(K)+b_4 \bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}+ \bigl(v_6(K) +b_{6}\bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_16.eps}}}+(\mbox{higher degree parts}).\end{aligned}$$ Here $b_{2i}$ denotes the modified Bernoulli numbers given by (\[eqn:modBer\]).
Except $v_{5}(K)$, these finite type invariants are written by using Conway polynomial and the Jones polynomials. Let $a_{2i}(K)$ be the coefficient of $z^{2i}$ in the Conway polynomial $\nabla_{K}(z)$ of $K$, and let $j_{n}(K)$ is the coefficient of $h^{n}$ in the Jones polynomial $V_{K}(e^{h})$ of $K$, putting the variable as $t=e^{h}$. Then we have the following (See Section \[section:sl2computation\] for proof).
\[lemma:AJ\]
1. $v_{2}(K)=-\frac{1}{2}a_{2}(K)$.
2. $v_{3}(K)=-\frac{1}{24}j_{3}(K)$.
3. $v_{4}(K)=-\frac{1}{2}a_{4}(K)-\frac{1}{24}a_{2}(K)+\frac{1}{4}a_{2}(K)^{2}$.
4. $w_{4}(K)=\frac{1}{96}j_{4}(K) + \frac{3}{32}a_{4}(K) - \frac{9}{2}a_{2}(K)^{2}$.
5. $v_{6}(K)=-\frac{1}{2}a_{6}(K)-\frac{1}{12}a_{4}(K)-\frac{1}{720}a_{2}(K)+\frac{1}{24}a_{2}(K)^{2}+\frac{1}{2}a_{2}(K)a_{4}(K)-\frac{1}{6}a_{2}(K)^{3}$.
Since the Jones polynomial is an integer coefficient polynomial, $j_{3}(K) \in 6{\mathbb{Z}}$ so $4v_{3}(K) \in {\mathbb{Z}}$. The degree three finite type invariant $v_3$ takes value $-\frac{1}{4}$ for a right-handed trefoil.
In general, the euler degree zero part of the Kontsevich invariant is wirtten by the Alexander polynomial, using the following formula (This is a consequence of Melvin-Morton-Rozansky conjecture [@bg]).
\[proposition:MMR\] Let $-\frac{1}{2} \log \Delta_{K}(e^{x})=\sum_{n=0}^{\infty} d_{2n}(K) x^{2n}$ where $\Delta_{K}(t)$ is the Alexander polynomial of $K$, normalized so that $\Delta_{K}(t)=\Delta_{K}(t^{-1})$, $\Delta_{K}(1)=1$. Then the euler degree zero part of the Kontsevich invariant is $$\exp_{\sqcup}\Bigl(\sum_{k=0}^{\infty} d_{2k}(K) \overbrace{\raisebox{-4.5mm}{\includegraphics*[width=10mm]{D_12m.eps}}}^{2n} \Bigr).$$ In particular, if $a_2(K)=a_{4}(K)=\cdots =a_{2m}(K)=0$ for some $m\geq0$ then $d_{2}(K)=d_4(K)=\cdots = d_{2m}(K)=0$ and $d_{2m+2}(K)=-\frac{1}{2}a_{2m+2}(K)$.
LMO invariant and rational surgery formula
------------------------------------------
The LMO invariant $\widehat{Z}^{LMO}(M)$ is an invariant of an oriented closed 3-manifold $M$ that takes value in $\mathcal{A}(\emptyset)$. Here we restrict our attention to the case $M=S^{3}(K,r)$ for $r \neq 0,\infty$. In particular, we will always assume that $M$ is a rational homology sphere.
To make computation simpler, we will use the following simplification. Let ${\raisebox{-1mm}{\includegraphics*[width=3mm]{A1.eps}}}$ be the theta-shaped Jacobi diagram which generates the degree one part of $\mathcal{A}(\emptyset)$.
Let $\mathcal{A}^{red}$ be the quotient of $\mathcal{A}(\emptyset)$ by the ideal generated by ${\raisebox{-1mm}{\includegraphics*[width=3mm]{A1.eps}}}$, and let $\pi: \mathcal{A}(\emptyset) \to \mathcal{A}^{red}$ be the quotient map. We call $\pi(\widehat{Z}^{LMO}(M)) \in \mathcal{A}^{red}$ the *reduced LMO invariant* and denote by $Z^{LMO}(M)$. By abuse of notation, we will simply refer the reduced LMO invariant simply as the LMO invariant. When we change the orientation, the (reduced) LMO invariant changes as $$Z^{LMO}_{n}(-M) = (-1)^{n}Z^{LMO}_n(M)$$ where $Z^{LMO}_{n}(M)$ denotes the degree $n$ part of the LMO invariant.
The low degree part of the (reduced) LMO invariant is written by $$Z^{LMO}(M)=1 + \lambda_{2}(M) {\raisebox{-2mm}{\includegraphics*[width=5mm]{A2.eps}}}+ \lambda_{3}(M) {\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}+(\mbox{higher degree parts}).$$ where $\lambda_2(M), \lambda_3(M) \in {\mathbb{C}}$ are finite type invariant of rational homology spheres. In the rest of arguments, unless otherwise specified, we will always work in $\mathcal{A}^{red}$. For example, we will always view the pairing $\langle D,D' \rangle$ so that it takes value in $\mathcal{A}^{red}$, by composing the quotient map $\pi$.
Using the simplification ${\raisebox{-1mm}{\includegraphics*[width=3mm]{A1.eps}}}=0$, the (reduced) LMO invariant of the 3-manifold obtained by rational Dehn surgery along a knot $K$ is given by the following simpler formula. Let [ ]{}be the strut, the Jacobi diagram homeomorphic to the interval.
\[theorem:rsurgery\] Let $K$ be a knot in $S^{3}$. Then the (reduced) LMO invariant of $p/q$-surgery along $K$ is given by $$Z^{LMO}(S^{3}(K,p/q))= \Bigl\langle Z^{\sf Wheel}(K) \sqcup \Omega_{q} {\, \raisebox{-1mm}{,} \,}\exp_{\sqcup}( -\frac{q}{2p}\raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}) \Bigr\rangle.$$
As an application of the rational surgery formula above, in [@bl Proposition 5.1] it is shown that the (reduced) LMO invariant of the lens space $L(p,q)$ is given by the $$\label{eqn:LMO-lens}
Z^{LMO}(L(p,q)) = \langle \Omega, \Omega^{-1} \sqcup \Omega_{p}\rangle$$
Thus the (reduced) LMO invariant of Lens space only depend on $p$. In particular, $$\label{eqn:LMO23-Lens}
\lambda_{2}(L(p,q))= \frac{1}{24}\left( \frac{1}{48p^{2}}-\frac{1}{48} \right),\quad Z^{LMO}_{2m+1}(L(p,q))= 0 \ (m\in {\mathbb{Z}}).$$
Proof of Theorems
=================
First of all we determine which part of the Kontsevich invariant contributes to the degree $n$ part of the (reduced) LMO invariants.
\[prop:degree\] The degree $n$ part of the LMO invariant for $S^{3}(K,p\slash q)$ is determined by the slope $p\slash q$ and the bigrading $(e,k)$ part of $Z^{\sigma}(K)_{e,k}$ with $e+ \frac{k}{2} \leq n$.
By definition of the pairing, for $D \in \mathcal{B}_{e,k}$ we have $\langle D, \exp_{\sqcup}(-\frac{q}{2p} { \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }) \rangle \in \mathcal{A}(\emptyset)_{e+\frac{k}{2}}$. Thus the degree $n$ part of the LMO invariant of $S^{3}(K,p\slash q)$ is determined by the bigrading $(e,k)$ part of $Z^{\sf Wheel}(K) \sqcup \Omega_{q}$, with $e+\frac{k}{2}=n$.
The bigrading $(e,k)$ part of $Z^{\sf Wheel}(K) \sqcup \Omega_{q}$ is determined by the bigrading $(e',k')$ part of $Z^{\sf Wheel}(K)$ with $(e',k') \in \{(e,k), (e,k-2),(e,k-4),\ldots \}$. Also, by definition of $\partial_{D}$, if $D \in \mathcal{B}_{e,k}$ and $D' \in \mathcal{B}_{e',k'}$, $\partial_{D}(D') \in \mathcal{B}_{e'+e+k,k'-k}$. This shows that the bigrading $(e,k)$ part of $Z^{\sf Wheel}(K)$ is determined by the bigrading $(e',k')$ part of $Z^{\sf Wheel}(K)$ with $(e',k') \in \{(e,k), (e-2,k+2),(e-4,k+4),\ldots \}$.
These observations show that the degree $n$ part of the LMO invariant of $S^{3}(K,p\slash q)$ is determined by the bigrading $(e,k)$ part of $Z^{\sigma}(K)_{e,k}$ with $e+ \frac{k}{2} \leq n$ (and the surgery slope $p\slash q$).
By Proposition \[prop:degree\], to compute the degree 2 and 3 part of the LMO invariant for $S^{3}(K,p\slash q)$, it is sufficient to consider the bigraging $(e,k)$ part of $Z^{\sigma}(K)$ for $e + \frac{k}{2} \leq 3$. As we have already seen, this is given by $$\begin{aligned}
Z^{\sigma}(K)&\!\!=\!\!&1 +\bigl(v_2(K) + b_2\bigr) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+ v_3(K){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ \frac{1}{2}\bigl(v_2(K) +b_{2}\bigr)^{2} {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+(v_{4}(K)+b_{4}) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}\\
& &+ w_4(K) {\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}+ v_3(K)(v_2(K)+b_2){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ v_5(K) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_24.eps}}}+ \frac{1}{6}\bigl(v_2(K) +b_{2}\bigr)^{3}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}\\
& &+\bigl( v_2(K)+b_2\bigr)\bigl( v_4(K)+b_4 \bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}+ \bigl(v_6(K) +b_{6}\bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_16.eps}}}+(\mbox{higher degree parts}).\end{aligned}$$ Here $b_2=\frac{1}{48}, b_{4}=-\frac{1}{5760}, b_{6}= \frac{1}{362880}$ are modified Bernoulli numbers. Since $$\begin{cases}
\partial_{\Omega^{-1}} \left( {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}\right) = {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}-2b_2{\raisebox{-2mm}{\includegraphics*[width=5mm]{A2.eps}}}, \
\partial_{\Omega^{-1}} \left({\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}\right) ={\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}-2 b_2{\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}, \\
\partial_{\Omega^{-1}}\left({\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}\right)= {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}- 8b_2{\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}+\mbox{(other parts),}\\
\partial_{\Omega^{-1}}\left({\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}\right)= {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}- 10b_2{\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}+ \mbox{(other parts)},
\end{cases}$$ the wheeled Kontsevich invariant is given by $$\begin{aligned}
Z^{\sf Wheel}(K)&= 1 +\bigl(v_2(K) + b_2\bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+ v_3(K){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ \frac{1}{2}\bigl(v_2(K)+b_{2}\bigr)^{2}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+(v_{4}(K)+b_{4}) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}\\
& \hspace{0.4cm} + w_4(K){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}+ v_3(K)(v_2(K)+b_2) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ v_5(K){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_24.eps}}}+ \frac{1}{6}\bigl(v_2(K) +b_{2}\bigr)^{3}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}\\
& \hspace{0.4cm} + \bigl( v_2(K)+b_2\bigr)\bigl( v_4(K)+b_4 \bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}+\bigl(v_6(K)+b_{6}\bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_16.eps}}}+\bigr(-2b_2 \bigr)\bigl(v_2(K)+ b_2\bigr){\raisebox{-2mm}{\includegraphics*[width=5mm]{A2.eps}}}\\
& \hspace{0.4cm}+ (-2b_2v_3(K)){\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}+ \bigl(-8b_2\frac{1}{2}\bigl(v_2(K) +b_{2}\bigr)^{2}-10b_2v_4(K)\bigr){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}+ (\mbox{other parts}).\end{aligned}$$
Thus $Z^{\sf Wheel}(K)\sqcup \Omega_{q}$ is equal to $$\begin{aligned}
& & Z^{\sf Wheel}(K)\sqcup \Omega_{q}\\
&=& 1 +\bigl(v_2(K) + b_2 + \frac{b_2}{q^{2}}\bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}+ v_3{\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}\\
& & + \Bigl\{ \frac{1}{2}\bigl(v_2(K) +b_{2}\bigr)^{2} + (v_2(K)+b_2)\frac{b_2}{q^{2}}+ \frac{1}{2}\frac{b_2^2}{q^{4}}\Bigr\} {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}\\
& & + (v_{4}(K)+b_{4}+\frac{b_{4}}{q^{4}}){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}+ \bigl(w_4(K)-4b_2\bigl(v_2(K) +b_{2}\bigr)^{2}-10b_2v_4(K)\bigr){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}\\
& & + \bigl(v_3(K)(v_2(K)+b_2) + v_3(K)\frac{b_2}{q^{2}}\bigr) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}+ v_5(K){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_24.eps}}}\\
& &+ \Bigl\{\frac{1}{6}\bigl(v_2(K)+b_{2}\bigr)^{3}+ \frac{1}{2}(v_2(K)+b_2)^{2}\frac{b_2}{q^{2}}+(v_{2}(K)+b_{2})\frac{1}{2}\frac{b_2^2}{q^{4}}+ \frac{1}{6}b_{2}^{3}q^{-6}\Bigr\}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}\\
& & + \Bigl\{ \bigl( v_2(K)+b_2\bigr)\bigl( v_4(K)+b_4 \bigr)+ (v_{4}(K)+b_{4})\frac{b_2}{q^{2}} + (v_{2}(K)+b_2)\frac{b_4}{q^{4}}+\frac{ b_{2}b_{4}}{q^{6}}\Bigr\}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}\\
& & +\bigl(v_6(K) +b_{6}+\frac{b_{6}}{q^{6}}\bigr){\raisebox{-2mm}{\includegraphics*[width=6mm]{D_16.eps}}}+ \bigr(-2b_2 \bigr)\bigl(v_2(K)+ b_2\bigr){\raisebox{-2mm}{\includegraphics*[width=5mm]{A2.eps}}}+ (-2b_2v_3(K)){\raisebox{-2mm}{\includegraphics*[width=4mm]{A_3.eps}}} + (\mbox{other parts})\end{aligned}$$ By direct computations (or, one use $\mathfrak{sl}_2$ weight system evaluation as we will use in Section \[section:sl2computation\]), the pairing with struts (in $\mathcal{A}^{red}$) are given by $$\begin{cases}
\langle {\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }\rangle = 2 {\raisebox{-2mm}{\includegraphics*[width=5mm]{A2.eps}}}\,, \quad
\langle {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^2 \rangle = 16 {\raisebox{-2mm}{\includegraphics*[width=5mm]{A2.eps}}}\,,\quad
\langle {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^2 \rangle = 20 {\raisebox{-2mm}{\includegraphics*[width=5mm]{A2.eps}}}\,,\\
\langle {\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^2 \rangle = 2 {\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}\,, \quad
\langle {\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^2 \rangle = 16 {\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}\,, \quad
\langle {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_24.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^2 \rangle = 20 {\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}\,, \\
\langle {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{3}\rangle = 384{\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}\,, \quad
\langle {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{3}\rangle= 480 {\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}\,,\\
\langle {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_16.eps}}}{\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{3}\rangle = 420 {\raisebox{-3mm}{\includegraphics*[width=5mm]{A_3.eps}}}.
\end{cases}$$ Consequently, we get $$\begin{aligned}
\lambda_2(S^{3}_{K}(p/q)) &=& v_{3}(K)\bigl(-\frac{q}{2p}\bigr)\cdot 2 \\
\nonumber & &+ \left( \frac{(v_2(K)+b_{2})^{2}}{2} + \frac{ (v_2(K)+b_2)b_2}{q^{2}} + \frac{b_2^{2}}{2q^{4}} \right)\frac{1}{2}\bigl(-\frac{q}{2p}\bigr)^{2}\cdot 16\\
\nonumber & &+\left( v_{4}(K)+b_{4}+\frac{b_{4}}{q^{4}} \right) \frac{1}{2}\bigl(-\frac{q}{2p}\bigr)^{2}\cdot 20 + \bigr(-2b_2 \bigr)\bigl(v_2(K)+ b_2\bigr).\end{aligned}$$ $$\begin{aligned}
\lambda_3(S^{3}(K,p/q)) &\!\! = \!\!&
\bigl(w_4(K)-4b_2\bigl(v_2(K) +b_{2}\bigr)^{2}-10b_2v_4(K)\bigr)\bigl( -\frac{q}{2p}\bigr) \cdot 2 \\
& & \hspace{-1.5cm}+ \left( (v_3(K)(v_2(K)+b_2) + \frac{v_3(K)b_2}{q^{2}}\right)\frac{1}{2}\bigl( -\frac{q}{2p}\bigr)^{2} \cdot 16 + v_{5}(K)\frac{1}{2}\bigl( -\frac{q}{2p}\bigr)^{2}\cdot 20\\
& &\hspace{-1.5cm} + \left( \frac{\bigl(v_2(K)+b_{2}\bigr)^{3}}{6}+ \frac{(v_2(K)+b_2)^{2}b_{2}}{2q^{2}}+\frac{(v_{2}(K)+b_{2})b_{2}^{2}}{2q^{4}}+ \frac{b_{2}^{3}}{6q^{6}}\right) \frac{1}{6}\bigl( -\frac{q}{2p}\bigr)^{3}\cdot 384\\
& &\hspace{-1.5cm} + \left( \bigl( v_2+b_2\bigr)\bigl( v_4+b_4 \bigr)+ \frac{(v_{4}(K)+b_{4})b_2}{q^{2}} + \frac{(v_{2}(K)+b_2)b_{4}}{q^{4}}+ \frac{b_{2}b_{4}}{q^{6}} \right) \frac{1}{6}\bigl( -\frac{q}{2p}\bigr)^{3}\!\!\cdot480\\
& &\hspace{-1.5cm} + \left(v_6(K)+b_{6}+\frac{b_{6}}{q^{6}}\right)\frac{1}{6}\bigl( -\frac{q}{2p}\bigr)^{3}\cdot 420 + (-2b_2v_3(K)).\end{aligned}$$ Thus we conclude $$\begin{aligned}
\lambda_2(S^{3}(K,p/q))-\lambda_{2}(L(p,q)) & \\
& \hspace{-3cm} =
\left( v_2(K)^{2} + \frac{1}{24}v_2(K)+\frac{5}{2}v_{4}(K)\right)\frac{q^{2}}{p^{2}}-v_3(K)\frac{q}{p} + \frac{v_2(K)}{24}\left( \frac{1}{p^{2}}-1\right) \\
& \hspace{-3cm} =
\left( \frac{7a_2(K)^2-a_2(K)-10a_{4}(K)}{8} \right)\frac{q^{2}}{p^{2}} -v_{3}(K)\frac{q}{p} + \frac{a_{2}(K)}{48}
\left(1-\frac{1}{p^{2}}\right),\end{aligned}$$ $$\begin{aligned}
\lambda_{3}(S^{3}(K,p/q))-\lambda_{3}(L(p,q)) & \\
& \hspace{-4cm} =
-\left(\frac{35}{4}v_6(K)+\frac{5}{24}v_4(K)+10v_2(K)v_4(K)+ \frac{4}{3}v_2(K)^{3}+\frac{1}{12}v_2(K)^{2}
\right)\frac{q^{3}}{p^{3}}\\
& \hspace{-3.5cm}
-\left( \frac{5}{24}v_4(K) +\frac{1}{288}v_2(K)+\frac{1}{12}v_2(K)^{2}\right)\frac{q}{p^{3}}\\
& \hspace{-3.5cm}
+ \left(\frac{5}{2}v_{5}(K)+ 2v_3(K)v_2(K)+\frac{1}{24}v_3(K)\right)\frac{q^2}{p^2} + \frac{v_3(K)}{24}\left(\frac{1}{p^2}-1\right)\\
&
\hspace{-3.5cm}- \left(w_4(K)-\frac{1}{12}v_2(K)^{2}-\frac{1}{288}v_2(K)-\frac{5}{24}v_4(K)\right)\frac{q}{p}\end{aligned}$$
\[remark:KKT\] In [@le Theorem7.1] Lescop proved a similar formula $$\label{eqn:Lescopformula}
\lambda_2^{KKT}(S^{3}(K,p\slash q)) = \lambda^{'' KKT}_{2}(K) \frac{q^{2}}{p^{2}} + w_3(K)\frac{q}{p} + c(p\slash q)a_{2}(K) + \lambda_{2}^{KKT}(L(p,q))$$ for the degree two part $\lambda_2^{KKT}$ of the Kontsevich-Kuperberg-Thurston universal finite type invariant $Z^{KKT}$, which is defined by configuration space integrals [@ko2; @kt]. For degree two part we have $\lambda_{2}^{KKT} = 2 \lambda_2$ (note that in Lescop uses the coefficient of the Jacobi diagram ${\raisebox{-2mm}{\includegraphics*[width=5mm]{A_2p.eps}}} = \frac{1}{2}{\raisebox{-2mm}{\includegraphics*[width=4mm]{A2.eps}}}$) so Theorem \[theorem:LMO23\] gives formulae of invariants in Lescop’s formula (\[eqn:Lescopformula\]), namely, $$\lambda^{'' KKT}_{2}(K)= \frac{7a_2(K)^2-a_2(K)-10a_{4}(K)}{4}, w_3(K)=-2v_3(K), c(p\slash q) = \frac{1}{24}-\frac{1}{24p^{2}}.$$
[$ $]{}\
(i,ii): By (\[eqn:NW\]), it is sufficient to consider the case $r'= - r$. Assume that $S^{3}(K,p/q) \cong \pm S^{3}(K,-p/q)$. Since $\lambda_2(M)=\lambda_2(-M)$, by Theorem \[theorem:LMO23\] $$\lambda_{2}(S^{3}(K,p/q)) - \lambda_{2}(\pm S^{3}(K,-p/q)) = -2 v_{3}(K) q\slash p =0$$ Therefore $v_{3}(K)=0$.\
(iii): Assume that $S^{3}(K,p/q) \cong \pm S^{3}(K, -p'/q')$. Since $H_{1}(S^{3}(K,p/q))\cong {\mathbb{Z}}\slash p{\mathbb{Z}}$, $p=p'$. By Theorem \[theorem:LMO23\] $$\begin{aligned}
0 &= & \lambda_{2}(S^{3}(K,p/q)) - \lambda_{2}(- S^{3}(K,p/q')) \\
& =& \left(\frac{7a_{2}(K)^{2}-a_{2}(K)-10a_{4}(K)}{8}\right)\frac{q^{2}-q'^{2}}{p^{2}} - v_{3}(K)\frac{q-q'}{p}.\end{aligned}$$ Since $q' \neq \pm q$, either $v_{3}(K)\neq 0$ (and $7a_{2}(K)^{2}-a_{2}(K)-10a_{4}(K)=0$), or, $$\frac{p}{q+q'} = \frac{7a_{2}(K)^{2}-a_{2}(K)-10a_{4}(K)}{8v_{3}(K)}.$$
(iv): Assume that $S^{3}(K,p/q) \cong \pm S^{3}(K', p/q)$. By Theorem \[theorem:LMO1\] (i), $a_{2}(K)=a_{2}(K')$. By Theorem \[theorem:LMO23\] $$\lambda_{2}(S^{3}(K,p/q)) - \lambda_{2}(S^{3}(K',p/q)) = \frac{5}{4}(a_{4}(K)-a_{4}(K'))\frac{q^{2}}{p^{2}} - (v_{3}(K)-v_{3}(K'))\frac{q}{p}=0 .$$
Assume that $S^{3}(K,p\slash q) \cong L(p',q')$. Then $|H_{1}(S^{3}(K,p\slash q);{\mathbb{Z}})|=p=p'=H_{1}(L(p',q));{\mathbb{Z}})$ so $p=p'$. By (\[eqn:LMO23-Lens\]) $\lambda_{2}(L(p,q))=\lambda_{2}(L(p,q'))$ so Theorem \[theorem:LMO23\] gives the desired equality.
(i): By (\[eqn:NW\]), it is sufficient to consider the case $r'= -r$. By Theorem \[theorem:LMO1\] (ii) and Corollary \[corollary:LMO2\] (i), $a_{2}(K)=v_2(K)=v_{3}(K)=0$. Thus by Theorem \[theorem:LMO23\] $$\begin{aligned}
0 &=& \lambda_{3}(S^{3}(K,p/q)) - \lambda_{3}(S^{3}(K,-p/q))\\
& =& -\left( \frac{35}{2}v_{6}(K) + \frac{5}{12} v_4(K)\right) \frac{q^{3}}{p^{3}} - \frac{5}{12}v_{4}(K) \frac{q}{p^{3}} - \left( 2w_{4}(K)-\frac{5}{12}v_{4}(K)\right) \frac{q}{p}.\end{aligned}$$ (ii): If $S^{3}_{K}(p/q)) \cong -S^{3}_{K}(-p/q)$ then by Corollary \[corollary:LMO2\] (ii) $v_{3}(K)=0$. Therefore by Theorem $ \lambda_{3}(S^{3}(K,p/q)) - \lambda_{3}(-S^{3}(K,-p/q)) = 5v_5(K)= 0$.
By Lemma \[lemma:AJ\] and Corollary \[corollary:LMO3\] (i), if $S^{3}(K,p\slash q) \cong S^{3}(K,-p\slash q)$ then we get $$\label{eqn:obstruction}
(19a_{4}(K)+j_4(K))p^{2} -10a_{4}(K)-(420a_{6}(K)+80a_{4}(K))q^{2}=0.$$
According to [@iw], the cosmetic surgery conjecture was confirmed for knots with less than or equal to 11 crossings, with 8 exceptions $$10_{33},10_{118},10_{146},11a_{91},11a_{138},11a_{285},11n_{86},11n_{157}$$ in the table KnotInfo [@cl].
For these knots, the values of $a_{4},j_{4}$ and $a_{6}$ are given as follows.
$10_{33}$ $10_{118}$ $10_{146}$ $11a_{91}$ $11a_{138}$ $11a_{285}$ $11n_{86}$ $11n_{157}$
--------- ----------- ------------ ------------ ------------ ------------- ------------- ------------ -------------
$a_{4}$ 4 2 2 0 2 2 -2 0
$j_{4}$ -12 -6 -6 0 -6 -6 6 0
$a_{6}$ 0 3 0 -2 -2 -2 -1 -1
Note that (\[eqn:obstruction\]) gives a diophantine equation of the form $ap^{2}-bq^{2}=c$, whose solvability can be checked algorithmically [@aa]. For these knots the equation (\[eqn:obstruction\]) has no integer solutions, except the case $K=10_{118}$ (The author uses the computer program at [@so]. In the case $K=10_{118}$ we get the equation $32p^2-20-1420q^{2}=0$ which has the solutions $p=20u+1065v$ and $q=3u-160v$, where $(u,v)$ are the solutions of Pell’s equation $u^2 - 2840v^2 = 1$.)
Next we proceed to see higher degree part. To use $C_{n}$-equvalence assumption, we observe the following.
\[lemma:KontCn\] If $K$ and $K'$ are $C_{n+1}$-equivalent, then for $e+\frac{k}{2} \leq n+1$, then $$(Z^{\sf Wheel}(K) \sqcup \Omega_{q} - Z^{\sf Wheel}(K) \sqcup \Omega_{q})_{e,k} = (Z^{\sigma}(K) \sqcup \Omega^{-1})_{e,k} - (Z^{\sigma}(K') \sqcup \Omega^{-1})_{e,k}$$ Similarly, if $K$ and $K'$ are odd $C_{n+1}$-equivalent, then for $e+\frac{k}{2} \leq n+1$ with odd $e+k$, $$(Z^{\sf Wheel}(K) \sqcup \Omega_{q} - Z^{\sf Wheel}(K) \sqcup \Omega_{q})_{e,k} = (Z^{\sigma}(K) \sqcup \Omega^{-1})_{e,k} - (Z^{\sigma}(K') \sqcup \Omega^{-1})_{e,k}$$
Since $K$ and $K'$ are $C_{n}$ equivalent $Z^{\sigma}(K)_{e',k'} = Z^{\sigma}(K')_{e',k'}$ if $e'+k' \leq n$. Since for the degree $d$ element $D$, the $\partial_{\Omega^{-1}}(D) = D+ (\mbox{degree} \geq d+2\mbox{ elements } )$, for $e+k \leq n+1$ $$Z^{\sf Wheel}(K)_{e,k}-Z^{\sf Wheel}(K')_{e,k} = (\partial_{\Omega^{-1}}(Z^{\sigma}(K)-Z^{\sigma}(K'))_{e,k} =
(Z^{\sigma}(K)-Z^{\sigma}(K'))_{e,k}.$$ Therefore for $e+k \leq n+1$ $$\begin{aligned}
(Z^{\sf Wheel}(K) \sqcup \Omega_{q} - Z^{\sf Wheel}(K') \sqcup \Omega_{q})_{e,k}& = & \left( (Z^{\sf Wheel}(K) - Z^{\sf Wheel}(K')) \sqcup \Omega_{q}\right)_{e,k}\\
& = & \left( (Z^{\sigma}(K)-Z^{\sigma}(K'))\sqcup \Omega_q \right)_{e,k}\\
& = & (Z^{\sigma}(K)-Z^{\sigma}(K'))_{e,k}\\
& = & (Z^{\sigma}(K) \sqcup \Omega^{-1})_{e,k} - (Z^{\sigma}(K') \sqcup \Omega^{-1})_{e,k} \end{aligned}$$
To see the latter assertion, we note that both $\partial_{\Omega^{-1}}$ and $\sqcup \Omega_{q}$ preserve the parity of $D$. Namely, if $D \in \mathcal{B}_{odd}$, where we denote by $\mathcal{B}_{odd}$ the odd degree part of $\mathcal{B}$, then $\partial_{\Omega^{-1}}(D), D\sqcup \Omega_{q} \in \mathcal{B}_{odd}$.
[$ $]{}\
(i): By Proposition \[prop:degree\], the degree $m+1$ part of the LMO invariant of $S^{3}(K,p\slash q)$ is determined by $(Z^{\sf Wheel}(K) \sqcup \Omega_{q})_{e,k}$ with $e+\frac{k}{2}=m+1$.
Since $K$ and $K'$ are $C_{2m+2}$-equivalent, for $e+\frac{k}{2}=m+1$ $$\begin{aligned}
(Z^{\sf Wheel}(K) \sqcup \Omega_{q} - Z^{\sf Wheel}(K') \sqcup \Omega_{q})_{e,k} &=&(Z^{\sigma}(K)\sqcup \Omega^{-1})_{e,k} - (Z^{\sigma}(K')\sqcup \Omega^{-1})_{e,k}\\
& & \hspace{-2cm}=
\begin{cases}
-\frac{1}{2}(a_{2m+2}(K)-a_{2m+2}(K'))\overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{2m+2} & (e,k)=(0,2m+2)\\
\quad 0 & \mbox{Otherwise},
\end{cases}\end{aligned}$$ by Lemma \[lemma:KontCn\] and Proposition \[proposition:MMR\]. Therefore $$\begin{aligned}
\label{eqn:LMO_2m+1}
Z^{LMO}_{m+1}(S^{3}(K,p\slash q))-Z^{LMO}_{m+1}(S^{3}(K',p\slash q)) & \\
& \nonumber \hspace{-6cm} = \left\langle -\frac{1}{2}(a_{2m+2}(K)-a_{2m+2}(K'))\overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{2m+2} {\, \raisebox{-3mm}{,} \,}\frac{1}{(m+1)!}\left(-\frac{q}{2p}\right)^{2m+1} \!\!\!\!\! { \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{m+1} \right\rangle \\
\nonumber&\ \hspace{-6cm} = - \frac{a_{2m+2}(K)-a_{2m+2}(K')}{2(m+1)!}\left(-\frac{q}{2p}\right)^{m+1}
\left\langle \overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{2m+2} {\, \raisebox{-3mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{m+1} \right\rangle {\, \raisebox{-3mm}{.} \,}\end{aligned}$$
As we will see in Lemma \[lemma:nonvanish\], $\left\langle \overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{2m+2} {\, \raisebox{-3mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{m+1} \right\rangle \neq 0$. This shows that $S^{3}(K,p\slash q) \cong S^{3}(K',p\slash q)$ implies $a_{2m+2}(K)=a_{2m+2}(K')$.\
(ii) By (\[eqn:LMO\_2m+1\]), when $K$ is $C_{4m+2}$-trivial, then $$Z^{LMO}_{2m+1}(S^{3}(K,p\slash q))-Z^{LMO}_{2m+1}(L(p,q)) = \left\langle -\frac{1}{2}a_{4m+2}(K)\overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{4m+2} {\, \raisebox{-3mm}{,} \,}\frac{1}{(2m+1)!}\left(-\frac{q}{2p}\right)^{2m+1} \!\!\!\!\!\! { \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{2m+1} \right\rangle {\, \raisebox{-3mm}{.} \,}$$ By (\[eqn:NW\]), if $S^{3}(K,p\slash q) \cong S^{3}(K,p'\slash q')$ then $\frac{p}{q}= - \frac{p'}{q'}$ and $Z^{LMO}_{2m+1}(L(p,q))-Z^{LMO}(L(p',q'))=0$ hence $$\begin{aligned}
0 &=& Z^{LMO}_{2m+1}(S^{3}(K,p\slash q))-Z^{LMO}_{2m+1}(S^{3}(K,-p\slash q))\\
& =& - \frac{a_{4m+2}(K)}{(2m+1)!}\left(-\frac{q}{2p}\right)^{2m+1} \left\langle\overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{4m+2} {\, \raisebox{-3mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{2m+1} \right\rangle {\, \raisebox{-3mm}{.} \,}\end{aligned}$$ Therefore $a_{4m+2}(K)=0$.
[$ $]{}\
(i): By the same argument as Theorem \[theorem:LMOhigh1\] (i), if $K$ and $K'$ are $C_{2m+1}$-equivalent for $e+\frac{k}{2}=m+1$, $$\begin{aligned}
(Z^{\sf Wheel}(K) \sqcup \Omega_{q} - Z^{\sf Wheel}(K') \sqcup \Omega_{q})_{e,k} &=& (Z^{\sigma}(K)\sqcup \Omega^{-1})_{e,k} - (Z^{\sigma}(K')\sqcup \Omega^{-1})_{e,k}\\
& & \hspace{-3cm}=
\begin{cases}
-\frac{1}{2}(a_{2m+2}(K)-a_{2m+2}(K'))\overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{2m+2} & (e,k)=(0,2m+2)\\
(Z^{\sigma}(K) \sqcup \Omega^{-1} - Z^{\sigma}(K) \sqcup \Omega^{-1})_{1,2m} & (e,k)=(1,2m)\\
0 & \mbox{Otherwise.}
\end{cases}\end{aligned}$$ Hence $$\begin{aligned}
Z^{LMO}_{m+1}(S^{3}(K,p\slash q))- Z^{LMO}_{m+1}(S^{3}(K',p\slash q)) & & \\
& &\hspace{-6cm} = \left\langle -\frac{1}{2}(a_{2m+2}(K)-a_{2m+2}(K'))\overbrace{\raisebox{-4.5mm}{\includegraphics*[width=9mm]{D_12m.eps}}}^{2m+2}
{\, \raisebox{-3mm}{,} \,}\frac{1}{(m+1)!}\left(-\frac{q}{2p}\right)^{m+1} \!\!\!\! { \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{m+1} \right\rangle \\
& & \hspace{-5cm}+ \left\langle (Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,2m} - (Z^{\sigma}(K')\sqcup \Omega^{-1})_{1,2m} {\, \raisebox{-3mm}{,} \,}\frac{1}{m!}\left(-\frac{q}{2p}\right)^{m}\!\!\!\! { \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{m} \right\rangle {\, \raisebox{-3mm}{.} \,}\end{aligned}$$ Thus, if $S^{3}(K,p\slash q) \cong S^{3}(K', p\slash q)$ and $a_{4m+4}(K)=a_{4m+4}(K')$ then $$\left\langle \Bigl( (Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,4m+2} - (Z^{\sigma}(K')\sqcup \Omega^{-1})_{1,4m+2} \Bigr), { \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{2m+1} \right\rangle =0.$$
(ii,iii) Assume that $K$ is odd $C_{4m+1}$-trivial. Since $$\begin{aligned}
Z^{LMO}_{2m+1}(S^{3}(K,p\slash q)) = \sum_{e=0}^{m} \left\langle (Z^{\sf Wheel}(K)\sqcup \Omega_{q})_{2e,4m+2-4e} {\, \raisebox{-3mm}{,} \,}\frac{1}{(2m+1-2e)!} \left( -\frac{q}{2p}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }\right)^{2m+1-2e} \right\rangle \\
+ \sum_{e=0}^{m} \left\langle (Z^{\sf Wheel}(K)\sqcup \Omega_{q})_{2e+1,4m-4e} {\, \raisebox{-3mm}{,} \,}\frac{1}{(2m-2e)!} \left( -\frac{q}{2p}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }\right)^{2m-2e} \right\rangle\end{aligned}$$ we get $$\begin{aligned}
&0= Z^{LMO}_{2m+1}(S^{3}(K,p\slash q))-Z^{LMO}_{2m+1}(-S^{3}(K,-p\slash q)) = Z^{LMO}_{2m+1}(S^{3}(K,p\slash q)) + Z^{LMO}_{2m+1}(S^{3}(K,-p\slash q))\\
& \hspace{0.5cm}= 2 \sum_{e=0}^{m} \left\langle (Z^{\sf Wheel}(K)\sqcup \Omega_{q})_{2e+1,4m-4e} {\, \raisebox{-3mm}{,} \,}\frac{1}{(2m-2e)!} \left( -\frac{q}{2p}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }\right)^{2m-2e} \right\rangle {\, \raisebox{-3mm}{.} \,}\end{aligned}$$ By Lemma \[lemma:KontCn\], $(Z^{\sf Wheel}(K)\sqcup \Omega_{q})_{2e+1,4m-4e} = 0$ unless $e=0$. Moreover for $e=0$ we have $$(Z^{\sf Wheel}(K)\sqcup \Omega_{q})_{1,4m}=(Z^{\sf Wheel}({\sf Unknot})\sqcup \Omega_q)_{1,4m} + (Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,4m} - (Z^{\sigma}({\sf Unknot})\sqcup \Omega^{-1})_{1,4m}.$$ Since for any $X \in \mathcal{B}$ and $q \in {\mathbb{Z}}\setminus\{0\}$, $(X\sqcup \Omega_q^{\pm 1})_{1,k}= X_{1,k}$ we have $$(Z^{\sf Wheel}({\sf Unknot})\sqcup \Omega_q)_{1,4m} = (Z^{\sigma}({\sf Unknot})\sqcup \Omega^{-1})_{1,4m}.$$ Thus we get $$(Z^{\sf Wheel}(K)\sqcup \Omega_{q})_{1,4m} = (Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,4m}.$$ Hence $$0= \frac{2}{(2m)!}\left(-\frac{q}{2p}\right)^{2m} \!\!\left\langle (Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,4m} {\, \raisebox{-1mm}{,} \,}{ \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}} }^{2m} \right\rangle .$$ (iii) is proved similarly.
Lemma \[lemma:nonvanish2\] in next Section shows that if $(Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,2m} \in K_{1,2m}$ then $j_{1,2m}(K)=0$.
Some $\mathfrak{sl}_2$ weight system computations {#section:sl2computation}
=================================================
In this section we use $(\mathfrak{sl}_2,V_n)$ weight system, which is a linear map $W_{(\mathfrak{sl_2},V_n)} : \mathcal{B} \ (\mbox{or, } \mathcal{A}(\emptyset))\rightarrow {\mathbb{C}}[[h]]$ that comes from the Lie algebra $\mathfrak{sl}_2$ and its $n$-dimensional irreducible representation, to confirm some assetions used in previous sections.
The image of $W_{\mathfrak{sl}_2,V_n}$ can be calculated recursively by the following relations [@cv]:
1. $W_{\mathfrak{sl_2}}(\raisebox{-2.5mm}{\includegraphics*[width=7mm]{I.eps}}) = 2h\Bigr(W_{\mathfrak{sl_2}}(\raisebox{-2.5mm}{\includegraphics*[width=7mm]{P.eps}})- W(\raisebox{-2.5mm}{\includegraphics*[width=7mm]{X.eps}})\Bigl)$
2. $W_{\mathfrak{sl_2}}(\raisebox{-2.5mm}{\includegraphics*[width=7mm]{bubble.eps}}) = 4h W_{\mathfrak{sl_2}}(\raisebox{-2.5mm}{\includegraphics*[width=7mm]{line.eps}})$.
3. $W_{\mathfrak{sl_2}}(\raisebox{-3mm}{\includegraphics*[width=7mm]{circ.eps}})=3$.
4. $W_{(\mathfrak{sl_2},V_n)}(D \sqcup \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}) = h \frac{n^{2}-1}{2}W_{(\mathfrak{sl_2},V_n)}(D)$.
Note that $\deg W_{(\mathfrak{sl_2},V_n)}(D) = \deg(D)$ and the relation (1)–(3) do not depend on $n$. Thus for $D \in \mathcal{A}(\emptyset)$, $ W_{(\mathfrak{sl_2},V_n)}(D)$ does not depend on $n$ so we will simply write by $W_{\mathfrak{sl_2}}(D)$.
By the definition of the weight system and the colored Jones polynomial as quantum $(\mathfrak{sl}_2, V_n)$ invariant, we have $$\label{eqn:Jones}
W_{(\mathfrak{sl_2},V_n)}(Z^{\sigma}(K)\sqcup \Omega^{-1}) = V_{n}(K;e^{-h})= \sum_{e\geq 0} \left(\sum_{k\geq 0} j_{e,k}(K) (nh)^{k}\right)h^{e}.$$ We remark that we put the variable $t$ in the colored Jones polynomial not $e^{h}$ but $e^{-h}$, due to the difference of normalization of the colored Jones polynomial and quantum $\mathfrak{sl}_2$ invariants.
We use this to check finite type invariants which we used can be written by Jones and Conway polynomials.
(1),(3) and (5) follow from Proposition \[proposition:MMR\] so we prove (2) and (4). The degree three and four part of $(Z^{\sigma}(K)\sqcup \Omega^{-1})$ are given $v_{3}(K){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}}$ and $\frac{1}{8}a_{2}(K)^{2}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}- \frac{1}{2}a_{4}(K) {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}}+ w_{4}(K){\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}}$, respectively. Thus by (\[eqn:Jones\]) applying $W_{(\mathfrak{sl_2},V_2)}$ we get $$\begin{aligned}
j_{3}(K)(-h)^{3} & = & v_{3}(K) W_{(\mathfrak{sl}_2, V_2)}({\raisebox{-2mm}{\includegraphics*[width=4mm]{D_22.eps}}})= 24v_{3}(K)h^{3} \\
j_{4}(K)(-h)^{4} & = & \frac{1}{8}a_{2}(K)^{2} W_{(\mathfrak{sl}_2, V_2)}({\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}{\raisebox{-2mm}{\includegraphics*[width=6mm]{D_12.eps}}}) - \frac{1}{2}a_{4}(K) W_{(\mathfrak{sl}_2, V_2)}( {\raisebox{-2mm}{\includegraphics*[width=6mm]{D_14.eps}}})+ w_{4}(K) W_{(\mathfrak{sl}_2, V_2)}({\raisebox{-2mm}{\includegraphics*[width=4mm]{D_32.eps}}})\\
& = &\frac{1}{8}a_{2}(K)^{2}36h^{4} - \frac{1}{2}a_{4}(K) 18h^{4}+ w_{4}(K)96h^{4}\\
& = & \left(\frac{9}{2}a_{2}(K)^{2} -9a_{4}(K) + 96w_{4}(K)\right) h^{4}\end{aligned}$$
For two Jacobi diagrams $D$ and $D'$ we write $D \equiv D'$ if $D$ is equal to $D'$ by using the $\mathfrak{sl}_2$ weight system relations (1)–(3) which are independent of $V_n$. By the $\mathfrak{sl}_2$ weight system relations (1)–(3) we remove all trivalent vertex of a Jacobi diagram when the number of univalent vertices are even.
\[lemma:nonvanish\] $W_{\mathfrak{sl}_2} \Bigl(\Bigl\langle \overbrace{\raisebox{-4mm}{\includegraphics*[width=10mm]{D_12m.eps}}}^{2m} {\, \raisebox{-3mm}{,} \,}\raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}^m \Bigr\rangle \Bigr) = 2(2h)^{m}(2m+1)!$. In particular, $\Bigl\langle \overbrace{\raisebox{-4mm}{\includegraphics*[width=10mm]{D_12m.eps}}}^{2m} {\, \raisebox{-3mm}{,} \,}\raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}^m \Bigr\rangle \neq 0$
First we observe that $$\overbrace{\raisebox{-4mm}{\includegraphics*[width=10mm]{D_12m.eps}}}^{2m} \equiv (2h)^{m}\sum_{{\mathbf{e}}=(e_1,\ldots,e_m) \in \{0,1\}^{m}} (-1)^{e_1+\cdots+e_m} D_{{\mathbf{e}}} \\$$ Here for ${\mathbf{e}}=(e_1,\ldots,e_m) \in \{0,1\}^{m}$, $D_{{\mathbf{e}}}$ denotes the Jacobi diagram\
$$\includegraphics*[width=70mm]{zu_d0.eps}$$ Then the pairing $ \left\langle D_{{\mathbf{e}}}, \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}^m \right\rangle = {\raisebox{-5mm}{\includegraphics*[width=20mm]{d1.eps}}}$ is given by $$\label{eqn:pairing0}
{\raisebox{-5mm}{\includegraphics*[width=25mm]{d1.eps}}} = \begin{cases} \includegraphics*[width=30mm]{d3.eps} & (e_{1}=e_2=\cdots=e_m=0)\\
\includegraphics*[width=25mm]{d2.eps} & (\text{otherwise})
\end{cases}$$ By definition, $W_{\mathfrak{sl}_2}\left({\raisebox{-3mm}{\includegraphics*[width=20mm]{d2.eps}}}\right)= W_{\mathfrak{sl}_2}\left(\Bigl\langle\raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}^{m}, \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}^{m} \Bigr\rangle \right) = (2m+1)!$ (see [@blt Lemma 6.1]). Therefore by (\[eqn:pairing0\]) $$W_{\mathfrak{sl}_2} \Bigl(\Bigl\langle \overbrace{\raisebox{-4mm}{\includegraphics*[width=10mm]{D_12m.eps}}}^{2n}{ \raisebox{-2mm},} \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}^m \Bigr\rangle \Bigr) = 2(2h)^{m}(2m+1)!$$
\[lemma:nonvanish2\] $$W_{\mathfrak{sl_2}}\left(\left\langle (Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,2m} {\raisebox{-2mm},} \
\raisebox{-1.8mm}{\includegraphics*[width=7mm]{strut.eps}}^{m}
\right\rangle \right) = 2^{m}h^{m+1} (2m+1)! j_{1,2m}(K)$$
Let us put $(Z^{\sigma}(K)\sqcup \Omega^{-1})_{e,2k} \equiv c_{e,2k}(K) h^{e+k} \raisebox{-1mm}{\includegraphics*[width=7mm]{strut.eps}}^{k}$. Then $$\begin{aligned}
W_{\mathfrak{sl}_2,V_{n}}((Z^{\sigma}(K)\sqcup \Omega^{-1})_{e,2k}) &= e_{e,2k}(K) h^{e+2k}\left(\frac{n^{2}-1}{2}\right)^{k} \\
&\hspace{-3.2cm} = \frac{c_{e,2k}(K)}{2^{k}} h^{e}(nh)^{2k} -\frac{c_{e,2k}(K)}{2^{k}}kh^{e+2}(nh)^{2k-2}+ \frac{c_{e,2k}(K)}{2^{k}}\binom{k}{2}h^{e+4}(nh)^{2k-4}-\cdots. \\\end{aligned}$$ By (\[eqn:Jones\]) we conclude $j_{1,2m}(K) = \frac{c_{1,2m}(K)}{2^{m}}$.Then the $\mathfrak{sl}_2$ weight system evaluation of the desired pairing is $$\begin{aligned}
W_{\mathfrak{sl_2}}\left( \left\langle(Z^{\sigma}(K)\sqcup \Omega^{-1})_{1,2m} {\raisebox{-2mm},}
\raisebox{-1.8mm}{\includegraphics*[width=7mm]{strut.eps}}^{m}
\right\rangle \right)& = & 2^{m}j_{1,2m}(K)h^{m+1} W_{\mathfrak{sl_2}}\left({\raisebox{-3mm}{\includegraphics*[width=20mm]{d2.eps}}}\right)\\
& = & 2^{m}h^{m+1} (2m+1)! j_{1,2m}(K)\end{aligned}$$
[1]{} T. Andreescu, and D. Andrica, [*Quadratic Diophantine equations.*]{} With a foreword by Preda Mihăilescu. Developments in Mathematics, 40. Springer, New York, 2015. xviii+211.
K. Baker and K. Motegi, [*Non-characterizing slopes for hyperbolic knots,*]{} arXiv:1601.01985.
D. Bar-Natan, [*On the Vassiliev knot invariants*]{}, Topology, **34**, (1995), 423–472.
D. Bar-Natan and S. Garoufalidis, [*On the Melvin-Morton-Rozansky conjecture*]{}, Invent. Math, **125**, (1996), 103–133.
D. Bar-Natan, S, Garoufalidis, L. Rozansky, and D. Thurston, [*The Arhus integral of rational homology 3-spheres. I. A highly non trivial flat connection on $S^{3}$,*]{} Selecta Math. (N.S.) **8** (2002), 315–339.
D. Bar-Natan, S, Garoufalidis, L. Rozansky, and D. Thurston, [*The Arhus integral of rational homology 3-spheres. II. Invariance and universality,*]{} Selecta Math. (N.S.) **8** (2002), 341–371.
D. Bar-Natan, S, Garoufalidis, L. Rozansky, Lev and D. Thurston, Dylan, [*The Arhus integral of rational homology 3-spheres. III. Relation with the Le-Murakami-Ohtsuki invariant,*]{} Selecta Math. (N.S.) **10** (2004), 305–324.
D. Bar-Natan, T. Le, and D. Thurston, [*Two applications of elementary knot theory to Lie algebras and Vassiliev invariants*]{}, Geom. Topol. **7** (2003), 1–31.
D. Bar-Natan, and R. Lawrence, [*A rational surgery formula for the LMO invariant,*]{} Israel J. Math. **140** (2004), 29–60.
S. Boyer and D. Lines, [*Surgery formulae for Casson’s invariant and extensions to homology lens spaces,*]{} J. Reine Angew. Math. **405** (1990), 181–220.
J. C. Cha and C. Livingston, KnotInfo: Table of knot invariants. <http://www.indiana.edu/knotinfo>
S.V. Chmutov and A. N. Varchenko, [*Remarks on the Vassiliev knot invariants coming from $sl_2$,*]{} Topology **36** (1997), 153–178.
M. Culler, C. Gordon, J. Luecke, and P. Shalen, [*Dehn surgery on knots,*]{} Ann. of Math. (2) **125**(1987), 237–300.
C. Gordon and J. Luecke, [*Knots are determined by their complements,*]{} J. Amer. Math. Soc. **2** (1989), 371–415.
M. Goussarov, [*Finite type invariants and n-equivalence of 3-manifolds,*]{} C. R. Acad. Sci. Paris Sér. I Math. **329** (1999), 517–522. S. Garoufalidis and L. Rozansky, [*The loop expansion of the Kontsevich integral, the null-move and S-equivalence,*]{} Topology **43** (2004), 1183–1210.
K. Habiro, [*Claspers and finite type invariants of links,*]{} Geom. Topol. **4** (2000), 1–83. K. Ichihara and Z. Wu, [*A note on Jones polynomial and cosmetic surgery,*]{} arXiv:1606.03372v1 Edited by Rob Kirby. AMS/IP Stud. Adv. Math., 2.2, Geometric topology (Athens, GA, 1993), 3–473, Amer. Math. Soc., Providence, RI, 1997.
M. Kontsevich, [*Vassiliev’s knot invariants,*]{} I. M. Gel’fand Seminar, Adv. Soviet Math. **16** (1993) 137–150.
M. Kontsevich, [*Feynman diagrams and low-dimensional topology,*]{} First European Congress of Mathematics, Vol. II (Paris, 1992), 97–121, Progr. Math., 120, Birkhäuser, Basel, 1994.
P. Kronheimer, T. Mrowka, P. Ozsváth and Z. Szabó, [*Monopoles and lens space surgeries,*]{} Ann. of Math. 165 (2007) 457–546
G. Kuperberg and D. Thurston, [*Perturbative 3-manifold invariants by cut-and-paste topology,*]{} arXiv:math.GT/9912167
T. Le, J. Murakami and T. Ohtsuki, [*On a universal perturbative invariant of 3-manifolds,*]{} Topology **37** (1998), 539–574,
T. Le, H. Murakami, J. Murakami and T. Ohtsuki, [*A three-manifold invariant via the Kontsevich integral*]{} Osaka J. Math. **36** (1999), 365–395.
C. Lescop, [*Surgery formulae for finite type invariants of rational homology 3-spheres,*]{} Algebr. Geom. Topol. **9** (2009), 979–1047.
P. Melvin and H. Morton, [*The coloured Jones function,*]{} Comm. Math. Phys **183** (1997) 291–306.
Y. Ni and Z. Wu, [*Cosmetic surgeries on knots in $S^{3}$,*]{} J. Reine Angew. Math. **706** (2015), 1-17.
Y. Ni and X. Zhang, [*Characterizing slopes for torus knots,*]{} Algebr. Geom. Topol. **14** (2014) 1249–1274.
Solving the diophantine equation $ax^2-by^2=c$, using the LMM method, in Keith Matthews’s webpage. available at <http://www.numbertheory.org/php/aa.html>
T. Ohtsuki, [*Quantum invariants,*]{} Series on Knots and Everything, 29. World Scientific Publishing Co., Inc., River Edge, NJ, 2002.
P. Ozsváth and Z. Szabó, [*The Dehn surgery characterization of the trefoil and the figure eight knot,*]{} arXiv:math.GT/0604079
P. Ozsváth and Z. Szabó, [*Knot Floer homology and rational surgeries,*]{} Algebr. Geom. Topol. **11** (2011) 1–68 .
K. Walker, [*An extension of Casson’s invariant,*]{} Annals of Mathematics Studies, 126. Princeton University Press, Princeton, NJ, 1992. vi+131
|
{
"pile_set_name": "ArXiv"
}
|
---
abstract: 'We study the representation theoretic results of the binary cubic generic Clifford algebra $\mathcal C$, which is an Artin-Schelter regular algebra of global dimension five. In particular, we show that $\mathcal C$ is a PI algebra of PI degree three and compute its point variety and discriminant ideals. As a consequence, we give a necessary and sufficient condition on a binary cubic form $f$ for the associated Clifford algebra $\mathcal C_f$ to be an Azumaya algebra.'
address:
- |
Department of Mathematics\
University of Pittsburgh, Pittsburgh, PA 15260
- |
Department of Mathematics\
Howard University, Washington, D.C. 20059
author:
- Linhong Wang
- 'Xingting Wang\*'
title: A note on generic Clifford algebras of binary cubic forms
---
[^1]
Introduction
============
Let $f$ be a form of degree $m$ in $n$ variables over a base field $k$. The Clifford algebra $\mathcal C_f$ associated to the form $f$ is defined to be an associative algebra $k\left\langle x_1,\dots,x_n\right\rangle$ subject to the relations $(a_1x_1+\cdots+a_nx_n)^m-f(a_1,\dots,a_n)$ for all $a_1,\dots,a_n\in k$. In [@CYZ], Chan-Young-Zhang defined the generic Clifford algebra (or more precisely universal Clifford algebra) $\mathcal C_{m, n}=k\left\langle x_1,\dots,x_n\right\rangle$ subject to the relations $(a_1x_1+\cdots+a_nx_n)^mx_j-x_j(a_1x_1+\cdots+a_nx_n)^m$ for all $a_1,\dots,a_n\in k$ and $1\le j\le n$. The work of $\S \ref {Discriminant}$ implies that the Clifford algebra $\mathcal C_f$ is a homomorphic image of the generic Clifford algebra $\mathcal C_{m,n}$, which paves a way to understand the representations of the former through the representations of the latter. Chan-Young-Zhang further showed that ([@CYZ Lemma 3.8]) when $m=2,n=2$, the binary quadratic generic Clifford algebra belongs to the nice family of connected graded algebras that are Artin-Schelter regular (see [@AS]).
In this note, we focus on the binary cubic generic Clifford algebra ($m=3,n=2$) and denote it by $\mathcal C$. In [@WW2012], all Artin-Schelter regular algebras of global dimension five that are generated by two generators with three generating relations are classified. The algebra $\mathcal C$ is listed as one of the type $\mathbf{A}$ algebras there, which is also proved to be strongly noetherian, Auslander regular and Cohen-Macaulay. In this note, we further show that $\mathcal C$ is indeed a polynomial identity (PI) algebra of PI degree three (see Theorem \[RepC\]). This and other known properties allow us to fully understand the representations of $\mathcal C$ by the well-developed representation theory of PI algebras (e.g. [@BG97; @BG2002; @BY]). As a consequence, we derive some ring-theoretic results of the binary cubic Clifford algebra $\mathcal C_f$, whose properties have been studied by Heerema [@Heer] and Haile [@Haile].
This work was inspired by the poster presentation given by Charlotte Ure at Algebra Extravaganza! at Temple University in July 2017. The authors also learned from Q.-S. Wu later that the binary cubic generic Clifford algebra has appeared as one of the type $\mathbf{A}$ algebras in the classification of global dimensional five Artin-Schelter regular algebras generated by two generators with three generating relations; see [@WW2012]. The authors are grateful for the correspondence from them. The authors also want to thank the referee for his/her careful reading and suggestions.
Background
==========
Let $k$ be an algebraically closed base field of characteristic not $2$ or $3$. The reader is referred (e.g.) to [@GW] and [@MR] for further background on the theory of PI rings. Recall that a ring $R$ with center $Z(R)$ is called *Azumaya* over $Z(R)$ if $R$ is a finitely generated projective $Z(R)$-module and the natural map $R\otimes_{Z(R)} R^{\rm{op}}\to \rm{End}_{Z(R)}(R)$ is an isomorphism. Let $\Lambda$ be a prime noetherian affine $k$-algebra finitely generated as a module over its center $Z$. Define $$\begin{aligned}
\mathcal A_\Lambda:&=\{\mathfrak m\in\, {\rm maxSpec}(Z)\, |\, \Lambda_\mathfrak m\, \text{is Azumaya over}\, Z_\mathfrak m\},\\
\mathcal S_\Lambda:&=\{\mathfrak m\in\, {\rm maxSpec}(Z)\, |\, \text{$Z_\mathfrak m$ is not regular}\}.\end{aligned}$$ The set $\mathcal A_\Lambda$ is called the [*Azumaya locus*]{} of $\Lambda$ over $Z$, and $\mathcal S_\Lambda$ is the *singular locus* of $\Lambda$ (or of $Z$).
[@BG2002 Theorem III.1.7][@BG97 Proposition 3.1, Lemma 3.3]\[PI\] Let $\Lambda$ be a prime noetherian affine $k$-algebra that is module-finite over its center $Z$.
- The maximum $k$-dimension of irreducible $\Lambda$-modules equals the PI degree of $\Lambda$.
- Let $S$ be an irreducible $\Lambda$-module, $P=\rm{Ann}_\Lambda(S)$, and $\mathfrak m=P\cap Z$. Thus $\dim_k(S)={\rm PI}$-${\rm deg}(\Lambda)$ if and only if $\Lambda_\mathfrak m$ is Azumaya over $Z_\mathfrak m$.
Moreover if ${\rm gldim}(\Lambda)<\infty$, then $\mathcal A_\Lambda$ is a nonempty open (and hence dense) subset of $\rm{maxSpec}(Z)\setminus \mathcal S_\Lambda$.
In [@CYZ], Chan-Young-Zhang defined the generic Clifford algebra and established some properties in the quadratic case. Recall that the [*binary cubic generic Clifford algebra*]{} $\mathcal C$ is defined as the quotient algebra of the free algebra $k\langle x,y\rangle$ subject to the relations: $$\begin{aligned}
\label{relationC}
x^3y-yx^3,\ x^2y^2+xyxy-yxyx-y^2x^2,\ xy^3-y^3x.\end{aligned}$$ Moreover, $\mathcal C$ is one of the type $\textbf{A}$ algebras studied in [@WW2012] (parameters $t=-1,l_2=1$).
[@WW2012 Theorem 5] The connected graded algebra $\mathcal C$ is Artin-Schelter regular of global dimension five. Moreover, $\mathcal C$ is an Auslander regular, Cohen-Macaulay, and strongly noetherian domain.
By the diamond lemma [@Be], we have that $$\left\{y^i(xy^2)^j(xy)^k(x^2y)^l x^m\, |\, i,j,k,l,m\in \mathbb N\right\}$$ is a $k$-linear basis for $\mathcal C$. Consequently, the Hilbert series of $\mathcal C$ is $\displaystyle{\frac{1}{(1-t)^5(1+t)(1+t+t^2)^2}}$.
\[Discriminant\] Let $f(u,v)=au^3+3bu^2v+3cuv^2+dv^3$ be a cubic form with $a,b,c,d\in k$. Its discriminant is given by $$\begin{aligned}
\label{D}
D:=\frac{1}{4}\left(ad-bc\right)^2-\left(ac-b^2\right)\left(bd-c^2\right).\end{aligned}$$ The [*Clifford algebra*]{} $\mathcal C_f$ of the cubic form $f$ is defined as $$\mathcal C_f:=k\langle x,y\rangle/I,$$ where $I$ is the ideal generated by the elements $(ux+vy)^3-f(u,v)$ for all $u,v\in k$. It is easy to see that $$\begin{aligned}
\label{relationCf}
I=\left(x^3-a,\ y^3-d,\ x^2y+xyx+yx^2-3b,\ y^2x+yxy+xy^2-3c\right).\end{aligned}$$ There is a natural surjection $\mathcal C\twoheadrightarrow \mathcal C_f$ by sending $x\mapsto x,y\mapsto y$ since the ideal generated by the defining relations in Eq. for $\mathcal C$ is contained in $I$.
[@Haile Theorem 1.1’, Corollary 1.2’]\[Clifford\] The Clifford algebra $\mathcal C_f$ is an Azumaya algebra of PI degree three if $D\neq 0$. In this case, its center is a Dedekind domain, isomorphic to the coordinate ring of the affine elliptic curve $u^2=v^3-27D$.
The center and singular locus of $\mathcal C$
=============================================
Consider the projective space $\mathbb P^3$ with homogenous coordinates $[z_0:z_1:z_2:z_3]$. For cubic binary forms, the discriminant projective variety is defined by $\Delta=0$, where (see Eq.) $$\Delta=\frac{1}{4}\left(z_0z_3-z_1z_2\right)^2-\left(z_0z_2-z_1^2\right)\left(z_1z_3-z_2^2\right).$$
\[SD\] The reduced variety of the singular locus of the discriminant projective variety is given by the twisted cubic curve $v: \mathbb P^1\to \mathbb P^3$ via $$v: [x_0:x_1]\mapsto [x_0^3:x_0^2x_1:x_0x_1^2:x_1^3]=[z_0:z_1:z_2:z_3].$$
It is well known (e.g. [@Hart Chapter I Theorem 5.1]) that the singular locus of $\Delta=0$ is the zero locus of the polynomials $$\begin{aligned}
\frac{\partial\Delta}{\partial z_0}&=-\frac{3}{2}z_1z_2z_3+\frac{1}{2}z_0z_3^2+z_2^3=0,\label{e1}\\
\frac{\partial\Delta}{\partial z_1}&=-\frac{3}{2}z_0z_2z_3-\frac{3}{2}z_1z_2^2+3z_1^2z_3=0,\label{e2}\\
\frac{\partial\Delta}{\partial z_2}&=-\frac{3}{2}z_0z_1z_3-\frac{3}{2}z_1^2z_2+3z_0z_2^2=0,\label{e3}\\
\frac{\partial\Delta}{\partial z_3}&=-\frac{3}{2}z_0z_1z_2+\frac{1}{2}z_0^2z_3+z_1^3=0.\label{e4}\end{aligned}$$ If $z_0=0$, then $z_1=z_2=0$ and $z_3\neq 0$, which corresponds to the point $v([0:1])$. Now let $z_0=1$. From Eq., we have $z_3=3z_1z_2-2z_1^3$. Substituting it into Eq., we obtain $(z_1^2-z_2)^2=0$ which implies that $z_2=z_1^2$ and $z_3=z_1^3$. One can check that $[1:z_1:z_1^2:z_1^3]=v([1:z_1])$ satisfies Eq.-Eq.. Hence the solutions are exactly given by the twisted cubic curve.
In the following, we denote by $Z$ the center of the generic Clifford algebra $\mathcal C$. We will describe the center $Z$ according to [@UK]. Consider the following central elements of $\mathcal C$, where $\omega\in k$ is a primitive third root of unity. $$\left\{
\begin{aligned}\label{Z}
z_0&=x^3\\
z_1&=\frac{1}{3}(x^2y+xyx+yx^2)\\
z_2&=\frac{1}{3}(y^2x+yxy+xy^2)\\
z_3&=y^3\\
z_4&=(yx-\omega xy)^3-\frac{3}{2}\omega(1-\omega)x^3y^3-\frac{9}{2}(1+2\omega^2)z_1z_2\\
z_5&=(xy)^2-y^2x^2=(yx)^2-x^2y^2.
\end{aligned}\right.$$ Define the formal discriminant element $\Delta$ in $\mathcal C$ as $$\begin{aligned}
\label{Delta}
\Delta=\frac{1}{4}\left(z_0z_3-z_1z_2\right)^2-\left(z_0z_2-z_1^2\right)\left(z_1z_3-z_2^2\right).\end{aligned}$$ Under the natural surjection $\mathcal C\twoheadrightarrow \mathcal C_f$, it follows that $(z_0,z_1,z_2,z_3)\mapsto (a,b,c,d)$. In particular, the formal discriminant element $\Delta$ maps to the discriminant $D\in k$ of the binary cubic form $f$ under the surjection $\mathcal C\twoheadrightarrow \mathcal C_f$.
[@UK]\[T:UK\] The generic Clifford algebra $\mathcal C$ is finitely generated as a module over its center $Z$. Moreover, the center $Z$ is generated by $(z_i)_{0\le i\le 5}$ subject to one relation $z_4^2=z_5^3-27\Delta$. Consequently, ${\rm maxSpec}(Z)$ is isomorphic to the coordinate ring of a relative quasiprojective curve over the 4-dimensional affine space $\mathbb A^4=\mathrm{maxSpec}(k[z_0,z_1,z_2,z_3])$ that is elliptic over an open subset of $\mathbb A^4$.
\[SC\] The reduced variety of the singular locus $\mathcal S_{\mathcal C}$ of $Z$ is an affine twisted cubic curve in $\mathbb A^4={\rm maxSpec}(k[z_0,z_1,z_2,z_3])$ defined by $$\mathcal S_C=\mathbb V\left(z_4,\, z_5,\, z_0z_3-z_1z_2,\, z_0z_2-z_1^2,\, z_1z_3-z_2^2\right).$$
Since the center $Z$ generated by $(z_i)_{0\le i\le 5}$ is subject to one relation $R:=z_4^2-z_5^3+27\Delta$, the singular locus $\mathcal S_\mathcal C$ of $Z$ is the zero locus of the derivatives $\frac{\partial R}{\partial z_i}$ for $0\le i\le 5$, or equivalently (if ignoring multiplicity) $z_4=z_5=0$ and $\frac{\partial \Delta}{\partial z_0}=\frac{\partial \Delta}{\partial z_1}=\frac{\partial \Delta}{\partial z_2}=\frac{\partial \Delta}{\partial z_3}=0$. Thus, the result follows from Lemma \[SD\].
Irreducible representations and Azumaya locus of $\mathcal C$
=============================================================
It is clear that the set of isomorphism classes of one-dimensional representations of $\mathcal C$ is bijective to $\mathbb A^2$ via the one-to-one correspondence ${\rm Ann}_\mathcal C(S)=(x-a,y-b)\leftrightarrow (a,b)\in \mathbb A^2$ for any $\mathcal C$-module $S$ with $\dim_k(S)=1$. The following result is crucial to our work.
\[dim2\] There are no two-dimensional irreducible representations over $\mathcal C$.
Suppose $S$ is any two-dimensional irreducible representation over $\mathcal C$. This implies that there is a surjective algebra map $\varphi: \mathcal C\twoheadrightarrow {\rm End}_k(S)={\rm M}_2(k)$. One sees easily that $\varphi(x)\neq 0$; otherwise $S$ can be viewed as a module over $\mathcal C/(x)\cong k[y]$, where all irreducible representations are one-dimensional. The same argument shows that $\varphi(x)$ is not a scalar multiple of the identity matrix in ${\rm M}_2(k)$. Since $x^3$ is central in $\mathcal C$, $\varphi(x^3)=\varphi(x)^3$ is a scalar multiple of the identity matrix. Therefore by a linear transformation of $S$ and a possible rescaling of the variables $x,y$ of $\mathcal C$, we can assume that $\varphi(x)=\begin{pmatrix} 1 & 0 \\ 0 & \omega \end{pmatrix}$ where $\omega$ is a primitive third root of unity. Now write $\varphi(y)=\begin{pmatrix} a & b \\ c & d \end{pmatrix}$ for some $a,b,c,d\in k$. When we apply $\varphi$ to the relations in $\mathcal C$, we obtain $$\begin{aligned}
\varphi(x^2y^2+(xy)^2-(yx)^2-y^2x^2)=\begin{pmatrix}0 & 3b(a-\omega ^2d) \\ -3c(a-\omega^2d) & 0 \end{pmatrix}=0.\end{aligned}$$ If $a\neq \omega^2d$, then we have $b=c=0$. So $\varphi(x)$ and $\varphi(y)$ are both diagonal. This shows that $S$ is not irreducible, which is a contradiction. Hence we have $a=\omega^2 d$. Thus, $\varphi$ sends the central element $y^3$ to $$\varphi(y^3)=\begin{pmatrix} a^3+(2+\omega)abc & b^2c \\ bc^2 &a^3+(1+2\omega)abc \end{pmatrix}.$$ Since $\varphi(y^3)$ is a scalar multiple of the identity matrix, we obtain $bc=0$ and $\varphi(y)$ is either upper or lower triangular. Again, it is a contradiction since $S$ is irreducible.
We denote by ${\rm Irr}\, \mathcal C$ the isomorphism classes of all irreducible representations over $\mathcal C$. For each possible integer $n\ge 1$, ${\rm Irr}_n\, \mathcal C$ denotes the isomorphism subclasses of all $n$-dimensional irreducible representations over $\mathcal C$. Recall $\mathcal S_\mathcal C\subset {\rm maxSpec}(Z)$ is the singular locus of $\mathcal C$ and we call its complement ${\rm maxSpec}(Z)\setminus \mathcal S_\mathcal C$ the *smooth locus* of $\mathcal C$. Note that there is a natural surjection $$\label{SurSpec}
\chi: {\rm Irr}\, \mathcal C\twoheadrightarrow Y:={\rm maxSpec}(Z)$$ via $S\mapsto {\rm Ann}_\mathcal C(S)\cap Z$ for any irreducible representation $S$ over $\mathcal C$.
\[RepC\] The following hold for the binary cubic generic Clifford algebra $\mathcal C$.
- $\mathcal C$ is a PI algebra of PI degree three.
- ${\rm Irr}\, \mathcal C={\rm Irr}_1\, \mathcal C\sqcup {\rm Irr}_3\, \mathcal C$.
- The map ${\rm Irr}_1\, \mathcal C\twoheadrightarrow S_\mathcal C$ is three to one.
- The map ${\rm Irr}_3\, \mathcal C\to {\rm maxSpec}(Z)\setminus \mathcal S_\mathcal C$ is one to one.
- The Azumaya locus of $\mathcal C$ coincides with the smooth locus of $\mathcal C$.
\(a) By Theorem \[T:UK\], we know $\mathcal C$ is module-finite over its center $Z$ and hence is PI by [@MR Corollary 13.1.13(iii)]. Now suppose $\mathcal C$ has PI degree $n$. Regarding Eq., we know $\chi({\rm Irr}_n\, \mathcal C)$ is an open dense subset of $Y$ by Theorem \[PI\]. Take the formal discriminant $\Delta$ of $\mathcal C$ as in Eq.. Note that $Y\setminus \mathbb V(\Delta)$ is another open dense subset of $Y$ since $Y$ is irreducible. So there exists a maximal ideal $$\mathfrak m=(z_0-a,z_1-b,z_2-c,z_3-d,z_4-e,z_5-f)\in \chi({\rm Irr}_n\, \mathcal C)\setminus \mathbb V(\Delta),$$ for some $a,b,c,d,e,f\in k$. Choose any irreducible representation $S$ over $\mathcal C$ such that $\chi(S)=\mathfrak m$ and $\dim_k (S)=n$. Let $C_f$ be the Clifford algebra associated to the binary cubic form $f(u,v)=au^3+3bu^2v+3cuv^2+dv^3$. Therefore, there are algebra surjections $$\mathcal C\twoheadrightarrow \mathcal C_f\twoheadrightarrow \mathcal C/{\mathfrak m}\, \mathcal C,$$ where the first surjection sends $\Delta$ to a nonzero discriminant $D$ in $C_f$ (see Eq.). By Theorem \[Clifford\], $\mathcal C_f$ is an Azumaya algebra of PI degree three. So every irreducible representation over $\mathcal C_f$ is three-dimensional and the same holds for its image $\mathcal C/{\mathfrak m}\,\mathcal C$ as well. Since $\mathfrak mS=0$, we can view $S$ as an irreducible representation over $\mathcal C/{\mathfrak m}\, \mathcal C$. This implies that $\dim_k(S)=n=3$.
\(b) follows from Theorem \[PI\] and Lemma \[dim2\] since $\mathcal C$ has PI degree three.
\(c) Let $S(a,b)=\mathcal C/(x-a,y-b)$ be any one-dimensional representation over $\mathcal C$ for some $a,b\in k$. One checks by Eq. that $$\chi\left(S(a,b)\right)=\left(z_0-a^3,\,z_1-a^2b,\,z_2-ab^2,\,z_3-b^3,\,z_4,\,z_5\right)=:\mathfrak n.$$ Also we have $\chi\left(S(\omega a, \omega b)\right)=\chi\left(S(\omega^2 a, \omega^2 b)\right)=\mathfrak n$. Suppose $\chi^{-1}(\mathfrak n)=\{S_1,S_2,\dots,S_t\}$. According to [@BY Theorem 3.1(f)], $$\dim_k(S_1)+\dim_k(S_2)+\cdots +\dim_k(S_t)\le {\rm PI}\text{-}{\rm deg}(\mathcal C)=3.$$ Hence $t\le 3$ and $\chi^{-1}(\mathfrak n)$ exactly contains three one-dimensional irreducible representations $S(a,b)$, $S(\omega a, \omega b)$ and $S(\omega^2 a, \omega^2 b)$. Moreover, $\chi$ maps ${\rm Irr}_1\, \mathcal C$ onto $\mathcal S_\mathcal C$ by Corollary \[SC\]. The result follows.
\(d) By (b), we have the surjection $\chi: \mathcal C={\rm Irr}_1\, \mathcal C\sqcup {\rm Irr}_3\, \mathcal C\twoheadrightarrow Y=\mathcal S_\mathcal C\sqcup(Y\setminus \mathcal S_\mathcal C)$. By (c), we have an induced surjection $\chi: {\rm Irr}_3\, \mathcal C\twoheadrightarrow Y\setminus \mathcal S_\mathcal C$. Note that $\mathcal C$ has PI degree three. By the same argument as in (c), $\chi$ yields a one-to-one correspondence between ${\rm Irr}_3\, \mathcal C$ and $Y\setminus \mathcal S_\mathcal C$.
\(e) is a consequence of (d) and (b) of Theorem \[PI\].
As recalled in Theorem \[Clifford\], Haile showed that the Clifford algebra $\mathcal C_f$ is an Azumaya algebra when the binary cubic form $f$ is nondegenerate or the discriminant $D\neq 0$ (see Eq.). In the following, we give a necessary and sufficient condition for $\mathcal C_f$ to be Azumaya. It turns out that $\mathcal C_f$ can be an Azumaya algebra even when $f$ is degenerate or $D=0$.
Let $k$ be a base field of characteristic different from $2$ or $3$. The Clifford algebra $\mathcal C_f$ associated to the binary cubic form $f(u,v)=au^3+3bu^2v+3cuv^2+dv^3$ is an Azumaya algebra (of [PI degree]{} three) if and only if the point $(a,b,c,d)$ does not lie on the affine twisted cubic curve in $\mathbb A^4$ described in Lemma \[SD\].
Without loss of generality, we can assume $k$ to be algebraically closed. Recall $Z=k[z_0,\dots,z_5]$, where $z_4^2=z_5^3-27 \Delta$, is the center of $\mathcal C$. It is easy to check that $Z$ has a $k$-basis $$\{z_0^{i_0}z_1^{i_1}z_2^{i_2}z_3^{i_3}z_4^{i_4}z_5^{i_5}\,|\, 0\le i_0,i_1,i_2,i_3,i_4, 0\le i_5\le 2\}.$$ As a consequence, it induces a projection $\pi: {\rm maxSpec}(Z)\twoheadrightarrow \mathbb A^4$ given by the natural inclusion $k[z_0,z_1,z_2,z_3]\hookrightarrow Z$. Therefore the natural surjection $\mathcal C\twoheadrightarrow \mathcal C_f$ factors through $\mathcal C/\mathfrak m\, \mathcal C$ for some $\mathfrak m\in {\rm maxSpec}(Z)$ if and only if $\mathfrak m\in \pi^{-1}(a,b,c,d)$. By Theorem \[RepC\], we know $$\mathcal C/\mathfrak m\,\mathcal C\, \cong\,
\begin{cases}
{\rm M}_3(k) & \mathfrak m\in{\rm maxSpec}(Z)\setminus S_\mathcal C\\
\text{a local algebra} & \mathfrak m\in S_\mathcal C
\end{cases}.$$ By Corollary \[SC\], $\pi^{-1}(a,b,c,d)\,\cap\, \mathcal S_\mathcal C\neq \emptyset$ if and only if $(a,b,c,d)$ lies in the affine twisted cubic curve in $\mathbb A^4$ given by Lemma \[SD\]. In this case, we know $\mathcal C_f$ has irreducible representations of both dimensions one and three. Otherwise, all the irreducible representations over $\mathcal C_f$ are of dimension three. Thus the result follows by the Artin-Procesi theorem on polynomial identities (see [@Artin Theorem 8.3]).
Point variety of $\mathcal C$
=============================
In noncommutative projective algebraic geometry (e.g. see [@ATV]), a [*point module*]{} $M=\bigoplus_{i\ge 0} M_i$ over $\mathcal C$ is a cyclic graded module with Hilbert series $\frac{1}{1-t}$, or namely $M$ is generated by $M_0$ and $\dim_k M_i=1$ for all $i\ge 0$. Note that every point module $M$ over $\mathcal C$ can be denoted by the following diagram $$\xymatrix{
M(p_0,p_1,p_2,\dots): &\underset{e_0}{\bullet}\ar@/^/[r]^-{x_0}\ar@/_/[r]_-{y_0} &\underset{e_1}{\bullet}\ar@/^/[r]^-{x_1}\ar@/_/[r]_-{y_1}& \underset{e_2}{\bullet}\ar@/^/[r]^-{x_2}\ar@/_/[r]_-{y_2}& \underset{e_3}{\bullet}\ar@/^/[r]^-{x_3}\ar@/_/[r]_-{y_3}& \underset{e_4}{\bullet}\ar@/^/[r]^-{x_4}\ar@/_/[r]_-{y_4} & \bullet &\cdots
}$$ where $M=\bigoplus_{i\ge 0} k\,e_i$ and $0\neq p_i=(x_i,y_i)\in k^2$ with the $\mathcal C$-action on $M$ given by $$\begin{aligned}
\label{ActionM}
xe_i=x_{i}e_{i+1},\quad ye_i=y_ie_{i+1}.\end{aligned}$$ It is straightforward to check that two point modules $M(p_0,p_1,\dots)$ and $M(q_0,q_1,\dots)$ are isomorphic as graded $\mathcal C$ modules if and only if $p_i=q_i$ in $\mathbb P^1$ for all $i$ or $(p_0,p_1,\dots)=(q_0,q_1,\dots)$ in $\mathbb P^1\times \mathbb P^1\times \cdots$. Therefore, the isomorphism classes of all point modules over $\mathcal C$ can be parametrized by a subvariety in $\mathbb P^\infty$, which is called the [*point variety*]{} of $\mathcal C$. Sometimes, the point variety in $\mathbb P^\infty$ is determined by a variety $X$ in $\mathbb P^r$ for some $r\ge 1$ and an automorphism $\sigma: X\to X$, where the point variety is determined by the natural embedding $X\hookrightarrow \mathbb P^\infty$ via $x\mapsto (x,\sigma(x),\sigma^2(x),\dots)$ for any $x\in X$. In such a case for simplicity, we say the point variety is given by the pair $(X,\sigma)$.
\[PS\] The point variety of $\mathcal C$ is given by the pair $(\mathbb P^1\times \mathbb P^1\times \mathbb P^1,\, {\rm id})$.
Let $M(p_0,p_1,\dots)$ be a point module. Taking any relation $f$ in Eq. of $\mathcal C$, we have $fe_0=0$. Therefore we obtain $$\left\{
\begin{aligned}
x_3x_2x_1y_0-y_3x_2x_1x_0=0\\
x_3x_2y_1y_0+x_3y_2x_1y_0-y_3x_2y_1x_0-y_3y_2x_1x_0=0\\
x_3y_2y_1y_0-y_3y_2y_1x_0=0
\end{aligned}\right.$$ Consider the above equations as a system of linear equations in terms of $x_3,y_3$ and rewrite it in the matrix form. $$\begin{bmatrix}
x_1x_2y_0 & -x_0x_1x_2\\
x_2y_0y_1+x_1y_0y_2& -x_0x_2y_1-x_1x_0y_2\\
y_0y_1y_2 & -x_0y_1y_2
\end{bmatrix}
\begin{bmatrix}
x_3\\y_3
\end{bmatrix}
=0.$$ It is easy to check that it has a solution $(x_3,y_3)=(x_0,y_0)$. We claim that it is the only solution up to a scalar multiple by showing that the coefficient matrix can not be identically zero. Suppose it is not true, then all the entries appearing in the $3\times 2$ matrix must be zero. Hence $$x_1x_2y_0=x_0x_1x_2=x_2y_0y_1+x_1y_0y_2= x_0x_2y_1+x_1x_0y_2=y_0y_1y_2=x_0y_1y_2=0.$$ One can check the equations have no solution. By repeating the above argument regarding the basis vector $e_i$ for all $i$, we see that $M(p_0,p_1,\dots)$ is well defined if and only if $p_i=p_{i+3}$ for all $i$. This yields our result.
Let $\underline{p}=(p_0,p_1,p_2)$ be a point in $\mathbb P^1\times \mathbb P^1\times \mathbb P^1$. According to Theorem \[PS\], the point variety over $\mathcal C$ is given by all $M(\underline{p})$ for some $\underline{p}\in \mathbb P^1\times \mathbb P^1\times \mathbb P^1$. Denote by $\Gamma=\{(p,p,p)\,|\, p\in \mathbb P^1\}\subset \mathbb P^1\times \mathbb P^1\times \mathbb P^1$ the diagonal part.
Let $M(\underline{p})$ be any point module over $\mathcal C$, and $S$ be a simple quotient of $M(\underline{p})$.
- If $\underline{p}\in \Gamma$, then $\dim_k(S)=1$.
- If $\underline{p}\not\in \Gamma$, then $S$ is either trivial or $\dim_k(S)=3$.
Let $M=M(p_0,p_1,p_2)=\bigoplus_{i\ge 0}k\, e_i$. We have $\mathcal C$ acts on $M$ as in .
\(a) Suppose $p_0=p_1=p_2=(a,b)\in \mathbb P^1$. It is direct to check that the central elements $\{z_i\}_{0\le i\le 5}$ in Eq. act on $M$ as follows. $$\begin{aligned}
z_0e_i=a^3e_{i+3},\ z_1e_i=a^2be_{i+3},\ z_2e_i=ab^2e_{i+3},\ z_3e_i=b^3e_{i+3},\ z_4e_i=z_5e_i=0.\end{aligned}$$ Hence $(z_0z_3-z_1z_2,z_0z_2-z_1^2,z_1z_3-z_2^2,z_4,z_5)\subset {\rm Ann}_\mathcal C(M)\subset {\rm Ann}_\mathcal C(S)$. By Theorem \[RepC\] and Corollary \[SC\], we have $\mathbb V({\rm Ann}_\mathcal C(S)\cap Z)\in \mathcal S_\mathcal C$. Hence $\dim_k S=1$.
\(b) According to Eq., we obtain the following actions on $M$. $$\begin{aligned}
(z_0z_3-z_1z_2)e_i=-\frac{a}{9}e_{i+6},\ (z_0z_2-z_1^2)e_i=-\frac{b}{18}e_{i+6},\ (z_1z_3-z_2^2)e_i=-\frac{c}{18}e_{i+6},\end{aligned}$$ where the coefficients $a,b,c$ are given by $$\begin{aligned}
a&\,=x_2y_2(x_1y_0-x_0y_1)^2+x_1y_1(x_2y_0-x_0y_2)^2+x_0y_0(x_2y_1-x_1y_2)^2\\
b&\,=x_0^2(x_2y_1-x_1y_2)^2+x_1^2(x_0y_2-x_2y_0)^2+x_2^2(x_1y_0-x_0y_1)^2\\
c&\,=y_0^2(x_2y_1-x_1y_2)^2+y_1^2(x_0y_2-x_2y_0)^2+y_2^2(x_1y_0-x_0y_1)^2.\end{aligned}$$ Here we write $p_i=(x_i,y_i)\in \mathbb P^1$ for all $i\ge 0$. Now suppose $\dim_kS=1$ and it is not trivial. By Theorem \[RepC\] and Corollary \[SC\], we have $(z_0z_3-z_1z_2,z_0z_2-z_1^2,z_1z_3-z_2^2)\subset {\rm Ann}_\mathcal C(S)$. If any of $a,b,c$ is not zero, it implies that $e_{i+6}=0$ in $S$ for all $i$. Thus $xe_5=ye_5=0$ in $S$. Hence $e_5=0$ in $S$ otherwise $S=ke_5$ would be trivial. Repeating the argument, we obtain all $e_i=0$ in $S$. It is absurd. So $a=b=c=0$. Now set the variables $$X:=(x_1y_0-x_0y_1)^2,\ Y:=(x_2y_0-x_0y_2)^2,\ Z:=(x_2y_1-x_1y_2)^2.$$ We write the above conditions as a system of linear equations in terms of $X,Y,Z$ as follows. $$\begin{aligned}
\label{abc}
\begin{bmatrix}
x_2y_2 & x_1y_1 & x_0y_0\\
x_2^2 & x_1^2&x_0^2\\
y_2^2&y_1^2&y_0^2
\end{bmatrix}
\begin{bmatrix}
X\\Y\\Z
\end{bmatrix}=0.\end{aligned}$$ We will show that Eq. has no solution whenever $\underline{p}\not\in \Gamma$, which yields that $\dim_k S>1$ if $S$ is not trivial. Hence part (b) follows from Theorem \[RepC\]. Note that $X=0$ implies that $p_0=p_1$, $Y=0$ implies that $p_0=p_2$, and $Z=0$ implies that $p_1=p_2$. Since $\underline{p}\not\in \Gamma$, $(X,Y, Z)$ is a nonzero solution for Eq.. So $$0={\rm det}
\begin{pmatrix}
x_2y_2 & x_1y_1 & x_0y_0\\
x_2^2 & x_1^2&x_0^2\\
y_2^2&y_1^2&y_0^2
\end{pmatrix}:=\gamma.$$ Direct computation shows that $\gamma^2=XYZ$. So $\mathbb V(\gamma)=\mathbb V(X)\,\cup\, \mathbb V(Y)\,\cup\, \mathbb V(Z)$. Without loss of generality, we assume that $\underline{p}\in \mathbb V(\gamma)\cap \mathbb V(X)$. Since $X=0$, we know $(x_0,y_0)=(x_1,y_1)\in \mathbb P^1$. Moreover, we can check that $Y=Z\neq 0$. But in such a case, $(X,Y,Z)$ does not satisfy Eq.. We obtain a contradiction, which proves our result.
Discriminant ideals of $\mathcal C$
===================================
Discriminant ideals of PI algebras play an important role in the study of a maximal orders [@Reiner], the automorphism and isomorphism problems for noncommutative algebras [@CPWZ], the Zariski cancellation problem for noncommutative algebras [@BZ], and the description of dimensions of irreducible representations [@BY].
Let $A$ be an algebra and $C\subseteq Z(A)$ be a central subalgebra. A *trace map* on $A$ is a nonzero map $\text{tr}: A\to C$ that is cyclic ($\text{tr}(xy)=\text{tr}(yx)$ for $x,y\in A$) and $C$-linear. For a positive integer $\ell$, the *$\ell$-th discriminant ideal* $D_\ell(A/C)$ and the *$\ell$-th modified discriminant ideal* ${MD}_\ell(A/C)$ of $A$ over $C$ are the ideals of $C$ with generating sets $$\left\{\text{det}([\text{tr}(y_iy_j)]_{i,j=1}^\ell)\,|\, y_1,\dots,y_\ell \in A\right\}$$ and $$\left\{\text{det}([\text{tr}(y_iy_j')]_{i,j=1}^\ell)\,|\, y_1,y_1',\dots,y_\ell,y_\ell' \in A\right\}.$$
Since we know the binary cubic generic Clifford algebra $\mathcal C$ is Auslander-regular, Cohen-Macaulay, and stably-free (it is connected graded), we can employ Stafford’s work [@Staff Theorem 2.10] to conclude that $\mathcal C$ is a maximal order in a central simple algebra and admits the reduced trace map $\text{tr}: \mathcal C\to Z$ (e.g. [@Reiner Section 9]). The next theorem describes the zero sets of the discriminant ideals of $\mathcal C$ which has PI degree three.
Let $\mathcal C$ be the binary cubic generic Clifford algebra of PI degree three with reduced trace map $\text{tr}: \mathcal C\to Z$. For all positive integers $\ell$, the zero sets of the $\ell$-th discriminant and $\ell$-th modified discriminant ideals of $\mathcal C$ over its center $Z$ coincide, $\mathbb V(D_\ell(\mathcal C/Z),{\rm tr})=\mathbb V(MD_\ell(\mathcal C/Z),{\rm tr})$; denote this set by $\mathbb V_\ell \subseteq Y:= {\rm maxSpec}(Z)$. The following hold:
1. $\mathbb V_\ell=\emptyset$ for $\ell\le 3$.
2. $\mathbb V_\ell=Y^{sing}={\mathcal S}_\mathcal C$ for $4\le \ell \le 9$.
3. $\mathbb V_\ell=Y$ for $\ell>9$.
For any $\mathfrak m\in \text{maxSpec}(Z)$, denote by $\text{Irr}_\mathfrak m(\mathcal C)$ the set of isomorphism classes of irreducible representations of $\mathcal C$ with central annihilator $\mathfrak m$. By [@BY Main Theorem (a),(e)], we have $$\mathbb V(D_\ell(\mathcal C/Z),\text{tr})= \mathbb V(MD_\ell(\mathcal C/Z),\text{tr})=\left\{\mathfrak m\in \text{maxSpec}(Z)\,\left |\, \sum_{V\in \text{Irr}_\mathfrak m(\mathcal C)} (\dim_k V)^2< \ell\right\}\right..$$ Hence, our result follows directly from Theorem \[RepC\].
[99]{} M. Artin, On Azumaya algebras and finite dimensional representations of rings, J. Algebra II (1969), 532–563. M. Artin and W. Schelter, Graded algebras of global dimension 3, Adv. Math. 66 (1987), 171–216. M. Artin, J. Tate, and M. Van den Bergh, Some algebras associated to automorphisms of elliptic curves, The Grothendieck Festschrift Vol. I, 33–85, Birkhäuser Boston, Boston, MA, 1990. J. Bell and J. J. Zhang, Zariski cancellation problem for noncommutative algebras, Selecta Math. (N.S.), 23 (2017), 1709–1737. G. Bergman, The diamond lemma for ring theory, Adv. Math. 109 (2) (1994), 228–287. K. A. Brown and K. R. Goodearl, Lectures on algebraic quantum groups, Advanced Courses in Mathematics. CRM Barcelona, Birkhäuser, Basel (2002). K. A. Brown and K. R. Goodearl, Homological aspects of noetherian PI Hopf algebras and irreducible modules of maximal dimension, J. Algebra, 198 (1997), 240–265. K. A. Brown and M. T. Yakimov, Azumaya loci and discriminant ideals of PI algebras, Preprint, 2017. Available at <http://arxiv.org/abs/1702.04305v4>. S. Ceken, J. H. Palmieri, Y.-H. Wang, and J. J. Zhang, The discriminant controls automorphism groups of noncommutative algebras, Adv. Math., 269 (2015), 551–584. K. Chan, A. A. Young, and J. J. Zhang, Discriminant formulas and applications, Algebra Number Theory, 10 (2016), 557–596. K. R. Goodearl and R. B. Warfield, Jr., An Introduction to Noncommutative Noetherian Rings, Second Edition, London Mathematical Society Student Texts 61, Cambridge University Press, Cambridge, 2004. D. E. Haile, On the Clifford algebra of a binary cubic form, Amer. J. Math., 106 (6) (1984), 1269–1280. R. Hartshorne, Algebraic Geometry, Berlin, Heidelberg, New York: Springer-Verlag, 1977. N. Heerema, An algebra determined by a binary cubic form. Duke Math. J., 21 (1954), 423–443. J. C. McConnell and J. C. Robson, Noncommutative Noetherian Rings Graduate Studies in Mathematics 30, American Mathematical Society, Providence, 2000. I. Reiner, Maximal orders, volume 28 of London Mathematical Society Monographs. New Series. The Clarendon Press, Oxford University Press, Oxford, 2003. Corrected reprint of the 1975 original, With a foreword by M. J. Taylor. J. T. Stafford, Auslander-regular algebras and maximal orders, J. London Math. Soc. (2), 50 (1994), 276–292. C. Ure and R. S. Kulkarni, The Generic Clifford Algebra, privately circulated, Algebra Extravaganza, July 2017, Temple University, Philadelphia. S.-Q. Wang and Q.-S. Wu, A class of AS-regular algebras of dimension five, J. of Algebra, 362 (2012), 117–144.
[^1]: \*corresponding author
|
{
"pile_set_name": "ArXiv"
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.